Load .env file automatically on startup

Add automatic .env file loading using python-dotenv. This searches
the current working directory and parent directories for a .env file
and loads environment variables from it.

Uses override=True so .env file values take precedence over existing
shell environment variables, which is the expected behavior when
running from a project directory.

.env.example

@@ -2,11 +2,23 @@
 # Copy this file to .env and fill in your values
 
 # LLM Configuration (Required)
+# Supported providers: openai, groq, ollama, gemini, anthropic, lmstudio
 HINDSIGHT_API_LLM_PROVIDER=openai
 HINDSIGHT_API_LLM_API_KEY=your-api-key-here
 HINDSIGHT_API_LLM_MODEL=o3-mini
 HINDSIGHT_API_LLM_BASE_URL=https://api.openai.com/v1
 
+# Example: Anthropic Claude configuration
+# HINDSIGHT_API_LLM_PROVIDER=anthropic
+# HINDSIGHT_API_LLM_API_KEY=your-anthropic-api-key
+# HINDSIGHT_API_LLM_MODEL=claude-sonnet-4-20250514
+
+# Example: LM Studio local configuration (Qwen 2.5 32B recommended)
+# HINDSIGHT_API_LLM_PROVIDER=lmstudio
+# HINDSIGHT_API_LLM_API_KEY=lmstudio
+# HINDSIGHT_API_LLM_BASE_URL=http://localhost:1234/v1
+# HINDSIGHT_API_LLM_MODEL=qwen2.5-32b-instruct
+
 # API Configuration (Optional)
 HINDSIGHT_API_HOST=0.0.0.0
 HINDSIGHT_API_PORT=8888

.githooks/pre-commit

@@ -0,0 +1,27 @@
+#!/bin/bash
+# Pre-commit hook - runs all scripts in scripts/hooks/
+
+set -e
+
+REPO_ROOT="$(git rev-parse --show-toplevel)"
+HOOKS_DIR="$REPO_ROOT/scripts/hooks"
+
+if [ ! -d "$HOOKS_DIR" ]; then
+    exit 0
+fi
+
+echo ""
+echo "=== Running pre-commit hooks ==="
+echo ""
+
+# Run all executable scripts in hooks directory
+for hook in "$HOOKS_DIR"/*.sh; do
+    if [ -x "$hook" ]; then
+        echo "[hook] $(basename "$hook")"
+        (cd "$REPO_ROOT" && "$hook")
+    fi
+done
+
+echo ""
+echo "=== Pre-commit hooks completed ==="
+echo ""

.github/workflows/deploy-docs.yml

@@ -27,7 +27,9 @@ jobs:
           node-version: 20
           cache: npm
           cache-dependency-path: package-lock.json
+      - uses: astral-sh/setup-uv@v4
       - run: npm ci --workspace=hindsight-docs
+      - run: uv run generate-llms-full
       - run: npm run build --workspace=hindsight-docs
       - uses: actions/upload-pages-artifact@v3
         with:

.github/workflows/release.yml

@@ -42,6 +42,10 @@ jobs:
       working-directory: ./hindsight-integrations/litellm
       run: uv build --out-dir dist
 
+    - name: Build hindsight-embed
+      working-directory: ./hindsight-embed
+      run: uv build --out-dir dist
+
     # Publish in order (client and api first, then hindsight-all which depends on them)
     - name: Publish hindsight-client to PyPI
       uses: pypa/gh-action-pypi-publish@release/v1
@@ -67,6 +71,12 @@ jobs:
         packages-dir: ./hindsight-integrations/litellm/dist
         skip-existing: true
 
+    - name: Publish hindsight-embed to PyPI
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        packages-dir: ./hindsight-embed/dist
+        skip-existing: true
+
     # Upload artifacts for GitHub release
     - name: Upload artifacts
       uses: actions/upload-artifact@v4
@@ -77,6 +87,7 @@ jobs:
           hindsight-api/dist/*
           hindsight/dist/*
           hindsight-integrations/litellm/dist/*
+          hindsight-embed/dist/*
         retention-days: 1
 
   release-typescript-client:
@@ -102,7 +113,18 @@ jobs:
 
     - name: Publish to npm
       working-directory: ./hindsight-clients/typescript
-      run: npm publish --access public
+      run: |
+        set +e
+        OUTPUT=$(npm publish --access public 2>&1)
+        EXIT_CODE=$?
+        echo "$OUTPUT"
+        if [ $EXIT_CODE -ne 0 ]; then
+          if echo "$OUTPUT" | grep -q "cannot publish over"; then
+            echo "Package version already published, skipping..."
+            exit 0
+          fi
+          exit $EXIT_CODE
+        fi
       env:
         NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
 
@@ -117,6 +139,65 @@ jobs:
         path: hindsight-clients/typescript/*.tgz
         retention-days: 1
 
+  release-control-plane:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+        registry-url: 'https://registry.npmjs.org'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
+
+    - name: Install dependencies
+      run: npm ci
+
+    - name: Build TypeScript client (dependency)
+      run: npm run build --workspace=hindsight-clients/typescript
+
+    - name: Fix platform-specific native modules
+      run: |
+        # npm ci installs from lockfile which may have wrong platform binaries
+        # Delete hoisted native modules and reinstall for current platform
+        rm -rf node_modules/lightningcss node_modules/@tailwindcss
+        npm install lightningcss @tailwindcss/postcss @tailwindcss/node
+
+    - name: Build
+      run: npm run build --workspace=hindsight-control-plane
+
+    - name: Publish to npm
+      working-directory: ./hindsight-control-plane
+      run: |
+        set +e
+        OUTPUT=$(npm publish --access public 2>&1)
+        EXIT_CODE=$?
+        echo "$OUTPUT"
+        if [ $EXIT_CODE -ne 0 ]; then
+          if echo "$OUTPUT" | grep -q "cannot publish over"; then
+            echo "Package version already published, skipping..."
+            exit 0
+          fi
+          exit $EXIT_CODE
+        fi
+      env:
+        NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+    - name: Pack for GitHub release
+      working-directory: ./hindsight-control-plane
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: control-plane
+        path: hindsight-control-plane/*.tgz
+        retention-days: 1
+
   release-rust-cli:
     runs-on: ${{ matrix.os }}
     strategy:
@@ -181,7 +262,7 @@ jobs:
     - name: Free Disk Space
       uses: jlumbroso/free-disk-space@main
       with:
-        tool-cache: false
+        tool-cache: true
         android: true
         dotnet: true
         haskell: true
@@ -206,7 +287,7 @@ jobs:
       id: get_version
       run: echo "VERSION=${GITHUB_REF#refs/tags/v}" >> $GITHUB_OUTPUT
 
-    - name: Extract metadata
+    - name: Extract metadata for release tags
       id: meta
       uses: docker/metadata-action@v5
       with:
@@ -217,7 +298,29 @@ jobs:
           type=semver,pattern={{major}},value=${{ steps.get_version.outputs.VERSION }}
           type=raw,value=latest
 
-    - name: Build and push
+    # TODO: Re-enable smoke test when disk space issue is resolved
+    # # Step 1: Build for local testing (single platform, no push)
+    # # This creates an identical image to what will be released, just for one platform
+    # - name: Build image for testing
+    #   uses: docker/build-push-action@v6
+    #   with:
+    #     context: .
+    #     file: docker/standalone/Dockerfile
+    #     target: ${{ matrix.target }}
+    #     push: false
+    #     load: true
+    #     tags: ${{ matrix.image_name }}:test
+    #     cache-from: type=gha
+    #     cache-to: type=gha,mode=max
+
+    # # Step 2: Test the image before pushing anything
+    # - name: Smoke test - verify container starts
+    #   env:
+    #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
+    #   run: ./scripts/docker-smoke-test.sh "${{ matrix.image_name }}:test" "${{ matrix.target }}"
+
+    # Build multi-platform and push to release tags
+    - name: Build and push release images
       uses: docker/build-push-action@v6
       with:
         context: .
@@ -263,7 +366,7 @@ jobs:
 
   create-github-release:
     runs-on: ubuntu-latest
-    needs: [release-python-packages, release-typescript-client, release-rust-cli, release-docker-images, release-helm-chart]
+    needs: [release-python-packages, release-typescript-client, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
     permissions:
       contents: write
 
@@ -286,6 +389,12 @@ jobs:
         name: typescript-client
         path: ./artifacts/typescript-client
 
+    - name: Download Control Plane
+      uses: actions/download-artifact@v4
+      with:
+        name: control-plane
+        path: ./artifacts/control-plane
+
     - name: Download Rust CLI (Linux)
       uses: actions/download-artifact@v4
       with:
@@ -318,8 +427,11 @@ jobs:
         cp artifacts/python-packages/hindsight-api/dist/* release-assets/ || true
         cp artifacts/python-packages/hindsight/dist/* release-assets/ || true
         cp artifacts/python-packages/hindsight-integrations/litellm/dist/* release-assets/ || true
+        cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
         # TypeScript client
         cp artifacts/typescript-client/*.tgz release-assets/ || true
+        # Control Plane
+        cp artifacts/control-plane/*.tgz release-assets/ || true
         # Rust CLI binaries
         cp artifacts/rust-cli-linux/hindsight-linux-amd64 release-assets/ || true
         cp artifacts/rust-cli-darwin-amd64/hindsight-darwin-amd64 release-assets/ || true

.github/workflows/test.yml

@@ -20,6 +20,8 @@ jobs:
             path: hindsight-api
           - name: hindsight-client
             path: hindsight-clients/python
+          - name: hindsight-embed
+            path: hindsight-embed
 
     steps:
     - uses: actions/checkout@v4
@@ -38,6 +40,29 @@ jobs:
       working-directory: ./${{ matrix.path }}
       run: uv build
 
+  build-api-python-versions:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ['3.11', '3.12', '3.13']
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v5
+      with:
+        python-version: ${{ matrix.python-version }}
+
+    - name: Build hindsight-api
+      working-directory: ./hindsight-api
+      run: uv build
+
   build-typescript-client:
     runs-on: ubuntu-latest
 
@@ -57,6 +82,58 @@ jobs:
     - name: Build TypeScript client
       run: npm run build --workspace=hindsight-clients/typescript
 
+  build-control-plane:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
+
+    - name: Install SDK dependencies
+      run: npm ci --workspace=hindsight-clients/typescript
+
+    - name: Build SDK
+      run: npm run build --workspace=hindsight-clients/typescript
+
+    # Install control plane deps and fix hoisted lightningcss binary
+    # lightningcss gets hoisted to root node_modules, so we need to reinstall it there
+    - name: Install Control Plane dependencies
+      run: |
+        npm install --workspace=hindsight-control-plane
+        rm -rf node_modules/lightningcss node_modules/@tailwindcss
+        npm install lightningcss @tailwindcss/postcss @tailwindcss/node
+
+    - name: Build Control Plane
+      run: npm run build --workspace=hindsight-control-plane
+
+    - name: Verify standalone build
+      run: |
+        test -f hindsight-control-plane/standalone/server.js || exit 1
+        test -d hindsight-control-plane/standalone/node_modules || exit 1
+        node hindsight-control-plane/bin/cli.js --help
+
+    - name: Smoke test - verify server starts
+      run: |
+        cd hindsight-control-plane
+        node bin/cli.js --port 9999 &
+        SERVER_PID=$!
+        sleep 5
+        if curl -sf http://localhost:9999 > /dev/null 2>&1; then
+          echo "Server started successfully"
+          kill $SERVER_PID 2>/dev/null || true
+          exit 0
+        else
+          echo "Server failed to respond"
+          kill $SERVER_PID 2>/dev/null || true
+          exit 1
+        fi
+
   build-docs:
     runs-on: ubuntu-latest
 
@@ -98,6 +175,90 @@ jobs:
       working-directory: hindsight-cli
       run: cargo build --release
 
+    - name: Upload CLI artifact
+      uses: actions/upload-artifact@v4
+      with:
+        name: hindsight-cli
+        path: hindsight-cli/target/release/hindsight
+        retention-days: 1
+
+  test-rust-cli:
+    runs-on: ubuntu-latest
+    needs: build-rust-cli
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      HINDSIGHT_API_URL: http://localhost:8888
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Download CLI artifact
+      uses: actions/download-artifact@v4
+      with:
+        name: hindsight-cli
+        path: /tmp/cli
+
+    - name: Make CLI executable
+      run: chmod +x /tmp/cli/hindsight
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Build API
+      working-directory: ./hindsight-api
+      run: uv build
+
+    - name: Install API dependencies
+      working-directory: ./hindsight-api
+      run: uv sync --no-install-project --index-strategy unsafe-best-match
+
+    - name: Create .env file
+      run: |
+        cat > .env << EOF
+        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
+        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
+        EOF
+
+    - name: Start API server
+      run: |
+        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
+        echo "Waiting for API server to be ready..."
+        for i in {1..60}; do
+          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
+            echo "API server is ready after ${i}s"
+            break
+          fi
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
+            cat /tmp/api-server.log
+            exit 1
+          fi
+          sleep 1
+        done
+
+    - name: Run CLI smoke test
+      run: |
+        HINDSIGHT_CLI=/tmp/cli/hindsight ./hindsight-cli/smoke-test.sh
+
+    - name: Show API server logs
+      if: always()
+      run: |
+        echo "=== API Server Logs ==="
+        cat /tmp/api-server.log || echo "No API server log found"
+
   lint-helm-chart:
     runs-on: ubuntu-latest
 
@@ -130,7 +291,7 @@ jobs:
     - name: Free Disk Space
       uses: jlumbroso/free-disk-space@main
       with:
-        tool-cache: false
+        tool-cache: true
         android: true
         dotnet: true
         haskell: true
@@ -148,6 +309,13 @@ jobs:
         file: docker/standalone/Dockerfile
         target: ${{ matrix.target }}
         push: false
+        load: false
+
+    # TODO: Re-enable smoke test when disk space issue is resolved
+    # - name: Smoke test - verify container starts
+    #   env:
+    #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
+    #   run: ./scripts/docker-smoke-test.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"
 
   test-api:
     runs-on: ubuntu-latest
@@ -472,4 +640,156 @@ jobs:
 
     - name: Run tests
       working-directory: ./hindsight-integrations/litellm
-      run: uv run pytest tests -v
+      run: uv run pytest tests -v
+
+  test-embed:
+    runs-on: ubuntu-latest
+    env:
+      HINDSIGHT_EMBED_LLM_PROVIDER: groq
+      HINDSIGHT_EMBED_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_EMBED_LLM_MODEL: openai/gpt-oss-20b
+      # Prefer CPU-only PyTorch in CI
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Install dependencies
+      working-directory: ./hindsight-embed
+      run: uv sync --index-strategy unsafe-best-match
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-embed-${{ hashFiles('hindsight-embed/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-embed-
+          ${{ runner.os }}-huggingface-
+
+    - name: Run smoke test
+      working-directory: ./hindsight-embed
+      run: ./test.sh
+
+  test-doc-examples:
+    runs-on: ubuntu-latest
+    needs: build-rust-cli
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      HINDSIGHT_API_URL: http://localhost:8888
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Download CLI artifact
+      uses: actions/download-artifact@v4
+      with:
+        name: hindsight-cli
+        path: /usr/local/bin
+
+    - name: Make CLI executable
+      run: chmod +x /usr/local/bin/hindsight
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
+
+    - name: Build and install API
+      working-directory: ./hindsight-api
+      run: |
+        uv build
+        uv sync --no-install-project --index-strategy unsafe-best-match
+
+    - name: Install Python client dependencies
+      working-directory: ./hindsight-clients/python
+      run: uv sync --extra test --index-strategy unsafe-best-match
+
+    - name: Install TypeScript client
+      run: |
+        npm ci --workspace=hindsight-clients/typescript
+        npm run build --workspace=hindsight-clients/typescript
+
+    - name: Create .env file
+      run: |
+        cat > .env << EOF
+        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
+        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
+        EOF
+
+    - name: Start API server
+      run: |
+        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
+        echo "Waiting for API server to be ready..."
+        for i in {1..60}; do
+          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
+            echo "API server is ready after ${i}s"
+            break
+          fi
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
+            cat /tmp/api-server.log
+            exit 1
+          fi
+          sleep 1
+        done
+
+    - name: Run Python doc examples
+      working-directory: ./hindsight-clients/python
+      run: |
+        for f in ../../hindsight-docs/examples/api/*.py; do
+          echo "Running $f..."
+          uv run python "$f"
+        done
+
+    - name: Run Node.js doc examples
+      run: |
+        for f in hindsight-docs/examples/api/*.mjs; do
+          echo "Running $f..."
+          node "$f"
+        done
+
+    - name: Configure CLI
+      run: hindsight configure --api-url http://localhost:8888
+
+    - name: Run CLI doc examples
+      run: |
+        for f in hindsight-docs/examples/api/*.sh; do
+          echo "Running $f..."
+          bash "$f"
+        done
+
+    - name: Show API server logs
+      if: always()
+      run: |
+        echo "=== API Server Logs ==="
+        cat /tmp/api-server.log || echo "No API server log found"

.gitignore

@@ -32,6 +32,9 @@ logs/
 
 .DS_Store
 
+# Generated docs files
+hindsight-docs/static/llms-full.txt
+
 
 hindsight-dev/benchmarks/locomo/results/
 hindsight-dev/benchmarks/longmemeval/results/

AGENTS.md

@@ -14,6 +14,7 @@ This document captures architectural decisions and coding conventions for the Hi
 hindsight/              # Python package for embedded usage
 hindsight-api/          # FastAPI server (core memory engine)
 hindsight-cli/          # Rust CLI client
+hindsight-embed/        # Embedded CLI (no server needed)
 hindsight-control-plane/ # Next.js admin UI
 hindsight-docs/         # Docusaurus documentation site
 hindsight-dev/          # Development tools and benchmarks
@@ -148,4 +149,5 @@ Note: The maintained wrapper `hindsight_client.py` and `README.md` are preserved
 
 # Branding
 ## Colors
-- Primary: gradient from #0074d9 to #009296  
+- Primary: gradient from #0074d9 to #009296
+

CLAUDE.md

@@ -0,0 +1,125 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+Hindsight is an agent memory system that provides long-term memory for AI agents using biomimetic data structures. It stores memories as World facts, Experiences, Opinions, and Observations across memory banks.
+
+## Development Commands
+
+### API Server (Python/FastAPI)
+```bash
+# Start API server (loads .env automatically)
+./scripts/dev/start-api.sh
+
+# Run tests
+cd hindsight-api && uv run pytest tests/
+
+# Run specific test file
+cd hindsight-api && uv run pytest tests/test_http_api_integration.py -v
+
+# Lint
+cd hindsight-api && uv run ruff check .
+```
+
+### Control Plane (Next.js)
+```bash
+./scripts/dev/start-control-plane.sh
+# Or manually:
+cd hindsight-control-plane && npm run dev
+```
+
+### Documentation Site (Docusaurus)
+```bash
+./scripts/dev/start-docs.sh
+```
+
+### Generating Clients/OpenAPI
+```bash
+# Regenerate OpenAPI spec after API changes
+./scripts/generate-openapi.sh
+
+# Regenerate all client SDKs (Python, TypeScript, Rust)
+./scripts/generate-clients.sh
+```
+
+### Benchmarks
+```bash
+./scripts/benchmarks/run-longmemeval.sh
+./scripts/benchmarks/run-locomo.sh
+./scripts/benchmarks/start-visualizer.sh  # View results at localhost:8001
+```
+
+## Architecture
+
+### Monorepo Structure
+- **hindsight-api/**: Core FastAPI server with memory engine (Python, uv)
+- **hindsight/**: Embedded Python bundle (hindsight-all package)
+- **hindsight-control-plane/**: Admin UI (Next.js, npm)
+- **hindsight-cli/**: CLI tool (Rust, cargo)
+- **hindsight-clients/**: Generated SDK clients (Python, TypeScript, Rust)
+- **hindsight-docs/**: Docusaurus documentation site
+- **hindsight-integrations/**: Framework integrations (LiteLLM, OpenAI)
+- **hindsight-dev/**: Development tools and benchmarks
+
+### Core Engine (hindsight-api/hindsight_api/engine/)
+- `memory_engine.py`: Main orchestrator for retain/recall/reflect operations
+- `llm_wrapper.py`: LLM abstraction supporting OpenAI, Anthropic, Gemini, Groq, Ollama, LM Studio
+- `embeddings.py`: Embedding generation (local or TEI)
+- `cross_encoder.py`: Reranking (local or TEI)
+- `entity_resolver.py`: Entity extraction and normalization
+- `query_analyzer.py`: Query intent analysis
+- `retain/`: Memory ingestion pipeline
+- `search/`: Multi-strategy retrieval (semantic, BM25, graph, temporal)
+
+### API Layer (hindsight-api/hindsight_api/api/)
+FastAPI routers for all endpoints. Main operations:
+- **Retain**: Store memories, extracts facts/entities/relationships
+- **Recall**: Retrieve memories via parallel search strategies + reranking
+- **Reflect**: Deep analysis forming new opinions/observations
+
+### Database
+PostgreSQL with pgvector. Schema managed via Alembic migrations in `hindsight-api/hindsight_api/alembic/`. Migrations run automatically on API startup.
+
+Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
+
+## Key Conventions
+
+### Memory Banks
+- Each bank is isolated (no cross-bank data access)
+- Banks have dispositions (skepticism, literalism, empathy traits 1-5) affecting reflect
+- Banks can have background context
+
+### API Design
+- All endpoints operate on a single bank per request
+- Multi-bank queries are client responsibility
+- Disposition traits only affect reflect, not recall
+
+### Python Style
+- Python 3.11+, type hints required
+- Async throughout (asyncpg, async FastAPI)
+- Pydantic models for request/response
+- Ruff for linting (line-length 120)
+
+### TypeScript Style
+- Next.js App Router for control plane
+- Tailwind CSS with shadcn/ui components
+
+## Environment Setup
+
+```bash
+cp .env.example .env
+# Edit .env with LLM API key
+
+# Python deps
+uv sync --directory hindsight-api/
+
+# Node deps (workspace)
+npm install
+```
+
+Required env vars:
+- `HINDSIGHT_API_LLM_PROVIDER`: openai, anthropic, gemini, groq, ollama, lmstudio
+- `HINDSIGHT_API_LLM_API_KEY`: Your API key
+- `HINDSIGHT_API_LLM_MODEL`: Model name (e.g., o3-mini, claude-sonnet-4-20250514)

README.md

@@ -2,14 +2,13 @@
 
 ![Hindsight Banner](./hindsight-docs/static/img/banner.svg)
 
-[Documentation](https://vectorize-io.github.io/hindsight) • [Paper](#coming-soon) • [Examples](https://github.com/vectorize-io/hindsight-cookbook)
+[Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://vectorize.io/hindsight/cloud)
 
 [![CI](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml)
-[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![PyPI - hindsight-api](https://img.shields.io/pypi/v/hindsight-api?label=hindsight-api)](https://pypi.org/project/hindsight-api/)
-[![PyPI - hindsight-client](https://img.shields.io/pypi/v/hindsight-client?label=hindsight-client)](https://pypi.org/project/hindsight-client/)
-[![npm - @vectorize-io/hindsight-client](https://img.shields.io/npm/v/@vectorize-io/hindsight-client)](https://www.npmjs.com/package/@vectorize-io/hindsight-client)
 [![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+![PyPI - Downloads](https://img.shields.io/pypi/dm/hindsight-api?label=PyPI)
+![NPM Downloads](https://img.shields.io/npm/dm/%40vectorize-io%2Fhindsight-client?logoColor=orange&label=NPM&color=blue&link=https%3A%2F%2Fwww.npmjs.com%2Fpackage%2F%40vectorize-io%2Fhindsight-client)
 
 
 </div>
@@ -18,7 +17,7 @@
 
 ## What is Hindsight?
 
-Hindsight™ is an agent memory system built to create smarter agents that learn over time. It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph.
+Hindsight™ is an agent memory system built to create smarter agents that learn over time. It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.
 
 Hindsight addresses common challenges that have frustrated AI engineers building agents to automate tasks and assist users with conversational interfaces. Many of these challenges stem directly from a lack of memory.
 
@@ -26,27 +25,48 @@ Hindsight addresses common challenges that have frustrated AI engineers building
 - **Hallucinations:** Long term memory can be seeded with external knowledge to ground agent behavior in reliable sources to augment training data.
 - **Cognitive Overload:** As workflows get complex, retrievals, tool calls, user messages and agent responses can grow to fill the context window leading to context rot. Short term memory optimization allows agents to reduce tokens and focus context by removing irrelevant details.
 
-## How Hindsight Works
+## How is Hindsight Different From Other Memory Systems?
 
 ![Overview](./hindsight-docs/static/img/hindsight-overview.webp)
 
-Hindsight organizes memory into four networks to mimic the way human memory works:
+Most agent memory implementation rely on basic vector search or sometimes use a knowledge graph. Hindsight uses biomimetic data structures to organize agent memories in a way that is more like how human memory works:
 
 - **World:** Facts about the world ("The stove gets hot")
 - **Experiences:** Agent's own experiences ("I touched the stove and it really hurt")
 - **Opinion:** Beliefs with confidence scores ("I shouldn't touch the stove again" - .99 confidence)
 - **Observation:** Complex mental models derived by reflecting on facts and experiences ("Curling irons, ovens, and fire are also hot. I shouldn't touch those either.")
 
+Memories in Hindsight are stored in banks (i.e. memory banks). When memories are added to Hindsight, they are pushed into either the world facts or experiences memory pathway. They are then represented as a combination of entities, relationships, and time series with sparse/dense vector representations to aid in later recall.
+
 Hindsight provides three simple methods to interact with the system:
 
 - **Retain:** Provide information to Hindsight that you want it to remember
 - **Recall:** Retrieve memories from Hindsight
 - **Reflect:** Reflect on memories and experiences to generate new observations and insights from existing memories.
 
-Memories in Hindsight are stored in banks (e.g. memory banks). When memories are retained, they are transformed to construct a series of search indexes, time series data, and entity/relationship graphs.
+### Agent Memory That Learns
+
+A key goal of Hindsight is to build agent memory that enables agents to learn and improve over time. This is the role of the `reflect` operation which provides the agent to form broader opinions and observations over time.
+
+For example, imagine a product support agent that is helping a user troubleshoot a problem. It uses a `search-documentation` tool it found on an MCP server. Later in the conversation, the agent discovers that the documentation returned from the tool wasn't for the product the user was asking about. The agent now has an experience in its memory bank. And just like humans, we want that agent to learn from its experience.
+
+As the agent gains more experiences, `reflect` allows the agent to form observations about what worked, what didn't, and what to do differently the next time it encounters a similar task.
 
 ---
 
+## Memory Performance & Accuracy
+
+Hindsight has achieved state-of-the-art performance on the LongMemEval benchmark, widely used to assess memory system performance across a variety of conversational
+AI scenarios. The current reported performance of Hindsight and other agent memory solutions as of December 2025 is shown here:
+
+![Overview](./hindsight-docs/static/img/hindsight-bench.jpg)
+
+The benchmark performance data for Hindsight and GPT-4o (full context) have been reproduced by research collaborators at the Virginia Tech [Sanghani Center for Artificial Intelligence and Data Analytics](https://sanghani.cs.vt.edu/) and The Washington Post. Other scores are self-reported by software vendors.
+
+A thorough examination of the techniques implemented in Hindsight and detailed breakdowns of benchmark performance are [available on arXiv](https://arxiv.org/abs/2512.12818). This research is currently being prepared for conference submission and the wider peer review process.
+
+The benchmark results from this research can be inspected in our [visual benchmark explorer](https://hindsight-benchmarks.vercel.app). As additional improvements are made to Hindsight, new benchmark data will be available for review using this same tool.
+
 ## Quick Start
 
 ### Docker (recommended)
@@ -61,6 +81,8 @@ docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
   ghcr.io/vectorize-io/hindsight:latest
 ```
 
+You can modify the LLM provider by setting `HINDSIGHT_API_LLM_PROVIDER`. Valid options are `openai`, `anthropic`, `gemini`, `groq`, `ollama`, and `lmstudio`. The documentation provides more details on [supported models](https://hindsight.vectorize.io/developer/models).
+
 API: http://localhost:8888  
 UI: http://localhost:9999
 
@@ -223,6 +245,10 @@ client.reflect(bank_id="my-bank", query="What should I know about Alice?")
 - [Slack](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
 - [GitHub Issues](https://github.com/vectorize-io/hindsight/issues)
 
+---
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=vectorize-io/hindsight&type=date&legend=top-left)](https://www.star-history.com/#vectorize-io/hindsight&type=date&legend=top-left)
 ---
 
 ## Contributing

docker/standalone/Dockerfile

@@ -2,16 +2,19 @@
 # Supports building API-only, Control Plane-only, or both
 #
 # Build args:
-#   INCLUDE_API=true/false  - Include API (default: true)
-#   INCLUDE_CP=true/false   - Include Control Plane (default: true)
+#   INCLUDE_API=true/false       - Include API (default: true)
+#   INCLUDE_CP=true/false        - Include Control Plane (default: true)
+#   PRELOAD_ML_MODELS=true/false - Pre-download ML models during build (default: true)
 #
 # Examples:
-#   docker build -t hindsight .                                    # Both (standalone)
-#   docker build -t hindsight-api --build-arg INCLUDE_CP=false .   # API only
-#   docker build -t hindsight-cp --build-arg INCLUDE_API=false .   # Control Plane only
+#   docker build -t hindsight .                                          # Both (standalone)
+#   docker build -t hindsight-api --build-arg INCLUDE_CP=false .         # API only
+#   docker build -t hindsight-cp --build-arg INCLUDE_API=false .         # Control Plane only
+#   docker build -t hindsight --build-arg PRELOAD_ML_MODELS=false .      # Skip ML model preload
 
 ARG INCLUDE_API=true
 ARG INCLUDE_CP=true
+ARG PRELOAD_ML_MODELS=true
 
 # =============================================================================
 # Stage: API Builder
@@ -60,8 +63,8 @@ WORKDIR /app
 COPY package.json package-lock.json ./
 COPY hindsight-clients/typescript/ ./hindsight-clients/typescript/
 
-# Install and build SDK using workspace
-RUN npm ci -w @vectorize-io/hindsight-client
+# Install and build SDK using workspace (--ignore-scripts skips git hooks setup)
+RUN npm ci --ignore-scripts -w @vectorize-io/hindsight-client
 RUN npm run build -w @vectorize-io/hindsight-client
 
 # =============================================================================
@@ -72,30 +75,48 @@ FROM node:20-slim AS cp-builder
 ARG INCLUDE_CP
 RUN if [ "$INCLUDE_CP" != "true" ]; then echo "Skipping CP build" && exit 0; fi
 
-WORKDIR /app
-
-# Copy built SDK
-COPY --from=sdk-builder /app/hindsight-clients/typescript /app/sdk
+# Create directory structure matching the monorepo layout
+# This is required because build:standalone script expects .next/standalone/memory-poc/hindsight-control-plane
+WORKDIR /app/memory-poc/hindsight-control-plane
 
 # Install Control Plane dependencies
 # Only copy package.json (not package-lock.json) to ensure npm installs
 # correct platform-specific native bindings for lightningcss/tailwindcss
 COPY hindsight-control-plane/package.json ./
+# Remove the file: dependency on SDK (we'll copy it directly later)
+RUN sed -i '/"@vectorize-io\/hindsight-client":/d' package.json
 RUN npm install
 
 # Copy Control Plane source (excluding node_modules via .dockerignore)
 COPY hindsight-control-plane/ ./
 # Remove package-lock.json to avoid conflicts with installed native bindings
-RUN rm -f package-lock.json
+# Also remove the file: dependency from package.json (restored by COPY above)
+RUN rm -f package-lock.json && sed -i '/"@vectorize-io\/hindsight-client":/d' package.json
 
-# Link SDK (temporary for build)
-RUN cd /app/sdk && npm link && cd /app && npm link @vectorize-io/hindsight-client
+# Copy built SDK directly into node_modules (more reliable than npm link in Docker)
+COPY --from=sdk-builder /app/hindsight-clients/typescript ./node_modules/@vectorize-io/hindsight-client
 
-# Build Control Plane
-RUN npm run build
+# Build Control Plane - run next build first, then custom standalone copy
+# (The build:standalone script expects a specific path structure that differs in Docker)
+RUN npm exec -- next build
 
-# Create public directory if it doesn't exist
-RUN mkdir -p public
+# Create standalone directory structure manually
+# Note: Must exclude node_modules from find to avoid wrong server.js from next/dist/experimental/testmode/
+# Note: Must explicitly copy .next since glob * doesn't match hidden directories
+RUN STANDALONE_ROOT=$(find .next/standalone -path '*/node_modules' -prune -o -name 'server.js' -print | head -1 | xargs dirname) && \
+    mkdir -p standalone && \
+    cp -r "$STANDALONE_ROOT"/* standalone/ && \
+    cp -r "$STANDALONE_ROOT"/.next standalone/.next && \
+    # Copy node_modules if separate from app dir (monorepo structure)
+    if [ -d ".next/standalone/node_modules" ] && [ "$STANDALONE_ROOT" != ".next/standalone" ]; then \
+      cp -r .next/standalone/node_modules standalone/node_modules; \
+    fi && \
+    cp -r .next/static standalone/.next/static && \
+    mkdir -p standalone/public && \
+    cp -r public/* standalone/public/ 2>/dev/null || true && \
+    # Verify required files exist
+    test -f standalone/server.js || (echo "ERROR: server.js missing!" && exit 1) && \
+    test -f standalone/.next/BUILD_ID || (echo "ERROR: BUILD_ID missing!" && exit 1)
 
 # =============================================================================
 # Stage: Final Image - API Only
@@ -104,14 +125,16 @@ FROM python:3.11-slim AS api-only
 
 WORKDIR /app
 
-# Install pg0 dependencies
+# Install pg0 dependencies (procps provides 'kill' command needed by pg0)
+# Note: libicu version varies by Debian version - try common versions in order
 RUN apt-get update && apt-get install -y \
     curl \
+    procps \
     libxml2 \
     libssl3 \
     libgssapi-krb5-2 \
     libossp-uuid16 \
-    && apt-get install -y libicu72 || apt-get install -y libicu74 || apt-get install -y libicu* \
+    && (apt-get install -y libicu72 2>/dev/null || apt-get install -y libicu74 2>/dev/null || apt-get install -y libicu76 2>/dev/null || true) \
     && rm -rf /var/lib/apt/lists/* \
     && pip install --no-cache-dir uv
 
@@ -139,14 +162,17 @@ ENV PG0_HOME=/home/hindsight/.pg0-cache
 
 ENV PG0_HOME=/home/hindsight/.pg0
 
-# Pre-download ML models to avoid runtime download
-RUN /app/api/.venv/bin/python -c "\
+# Pre-download ML models to avoid runtime download (conditional)
+ARG PRELOAD_ML_MODELS
+RUN if [ "$PRELOAD_ML_MODELS" = "true" ]; then \
+    /app/api/.venv/bin/python -c "\
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Models cached successfully')"
+print('Models cached successfully')"; \
+    else echo "Skipping ML model preload"; fi
 
 EXPOSE 8888
 
@@ -171,9 +197,9 @@ COPY --from=sdk-builder /app/hindsight-clients/typescript /app/sdk
 
 # Copy Control Plane standalone build
 WORKDIR /app/control-plane
-COPY --from=cp-builder /app/.next/standalone ./
-COPY --from=cp-builder /app/.next/static ./.next/static
-COPY --from=cp-builder /app/public ./public
+COPY --from=cp-builder /app/memory-poc/hindsight-control-plane/standalone ./
+COPY --from=cp-builder /app/memory-poc/hindsight-control-plane/.next/static ./.next/static
+COPY --from=cp-builder /app/memory-poc/hindsight-control-plane/public ./public
 
 WORKDIR /app
 
@@ -200,14 +226,16 @@ FROM python:3.11-slim AS standalone
 
 WORKDIR /app
 
-# Install Node.js, curl, uv, and pg0 dependencies
+# Install Node.js, curl, uv, and pg0 dependencies (procps provides 'kill' command needed by pg0)
+# Note: libicu version varies by Debian version - try common versions in order
 RUN apt-get update && apt-get install -y \
     curl \
+    procps \
     libxml2 \
     libssl3 \
     libgssapi-krb5-2 \
     libossp-uuid16 \
-    && apt-get install -y libicu72 || apt-get install -y libicu74 || apt-get install -y libicu* \
+    && (apt-get install -y libicu72 2>/dev/null || apt-get install -y libicu74 2>/dev/null || apt-get install -y libicu76 2>/dev/null || true) \
     && curl -fsSL https://deb.nodesource.com/setup_20.x | bash - \
     && apt-get install -y nodejs \
     && rm -rf /var/lib/apt/lists/* \
@@ -224,9 +252,9 @@ COPY --from=sdk-builder /app/hindsight-clients/typescript /app/sdk
 
 # Copy Control Plane standalone build
 WORKDIR /app/control-plane
-COPY --from=cp-builder /app/.next/standalone ./
-COPY --from=cp-builder /app/.next/static ./.next/static
-COPY --from=cp-builder /app/public ./public
+COPY --from=cp-builder /app/memory-poc/hindsight-control-plane/standalone ./
+COPY --from=cp-builder /app/memory-poc/hindsight-control-plane/.next/static ./.next/static
+COPY --from=cp-builder /app/memory-poc/hindsight-control-plane/public ./public
 
 WORKDIR /app
 
@@ -255,14 +283,17 @@ print('PostgreSQL pre-cached to PG0_HOME')" || echo "Pre-download skipped"
 
 ENV PG0_HOME=/home/hindsight/.pg0
 
-# Pre-download ML models to avoid runtime download
-RUN /app/api/.venv/bin/python -c "\
+# Pre-download ML models to avoid runtime download (conditional)
+ARG PRELOAD_ML_MODELS
+RUN if [ "$PRELOAD_ML_MODELS" = "true" ]; then \
+    /app/api/.venv/bin/python -c "\
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Models cached successfully')"
+print('Models cached successfully')"; \
+    else echo "Skipping ML model preload"; fi
 
 EXPOSE 8888 9999
 

docker/standalone/start-all.sh

@@ -23,7 +23,8 @@ PIDS=()
 # Start API if enabled
 if [ "$ENABLE_API" = "true" ]; then
     cd /app/api
-    hindsight-api 2>&1 | sed -u 's/^/[api] /' &
+    # Run API directly - Python's PYTHONUNBUFFERED=1 handles output buffering
+    hindsight-api &
     API_PID=$!
     PIDS+=($API_PID)
 
@@ -42,7 +43,7 @@ fi
 if [ "$ENABLE_CP" = "true" ]; then
     echo "🎛️  Starting Control Plane..."
     cd /app/control-plane
-    PORT=9999 node server.js 2>&1 | grep -v -E "^[[:space:]]*(▲|✓|-|$)" | sed -u 's/^/[control-plane] /' &
+    PORT=9999 node server.js &
     CP_PID=$!
     PIDS+=($CP_PID)
 else

helm/hindsight/Chart.yaml

@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.1.5
-appVersion: "0.1.5"
+version: 0.1.16
+appVersion: "0.1.16"
 keywords:
   - ai
   - memory

hindsight-api/README.md

@@ -1 +1,137 @@
-# Memory 
+# Hindsight API
+
+**Memory System for AI Agents** — Temporal + Semantic + Entity Memory Architecture using PostgreSQL with pgvector.
+
+Hindsight gives AI agents persistent memory that works like human memory: it stores facts, tracks entities and relationships, handles temporal reasoning ("what happened last spring?"), and forms opinions based on configurable disposition traits.
+
+## Installation
+
+```bash
+pip install hindsight-api
+```
+
+## Quick Start
+
+### Run the Server
+
+```bash
+# Set your LLM provider
+export HINDSIGHT_API_LLM_PROVIDER=openai
+export HINDSIGHT_API_LLM_API_KEY=sk-xxxxxxxxxxxx
+
+# Start the server (uses embedded PostgreSQL by default)
+hindsight-api
+```
+
+The server starts at http://localhost:8888 with:
+- REST API for memory operations
+- MCP server at `/mcp` for tool-use integration
+
+### Use the Python API
+
+```python
+from hindsight_api import MemoryEngine
+
+# Create and initialize the memory engine
+memory = MemoryEngine()
+await memory.initialize()
+
+# Create a memory bank for your agent
+bank = await memory.create_memory_bank(
+    name="my-assistant",
+    background="A helpful coding assistant"
+)
+
+# Store a memory
+await memory.retain(
+    memory_bank_id=bank.id,
+    content="The user prefers Python for data science projects"
+)
+
+# Recall memories
+results = await memory.recall(
+    memory_bank_id=bank.id,
+    query="What programming language does the user prefer?"
+)
+
+# Reflect with reasoning
+response = await memory.reflect(
+    memory_bank_id=bank.id,
+    query="Should I recommend Python or R for this ML project?"
+)
+```
+
+## CLI Options
+
+```bash
+hindsight-api --help
+
+# Common options
+hindsight-api --port 9000          # Custom port (default: 8888)
+hindsight-api --host 127.0.0.1     # Bind to localhost only
+hindsight-api --workers 4          # Multiple worker processes
+hindsight-api --log-level debug    # Verbose logging
+```
+
+## Configuration
+
+Configure via environment variables:
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `HINDSIGHT_API_DATABASE_URL` | PostgreSQL connection string | `pg0` (embedded) |
+| `HINDSIGHT_API_LLM_PROVIDER` | `openai`, `anthropic`, `gemini`, `groq`, `ollama`, `lmstudio` | `openai` |
+| `HINDSIGHT_API_LLM_API_KEY` | API key for LLM provider | - |
+| `HINDSIGHT_API_LLM_MODEL` | Model name | `gpt-4o-mini` |
+| `HINDSIGHT_API_HOST` | Server bind address | `0.0.0.0` |
+| `HINDSIGHT_API_PORT` | Server port | `8888` |
+
+### Example with External PostgreSQL
+
+```bash
+export HINDSIGHT_API_DATABASE_URL=postgresql://user:pass@localhost:5432/hindsight
+export HINDSIGHT_API_LLM_PROVIDER=groq
+export HINDSIGHT_API_LLM_API_KEY=gsk_xxxxxxxxxxxx
+
+hindsight-api
+```
+
+## Docker
+
+```bash
+docker run --rm -it -p 8888:8888 \
+  -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
+  -v $HOME/.hindsight-docker:/home/hindsight/.pg0 \
+  ghcr.io/vectorize-io/hindsight:latest
+```
+
+## MCP Server
+
+For local MCP integration without running the full API server:
+
+```bash
+hindsight-local-mcp
+```
+
+This runs a stdio-based MCP server that can be used directly with MCP-compatible clients.
+
+## Key Features
+
+- **Multi-Strategy Retrieval (TEMPR)** — Semantic, keyword, graph, and temporal search combined with RRF fusion
+- **Entity Graph** — Automatic entity extraction and relationship tracking
+- **Temporal Reasoning** — Native support for time-based queries
+- **Disposition Traits** — Configurable skepticism, literalism, and empathy influence opinion formation
+- **Three Memory Types** — World facts, bank actions, and formed opinions with confidence scores
+
+## Documentation
+
+Full documentation: [https://hindsight.vectorize.io](https://hindsight.vectorize.io)
+
+- [Installation Guide](https://hindsight.vectorize.io/developer/installation)
+- [Configuration Reference](https://hindsight.vectorize.io/developer/configuration)
+- [API Reference](https://hindsight.vectorize.io/api-reference)
+- [Python SDK](https://hindsight.vectorize.io/sdks/python)
+
+## License
+
+Apache 2.0

hindsight-api/hindsight_api/__init__.py

@@ -3,26 +3,29 @@ Memory System for AI Agents.
 
 Temporal + Semantic Memory Architecture using PostgreSQL with pgvector.
 """
+
+from .config import HindsightConfig, get_config
+from .engine.cross_encoder import CrossEncoderModel, LocalSTCrossEncoder, RemoteTEICrossEncoder
+from .engine.embeddings import Embeddings, LocalSTEmbeddings, RemoteTEIEmbeddings
+from .engine.llm_wrapper import LLMConfig
 from .engine.memory_engine import MemoryEngine
 from .engine.search.trace import (
-    SearchTrace,
-    QueryInfo,
     EntryPoint,
-    NodeVisit,
-    WeightComponents,
     LinkInfo,
+    NodeVisit,
     PruningDecision,
-    SearchSummary,
+    QueryInfo,
     SearchPhaseMetrics,
+    SearchSummary,
+    SearchTrace,
+    WeightComponents,
 )
 from .engine.search.tracer import SearchTracer
-from .engine.embeddings import Embeddings, LocalSTEmbeddings, RemoteTEIEmbeddings
-from .engine.cross_encoder import CrossEncoderModel, LocalSTCrossEncoder, RemoteTEICrossEncoder
-from .engine.llm_wrapper import LLMConfig
-from .config import HindsightConfig, get_config
+from .models import RequestContext
 
 __all__ = [
     "MemoryEngine",
+    "RequestContext",
     "HindsightConfig",
     "get_config",
     "SearchTrace",

hindsight-api/hindsight_api/alembic/env.py

@@ -2,20 +2,19 @@
 Alembic environment configuration for SQLAlchemy with pgvector.
 Uses synchronous psycopg2 driver for migrations to avoid pgbouncer issues.
 """
+
 import logging
 import os
-import sys
 from pathlib import Path
 
-from sqlalchemy import pool, engine_from_config
-from sqlalchemy.engine import Connection
-
 from alembic import context
 from dotenv import load_dotenv
+from sqlalchemy import engine_from_config, pool
 
 # Import your models here
 from hindsight_api.models import Base
 
+
 # Load environment variables based on HINDSIGHT_API_DATABASE_URL env var or default to local
 def load_env():
     """Load environment variables from .env"""
@@ -30,6 +29,7 @@ def load_env():
     if env_file.exists():
         load_dotenv(env_file)
 
+
 load_env()
 
 # this is the Alembic Config object, which provides
@@ -109,6 +109,9 @@ def run_migrations_online() -> None:
 
     get_database_url()  # Process and set the database URL in config
 
+    # Check if we're targeting a specific schema (for multi-tenant isolation)
+    target_schema = config.get_main_option("target_schema")
+
     connectable = engine_from_config(
         config.get_section(config.config_ini_section, {}),
         prefix="sqlalchemy.",
@@ -121,17 +124,34 @@ def run_migrations_online() -> None:
     def set_read_write_mode(dbapi_connection, connection_record):
         cursor = dbapi_connection.cursor()
         cursor.execute("SET SESSION CHARACTERISTICS AS TRANSACTION READ WRITE")
+        # If targeting a specific schema, set search_path
+        # Include public in search_path for access to shared extensions (pgvector)
+        if target_schema:
+            cursor.execute(f'CREATE SCHEMA IF NOT EXISTS "{target_schema}"')
+            cursor.execute(f'SET search_path TO "{target_schema}", public')
         cursor.close()
 
     with connectable.connect() as connection:
         # Also explicitly set read-write mode on this connection
         connection.execute(text("SET SESSION CHARACTERISTICS AS TRANSACTION READ WRITE"))
+
+        # If targeting a specific schema, set search_path
+        # Include public in search_path for access to shared extensions (pgvector)
+        if target_schema:
+            connection.execute(text(f'CREATE SCHEMA IF NOT EXISTS "{target_schema}"'))
+            connection.execute(text(f'SET search_path TO "{target_schema}", public'))
+
         connection.commit()  # Commit the SET command
 
-        context.configure(
-            connection=connection,
-            target_metadata=target_metadata
-        )
+        # Configure context with version_table_schema if using a specific schema
+        context_opts = {
+            "connection": connection,
+            "target_metadata": target_metadata,
+        }
+        if target_schema:
+            context_opts["version_table_schema"] = target_schema
+
+        context.configure(**context_opts)
 
         with context.begin_transaction():
             context.run_migrations()

hindsight-api/hindsight_api/alembic/versions/5a366d414dce_initial_schema.py

@@ -5,120 +5,150 @@ Revises:
 Create Date: 2025-11-27 11:54:19.228030
 
 """
-from typing import Sequence, Union
 
-from alembic import op
+from collections.abc import Sequence
+
 import sqlalchemy as sa
-from sqlalchemy.dialects import postgresql
+from alembic import op
 from pgvector.sqlalchemy import Vector
-
+from sqlalchemy.dialects import postgresql
 
 # revision identifiers, used by Alembic.
-revision: str = '5a366d414dce'
-down_revision: Union[str, Sequence[str], None] = None
-branch_labels: Union[str, Sequence[str], None] = None
-depends_on: Union[str, Sequence[str], None] = None
+revision: str = "5a366d414dce"
+down_revision: str | Sequence[str] | None = None
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
 
 
 def upgrade() -> None:
     """Upgrade schema - create all tables from scratch."""
 
     # Enable required extensions
-    op.execute('CREATE EXTENSION IF NOT EXISTS vector')
+    op.execute("CREATE EXTENSION IF NOT EXISTS vector")
 
     # Create banks table
     op.create_table(
-        'banks',
-        sa.Column('bank_id', sa.Text(), nullable=False),
-        sa.Column('name', sa.Text(), nullable=True),
-        sa.Column('personality', postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False),
-        sa.Column('background', sa.Text(), nullable=True),
-        sa.Column('created_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.Column('updated_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.PrimaryKeyConstraint('bank_id', name=op.f('pk_banks'))
+        "banks",
+        sa.Column("bank_id", sa.Text(), nullable=False),
+        sa.Column("name", sa.Text(), nullable=True),
+        sa.Column(
+            "personality",
+            postgresql.JSONB(astext_type=sa.Text()),
+            server_default=sa.text("'{}'::jsonb"),
+            nullable=False,
+        ),
+        sa.Column("background", sa.Text(), nullable=True),
+        sa.Column("created_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.PrimaryKeyConstraint("bank_id", name=op.f("pk_banks")),
     )
 
     # Create documents table
     op.create_table(
-        'documents',
-        sa.Column('id', sa.Text(), nullable=False),
-        sa.Column('bank_id', sa.Text(), nullable=False),
-        sa.Column('original_text', sa.Text(), nullable=True),
-        sa.Column('content_hash', sa.Text(), nullable=True),
-        sa.Column('metadata', postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False),
-        sa.Column('created_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.Column('updated_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.PrimaryKeyConstraint('id', 'bank_id', name=op.f('pk_documents'))
+        "documents",
+        sa.Column("id", sa.Text(), nullable=False),
+        sa.Column("bank_id", sa.Text(), nullable=False),
+        sa.Column("original_text", sa.Text(), nullable=True),
+        sa.Column("content_hash", sa.Text(), nullable=True),
+        sa.Column(
+            "metadata", postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False
+        ),
+        sa.Column("created_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.PrimaryKeyConstraint("id", "bank_id", name=op.f("pk_documents")),
     )
-    op.create_index('idx_documents_bank_id', 'documents', ['bank_id'])
-    op.create_index('idx_documents_content_hash', 'documents', ['content_hash'])
+    op.create_index("idx_documents_bank_id", "documents", ["bank_id"])
+    op.create_index("idx_documents_content_hash", "documents", ["content_hash"])
 
     # Create async_operations table
     op.create_table(
-        'async_operations',
-        sa.Column('operation_id', postgresql.UUID(as_uuid=True), server_default=sa.text('gen_random_uuid()'), nullable=False),
-        sa.Column('bank_id', sa.Text(), nullable=False),
-        sa.Column('operation_type', sa.Text(), nullable=False),
-        sa.Column('status', sa.Text(), server_default='pending', nullable=False),
-        sa.Column('created_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.Column('updated_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.Column('completed_at', postgresql.TIMESTAMP(timezone=True), nullable=True),
-        sa.Column('error_message', sa.Text(), nullable=True),
-        sa.Column('result_metadata', postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False),
-        sa.PrimaryKeyConstraint('operation_id', name=op.f('pk_async_operations')),
-        sa.CheckConstraint("status IN ('pending', 'processing', 'completed', 'failed')", name='async_operations_status_check')
+        "async_operations",
+        sa.Column(
+            "operation_id", postgresql.UUID(as_uuid=True), server_default=sa.text("gen_random_uuid()"), nullable=False
+        ),
+        sa.Column("bank_id", sa.Text(), nullable=False),
+        sa.Column("operation_type", sa.Text(), nullable=False),
+        sa.Column("status", sa.Text(), server_default="pending", nullable=False),
+        sa.Column("created_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("completed_at", postgresql.TIMESTAMP(timezone=True), nullable=True),
+        sa.Column("error_message", sa.Text(), nullable=True),
+        sa.Column(
+            "result_metadata",
+            postgresql.JSONB(astext_type=sa.Text()),
+            server_default=sa.text("'{}'::jsonb"),
+            nullable=False,
+        ),
+        sa.PrimaryKeyConstraint("operation_id", name=op.f("pk_async_operations")),
+        sa.CheckConstraint(
+            "status IN ('pending', 'processing', 'completed', 'failed')", name="async_operations_status_check"
+        ),
     )
-    op.create_index('idx_async_operations_bank_id', 'async_operations', ['bank_id'])
-    op.create_index('idx_async_operations_status', 'async_operations', ['status'])
-    op.create_index('idx_async_operations_bank_status', 'async_operations', ['bank_id', 'status'])
+    op.create_index("idx_async_operations_bank_id", "async_operations", ["bank_id"])
+    op.create_index("idx_async_operations_status", "async_operations", ["status"])
+    op.create_index("idx_async_operations_bank_status", "async_operations", ["bank_id", "status"])
 
     # Create entities table
     op.create_table(
-        'entities',
-        sa.Column('id', postgresql.UUID(as_uuid=True), server_default=sa.text('gen_random_uuid()'), nullable=False),
-        sa.Column('canonical_name', sa.Text(), nullable=False),
-        sa.Column('bank_id', sa.Text(), nullable=False),
-        sa.Column('metadata', postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False),
-        sa.Column('first_seen', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.Column('last_seen', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.Column('mention_count', sa.Integer(), server_default='1', nullable=False),
-        sa.PrimaryKeyConstraint('id', name=op.f('pk_entities'))
+        "entities",
+        sa.Column("id", postgresql.UUID(as_uuid=True), server_default=sa.text("gen_random_uuid()"), nullable=False),
+        sa.Column("canonical_name", sa.Text(), nullable=False),
+        sa.Column("bank_id", sa.Text(), nullable=False),
+        sa.Column(
+            "metadata", postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False
+        ),
+        sa.Column("first_seen", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("last_seen", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("mention_count", sa.Integer(), server_default="1", nullable=False),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_entities")),
     )
-    op.create_index('idx_entities_bank_id', 'entities', ['bank_id'])
-    op.create_index('idx_entities_canonical_name', 'entities', ['canonical_name'])
-    op.create_index('idx_entities_bank_name', 'entities', ['bank_id', 'canonical_name'])
+    op.create_index("idx_entities_bank_id", "entities", ["bank_id"])
+    op.create_index("idx_entities_canonical_name", "entities", ["canonical_name"])
+    op.create_index("idx_entities_bank_name", "entities", ["bank_id", "canonical_name"])
     # Create unique index on (bank_id, LOWER(canonical_name)) for entity resolution
-    op.execute('CREATE UNIQUE INDEX idx_entities_bank_lower_name ON entities (bank_id, LOWER(canonical_name))')
+    op.execute("CREATE UNIQUE INDEX idx_entities_bank_lower_name ON entities (bank_id, LOWER(canonical_name))")
 
     # Create memory_units table
     op.create_table(
-        'memory_units',
-        sa.Column('id', postgresql.UUID(as_uuid=True), server_default=sa.text('gen_random_uuid()'), nullable=False),
-        sa.Column('bank_id', sa.Text(), nullable=False),
-        sa.Column('document_id', sa.Text(), nullable=True),
-        sa.Column('text', sa.Text(), nullable=False),
-        sa.Column('embedding', Vector(384), nullable=True),
-        sa.Column('context', sa.Text(), nullable=True),
-        sa.Column('event_date', postgresql.TIMESTAMP(timezone=True), nullable=False),
-        sa.Column('occurred_start', postgresql.TIMESTAMP(timezone=True), nullable=True),
-        sa.Column('occurred_end', postgresql.TIMESTAMP(timezone=True), nullable=True),
-        sa.Column('mentioned_at', postgresql.TIMESTAMP(timezone=True), nullable=True),
-        sa.Column('fact_type', sa.Text(), server_default='world', nullable=False),
-        sa.Column('confidence_score', sa.Float(), nullable=True),
-        sa.Column('access_count', sa.Integer(), server_default='0', nullable=False),
-        sa.Column('metadata', postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False),
-        sa.Column('created_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.Column('updated_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.ForeignKeyConstraint(['document_id', 'bank_id'], ['documents.id', 'documents.bank_id'], name='memory_units_document_fkey', ondelete='CASCADE'),
-        sa.PrimaryKeyConstraint('id', name=op.f('pk_memory_units')),
-        sa.CheckConstraint("fact_type IN ('world', 'bank', 'opinion', 'observation')", name='memory_units_fact_type_check'),
-        sa.CheckConstraint("confidence_score IS NULL OR (confidence_score >= 0.0 AND confidence_score <= 1.0)", name='memory_units_confidence_range_check'),
+        "memory_units",
+        sa.Column("id", postgresql.UUID(as_uuid=True), server_default=sa.text("gen_random_uuid()"), nullable=False),
+        sa.Column("bank_id", sa.Text(), nullable=False),
+        sa.Column("document_id", sa.Text(), nullable=True),
+        sa.Column("text", sa.Text(), nullable=False),
+        sa.Column("embedding", Vector(384), nullable=True),
+        sa.Column("context", sa.Text(), nullable=True),
+        sa.Column("event_date", postgresql.TIMESTAMP(timezone=True), nullable=False),
+        sa.Column("occurred_start", postgresql.TIMESTAMP(timezone=True), nullable=True),
+        sa.Column("occurred_end", postgresql.TIMESTAMP(timezone=True), nullable=True),
+        sa.Column("mentioned_at", postgresql.TIMESTAMP(timezone=True), nullable=True),
+        sa.Column("fact_type", sa.Text(), server_default="world", nullable=False),
+        sa.Column("confidence_score", sa.Float(), nullable=True),
+        sa.Column("access_count", sa.Integer(), server_default="0", nullable=False),
+        sa.Column(
+            "metadata", postgresql.JSONB(astext_type=sa.Text()), server_default=sa.text("'{}'::jsonb"), nullable=False
+        ),
+        sa.Column("created_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["document_id", "bank_id"],
+            ["documents.id", "documents.bank_id"],
+            name="memory_units_document_fkey",
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_memory_units")),
+        sa.CheckConstraint(
+            "fact_type IN ('world', 'bank', 'opinion', 'observation')", name="memory_units_fact_type_check"
+        ),
+        sa.CheckConstraint(
+            "confidence_score IS NULL OR (confidence_score >= 0.0 AND confidence_score <= 1.0)",
+            name="memory_units_confidence_range_check",
+        ),
         sa.CheckConstraint(
             "(fact_type = 'opinion' AND confidence_score IS NOT NULL) OR "
             "(fact_type = 'observation') OR "
             "(fact_type NOT IN ('opinion', 'observation') AND confidence_score IS NULL)",
-            name='confidence_score_fact_type_check'
-        )
+            name="confidence_score_fact_type_check",
+        ),
     )
 
     # Add search_vector column for full-text search
@@ -128,18 +158,41 @@ def upgrade() -> None:
         GENERATED ALWAYS AS (to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))) STORED
     """)
 
-    op.create_index('idx_memory_units_bank_id', 'memory_units', ['bank_id'])
-    op.create_index('idx_memory_units_document_id', 'memory_units', ['document_id'])
-    op.create_index('idx_memory_units_event_date', 'memory_units', [sa.text('event_date DESC')])
-    op.create_index('idx_memory_units_bank_date', 'memory_units', ['bank_id', sa.text('event_date DESC')])
-    op.create_index('idx_memory_units_access_count', 'memory_units', [sa.text('access_count DESC')])
-    op.create_index('idx_memory_units_fact_type', 'memory_units', ['fact_type'])
-    op.create_index('idx_memory_units_bank_fact_type', 'memory_units', ['bank_id', 'fact_type'])
-    op.create_index('idx_memory_units_bank_type_date', 'memory_units', ['bank_id', 'fact_type', sa.text('event_date DESC')])
-    op.create_index('idx_memory_units_opinion_confidence', 'memory_units', ['bank_id', sa.text('confidence_score DESC')], postgresql_where=sa.text("fact_type = 'opinion'"))
-    op.create_index('idx_memory_units_opinion_date', 'memory_units', ['bank_id', sa.text('event_date DESC')], postgresql_where=sa.text("fact_type = 'opinion'"))
-    op.create_index('idx_memory_units_observation_date', 'memory_units', ['bank_id', sa.text('event_date DESC')], postgresql_where=sa.text("fact_type = 'observation'"))
-    op.create_index('idx_memory_units_embedding', 'memory_units', ['embedding'], postgresql_using='hnsw', postgresql_ops={'embedding': 'vector_cosine_ops'})
+    op.create_index("idx_memory_units_bank_id", "memory_units", ["bank_id"])
+    op.create_index("idx_memory_units_document_id", "memory_units", ["document_id"])
+    op.create_index("idx_memory_units_event_date", "memory_units", [sa.text("event_date DESC")])
+    op.create_index("idx_memory_units_bank_date", "memory_units", ["bank_id", sa.text("event_date DESC")])
+    op.create_index("idx_memory_units_access_count", "memory_units", [sa.text("access_count DESC")])
+    op.create_index("idx_memory_units_fact_type", "memory_units", ["fact_type"])
+    op.create_index("idx_memory_units_bank_fact_type", "memory_units", ["bank_id", "fact_type"])
+    op.create_index(
+        "idx_memory_units_bank_type_date", "memory_units", ["bank_id", "fact_type", sa.text("event_date DESC")]
+    )
+    op.create_index(
+        "idx_memory_units_opinion_confidence",
+        "memory_units",
+        ["bank_id", sa.text("confidence_score DESC")],
+        postgresql_where=sa.text("fact_type = 'opinion'"),
+    )
+    op.create_index(
+        "idx_memory_units_opinion_date",
+        "memory_units",
+        ["bank_id", sa.text("event_date DESC")],
+        postgresql_where=sa.text("fact_type = 'opinion'"),
+    )
+    op.create_index(
+        "idx_memory_units_observation_date",
+        "memory_units",
+        ["bank_id", sa.text("event_date DESC")],
+        postgresql_where=sa.text("fact_type = 'observation'"),
+    )
+    op.create_index(
+        "idx_memory_units_embedding",
+        "memory_units",
+        ["embedding"],
+        postgresql_using="hnsw",
+        postgresql_ops={"embedding": "vector_cosine_ops"},
+    )
 
     # Create BM25 full-text search index on search_vector
     op.execute("""
@@ -158,116 +211,149 @@ def upgrade() -> None:
         FROM memory_units
     """)
 
-    op.create_index('idx_memory_units_bm25_bank', 'memory_units_bm25', ['bank_id'])
-    op.create_index('idx_memory_units_bm25_text_vector', 'memory_units_bm25', ['text_vector'], postgresql_using='gin')
+    op.create_index("idx_memory_units_bm25_bank", "memory_units_bm25", ["bank_id"])
+    op.create_index("idx_memory_units_bm25_text_vector", "memory_units_bm25", ["text_vector"], postgresql_using="gin")
 
     # Create entity_cooccurrences table
     op.create_table(
-        'entity_cooccurrences',
-        sa.Column('entity_id_1', postgresql.UUID(as_uuid=True), nullable=False),
-        sa.Column('entity_id_2', postgresql.UUID(as_uuid=True), nullable=False),
-        sa.Column('cooccurrence_count', sa.Integer(), server_default='1', nullable=False),
-        sa.Column('last_cooccurred', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.ForeignKeyConstraint(['entity_id_1'], ['entities.id'], name=op.f('fk_entity_cooccurrences_entity_id_1_entities'), ondelete='CASCADE'),
-        sa.ForeignKeyConstraint(['entity_id_2'], ['entities.id'], name=op.f('fk_entity_cooccurrences_entity_id_2_entities'), ondelete='CASCADE'),
-        sa.PrimaryKeyConstraint('entity_id_1', 'entity_id_2', name=op.f('pk_entity_cooccurrences')),
-        sa.CheckConstraint('entity_id_1 < entity_id_2', name='entity_cooccurrence_order_check')
+        "entity_cooccurrences",
+        sa.Column("entity_id_1", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.Column("entity_id_2", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.Column("cooccurrence_count", sa.Integer(), server_default="1", nullable=False),
+        sa.Column(
+            "last_cooccurred", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False
+        ),
+        sa.ForeignKeyConstraint(
+            ["entity_id_1"],
+            ["entities.id"],
+            name=op.f("fk_entity_cooccurrences_entity_id_1_entities"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["entity_id_2"],
+            ["entities.id"],
+            name=op.f("fk_entity_cooccurrences_entity_id_2_entities"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("entity_id_1", "entity_id_2", name=op.f("pk_entity_cooccurrences")),
+        sa.CheckConstraint("entity_id_1 < entity_id_2", name="entity_cooccurrence_order_check"),
     )
-    op.create_index('idx_entity_cooccurrences_entity1', 'entity_cooccurrences', ['entity_id_1'])
-    op.create_index('idx_entity_cooccurrences_entity2', 'entity_cooccurrences', ['entity_id_2'])
-    op.create_index('idx_entity_cooccurrences_count', 'entity_cooccurrences', [sa.text('cooccurrence_count DESC')])
+    op.create_index("idx_entity_cooccurrences_entity1", "entity_cooccurrences", ["entity_id_1"])
+    op.create_index("idx_entity_cooccurrences_entity2", "entity_cooccurrences", ["entity_id_2"])
+    op.create_index("idx_entity_cooccurrences_count", "entity_cooccurrences", [sa.text("cooccurrence_count DESC")])
 
     # Create memory_links table
     op.create_table(
-        'memory_links',
-        sa.Column('from_unit_id', postgresql.UUID(as_uuid=True), nullable=False),
-        sa.Column('to_unit_id', postgresql.UUID(as_uuid=True), nullable=False),
-        sa.Column('link_type', sa.Text(), nullable=False),
-        sa.Column('entity_id', postgresql.UUID(as_uuid=True), nullable=True),
-        sa.Column('weight', sa.Float(), server_default='1.0', nullable=False),
-        sa.Column('created_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.ForeignKeyConstraint(['entity_id'], ['entities.id'], name=op.f('fk_memory_links_entity_id_entities'), ondelete='CASCADE'),
-        sa.ForeignKeyConstraint(['from_unit_id'], ['memory_units.id'], name=op.f('fk_memory_links_from_unit_id_memory_units'), ondelete='CASCADE'),
-        sa.ForeignKeyConstraint(['to_unit_id'], ['memory_units.id'], name=op.f('fk_memory_links_to_unit_id_memory_units'), ondelete='CASCADE'),
-        sa.CheckConstraint("link_type IN ('temporal', 'semantic', 'entity', 'causes', 'caused_by', 'enables', 'prevents')", name='memory_links_link_type_check'),
-        sa.CheckConstraint('weight >= 0.0 AND weight <= 1.0', name='memory_links_weight_check')
+        "memory_links",
+        sa.Column("from_unit_id", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.Column("to_unit_id", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.Column("link_type", sa.Text(), nullable=False),
+        sa.Column("entity_id", postgresql.UUID(as_uuid=True), nullable=True),
+        sa.Column("weight", sa.Float(), server_default="1.0", nullable=False),
+        sa.Column("created_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["entity_id"], ["entities.id"], name=op.f("fk_memory_links_entity_id_entities"), ondelete="CASCADE"
+        ),
+        sa.ForeignKeyConstraint(
+            ["from_unit_id"],
+            ["memory_units.id"],
+            name=op.f("fk_memory_links_from_unit_id_memory_units"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["to_unit_id"],
+            ["memory_units.id"],
+            name=op.f("fk_memory_links_to_unit_id_memory_units"),
+            ondelete="CASCADE",
+        ),
+        sa.CheckConstraint(
+            "link_type IN ('temporal', 'semantic', 'entity', 'causes', 'caused_by', 'enables', 'prevents')",
+            name="memory_links_link_type_check",
+        ),
+        sa.CheckConstraint("weight >= 0.0 AND weight <= 1.0", name="memory_links_weight_check"),
     )
     # Create unique constraint using COALESCE for nullable entity_id
-    op.execute("CREATE UNIQUE INDEX idx_memory_links_unique ON memory_links (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid))")
-    op.create_index('idx_memory_links_from_unit', 'memory_links', ['from_unit_id'])
-    op.create_index('idx_memory_links_to_unit', 'memory_links', ['to_unit_id'])
-    op.create_index('idx_memory_links_entity', 'memory_links', ['entity_id'])
-    op.create_index('idx_memory_links_link_type', 'memory_links', ['link_type'])
+    op.execute(
+        "CREATE UNIQUE INDEX idx_memory_links_unique ON memory_links (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid))"
+    )
+    op.create_index("idx_memory_links_from_unit", "memory_links", ["from_unit_id"])
+    op.create_index("idx_memory_links_to_unit", "memory_links", ["to_unit_id"])
+    op.create_index("idx_memory_links_entity", "memory_links", ["entity_id"])
+    op.create_index("idx_memory_links_link_type", "memory_links", ["link_type"])
 
     # Create unit_entities table
     op.create_table(
-        'unit_entities',
-        sa.Column('unit_id', postgresql.UUID(as_uuid=True), nullable=False),
-        sa.Column('entity_id', postgresql.UUID(as_uuid=True), nullable=False),
-        sa.ForeignKeyConstraint(['entity_id'], ['entities.id'], name=op.f('fk_unit_entities_entity_id_entities'), ondelete='CASCADE'),
-        sa.ForeignKeyConstraint(['unit_id'], ['memory_units.id'], name=op.f('fk_unit_entities_unit_id_memory_units'), ondelete='CASCADE'),
-        sa.PrimaryKeyConstraint('unit_id', 'entity_id', name=op.f('pk_unit_entities'))
+        "unit_entities",
+        sa.Column("unit_id", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.Column("entity_id", postgresql.UUID(as_uuid=True), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["entity_id"], ["entities.id"], name=op.f("fk_unit_entities_entity_id_entities"), ondelete="CASCADE"
+        ),
+        sa.ForeignKeyConstraint(
+            ["unit_id"], ["memory_units.id"], name=op.f("fk_unit_entities_unit_id_memory_units"), ondelete="CASCADE"
+        ),
+        sa.PrimaryKeyConstraint("unit_id", "entity_id", name=op.f("pk_unit_entities")),
     )
-    op.create_index('idx_unit_entities_unit', 'unit_entities', ['unit_id'])
-    op.create_index('idx_unit_entities_entity', 'unit_entities', ['entity_id'])
+    op.create_index("idx_unit_entities_unit", "unit_entities", ["unit_id"])
+    op.create_index("idx_unit_entities_entity", "unit_entities", ["entity_id"])
 
 
 def downgrade() -> None:
     """Downgrade schema - drop all tables."""
 
     # Drop tables in reverse dependency order
-    op.drop_index('idx_unit_entities_entity', table_name='unit_entities')
-    op.drop_index('idx_unit_entities_unit', table_name='unit_entities')
-    op.drop_table('unit_entities')
+    op.drop_index("idx_unit_entities_entity", table_name="unit_entities")
+    op.drop_index("idx_unit_entities_unit", table_name="unit_entities")
+    op.drop_table("unit_entities")
 
-    op.drop_index('idx_memory_links_link_type', table_name='memory_links')
-    op.drop_index('idx_memory_links_entity', table_name='memory_links')
-    op.drop_index('idx_memory_links_to_unit', table_name='memory_links')
-    op.drop_index('idx_memory_links_from_unit', table_name='memory_links')
-    op.execute('DROP INDEX IF EXISTS idx_memory_links_unique')
-    op.drop_table('memory_links')
+    op.drop_index("idx_memory_links_link_type", table_name="memory_links")
+    op.drop_index("idx_memory_links_entity", table_name="memory_links")
+    op.drop_index("idx_memory_links_to_unit", table_name="memory_links")
+    op.drop_index("idx_memory_links_from_unit", table_name="memory_links")
+    op.execute("DROP INDEX IF EXISTS idx_memory_links_unique")
+    op.drop_table("memory_links")
 
-    op.drop_index('idx_entity_cooccurrences_count', table_name='entity_cooccurrences')
-    op.drop_index('idx_entity_cooccurrences_entity2', table_name='entity_cooccurrences')
-    op.drop_index('idx_entity_cooccurrences_entity1', table_name='entity_cooccurrences')
-    op.drop_table('entity_cooccurrences')
+    op.drop_index("idx_entity_cooccurrences_count", table_name="entity_cooccurrences")
+    op.drop_index("idx_entity_cooccurrences_entity2", table_name="entity_cooccurrences")
+    op.drop_index("idx_entity_cooccurrences_entity1", table_name="entity_cooccurrences")
+    op.drop_table("entity_cooccurrences")
 
     # Drop BM25 materialized view and index
-    op.drop_index('idx_memory_units_bm25_text_vector', table_name='memory_units_bm25')
-    op.drop_index('idx_memory_units_bm25_bank', table_name='memory_units_bm25')
-    op.execute('DROP MATERIALIZED VIEW IF EXISTS memory_units_bm25')
+    op.drop_index("idx_memory_units_bm25_text_vector", table_name="memory_units_bm25")
+    op.drop_index("idx_memory_units_bm25_bank", table_name="memory_units_bm25")
+    op.execute("DROP MATERIALIZED VIEW IF EXISTS memory_units_bm25")
 
-    op.drop_index('idx_memory_units_embedding', table_name='memory_units')
-    op.drop_index('idx_memory_units_observation_date', table_name='memory_units')
-    op.drop_index('idx_memory_units_opinion_date', table_name='memory_units')
-    op.drop_index('idx_memory_units_opinion_confidence', table_name='memory_units')
-    op.drop_index('idx_memory_units_bank_type_date', table_name='memory_units')
-    op.drop_index('idx_memory_units_bank_fact_type', table_name='memory_units')
-    op.drop_index('idx_memory_units_fact_type', table_name='memory_units')
-    op.drop_index('idx_memory_units_access_count', table_name='memory_units')
-    op.drop_index('idx_memory_units_bank_date', table_name='memory_units')
-    op.drop_index('idx_memory_units_event_date', table_name='memory_units')
-    op.drop_index('idx_memory_units_document_id', table_name='memory_units')
-    op.drop_index('idx_memory_units_bank_id', table_name='memory_units')
-    op.execute('DROP INDEX IF EXISTS idx_memory_units_text_search')
-    op.drop_table('memory_units')
+    op.drop_index("idx_memory_units_embedding", table_name="memory_units")
+    op.drop_index("idx_memory_units_observation_date", table_name="memory_units")
+    op.drop_index("idx_memory_units_opinion_date", table_name="memory_units")
+    op.drop_index("idx_memory_units_opinion_confidence", table_name="memory_units")
+    op.drop_index("idx_memory_units_bank_type_date", table_name="memory_units")
+    op.drop_index("idx_memory_units_bank_fact_type", table_name="memory_units")
+    op.drop_index("idx_memory_units_fact_type", table_name="memory_units")
+    op.drop_index("idx_memory_units_access_count", table_name="memory_units")
+    op.drop_index("idx_memory_units_bank_date", table_name="memory_units")
+    op.drop_index("idx_memory_units_event_date", table_name="memory_units")
+    op.drop_index("idx_memory_units_document_id", table_name="memory_units")
+    op.drop_index("idx_memory_units_bank_id", table_name="memory_units")
+    op.execute("DROP INDEX IF EXISTS idx_memory_units_text_search")
+    op.drop_table("memory_units")
 
-    op.execute('DROP INDEX IF EXISTS idx_entities_bank_lower_name')
-    op.drop_index('idx_entities_bank_name', table_name='entities')
-    op.drop_index('idx_entities_canonical_name', table_name='entities')
-    op.drop_index('idx_entities_bank_id', table_name='entities')
-    op.drop_table('entities')
+    op.execute("DROP INDEX IF EXISTS idx_entities_bank_lower_name")
+    op.drop_index("idx_entities_bank_name", table_name="entities")
+    op.drop_index("idx_entities_canonical_name", table_name="entities")
+    op.drop_index("idx_entities_bank_id", table_name="entities")
+    op.drop_table("entities")
 
-    op.drop_index('idx_async_operations_bank_status', table_name='async_operations')
-    op.drop_index('idx_async_operations_status', table_name='async_operations')
-    op.drop_index('idx_async_operations_bank_id', table_name='async_operations')
-    op.drop_table('async_operations')
+    op.drop_index("idx_async_operations_bank_status", table_name="async_operations")
+    op.drop_index("idx_async_operations_status", table_name="async_operations")
+    op.drop_index("idx_async_operations_bank_id", table_name="async_operations")
+    op.drop_table("async_operations")
 
-    op.drop_index('idx_documents_content_hash', table_name='documents')
-    op.drop_index('idx_documents_bank_id', table_name='documents')
-    op.drop_table('documents')
+    op.drop_index("idx_documents_content_hash", table_name="documents")
+    op.drop_index("idx_documents_bank_id", table_name="documents")
+    op.drop_table("documents")
 
-    op.drop_table('banks')
+    op.drop_table("banks")
 
     # Drop extensions (optional - comment out if you want to keep them)
     # op.execute('DROP EXTENSION IF EXISTS vector')

hindsight-api/hindsight_api/alembic/versions/b7c4d8e9f1a2_add_chunks_table.py

@@ -5,18 +5,18 @@ Revises: 5a366d414dce
 Create Date: 2025-11-28 00:00:00.000000
 
 """
-from typing import Sequence, Union
 
-from alembic import op
+from collections.abc import Sequence
+
 import sqlalchemy as sa
+from alembic import op
 from sqlalchemy.dialects import postgresql
 
-
 # revision identifiers, used by Alembic.
-revision: str = 'b7c4d8e9f1a2'
-down_revision: Union[str, Sequence[str], None] = '5a366d414dce'
-branch_labels: Union[str, Sequence[str], None] = None
-depends_on: Union[str, Sequence[str], None] = None
+revision: str = "b7c4d8e9f1a2"
+down_revision: str | Sequence[str] | None = "5a366d414dce"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
 
 
 def upgrade() -> None:
@@ -24,47 +24,47 @@ def upgrade() -> None:
 
     # Create chunks table with single text PK (bank_id_document_id_chunk_index)
     op.create_table(
-        'chunks',
-        sa.Column('chunk_id', sa.Text(), nullable=False),
-        sa.Column('document_id', sa.Text(), nullable=False),
-        sa.Column('bank_id', sa.Text(), nullable=False),
-        sa.Column('chunk_index', sa.Integer(), nullable=False),
-        sa.Column('chunk_text', sa.Text(), nullable=False),
-        sa.Column('created_at', postgresql.TIMESTAMP(timezone=True), server_default=sa.text('now()'), nullable=False),
-        sa.ForeignKeyConstraint(['document_id', 'bank_id'], ['documents.id', 'documents.bank_id'], name='chunks_document_fkey', ondelete='CASCADE'),
-        sa.PrimaryKeyConstraint('chunk_id', name=op.f('pk_chunks'))
+        "chunks",
+        sa.Column("chunk_id", sa.Text(), nullable=False),
+        sa.Column("document_id", sa.Text(), nullable=False),
+        sa.Column("bank_id", sa.Text(), nullable=False),
+        sa.Column("chunk_index", sa.Integer(), nullable=False),
+        sa.Column("chunk_text", sa.Text(), nullable=False),
+        sa.Column("created_at", postgresql.TIMESTAMP(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["document_id", "bank_id"],
+            ["documents.id", "documents.bank_id"],
+            name="chunks_document_fkey",
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("chunk_id", name=op.f("pk_chunks")),
     )
 
     # Add indexes for efficient queries
-    op.create_index('idx_chunks_document_id', 'chunks', ['document_id'])
-    op.create_index('idx_chunks_bank_id', 'chunks', ['bank_id'])
+    op.create_index("idx_chunks_document_id", "chunks", ["document_id"])
+    op.create_index("idx_chunks_bank_id", "chunks", ["bank_id"])
 
     # Add chunk_id column to memory_units (nullable, as existing records won't have chunks)
-    op.add_column('memory_units', sa.Column('chunk_id', sa.Text(), nullable=True))
+    op.add_column("memory_units", sa.Column("chunk_id", sa.Text(), nullable=True))
 
     # Add foreign key constraint to chunks table
     op.create_foreign_key(
-        'memory_units_chunk_fkey',
-        'memory_units',
-        'chunks',
-        ['chunk_id'],
-        ['chunk_id'],
-        ondelete='SET NULL'
+        "memory_units_chunk_fkey", "memory_units", "chunks", ["chunk_id"], ["chunk_id"], ondelete="SET NULL"
     )
 
     # Add index on chunk_id for efficient lookups
-    op.create_index('idx_memory_units_chunk_id', 'memory_units', ['chunk_id'])
+    op.create_index("idx_memory_units_chunk_id", "memory_units", ["chunk_id"])
 
 
 def downgrade() -> None:
     """Remove chunks table and chunk_id from memory_units."""
 
     # Drop index and foreign key from memory_units
-    op.drop_index('idx_memory_units_chunk_id', table_name='memory_units')
-    op.drop_constraint('memory_units_chunk_fkey', 'memory_units', type_='foreignkey')
-    op.drop_column('memory_units', 'chunk_id')
+    op.drop_index("idx_memory_units_chunk_id", table_name="memory_units")
+    op.drop_constraint("memory_units_chunk_fkey", "memory_units", type_="foreignkey")
+    op.drop_column("memory_units", "chunk_id")
 
     # Drop chunks table indexes and table
-    op.drop_index('idx_chunks_bank_id', table_name='chunks')
-    op.drop_index('idx_chunks_document_id', table_name='chunks')
-    op.drop_table('chunks')
+    op.drop_index("idx_chunks_bank_id", table_name="chunks")
+    op.drop_index("idx_chunks_document_id", table_name="chunks")
+    op.drop_table("chunks")

hindsight-api/hindsight_api/alembic/versions/c8e5f2a3b4d1_add_retain_params_to_documents.py

@@ -5,35 +5,35 @@ Revises: b7c4d8e9f1a2
 Create Date: 2025-12-02 00:00:00.000000
 
 """
-from typing import Sequence, Union
 
-from alembic import op
+from collections.abc import Sequence
+
 import sqlalchemy as sa
+from alembic import op
 from sqlalchemy.dialects import postgresql
 
-
 # revision identifiers, used by Alembic.
-revision: str = 'c8e5f2a3b4d1'
-down_revision: Union[str, Sequence[str], None] = 'b7c4d8e9f1a2'
-branch_labels: Union[str, Sequence[str], None] = None
-depends_on: Union[str, Sequence[str], None] = None
+revision: str = "c8e5f2a3b4d1"
+down_revision: str | Sequence[str] | None = "b7c4d8e9f1a2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
 
 
 def upgrade() -> None:
     """Add retain_params JSONB column to documents table."""
 
     # Add retain_params column to store parameters passed during retain
-    op.add_column('documents', sa.Column('retain_params', postgresql.JSONB(), nullable=True))
+    op.add_column("documents", sa.Column("retain_params", postgresql.JSONB(), nullable=True))
 
     # Add index for efficient queries on retain_params
-    op.create_index('idx_documents_retain_params', 'documents', ['retain_params'], postgresql_using='gin')
+    op.create_index("idx_documents_retain_params", "documents", ["retain_params"], postgresql_using="gin")
 
 
 def downgrade() -> None:
     """Remove retain_params column from documents table."""
 
     # Drop index
-    op.drop_index('idx_documents_retain_params', table_name='documents')
+    op.drop_index("idx_documents_retain_params", table_name="documents")
 
     # Drop column
-    op.drop_column('documents', 'retain_params')
+    op.drop_column("documents", "retain_params")

hindsight-api/hindsight_api/alembic/versions/d9f6a3b4c5e2_rename_bank_to_interactions.py

@@ -5,44 +5,49 @@ Revises: c8e5f2a3b4d1
 Create Date: 2024-12-04 15:00:00.000000
 
 """
-from alembic import op
-import sqlalchemy as sa
 
+from alembic import context, op
 
 # revision identifiers, used by Alembic.
-revision = 'd9f6a3b4c5e2'
-down_revision = 'c8e5f2a3b4d1'
+revision = "d9f6a3b4c5e2"
+down_revision = "c8e5f2a3b4d1"
 branch_labels = None
 depends_on = None
 
 
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (e.g., 'tenant_x.' or '' for public)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
 def upgrade():
+    schema = _get_schema_prefix()
+
     # Drop old check constraint FIRST (before updating data)
-    op.drop_constraint('memory_units_fact_type_check', 'memory_units', type_='check')
+    op.drop_constraint("memory_units_fact_type_check", "memory_units", type_="check")
 
     # Update existing 'bank' values to 'experience'
-    op.execute("UPDATE memory_units SET fact_type = 'experience' WHERE fact_type = 'bank'")
+    op.execute(f"UPDATE {schema}memory_units SET fact_type = 'experience' WHERE fact_type = 'bank'")
     # Also update any 'interactions' values (in case of partial migration)
-    op.execute("UPDATE memory_units SET fact_type = 'experience' WHERE fact_type = 'interactions'")
+    op.execute(f"UPDATE {schema}memory_units SET fact_type = 'experience' WHERE fact_type = 'interactions'")
 
     # Create new check constraint with 'experience' instead of 'bank'
     op.create_check_constraint(
-        'memory_units_fact_type_check',
-        'memory_units',
-        "fact_type IN ('world', 'experience', 'opinion', 'observation')"
+        "memory_units_fact_type_check", "memory_units", "fact_type IN ('world', 'experience', 'opinion', 'observation')"
     )
 
 
 def downgrade():
+    schema = _get_schema_prefix()
+
     # Drop new check constraint FIRST
-    op.drop_constraint('memory_units_fact_type_check', 'memory_units', type_='check')
+    op.drop_constraint("memory_units_fact_type_check", "memory_units", type_="check")
 
     # Update 'experience' back to 'bank'
-    op.execute("UPDATE memory_units SET fact_type = 'bank' WHERE fact_type = 'experience'")
+    op.execute(f"UPDATE {schema}memory_units SET fact_type = 'bank' WHERE fact_type = 'experience'")
 
     # Recreate old check constraint
     op.create_check_constraint(
-        'memory_units_fact_type_check',
-        'memory_units',
-        "fact_type IN ('world', 'bank', 'opinion', 'observation')"
+        "memory_units_fact_type_check", "memory_units", "fact_type IN ('world', 'bank', 'opinion', 'observation')"
     )

hindsight-api/hindsight_api/alembic/versions/e0a1b2c3d4e5_disposition_to_3_traits.py

@@ -8,22 +8,49 @@ Migrate disposition traits from Big Five (openness, conscientiousness, extravers
 agreeableness, neuroticism, bias_strength with 0-1 float values) to the new 3-trait
 system (skepticism, literalism, empathy with 1-5 integer values).
 """
-from typing import Sequence, Union
 
-from alembic import op
+from collections.abc import Sequence
+
 import sqlalchemy as sa
-
+from alembic import context, op
 
 # revision identifiers, used by Alembic.
-revision: str = 'e0a1b2c3d4e5'
-down_revision: Union[str, Sequence[str], None] = 'rename_personality'
-branch_labels: Union[str, Sequence[str], None] = None
-depends_on: Union[str, Sequence[str], None] = None
+revision: str = "e0a1b2c3d4e5"
+down_revision: str | Sequence[str] | None = "rename_personality"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (e.g., 'tenant_x.' or '' for public)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def _get_target_schema() -> str:
+    """Get the target schema name (tenant schema or 'public')."""
+    schema = context.config.get_main_option("target_schema")
+    return schema if schema else "public"
 
 
 def upgrade() -> None:
     """Convert Big Five disposition to 3-trait disposition."""
     conn = op.get_bind()
+    schema = _get_schema_prefix()
+    target_schema = _get_target_schema()
+
+    # Check if disposition column exists (should have been created by previous migration)
+    result = conn.execute(
+        sa.text("""
+        SELECT column_name
+        FROM information_schema.columns
+        WHERE table_schema = :schema AND table_name = 'banks' AND column_name = 'disposition'
+    """),
+        {"schema": target_schema},
+    )
+    if not result.fetchone():
+        # Column doesn't exist yet (shouldn't happen but be safe)
+        return
 
     # Update all existing banks to use the new disposition format
     # Convert from old format to new format with reasonable mappings:
@@ -31,32 +58,54 @@ def upgrade() -> None:
     # - literalism: derived from conscientiousness (detail-oriented people are more literal)
     # - empathy: derived from agreeableness + inverse of neuroticism
     # Default all to 3 (neutral) for simplicity
-    conn.execute(sa.text("""
-        UPDATE banks
-        SET disposition = '{"skepticism": 3, "literalism": 3, "empathy": 3}'::jsonb
+    conn.execute(
+        sa.text(f"""
+        UPDATE {schema}banks
+        SET disposition = '{{"skepticism": 3, "literalism": 3, "empathy": 3}}'::jsonb
         WHERE disposition IS NOT NULL
-    """))
+    """)
+    )
 
     # Update the default for new banks
-    conn.execute(sa.text("""
-        ALTER TABLE banks
-        ALTER COLUMN disposition SET DEFAULT '{"skepticism": 3, "literalism": 3, "empathy": 3}'::jsonb
-    """))
+    conn.execute(
+        sa.text(f"""
+        ALTER TABLE {schema}banks
+        ALTER COLUMN disposition SET DEFAULT '{{"skepticism": 3, "literalism": 3, "empathy": 3}}'::jsonb
+    """)
+    )
 
 
 def downgrade() -> None:
     """Convert back to Big Five disposition."""
     conn = op.get_bind()
+    schema = _get_schema_prefix()
+    target_schema = _get_target_schema()
+
+    # Check if disposition column exists
+    result = conn.execute(
+        sa.text("""
+        SELECT column_name
+        FROM information_schema.columns
+        WHERE table_schema = :schema AND table_name = 'banks' AND column_name = 'disposition'
+    """),
+        {"schema": target_schema},
+    )
+    if not result.fetchone():
+        return
 
     # Revert to Big Five format with default values
-    conn.execute(sa.text("""
-        UPDATE banks
-        SET disposition = '{"openness": 0.5, "conscientiousness": 0.5, "extraversion": 0.5, "agreeableness": 0.5, "neuroticism": 0.5, "bias_strength": 0.5}'::jsonb
+    conn.execute(
+        sa.text(f"""
+        UPDATE {schema}banks
+        SET disposition = '{{"openness": 0.5, "conscientiousness": 0.5, "extraversion": 0.5, "agreeableness": 0.5, "neuroticism": 0.5, "bias_strength": 0.5}}'::jsonb
         WHERE disposition IS NOT NULL
-    """))
+    """)
+    )
 
     # Update the default for new banks
-    conn.execute(sa.text("""
-        ALTER TABLE banks
-        ALTER COLUMN disposition SET DEFAULT '{"openness": 0.5, "conscientiousness": 0.5, "extraversion": 0.5, "agreeableness": 0.5, "neuroticism": 0.5, "bias_strength": 0.5}'::jsonb
-    """))
+    conn.execute(
+        sa.text(f"""
+        ALTER TABLE {schema}banks
+        ALTER COLUMN disposition SET DEFAULT '{{"openness": 0.5, "conscientiousness": 0.5, "extraversion": 0.5, "agreeableness": 0.5, "neuroticism": 0.5, "bias_strength": 0.5}}'::jsonb
+    """)
+    )

hindsight-api/hindsight_api/alembic/versions/rename_personality_to_disposition.py

@@ -5,61 +5,81 @@ Revises: d9f6a3b4c5e2
 Create Date: 2024-12-04
 
 """
-from typing import Sequence, Union
 
-from alembic import op
+from collections.abc import Sequence
+
 import sqlalchemy as sa
+from alembic import context, op
 from sqlalchemy.dialects import postgresql
 
-
 # revision identifiers, used by Alembic.
-revision: str = 'rename_personality'
-down_revision: Union[str, Sequence[str], None] = 'd9f6a3b4c5e2'
-branch_labels: Union[str, Sequence[str], None] = None
-depends_on: Union[str, Sequence[str], None] = None
+revision: str = "rename_personality"
+down_revision: str | Sequence[str] | None = "d9f6a3b4c5e2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_target_schema() -> str:
+    """Get the target schema name (tenant schema or 'public')."""
+    schema = context.config.get_main_option("target_schema")
+    return schema if schema else "public"
 
 
 def upgrade() -> None:
     """Rename personality column to disposition in banks table (if it exists)."""
     conn = op.get_bind()
+    target_schema = _get_target_schema()
 
     # Check if 'personality' column exists (old database)
-    result = conn.execute(sa.text("""
+    result = conn.execute(
+        sa.text("""
         SELECT column_name
         FROM information_schema.columns
-        WHERE table_name = 'banks' AND column_name = 'personality'
-    """))
+        WHERE table_schema = :schema AND table_name = 'banks' AND column_name = 'personality'
+    """),
+        {"schema": target_schema},
+    )
     has_personality = result.fetchone() is not None
 
     # Check if 'disposition' column exists (new database)
-    result = conn.execute(sa.text("""
+    result = conn.execute(
+        sa.text("""
         SELECT column_name
         FROM information_schema.columns
-        WHERE table_name = 'banks' AND column_name = 'disposition'
-    """))
+        WHERE table_schema = :schema AND table_name = 'banks' AND column_name = 'disposition'
+    """),
+        {"schema": target_schema},
+    )
     has_disposition = result.fetchone() is not None
 
     if has_personality and not has_disposition:
         # Old database: rename personality -> disposition
-        op.alter_column('banks', 'personality', new_column_name='disposition')
+        op.alter_column("banks", "personality", new_column_name="disposition")
     elif not has_personality and not has_disposition:
         # Neither exists (shouldn't happen, but be safe): add disposition column
-        op.add_column('banks', sa.Column(
-            'disposition',
-            postgresql.JSONB(astext_type=sa.Text()),
-            server_default=sa.text("'{}'::jsonb"),
-            nullable=False
-        ))
+        op.add_column(
+            "banks",
+            sa.Column(
+                "disposition",
+                postgresql.JSONB(astext_type=sa.Text()),
+                server_default=sa.text("'{}'::jsonb"),
+                nullable=False,
+            ),
+        )
     # else: disposition already exists, nothing to do
 
 
 def downgrade() -> None:
     """Revert disposition column back to personality."""
     conn = op.get_bind()
-    result = conn.execute(sa.text("""
+    target_schema = _get_target_schema()
+    result = conn.execute(
+        sa.text("""
         SELECT column_name
         FROM information_schema.columns
-        WHERE table_name = 'banks' AND column_name = 'disposition'
-    """))
+        WHERE table_schema = :schema AND table_name = 'banks' AND column_name = 'disposition'
+    """),
+        {"schema": target_schema},
+    )
     if result.fetchone():
-        op.alter_column('banks', 'disposition', new_column_name='personality')
+        op.alter_column("banks", "disposition", new_column_name="personality")

hindsight-api/hindsight_api/api/__init__.py

@@ -3,8 +3,11 @@ Unified API module for Hindsight.
 
 Provides both HTTP REST API and MCP (Model Context Protocol) server.
 """
+
 import logging
+from contextlib import asynccontextmanager
 from typing import Optional
+
 from fastapi import FastAPI
 
 from hindsight_api import MemoryEngine
@@ -17,7 +20,7 @@ def create_app(
     http_api_enabled: bool = True,
     mcp_api_enabled: bool = False,
     mcp_mount_path: str = "/mcp",
-    initialize_memory: bool = True
+    initialize_memory: bool = True,
 ) -> FastAPI:
     """
     Create and configure the unified Hindsight API application.
@@ -43,49 +46,70 @@ def create_app(
         # Both HTTP and MCP
         app = create_app(memory, mcp_api_enabled=True)
     """
+    mcp_app = None
+
+    # Create MCP app first if enabled (we need its lifespan for chaining)
+    if mcp_api_enabled:
+        try:
+            from .mcp import create_mcp_app
+
+            mcp_app = create_mcp_app(memory=memory)
+        except ImportError as e:
+            logger.error(f"MCP server requested but dependencies not available: {e}")
+            logger.error("Install with: pip install hindsight-api[mcp]")
+            raise
 
     # Import and create HTTP API if enabled
     if http_api_enabled:
         from .http import create_app as create_http_app
-        app = create_http_app(
-            memory=memory,
-            initialize_memory=initialize_memory
-        )
+
+        app = create_http_app(memory=memory, initialize_memory=initialize_memory)
         logger.info("HTTP REST API enabled")
     else:
         # Create minimal FastAPI app
         app = FastAPI(title="Hindsight API", version="0.0.7")
         logger.info("HTTP REST API disabled")
 
-    # Mount MCP server if enabled
-    if mcp_api_enabled:
-        try:
-            from .mcp import create_mcp_app
+    # Mount MCP server and chain its lifespan if enabled
+    if mcp_app is not None:
+        # Get the MCP app's underlying Starlette app for lifespan access
+        mcp_starlette_app = mcp_app.mcp_app
 
-            # Create MCP app with dynamic bank_id support
-            # Supports: /mcp/{bank_id}/sse (bank-specific SSE endpoint)
-            mcp_app = create_mcp_app(memory=memory)
-            app.mount(mcp_mount_path, mcp_app)
-            logger.info(f"MCP server enabled at {mcp_mount_path}/{{bank_id}}/sse")
-        except ImportError as e:
-            logger.error(f"MCP server requested but dependencies not available: {e}")
-            logger.error("Install with: pip install hindsight-api[mcp]")
-            raise
+        # Store the original lifespan
+        original_lifespan = app.router.lifespan_context
+
+        @asynccontextmanager
+        async def chained_lifespan(app_instance: FastAPI):
+            """Chain the MCP lifespan with the main app lifespan."""
+            # Start MCP lifespan first
+            async with mcp_starlette_app.router.lifespan_context(mcp_starlette_app):
+                logger.info("MCP lifespan started")
+                # Then start the original app lifespan
+                async with original_lifespan(app_instance):
+                    yield
+            logger.info("MCP lifespan stopped")
+
+        # Replace the app's lifespan with the chained version
+        app.router.lifespan_context = chained_lifespan
+
+        # Mount the MCP middleware
+        app.mount(mcp_mount_path, mcp_app)
+        logger.info(f"MCP server enabled at {mcp_mount_path}/{{bank_id}}/mcp")
 
     return app
 
 
 # Re-export commonly used items for backwards compatibility
 from .http import (
-    RecallRequest,
-    RecallResult,
-    RecallResponse,
-    MemoryItem,
-    RetainRequest,
-    ReflectRequest,
-    ReflectResponse,
     CreateBankRequest,
     DispositionTraits,
+    MemoryItem,
+    RecallRequest,
+    RecallResponse,
+    RecallResult,
+    ReflectRequest,
+    ReflectResponse,
+    RetainRequest,
 )
 
 __all__ = [

hindsight-api/hindsight_api/api/mcp.py

@@ -4,27 +4,34 @@ import json
 import logging
 import os
 from contextvars import ContextVar
-from typing import Optional
 
 from fastmcp import FastMCP
+
 from hindsight_api import MemoryEngine
 from hindsight_api.engine.response_models import VALID_RECALL_FACT_TYPES
+from hindsight_api.models import RequestContext
 
 # Configure logging from HINDSIGHT_API_LOG_LEVEL environment variable
 _log_level_str = os.environ.get("HINDSIGHT_API_LOG_LEVEL", "info").lower()
-_log_level_map = {"critical": logging.CRITICAL, "error": logging.ERROR, "warning": logging.WARNING,
-                  "info": logging.INFO, "debug": logging.DEBUG, "trace": logging.DEBUG}
+_log_level_map = {
+    "critical": logging.CRITICAL,
+    "error": logging.ERROR,
+    "warning": logging.WARNING,
+    "info": logging.INFO,
+    "debug": logging.DEBUG,
+    "trace": logging.DEBUG,
+}
 logging.basicConfig(
     level=_log_level_map.get(_log_level_str, logging.INFO),
-    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s"
+    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
 )
 logger = logging.getLogger(__name__)
 
 # Context variable to hold the current bank_id from the URL path
-_current_bank_id: ContextVar[Optional[str]] = ContextVar("current_bank_id", default=None)
+_current_bank_id: ContextVar[str | None] = ContextVar("current_bank_id", default=None)
 
 
-def get_current_bank_id() -> Optional[str]:
+def get_current_bank_id() -> str | None:
     """Get the current bank_id from context (set from URL path)."""
     return _current_bank_id.get()
 
@@ -61,9 +68,10 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
         """
         try:
             bank_id = get_current_bank_id()
-            await memory.put_batch_async(
-                bank_id=bank_id,
-                contents=[{"content": content, "context": context}]
+            if bank_id is None:
+                return "Error: No bank_id configured"
+            await memory.retain_batch_async(
+                bank_id=bank_id, contents=[{"content": content, "context": context}], request_context=RequestContext()
             )
             return "Memory stored successfully"
         except Exception as e:
@@ -87,12 +95,16 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
         """
         try:
             bank_id = get_current_bank_id()
+            if bank_id is None:
+                return "Error: No bank_id configured"
             from hindsight_api.engine.memory_engine import Budget
+
             search_result = await memory.recall_async(
                 bank_id=bank_id,
                 query=query,
                 fact_type=list(VALID_RECALL_FACT_TYPES),
-                budget=Budget.LOW
+                budget=Budget.LOW,
+                request_context=RequestContext(),
             )
 
             results = [
@@ -101,7 +113,7 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
                     "text": fact.text,
                     "type": fact.fact_type,
                     "context": fact.context,
-                    "event_date": fact.event_date,
+                    "occurred_start": fact.occurred_start,
                 }
                 for fact in search_result.results[:max_results]
             ]
@@ -133,7 +145,7 @@ class MCPMiddleware:
         # Strip any mount prefix (e.g., /mcp) that FastAPI might not have stripped
         root_path = scope.get("root_path", "")
         if root_path and path.startswith(root_path):
-            path = path[len(root_path):] or "/"
+            path = path[len(root_path) :] or "/"
 
         # Also handle case where mount path wasn't stripped (e.g., /mcp/...)
         if path.startswith("/mcp/"):
@@ -169,10 +181,7 @@ class MCPMiddleware:
                     body = message.get("body", b"")
                     if body and b"/messages" in body:
                         # Rewrite /messages to /{bank_id}/messages in SSE endpoint event
-                        body = body.replace(
-                            b"data: /messages",
-                            f"data: /{bank_id}/messages".encode()
-                        )
+                        body = body.replace(b"data: /messages", f"data: /{bank_id}/messages".encode())
                         message = {**message, "body": body}
                 await send(message)
 
@@ -183,15 +192,19 @@ class MCPMiddleware:
     async def _send_error(self, send, status: int, message: str):
         """Send an error response."""
         body = json.dumps({"error": message}).encode()
-        await send({
-            "type": "http.response.start",
-            "status": status,
-            "headers": [(b"content-type", b"application/json")],
-        })
-        await send({
-            "type": "http.response.body",
-            "body": body,
-        })
+        await send(
+            {
+                "type": "http.response.start",
+                "status": status,
+                "headers": [(b"content-type", b"application/json")],
+            }
+        )
+        await send(
+            {
+                "type": "http.response.body",
+                "body": body,
+            }
+        )
 
 
 def create_mcp_app(memory: MemoryEngine):

hindsight-api/hindsight_api/banner.py

@@ -6,7 +6,7 @@ Shows the logo and tagline with gradient colors.
 
 # Gradient colors: #0074d9 -> #009296
 GRADIENT_START = (0, 116, 217)  # #0074d9
-GRADIENT_END = (0, 146, 150)    # #009296
+GRADIENT_END = (0, 146, 150)  # #009296
 
 # Pre-generated logo (generated by test-logo.py)
 LOGO = """\
@@ -31,8 +31,8 @@ def gradient_text(text: str, start: tuple = GRADIENT_START, end: tuple = GRADIEN
     result = []
     length = len(text)
     for i, char in enumerate(text):
-        if char == ' ':
-            result.append(' ')
+        if char == " ":
+            result.append(" ")
         else:
             t = i / max(length - 1, 1)
             r, g, b = _interpolate_color(start, end, t)
@@ -74,9 +74,16 @@ def dim(text: str) -> str:
     return f"\033[38;2;128;128;128m{text}\033[0m"
 
 
-def print_startup_info(host: str, port: int, database_url: str, llm_provider: str,
-                       llm_model: str, embeddings_provider: str, reranker_provider: str,
-                       mcp_enabled: bool = False):
+def print_startup_info(
+    host: str,
+    port: int,
+    database_url: str,
+    llm_provider: str,
+    llm_model: str,
+    embeddings_provider: str,
+    reranker_provider: str,
+    mcp_enabled: bool = False,
+):
     """Print styled startup information."""
     print(color_start("Starting Hindsight API..."))
     print(f"  {dim('URL:')} {color(f'http://{host}:{port}', 0.2)}")

hindsight-api/hindsight_api/config.py

@@ -3,10 +3,15 @@ Centralized configuration for Hindsight API.
 
 All environment variables and their defaults are defined here.
 """
+
+import logging
 import os
 from dataclasses import dataclass
-from typing import Optional
-import logging
+
+from dotenv import find_dotenv, load_dotenv
+
+# Load .env file, searching current and parent directories (overrides existing env vars)
+load_dotenv(find_dotenv(usecwd=True), override=True)
 
 logger = logging.getLogger(__name__)
 
@@ -16,6 +21,8 @@ ENV_LLM_PROVIDER = "HINDSIGHT_API_LLM_PROVIDER"
 ENV_LLM_API_KEY = "HINDSIGHT_API_LLM_API_KEY"
 ENV_LLM_MODEL = "HINDSIGHT_API_LLM_MODEL"
 ENV_LLM_BASE_URL = "HINDSIGHT_API_LLM_BASE_URL"
+ENV_LLM_MAX_CONCURRENT = "HINDSIGHT_API_LLM_MAX_CONCURRENT"
+ENV_LLM_TIMEOUT = "HINDSIGHT_API_LLM_TIMEOUT"
 
 ENV_EMBEDDINGS_PROVIDER = "HINDSIGHT_API_EMBEDDINGS_PROVIDER"
 ENV_EMBEDDINGS_LOCAL_MODEL = "HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL"
@@ -30,11 +37,23 @@ ENV_PORT = "HINDSIGHT_API_PORT"
 ENV_LOG_LEVEL = "HINDSIGHT_API_LOG_LEVEL"
 ENV_MCP_ENABLED = "HINDSIGHT_API_MCP_ENABLED"
 ENV_GRAPH_RETRIEVER = "HINDSIGHT_API_GRAPH_RETRIEVER"
+ENV_MCP_LOCAL_BANK_ID = "HINDSIGHT_API_MCP_LOCAL_BANK_ID"
+ENV_MCP_INSTRUCTIONS = "HINDSIGHT_API_MCP_INSTRUCTIONS"
+
+# Observation thresholds
+ENV_OBSERVATION_MIN_FACTS = "HINDSIGHT_API_OBSERVATION_MIN_FACTS"
+ENV_OBSERVATION_TOP_ENTITIES = "HINDSIGHT_API_OBSERVATION_TOP_ENTITIES"
+
+# Optimization flags
+ENV_SKIP_LLM_VERIFICATION = "HINDSIGHT_API_SKIP_LLM_VERIFICATION"
+ENV_LAZY_RERANKER = "HINDSIGHT_API_LAZY_RERANKER"
 
 # Default values
 DEFAULT_DATABASE_URL = "pg0"
 DEFAULT_LLM_PROVIDER = "openai"
 DEFAULT_LLM_MODEL = "gpt-5-mini"
+DEFAULT_LLM_MAX_CONCURRENT = 32
+DEFAULT_LLM_TIMEOUT = 120.0  # seconds
 
 DEFAULT_EMBEDDINGS_PROVIDER = "local"
 DEFAULT_EMBEDDINGS_LOCAL_MODEL = "BAAI/bge-small-en-v1.5"
@@ -47,6 +66,31 @@ DEFAULT_PORT = 8888
 DEFAULT_LOG_LEVEL = "info"
 DEFAULT_MCP_ENABLED = True
 DEFAULT_GRAPH_RETRIEVER = "bfs"  # Options: "bfs", "mpfp"
+DEFAULT_MCP_LOCAL_BANK_ID = "mcp"
+
+# Observation thresholds
+DEFAULT_OBSERVATION_MIN_FACTS = 5  # Min facts required to generate entity observations
+DEFAULT_OBSERVATION_TOP_ENTITIES = 5  # Max entities to process per retain batch
+
+# Default MCP tool descriptions (can be customized via env vars)
+DEFAULT_MCP_RETAIN_DESCRIPTION = """Store important information to long-term memory.
+
+Use this tool PROACTIVELY whenever the user shares:
+- Personal facts, preferences, or interests
+- Important events or milestones
+- User history, experiences, or background
+- Decisions, opinions, or stated preferences
+- Goals, plans, or future intentions
+- Relationships or people mentioned
+- Work context, projects, or responsibilities"""
+
+DEFAULT_MCP_RECALL_DESCRIPTION = """Search memories to provide personalized, context-aware responses.
+
+Use this tool PROACTIVELY to:
+- Check user's preferences before making suggestions
+- Recall user's history to provide continuity
+- Remember user's goals and context
+- Personalize responses based on past interactions"""
 
 # Required embedding dimension for database schema
 EMBEDDING_DIMENSION = 384
@@ -61,19 +105,21 @@ class HindsightConfig:
 
     # LLM
     llm_provider: str
-    llm_api_key: Optional[str]
+    llm_api_key: str | None
     llm_model: str
-    llm_base_url: Optional[str]
+    llm_base_url: str | None
+    llm_max_concurrent: int
+    llm_timeout: float
 
     # Embeddings
     embeddings_provider: str
     embeddings_local_model: str
-    embeddings_tei_url: Optional[str]
+    embeddings_tei_url: str | None
 
     # Reranker
     reranker_provider: str
     reranker_local_model: str
-    reranker_tei_url: Optional[str]
+    reranker_tei_url: str | None
 
     # Server
     host: str
@@ -84,37 +130,50 @@ class HindsightConfig:
     # Recall
     graph_retriever: str
 
+    # Observation thresholds
+    observation_min_facts: int
+    observation_top_entities: int
+
+    # Optimization flags
+    skip_llm_verification: bool
+    lazy_reranker: bool
+
     @classmethod
     def from_env(cls) -> "HindsightConfig":
         """Create configuration from environment variables."""
         return cls(
             # Database
             database_url=os.getenv(ENV_DATABASE_URL, DEFAULT_DATABASE_URL),
-
             # LLM
             llm_provider=os.getenv(ENV_LLM_PROVIDER, DEFAULT_LLM_PROVIDER),
             llm_api_key=os.getenv(ENV_LLM_API_KEY),
             llm_model=os.getenv(ENV_LLM_MODEL, DEFAULT_LLM_MODEL),
             llm_base_url=os.getenv(ENV_LLM_BASE_URL) or None,
-
+            llm_max_concurrent=int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_CONCURRENT))),
+            llm_timeout=float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT))),
             # Embeddings
             embeddings_provider=os.getenv(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER),
             embeddings_local_model=os.getenv(ENV_EMBEDDINGS_LOCAL_MODEL, DEFAULT_EMBEDDINGS_LOCAL_MODEL),
             embeddings_tei_url=os.getenv(ENV_EMBEDDINGS_TEI_URL),
-
             # Reranker
             reranker_provider=os.getenv(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER),
             reranker_local_model=os.getenv(ENV_RERANKER_LOCAL_MODEL, DEFAULT_RERANKER_LOCAL_MODEL),
             reranker_tei_url=os.getenv(ENV_RERANKER_TEI_URL),
-
             # Server
             host=os.getenv(ENV_HOST, DEFAULT_HOST),
             port=int(os.getenv(ENV_PORT, DEFAULT_PORT)),
             log_level=os.getenv(ENV_LOG_LEVEL, DEFAULT_LOG_LEVEL),
             mcp_enabled=os.getenv(ENV_MCP_ENABLED, str(DEFAULT_MCP_ENABLED)).lower() == "true",
-
             # Recall
             graph_retriever=os.getenv(ENV_GRAPH_RETRIEVER, DEFAULT_GRAPH_RETRIEVER),
+            # Optimization flags
+            skip_llm_verification=os.getenv(ENV_SKIP_LLM_VERIFICATION, "false").lower() == "true",
+            lazy_reranker=os.getenv(ENV_LAZY_RERANKER, "false").lower() == "true",
+            # Observation thresholds
+            observation_min_facts=int(os.getenv(ENV_OBSERVATION_MIN_FACTS, str(DEFAULT_OBSERVATION_MIN_FACTS))),
+            observation_top_entities=int(
+                os.getenv(ENV_OBSERVATION_TOP_ENTITIES, str(DEFAULT_OBSERVATION_TOP_ENTITIES))
+            ),
         )
 
     def get_llm_base_url(self) -> str:
@@ -127,6 +186,8 @@ class HindsightConfig:
             return "https://api.groq.com/openai/v1"
         elif provider == "ollama":
             return "http://localhost:11434/v1"
+        elif provider == "lmstudio":
+            return "http://localhost:1234/v1"
         else:
             return ""
 
@@ -146,7 +207,8 @@ class HindsightConfig:
         """Configure Python logging based on the log level."""
         logging.basicConfig(
             level=self.get_python_log_level(),
-            format="%(asctime)s - %(levelname)s - %(name)s - %(message)s"
+            format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+            force=True,  # Override any existing configuration
         )
 
     def log_config(self) -> None:

hindsight-api/hindsight_api/daemon.py

@@ -0,0 +1,204 @@
+"""
+Daemon mode support for Hindsight API.
+
+Provides idle timeout and lockfile management for running as a background daemon.
+"""
+
+import asyncio
+import fcntl
+import logging
+import os
+import sys
+import time
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+# Default daemon configuration
+DEFAULT_DAEMON_PORT = 8889
+DEFAULT_IDLE_TIMEOUT = 0  # 0 = no auto-exit (hindsight-embed passes its own timeout)
+LOCKFILE_PATH = Path.home() / ".hindsight" / "daemon.lock"
+DAEMON_LOG_PATH = Path.home() / ".hindsight" / "daemon.log"
+
+
+class IdleTimeoutMiddleware:
+    """ASGI middleware that tracks activity and exits after idle timeout."""
+
+    def __init__(self, app, idle_timeout: int = DEFAULT_IDLE_TIMEOUT):
+        self.app = app
+        self.idle_timeout = idle_timeout
+        self.last_activity = time.time()
+        self._checker_task = None
+
+    async def __call__(self, scope, receive, send):
+        # Update activity timestamp on each request
+        self.last_activity = time.time()
+        await self.app(scope, receive, send)
+
+    def start_idle_checker(self):
+        """Start the background task that checks for idle timeout."""
+        self._checker_task = asyncio.create_task(self._check_idle())
+
+    async def _check_idle(self):
+        """Background task that exits the process after idle timeout."""
+        # If idle_timeout is 0, don't auto-exit
+        if self.idle_timeout <= 0:
+            return
+
+        while True:
+            await asyncio.sleep(30)  # Check every 30 seconds
+            idle_time = time.time() - self.last_activity
+            if idle_time > self.idle_timeout:
+                logger.info(f"Idle timeout reached ({self.idle_timeout}s), shutting down daemon")
+                # Give a moment for any in-flight requests
+                await asyncio.sleep(1)
+                os._exit(0)
+
+
+class DaemonLock:
+    """
+    File-based lock to prevent multiple daemon instances.
+
+    Uses fcntl.flock for atomic locking on Unix systems.
+    """
+
+    def __init__(self, lockfile: Path = LOCKFILE_PATH):
+        self.lockfile = lockfile
+        self._fd = None
+
+    def acquire(self) -> bool:
+        """
+        Try to acquire the daemon lock.
+
+        Returns True if lock acquired, False if another daemon is running.
+        """
+        self.lockfile.parent.mkdir(parents=True, exist_ok=True)
+
+        try:
+            self._fd = open(self.lockfile, "w")
+            fcntl.flock(self._fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+            # Write PID for debugging
+            self._fd.write(str(os.getpid()))
+            self._fd.flush()
+            return True
+        except (IOError, OSError):
+            # Lock is held by another process
+            if self._fd:
+                self._fd.close()
+                self._fd = None
+            return False
+
+    def release(self):
+        """Release the daemon lock."""
+        if self._fd:
+            try:
+                fcntl.flock(self._fd.fileno(), fcntl.LOCK_UN)
+                self._fd.close()
+            except Exception:
+                pass
+            finally:
+                self._fd = None
+                # Remove lockfile
+                try:
+                    self.lockfile.unlink()
+                except Exception:
+                    pass
+
+    def is_locked(self) -> bool:
+        """Check if the lock is held by another process."""
+        if not self.lockfile.exists():
+            return False
+
+        try:
+            fd = open(self.lockfile, "r")
+            fcntl.flock(fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+            # We got the lock, so no one else has it
+            fcntl.flock(fd.fileno(), fcntl.LOCK_UN)
+            fd.close()
+            return False
+        except (IOError, OSError):
+            return True
+
+    def get_pid(self) -> int | None:
+        """Get the PID of the daemon holding the lock."""
+        if not self.lockfile.exists():
+            return None
+        try:
+            with open(self.lockfile, "r") as f:
+                return int(f.read().strip())
+        except (ValueError, IOError):
+            return None
+
+
+def daemonize():
+    """
+    Fork the current process into a background daemon.
+
+    Uses double-fork technique to properly detach from terminal.
+    """
+    # First fork
+    pid = os.fork()
+    if pid > 0:
+        # Parent exits
+        sys.exit(0)
+
+    # Create new session
+    os.setsid()
+
+    # Second fork to prevent zombie processes
+    pid = os.fork()
+    if pid > 0:
+        sys.exit(0)
+
+    # Redirect standard file descriptors to log file
+    DAEMON_LOG_PATH.parent.mkdir(parents=True, exist_ok=True)
+
+    sys.stdout.flush()
+    sys.stderr.flush()
+
+    # Redirect stdin to /dev/null
+    with open("/dev/null", "r") as devnull:
+        os.dup2(devnull.fileno(), sys.stdin.fileno())
+
+    # Redirect stdout/stderr to log file
+    log_fd = open(DAEMON_LOG_PATH, "a")
+    os.dup2(log_fd.fileno(), sys.stdout.fileno())
+    os.dup2(log_fd.fileno(), sys.stderr.fileno())
+
+
+def check_daemon_running(port: int = DEFAULT_DAEMON_PORT) -> bool:
+    """Check if a daemon is running and responsive on the given port."""
+    import socket
+
+    try:
+        sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        sock.settimeout(1)
+        result = sock.connect_ex(("127.0.0.1", port))
+        sock.close()
+        return result == 0
+    except Exception:
+        return False
+
+
+def stop_daemon(port: int = DEFAULT_DAEMON_PORT) -> bool:
+    """Stop a running daemon by sending SIGTERM to the process."""
+    lock = DaemonLock()
+    pid = lock.get_pid()
+
+    if pid is None:
+        return False
+
+    try:
+        import signal
+
+        os.kill(pid, signal.SIGTERM)
+        # Wait for process to exit
+        for _ in range(50):  # Wait up to 5 seconds
+            time.sleep(0.1)
+            try:
+                os.kill(pid, 0)  # Check if process exists
+            except OSError:
+                return True  # Process exited
+        return False
+    except OSError:
+        return False

hindsight-api/hindsight_api/engine/__init__.py

@@ -7,24 +7,30 @@ This package contains all the implementation details of the memory engine:
 - Supporting modules: embeddings, cross_encoder, entity_resolver, etc.
 """
 
-from .memory_engine import MemoryEngine
+from .cross_encoder import CrossEncoderModel, LocalSTCrossEncoder, RemoteTEICrossEncoder
 from .db_utils import acquire_with_retry
 from .embeddings import Embeddings, LocalSTEmbeddings, RemoteTEIEmbeddings
-from .cross_encoder import CrossEncoderModel, LocalSTCrossEncoder, RemoteTEICrossEncoder
+from .llm_wrapper import LLMConfig
+from .memory_engine import (
+    MemoryEngine,
+    UnqualifiedTableError,
+    fq_table,
+    get_current_schema,
+    validate_sql_schema,
+)
+from .response_models import MemoryFact, RecallResult, ReflectResult
 from .search.trace import (
-    SearchTrace,
-    QueryInfo,
     EntryPoint,
-    NodeVisit,
-    WeightComponents,
     LinkInfo,
+    NodeVisit,
     PruningDecision,
-    SearchSummary,
+    QueryInfo,
     SearchPhaseMetrics,
+    SearchSummary,
+    SearchTrace,
+    WeightComponents,
 )
 from .search.tracer import SearchTracer
-from .llm_wrapper import LLMConfig
-from .response_models import RecallResult, ReflectResult, MemoryFact
 
 __all__ = [
     "MemoryEngine",
@@ -49,4 +55,9 @@ __all__ = [
     "RecallResult",
     "ReflectResult",
     "MemoryFact",
+    # Schema safety utilities
+    "fq_table",
+    "get_current_schema",
+    "validate_sql_schema",
+    "UnqualifiedTableError",
 ]

hindsight-api/hindsight_api/engine/cross_encoder.py

@@ -5,19 +5,19 @@ Provides an interface for reranking with different backends.
 
 Configuration via environment variables - see hindsight_api.config for all env var names.
 """
-from abc import ABC, abstractmethod
-from typing import List, Tuple, Optional
+
 import logging
 import os
+from abc import ABC, abstractmethod
 
 import httpx
 
 from ..config import (
-    ENV_RERANKER_PROVIDER,
-    ENV_RERANKER_LOCAL_MODEL,
-    ENV_RERANKER_TEI_URL,
-    DEFAULT_RERANKER_PROVIDER,
     DEFAULT_RERANKER_LOCAL_MODEL,
+    DEFAULT_RERANKER_PROVIDER,
+    ENV_RERANKER_LOCAL_MODEL,
+    ENV_RERANKER_PROVIDER,
+    ENV_RERANKER_TEI_URL,
 )
 
 logger = logging.getLogger(__name__)
@@ -47,7 +47,7 @@ class CrossEncoderModel(ABC):
         pass
 
     @abstractmethod
-    def predict(self, pairs: List[Tuple[str, str]]) -> List[float]:
+    def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
         """
         Score query-document pairs for relevance.
 
@@ -72,7 +72,7 @@ class LocalSTCrossEncoder(CrossEncoderModel):
     - Trained for passage re-ranking
     """
 
-    def __init__(self, model_name: Optional[str] = None):
+    def __init__(self, model_name: str | None = None):
         """
         Initialize local SentenceTransformers cross-encoder.
 
@@ -104,7 +104,7 @@ class LocalSTCrossEncoder(CrossEncoderModel):
         self._model = CrossEncoder(self.model_name)
         logger.info("Reranker: local provider initialized")
 
-    def predict(self, pairs: List[Tuple[str, str]]) -> List[float]:
+    def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
         """
         Score query-document pairs for relevance.
 
@@ -117,7 +117,7 @@ class LocalSTCrossEncoder(CrossEncoderModel):
         if self._model is None:
             raise RuntimeError("Reranker not initialized. Call initialize() first.")
         scores = self._model.predict(pairs, show_progress_bar=False)
-        return scores.tolist() if hasattr(scores, 'tolist') else list(scores)
+        return scores.tolist() if hasattr(scores, "tolist") else list(scores)
 
 
 class RemoteTEICrossEncoder(CrossEncoderModel):
@@ -153,8 +153,8 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
         self.batch_size = batch_size
         self.max_retries = max_retries
         self.retry_delay = retry_delay
-        self._client: Optional[httpx.Client] = None
-        self._model_id: Optional[str] = None
+        self._client: httpx.Client | None = None
+        self._model_id: str | None = None
 
     @property
     def provider_name(self) -> str:
@@ -163,6 +163,7 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
     def _request_with_retry(self, method: str, url: str, **kwargs) -> httpx.Response:
         """Make an HTTP request with automatic retries on transient errors."""
         import time
+
         last_error = None
         delay = self.retry_delay
 
@@ -177,14 +178,18 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
             except (httpx.ConnectError, httpx.ReadTimeout, httpx.WriteTimeout) as e:
                 last_error = e
                 if attempt < self.max_retries:
-                    logger.warning(f"TEI request failed (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s...")
+                    logger.warning(
+                        f"TEI request failed (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s..."
+                    )
                     time.sleep(delay)
                     delay *= 2  # Exponential backoff
             except httpx.HTTPStatusError as e:
                 # Retry on 5xx server errors
                 if e.response.status_code >= 500 and attempt < self.max_retries:
                     last_error = e
-                    logger.warning(f"TEI server error (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s...")
+                    logger.warning(
+                        f"TEI server error (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s..."
+                    )
                     time.sleep(delay)
                     delay *= 2
                 else:
@@ -209,7 +214,7 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
         except httpx.HTTPError as e:
             raise RuntimeError(f"Failed to connect to TEI server at {self.base_url}: {e}")
 
-    def predict(self, pairs: List[Tuple[str, str]]) -> List[float]:
+    def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
         """
         Score query-document pairs using the remote TEI reranker.
 
@@ -229,7 +234,7 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
 
         # Process in batches
         for i in range(0, len(pairs), self.batch_size):
-            batch = pairs[i:i + self.batch_size]
+            batch = pairs[i : i + self.batch_size]
 
             # TEI rerank endpoint expects query and texts separately
             # All pairs in a batch should have the same query for optimal performance
@@ -287,15 +292,11 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
     if provider == "tei":
         url = os.environ.get(ENV_RERANKER_TEI_URL)
         if not url:
-            raise ValueError(
-                f"{ENV_RERANKER_TEI_URL} is required when {ENV_RERANKER_PROVIDER} is 'tei'"
-            )
+            raise ValueError(f"{ENV_RERANKER_TEI_URL} is required when {ENV_RERANKER_PROVIDER} is 'tei'")
         return RemoteTEICrossEncoder(base_url=url)
     elif provider == "local":
         model = os.environ.get(ENV_RERANKER_LOCAL_MODEL)
         model_name = model or DEFAULT_RERANKER_LOCAL_MODEL
         return LocalSTCrossEncoder(model_name=model_name)
     else:
-        raise ValueError(
-            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei'"
-        )
+        raise ValueError(f"Unknown reranker provider: {provider}. Supported: 'local', 'tei'")

hindsight-api/hindsight_api/engine/db_utils.py

@@ -1,9 +1,11 @@
 """
 Database utility functions for connection management with retry logic.
 """
+
 import asyncio
 import logging
 from contextlib import asynccontextmanager
+
 import asyncpg
 
 logger = logging.getLogger(__name__)
@@ -54,16 +56,14 @@ async def retry_with_backoff(
         except retryable_exceptions as e:
             last_exception = e
             if attempt < max_retries:
-                delay = min(base_delay * (2 ** attempt), max_delay)
+                delay = min(base_delay * (2**attempt), max_delay)
                 logger.warning(
                     f"Database operation failed (attempt {attempt + 1}/{max_retries + 1}): {e}. "
                     f"Retrying in {delay:.1f}s..."
                 )
                 await asyncio.sleep(delay)
             else:
-                logger.error(
-                    f"Database operation failed after {max_retries + 1} attempts: {e}"
-                )
+                logger.error(f"Database operation failed after {max_retries + 1} attempts: {e}")
     raise last_exception
 
 
@@ -83,6 +83,7 @@ async def acquire_with_retry(pool: asyncpg.Pool, max_retries: int = DEFAULT_MAX_
     Yields:
         An asyncpg connection
     """
+
     async def acquire():
         return await pool.acquire()
 

hindsight-api/hindsight_api/engine/embeddings.py

@@ -8,20 +8,20 @@ the database schema (pgvector column defined as vector(384)).
 
 Configuration via environment variables - see hindsight_api.config for all env var names.
 """
-from abc import ABC, abstractmethod
-from typing import List, Optional
+
 import logging
 import os
+from abc import ABC, abstractmethod
 
 import httpx
 
 from ..config import (
-    ENV_EMBEDDINGS_PROVIDER,
-    ENV_EMBEDDINGS_LOCAL_MODEL,
-    ENV_EMBEDDINGS_TEI_URL,
-    DEFAULT_EMBEDDINGS_PROVIDER,
     DEFAULT_EMBEDDINGS_LOCAL_MODEL,
+    DEFAULT_EMBEDDINGS_PROVIDER,
     EMBEDDING_DIMENSION,
+    ENV_EMBEDDINGS_LOCAL_MODEL,
+    ENV_EMBEDDINGS_PROVIDER,
+    ENV_EMBEDDINGS_TEI_URL,
 )
 
 logger = logging.getLogger(__name__)
@@ -52,7 +52,7 @@ class Embeddings(ABC):
         pass
 
     @abstractmethod
-    def encode(self, texts: List[str]) -> List[List[float]]:
+    def encode(self, texts: list[str]) -> list[list[float]]:
         """
         Generate 384-dimensional embeddings for a list of texts.
 
@@ -75,7 +75,7 @@ class LocalSTEmbeddings(Embeddings):
     embeddings matching the database schema.
     """
 
-    def __init__(self, model_name: Optional[str] = None):
+    def __init__(self, model_name: str | None = None):
         """
         Initialize local SentenceTransformers embeddings.
 
@@ -123,7 +123,7 @@ class LocalSTEmbeddings(Embeddings):
 
         logger.info(f"Embeddings: local provider initialized (dim: {model_dim})")
 
-    def encode(self, texts: List[str]) -> List[List[float]]:
+    def encode(self, texts: list[str]) -> list[list[float]]:
         """
         Generate 384-dimensional embeddings for a list of texts.
 
@@ -172,8 +172,8 @@ class RemoteTEIEmbeddings(Embeddings):
         self.batch_size = batch_size
         self.max_retries = max_retries
         self.retry_delay = retry_delay
-        self._client: Optional[httpx.Client] = None
-        self._model_id: Optional[str] = None
+        self._client: httpx.Client | None = None
+        self._model_id: str | None = None
 
     @property
     def provider_name(self) -> str:
@@ -182,6 +182,7 @@ class RemoteTEIEmbeddings(Embeddings):
     def _request_with_retry(self, method: str, url: str, **kwargs) -> httpx.Response:
         """Make an HTTP request with automatic retries on transient errors."""
         import time
+
         last_error = None
         delay = self.retry_delay
 
@@ -196,14 +197,18 @@ class RemoteTEIEmbeddings(Embeddings):
             except (httpx.ConnectError, httpx.ReadTimeout, httpx.WriteTimeout) as e:
                 last_error = e
                 if attempt < self.max_retries:
-                    logger.warning(f"TEI request failed (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s...")
+                    logger.warning(
+                        f"TEI request failed (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s..."
+                    )
                     time.sleep(delay)
                     delay *= 2  # Exponential backoff
             except httpx.HTTPStatusError as e:
                 # Retry on 5xx server errors
                 if e.response.status_code >= 500 and attempt < self.max_retries:
                     last_error = e
-                    logger.warning(f"TEI server error (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s...")
+                    logger.warning(
+                        f"TEI server error (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s..."
+                    )
                     time.sleep(delay)
                     delay *= 2
                 else:
@@ -228,7 +233,7 @@ class RemoteTEIEmbeddings(Embeddings):
         except httpx.HTTPError as e:
             raise RuntimeError(f"Failed to connect to TEI server at {self.base_url}: {e}")
 
-    def encode(self, texts: List[str]) -> List[List[float]]:
+    def encode(self, texts: list[str]) -> list[list[float]]:
         """
         Generate embeddings using the remote TEI server.
 
@@ -248,7 +253,7 @@ class RemoteTEIEmbeddings(Embeddings):
 
         # Process in batches
         for i in range(0, len(texts), self.batch_size):
-            batch = texts[i:i + self.batch_size]
+            batch = texts[i : i + self.batch_size]
 
             try:
                 response = self._request_with_retry(
@@ -278,15 +283,11 @@ def create_embeddings_from_env() -> Embeddings:
     if provider == "tei":
         url = os.environ.get(ENV_EMBEDDINGS_TEI_URL)
         if not url:
-            raise ValueError(
-                f"{ENV_EMBEDDINGS_TEI_URL} is required when {ENV_EMBEDDINGS_PROVIDER} is 'tei'"
-            )
+            raise ValueError(f"{ENV_EMBEDDINGS_TEI_URL} is required when {ENV_EMBEDDINGS_PROVIDER} is 'tei'")
         return RemoteTEIEmbeddings(base_url=url)
     elif provider == "local":
         model = os.environ.get(ENV_EMBEDDINGS_LOCAL_MODEL)
         model_name = model or DEFAULT_EMBEDDINGS_LOCAL_MODEL
         return LocalSTEmbeddings(model_name=model_name)
     else:
-        raise ValueError(
-            f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei'"
-        )
+        raise ValueError(f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei'")

hindsight-api/hindsight_api/engine/entity_resolver.py

@@ -4,12 +4,14 @@ Entity extraction and resolution for memory system.
 Uses spaCy for entity extraction and implements resolution logic
 to disambiguate entities across memory units.
 """
-import asyncpg
-from typing import List, Dict, Optional, Set, Any
-from difflib import SequenceMatcher
-from datetime import datetime, timezone
-from .db_utils import acquire_with_retry
 
+from datetime import UTC, datetime
+from difflib import SequenceMatcher
+
+import asyncpg
+
+from .db_utils import acquire_with_retry
+from .memory_engine import fq_table
 
 # Load spaCy model (singleton)
 _nlp = None
@@ -32,11 +34,11 @@ class EntityResolver:
     async def resolve_entities_batch(
         self,
         bank_id: str,
-        entities_data: List[Dict],
+        entities_data: list[dict],
         context: str,
         unit_event_date,
         conn=None,
-    ) -> List[str]:
+    ) -> list[str]:
         """
         Resolve multiple entities in batch (MUCH faster than sequential).
 
@@ -62,36 +64,38 @@ class EntityResolver:
         else:
             return await self._resolve_entities_batch_impl(conn, bank_id, entities_data, context, unit_event_date)
 
-    async def _resolve_entities_batch_impl(self, conn, bank_id: str, entities_data: List[Dict], context: str, unit_event_date) -> List[str]:
+    async def _resolve_entities_batch_impl(
+        self, conn, bank_id: str, entities_data: list[dict], context: str, unit_event_date
+    ) -> list[str]:
         # Query ALL candidates for this bank
         all_entities = await conn.fetch(
-            """
+            f"""
             SELECT canonical_name, id, metadata, last_seen, mention_count
-            FROM entities
+            FROM {fq_table("entities")}
             WHERE bank_id = $1
             """,
-            bank_id
+            bank_id,
         )
 
         # Build entity ID to name mapping for co-occurrence lookups
-        entity_id_to_name = {row['id']: row['canonical_name'].lower() for row in all_entities}
+        entity_id_to_name = {row["id"]: row["canonical_name"].lower() for row in all_entities}
 
         # Query ALL co-occurrences for this bank's entities in one query
         # This builds a map of entity_id -> set of co-occurring entity names
         all_cooccurrences = await conn.fetch(
-            """
+            f"""
             SELECT ec.entity_id_1, ec.entity_id_2, ec.cooccurrence_count
-            FROM entity_cooccurrences ec
-            WHERE ec.entity_id_1 IN (SELECT id FROM entities WHERE bank_id = $1)
-               OR ec.entity_id_2 IN (SELECT id FROM entities WHERE bank_id = $1)
+            FROM {fq_table("entity_cooccurrences")} ec
+            WHERE ec.entity_id_1 IN (SELECT id FROM {fq_table("entities")} WHERE bank_id = $1)
+               OR ec.entity_id_2 IN (SELECT id FROM {fq_table("entities")} WHERE bank_id = $1)
             """,
-            bank_id
+            bank_id,
         )
 
         # Build co-occurrence map: entity_id -> set of co-occurring entity names (lowercase)
-        cooccurrence_map: Dict[str, Set[str]] = {}
+        cooccurrence_map: dict[str, set[str]] = {}
         for row in all_cooccurrences:
-            eid1, eid2 = row['entity_id_1'], row['entity_id_2']
+            eid1, eid2 = row["entity_id_1"], row["entity_id_2"]
             # Add both directions
             if eid1 not in cooccurrence_map:
                 cooccurrence_map[eid1] = set()
@@ -105,22 +109,24 @@ class EntityResolver:
 
         # Build candidate map for each entity text
         all_candidates = {}  # Maps entity_text -> list of candidates
-        entity_texts = list(set(e['text'] for e in entities_data))
+        entity_texts = list(set(e["text"] for e in entities_data))
 
         for entity_text in entity_texts:
             matching = []
             entity_text_lower = entity_text.lower()
             for row in all_entities:
-                canonical_name = row['canonical_name']
-                ent_id = row['id']
-                metadata = row['metadata']
-                last_seen = row['last_seen']
-                mention_count = row['mention_count']
+                canonical_name = row["canonical_name"]
+                ent_id = row["id"]
+                metadata = row["metadata"]
+                last_seen = row["last_seen"]
+                mention_count = row["mention_count"]
                 canonical_lower = canonical_name.lower()
                 # Match if exact or substring match
-                if (entity_text_lower == canonical_lower or
-                    entity_text_lower in canonical_lower or
-                    canonical_lower in entity_text_lower):
+                if (
+                    entity_text_lower == canonical_lower
+                    or entity_text_lower in canonical_lower
+                    or canonical_lower in entity_text_lower
+                ):
                     matching.append((ent_id, canonical_name, metadata, last_seen, mention_count))
             all_candidates[entity_text] = matching
 
@@ -130,10 +136,10 @@ class EntityResolver:
         entities_to_create = []  # (idx, entity_data, event_date)
 
         for idx, entity_data in enumerate(entities_data):
-            entity_text = entity_data['text']
-            nearby_entities = entity_data.get('nearby_entities', [])
+            entity_text = entity_data["text"]
+            nearby_entities = entity_data.get("nearby_entities", [])
             # Use per-entity date if available, otherwise fall back to batch-level date
-            entity_event_date = entity_data.get('event_date', unit_event_date)
+            entity_event_date = entity_data.get("event_date", unit_event_date)
 
             candidates = all_candidates.get(entity_text, [])
 
@@ -146,17 +152,13 @@ class EntityResolver:
             best_candidate = None
             best_score = 0.0
 
-            nearby_entity_set = {e['text'].lower() for e in nearby_entities if e['text'] != entity_text}
+            nearby_entity_set = {e["text"].lower() for e in nearby_entities if e["text"] != entity_text}
 
             for candidate_id, canonical_name, metadata, last_seen, mention_count in candidates:
                 score = 0.0
 
                 # 1. Name similarity (0-0.5)
-                name_similarity = SequenceMatcher(
-                    None,
-                    entity_text.lower(),
-                    canonical_name.lower()
-                ).ratio()
+                name_similarity = SequenceMatcher(None, entity_text.lower(), canonical_name.lower()).ratio()
                 score += name_similarity * 0.5
 
                 # 2. Co-occurring entities (0-0.3)
@@ -169,8 +171,10 @@ class EntityResolver:
                 # 3. Temporal proximity (0-0.2)
                 if last_seen and entity_event_date:
                     # Normalize timezone awareness for comparison
-                    event_date_utc = entity_event_date if entity_event_date.tzinfo else entity_event_date.replace(tzinfo=timezone.utc)
-                    last_seen_utc = last_seen if last_seen.tzinfo else last_seen.replace(tzinfo=timezone.utc)
+                    event_date_utc = (
+                        entity_event_date if entity_event_date.tzinfo else entity_event_date.replace(tzinfo=UTC)
+                    )
+                    last_seen_utc = last_seen if last_seen.tzinfo else last_seen.replace(tzinfo=UTC)
                     days_diff = abs((event_date_utc - last_seen_utc).total_seconds() / 86400)
                     if days_diff < 7:
                         temporal_score = max(0, 1.0 - (days_diff / 7))
@@ -192,13 +196,13 @@ class EntityResolver:
         # Batch update existing entities
         if entities_to_update:
             await conn.executemany(
-                """
-                UPDATE entities SET
+                f"""
+                UPDATE {fq_table("entities")} SET
                     mention_count = mention_count + 1,
                     last_seen = $2
                 WHERE id = $1::uuid
                 """,
-                entities_to_update
+                entities_to_update,
             )
 
         # Batch create new entities using COPY + INSERT for maximum speed
@@ -208,7 +212,7 @@ class EntityResolver:
             # For duplicates, we only insert once and reuse the ID
             unique_entities = {}  # lowercase_name -> (entity_data, event_date, [indices])
             for idx, entity_data, event_date in entities_to_create:
-                name_lower = entity_data['text'].lower()
+                name_lower = entity_data["text"].lower()
                 if name_lower not in unique_entities:
                     unique_entities[name_lower] = (entity_data, event_date, [idx])
                 else:
@@ -222,31 +226,31 @@ class EntityResolver:
             indices_map = []  # Maps result index -> list of original indices
 
             for name_lower, (entity_data, event_date, indices) in unique_entities.items():
-                entity_names.append(entity_data['text'])
+                entity_names.append(entity_data["text"])
                 entity_dates.append(event_date)
                 indices_map.append(indices)
 
             # Batch INSERT ... ON CONFLICT with RETURNING
             # This is much faster than individual inserts
             rows = await conn.fetch(
-                """
-                INSERT INTO entities (bank_id, canonical_name, first_seen, last_seen, mention_count)
+                f"""
+                INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
                 SELECT $1, name, event_date, event_date, 1
                 FROM unnest($2::text[], $3::timestamptz[]) AS t(name, event_date)
                 ON CONFLICT (bank_id, LOWER(canonical_name))
                 DO UPDATE SET
-                    mention_count = entities.mention_count + 1,
+                    mention_count = {fq_table("entities")}.mention_count + 1,
                     last_seen = EXCLUDED.last_seen
                 RETURNING id
                 """,
                 bank_id,
                 entity_names,
-                entity_dates
+                entity_dates,
             )
 
             # Map returned IDs back to original indices
             for result_idx, row in enumerate(rows):
-                entity_id = row['id']
+                entity_id = row["id"]
                 for original_idx in indices_map[result_idx]:
                     entity_ids[original_idx] = entity_id
 
@@ -257,7 +261,7 @@ class EntityResolver:
         bank_id: str,
         entity_text: str,
         context: str,
-        nearby_entities: List[Dict],
+        nearby_entities: list[dict],
         unit_event_date,
     ) -> str:
         """
@@ -276,9 +280,9 @@ class EntityResolver:
         async with acquire_with_retry(self.pool) as conn:
             # Find candidate entities with similar name
             candidates = await conn.fetch(
-                """
+                f"""
                 SELECT id, canonical_name, metadata, last_seen
-                FROM entities
+                FROM {fq_table("entities")}
                 WHERE bank_id = $1
                   AND (
                     canonical_name ILIKE $2
@@ -287,14 +291,14 @@ class EntityResolver:
                   )
                 ORDER BY mention_count DESC
                 """,
-                bank_id, entity_text, f"%{entity_text}%"
+                bank_id,
+                entity_text,
+                f"%{entity_text}%",
             )
 
             if not candidates:
                 # New entity - create it
-                return await self._create_entity(
-                    conn, bank_id, entity_text, unit_event_date
-                )
+                return await self._create_entity(conn, bank_id, entity_text, unit_event_date)
 
             # Score candidates based on:
             # 1. Name similarity
@@ -306,31 +310,27 @@ class EntityResolver:
             best_score = 0.0
             best_name_similarity = 0.0
 
-            nearby_entity_set = {e['text'].lower() for e in nearby_entities if e['text'] != entity_text}
+            nearby_entity_set = {e["text"].lower() for e in nearby_entities if e["text"] != entity_text}
 
             for row in candidates:
-                candidate_id = row['id']
-                canonical_name = row['canonical_name']
-                metadata = row['metadata']
-                last_seen = row['last_seen']
+                candidate_id = row["id"]
+                canonical_name = row["canonical_name"]
+                metadata = row["metadata"]
+                last_seen = row["last_seen"]
                 score = 0.0
 
                 # 1. Name similarity (0-1)
-                name_similarity = SequenceMatcher(
-                    None,
-                    entity_text.lower(),
-                    canonical_name.lower()
-                ).ratio()
+                name_similarity = SequenceMatcher(None, entity_text.lower(), canonical_name.lower()).ratio()
                 score += name_similarity * 0.5
 
                 # 2. Co-occurring entities (0-0.5)
                 # Get entities that co-occurred with this candidate before
                 # Use the materialized co-occurrence cache for fast lookup
                 co_entity_rows = await conn.fetch(
-                    """
+                    f"""
                     SELECT e.canonical_name, ec.cooccurrence_count
-                    FROM entity_cooccurrences ec
-                    JOIN entities e ON (
+                    FROM {fq_table("entity_cooccurrences")} ec
+                    JOIN {fq_table("entities")} e ON (
                         CASE
                             WHEN ec.entity_id_1 = $1 THEN ec.entity_id_2
                             WHEN ec.entity_id_2 = $1 THEN ec.entity_id_1
@@ -338,9 +338,9 @@ class EntityResolver:
                     )
                     WHERE ec.entity_id_1 = $1 OR ec.entity_id_2 = $1
                     """,
-                    candidate_id
+                    candidate_id,
                 )
-                co_entities = {r['canonical_name'].lower() for r in co_entity_rows}
+                co_entities = {r["canonical_name"].lower() for r in co_entity_rows}
 
                 # Check overlap with nearby entities
                 overlap = len(nearby_entity_set & co_entities)
@@ -366,20 +366,19 @@ class EntityResolver:
             if best_score > threshold:
                 # Update entity
                 await conn.execute(
-                    """
-                    UPDATE entities
+                    f"""
+                    UPDATE {fq_table("entities")}
                     SET mention_count = mention_count + 1,
                         last_seen = $1
                     WHERE id = $2
                     """,
-                    unit_event_date, best_candidate
+                    unit_event_date,
+                    best_candidate,
                 )
                 return best_candidate
             else:
                 # Not confident - create new entity
-                return await self._create_entity(
-                    conn, bank_id, entity_text, unit_event_date
-                )
+                return await self._create_entity(conn, bank_id, entity_text, unit_event_date)
 
     async def _create_entity(
         self,
@@ -404,16 +403,19 @@ class EntityResolver:
             Entity ID
         """
         entity_id = await conn.fetchval(
-            """
-            INSERT INTO entities (bank_id, canonical_name, first_seen, last_seen, mention_count)
+            f"""
+            INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
             VALUES ($1, $2, $3, $4, 1)
             ON CONFLICT (bank_id, LOWER(canonical_name))
             DO UPDATE SET
-                mention_count = entities.mention_count + 1,
+                mention_count = {fq_table("entities")}.mention_count + 1,
                 last_seen = EXCLUDED.last_seen
             RETURNING id
             """,
-            bank_id, entity_text, event_date, event_date
+            bank_id,
+            entity_text,
+            event_date,
+            event_date,
         )
         return entity_id
 
@@ -429,25 +431,27 @@ class EntityResolver:
         async with acquire_with_retry(self.pool) as conn:
             # Insert unit-entity link
             await conn.execute(
-                """
-                INSERT INTO unit_entities (unit_id, entity_id)
+                f"""
+                INSERT INTO {fq_table("unit_entities")} (unit_id, entity_id)
                 VALUES ($1, $2)
                 ON CONFLICT DO NOTHING
                 """,
-                unit_id, entity_id
+                unit_id,
+                entity_id,
             )
 
             # Update co-occurrence cache: find other entities in this unit
             rows = await conn.fetch(
-                """
+                f"""
                 SELECT entity_id
-                FROM unit_entities
+                FROM {fq_table("unit_entities")}
                 WHERE unit_id = $1 AND entity_id != $2
                 """,
-                unit_id, entity_id
+                unit_id,
+                entity_id,
             )
 
-            other_entities = [row['entity_id'] for row in rows]
+            other_entities = [row["entity_id"] for row in rows]
 
             # Update co-occurrences for each pair
             for other_entity_id in other_entities:
@@ -469,18 +473,19 @@ class EntityResolver:
             entity_id_1, entity_id_2 = entity_id_2, entity_id_1
 
         await conn.execute(
-            """
-            INSERT INTO entity_cooccurrences (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
+            f"""
+            INSERT INTO {fq_table("entity_cooccurrences")} (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
             VALUES ($1, $2, 1, NOW())
             ON CONFLICT (entity_id_1, entity_id_2)
             DO UPDATE SET
-                cooccurrence_count = entity_cooccurrences.cooccurrence_count + 1,
+                cooccurrence_count = {fq_table("entity_cooccurrences")}.cooccurrence_count + 1,
                 last_cooccurred = NOW()
             """,
-            entity_id_1, entity_id_2
+            entity_id_1,
+            entity_id_2,
         )
 
-    async def link_units_to_entities_batch(self, unit_entity_pairs: List[tuple[str, str]], conn=None):
+    async def link_units_to_entities_batch(self, unit_entity_pairs: list[tuple[str, str]], conn=None):
         """
         Link multiple memory units to entities in batch (MUCH faster than sequential).
 
@@ -499,15 +504,15 @@ class EntityResolver:
         else:
             return await self._link_units_to_entities_batch_impl(conn, unit_entity_pairs)
 
-    async def _link_units_to_entities_batch_impl(self, conn, unit_entity_pairs: List[tuple[str, str]]):
+    async def _link_units_to_entities_batch_impl(self, conn, unit_entity_pairs: list[tuple[str, str]]):
         # Batch insert all unit-entity links
         await conn.executemany(
-            """
-            INSERT INTO unit_entities (unit_id, entity_id)
+            f"""
+            INSERT INTO {fq_table("unit_entities")} (unit_id, entity_id)
             VALUES ($1, $2)
             ON CONFLICT DO NOTHING
             """,
-            unit_entity_pairs
+            unit_entity_pairs,
         )
 
         # Build map of unit -> entities for co-occurrence calculation
@@ -524,7 +529,7 @@ class EntityResolver:
             entity_list = list(entity_ids)  # Convert set to list for iteration
             # For each pair of entities in this unit, create co-occurrence
             for i, entity_id_1 in enumerate(entity_list):
-                for entity_id_2 in entity_list[i+1:]:
+                for entity_id_2 in entity_list[i + 1 :]:
                     # Skip if same entity (shouldn't happen with set, but be safe)
                     if entity_id_1 == entity_id_2:
                         continue
@@ -535,20 +540,20 @@ class EntityResolver:
 
         # Batch update co-occurrences
         if cooccurrence_pairs:
-            now = datetime.now(timezone.utc)
+            now = datetime.now(UTC)
             await conn.executemany(
-                """
-                INSERT INTO entity_cooccurrences (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
+                f"""
+                INSERT INTO {fq_table("entity_cooccurrences")} (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
                 VALUES ($1, $2, $3, $4)
                 ON CONFLICT (entity_id_1, entity_id_2)
                 DO UPDATE SET
-                    cooccurrence_count = entity_cooccurrences.cooccurrence_count + 1,
+                    cooccurrence_count = {fq_table("entity_cooccurrences")}.cooccurrence_count + 1,
                     last_cooccurred = EXCLUDED.last_cooccurred
                 """,
-                [(e1, e2, 1, now) for e1, e2 in cooccurrence_pairs]
+                [(e1, e2, 1, now) for e1, e2 in cooccurrence_pairs],
             )
 
-    async def get_units_by_entity(self, entity_id: str, limit: int = 100) -> List[str]:
+    async def get_units_by_entity(self, entity_id: str, limit: int = 100) -> list[str]:
         """
         Get all units that mention an entity.
 
@@ -561,22 +566,23 @@ class EntityResolver:
         """
         async with acquire_with_retry(self.pool) as conn:
             rows = await conn.fetch(
-                """
+                f"""
                 SELECT unit_id
-                FROM unit_entities
+                FROM {fq_table("unit_entities")}
                 WHERE entity_id = $1
                 ORDER BY unit_id
                 LIMIT $2
                 """,
-                entity_id, limit
+                entity_id,
+                limit,
             )
-            return [row['unit_id'] for row in rows]
+            return [row["unit_id"] for row in rows]
 
     async def get_entity_by_text(
         self,
         bank_id: str,
         entity_text: str,
-    ) -> Optional[str]:
+    ) -> str | None:
         """
         Find an entity by text (for query resolution).
 
@@ -589,14 +595,15 @@ class EntityResolver:
         """
         async with acquire_with_retry(self.pool) as conn:
             row = await conn.fetchrow(
-                """
-                SELECT id FROM entities
+                f"""
+                SELECT id FROM {fq_table("entities")}
                 WHERE bank_id = $1
                   AND canonical_name ILIKE $2
                 ORDER BY mention_count DESC
                 LIMIT 1
                 """,
-                bank_id, entity_text
+                bank_id,
+                entity_text,
             )
 
-            return row['id'] if row else None
+            return row["id"] if row else None

hindsight-api/hindsight_api/engine/interface.py

@@ -0,0 +1,596 @@
+"""Abstract interface for MemoryEngine public methods.
+
+This module defines the public API that HTTP endpoints and extensions should use
+to interact with the memory system. All methods require a RequestContext for
+authentication when a TenantExtension is configured.
+"""
+
+from abc import ABC, abstractmethod
+from datetime import datetime
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from hindsight_api.engine.memory_engine import Budget
+    from hindsight_api.engine.response_models import RecallResult, ReflectResult
+    from hindsight_api.models import RequestContext
+
+
+class MemoryEngineInterface(ABC):
+    """
+    Abstract interface for the Memory Engine.
+
+    This defines the public API that should be used by HTTP endpoints and extensions.
+    All methods require a RequestContext for authentication.
+    """
+
+    # =========================================================================
+    # Health & Status
+    # =========================================================================
+
+    @abstractmethod
+    async def health_check(self) -> dict:
+        """
+        Check the health of the memory system.
+
+        Returns:
+            Dict with 'status' key ('healthy' or 'unhealthy') and additional info.
+        """
+        ...
+
+    # =========================================================================
+    # Core Memory Operations
+    # =========================================================================
+
+    @abstractmethod
+    async def retain_batch_async(
+        self,
+        bank_id: str,
+        contents: list[dict[str, Any]],
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Retain a batch of memory items.
+
+        Args:
+            bank_id: The memory bank ID.
+            contents: List of content dicts with 'content', optional 'event_date',
+                     'context', 'metadata', 'document_id'.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with processing results.
+        """
+        ...
+
+    @abstractmethod
+    async def recall_async(
+        self,
+        bank_id: str,
+        query: str,
+        *,
+        budget: "Budget | None" = None,
+        max_tokens: int = 4096,
+        enable_trace: bool = False,
+        fact_type: list[str] | None = None,
+        question_date: datetime | None = None,
+        include_entities: bool = False,
+        max_entity_tokens: int = 500,
+        include_chunks: bool = False,
+        max_chunk_tokens: int = 8192,
+        request_context: "RequestContext",
+    ) -> "RecallResult":
+        """
+        Recall memories relevant to a query.
+
+        Args:
+            bank_id: The memory bank ID.
+            query: The search query.
+            budget: Search budget (LOW, MID, HIGH).
+            max_tokens: Maximum tokens in response.
+            enable_trace: Include trace information.
+            fact_type: Filter by fact types.
+            question_date: Context date for temporal relevance.
+            include_entities: Include entity observations.
+            max_entity_tokens: Max tokens for entity observations.
+            include_chunks: Include raw chunks.
+            max_chunk_tokens: Max tokens for chunks.
+            request_context: Request context for authentication.
+
+        Returns:
+            RecallResult with matching memories.
+        """
+        ...
+
+    @abstractmethod
+    async def reflect_async(
+        self,
+        bank_id: str,
+        query: str,
+        *,
+        budget: "Budget | None" = None,
+        context: str | None = None,
+        max_tokens: int = 4096,
+        response_schema: dict | None = None,
+        request_context: "RequestContext",
+    ) -> "ReflectResult":
+        """
+        Reflect on a query and generate a thoughtful response.
+
+        Args:
+            bank_id: The memory bank ID.
+            query: The question to reflect on.
+            budget: Search budget for retrieving context.
+            context: Additional context for the reflection.
+            max_tokens: Maximum tokens for the response.
+            response_schema: Optional JSON Schema for structured output.
+            request_context: Request context for authentication.
+
+        Returns:
+            ReflectResult with generated response and supporting facts.
+        """
+        ...
+
+    # =========================================================================
+    # Bank Management
+    # =========================================================================
+
+    @abstractmethod
+    async def list_banks(
+        self,
+        *,
+        request_context: "RequestContext",
+    ) -> list[dict[str, Any]]:
+        """
+        List all memory banks.
+
+        Args:
+            request_context: Request context for authentication.
+
+        Returns:
+            List of bank info dicts.
+        """
+        ...
+
+    @abstractmethod
+    async def get_bank_profile(
+        self,
+        bank_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Get bank profile including disposition and background.
+
+        Args:
+            bank_id: The memory bank ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            Bank profile dict.
+        """
+        ...
+
+    @abstractmethod
+    async def update_bank_disposition(
+        self,
+        bank_id: str,
+        disposition: dict[str, int],
+        *,
+        request_context: "RequestContext",
+    ) -> None:
+        """
+        Update bank disposition traits.
+
+        Args:
+            bank_id: The memory bank ID.
+            disposition: Dict with trait values.
+            request_context: Request context for authentication.
+        """
+        ...
+
+    @abstractmethod
+    async def merge_bank_background(
+        self,
+        bank_id: str,
+        new_info: str,
+        *,
+        update_disposition: bool = True,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Merge new background information into bank profile.
+
+        Args:
+            bank_id: The memory bank ID.
+            new_info: New background information to merge.
+            update_disposition: Whether to infer disposition from background.
+            request_context: Request context for authentication.
+
+        Returns:
+            Updated background info.
+        """
+        ...
+
+    @abstractmethod
+    async def delete_bank(
+        self,
+        bank_id: str,
+        *,
+        fact_type: str | None = None,
+        request_context: "RequestContext",
+    ) -> dict[str, int]:
+        """
+        Delete a bank or its memories.
+
+        Args:
+            bank_id: The memory bank ID.
+            fact_type: If specified, only delete memories of this type.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with deletion counts.
+        """
+        ...
+
+    # =========================================================================
+    # Memory Units
+    # =========================================================================
+
+    @abstractmethod
+    async def list_memory_units(
+        self,
+        bank_id: str,
+        *,
+        fact_type: str | None = None,
+        search_query: str | None = None,
+        limit: int = 100,
+        offset: int = 0,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        List memory units with pagination.
+
+        Args:
+            bank_id: The memory bank ID.
+            fact_type: Filter by fact type.
+            search_query: Full-text search query.
+            limit: Maximum results.
+            offset: Pagination offset.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with 'items', 'total', 'limit', 'offset'.
+        """
+        ...
+
+    @abstractmethod
+    async def delete_memory_unit(
+        self,
+        unit_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Delete a specific memory unit.
+
+        Args:
+            unit_id: The memory unit ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            Deletion result.
+        """
+        ...
+
+    @abstractmethod
+    async def get_graph_data(
+        self,
+        bank_id: str,
+        *,
+        fact_type: str | None = None,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Get graph data for visualization.
+
+        Args:
+            bank_id: The memory bank ID.
+            fact_type: Filter by fact type.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with nodes, edges, table_rows, total_units.
+        """
+        ...
+
+    # =========================================================================
+    # Documents
+    # =========================================================================
+
+    @abstractmethod
+    async def list_documents(
+        self,
+        bank_id: str,
+        *,
+        search_query: str | None = None,
+        limit: int = 100,
+        offset: int = 0,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        List documents with pagination.
+
+        Args:
+            bank_id: The memory bank ID.
+            search_query: Search query.
+            limit: Maximum results.
+            offset: Pagination offset.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with 'items', 'total', 'limit', 'offset'.
+        """
+        ...
+
+    @abstractmethod
+    async def get_document(
+        self,
+        document_id: str,
+        bank_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any] | None:
+        """
+        Get a specific document.
+
+        Args:
+            document_id: The document ID.
+            bank_id: The memory bank ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            Document dict or None if not found.
+        """
+        ...
+
+    @abstractmethod
+    async def delete_document(
+        self,
+        document_id: str,
+        bank_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, int]:
+        """
+        Delete a document and its memory units.
+
+        Args:
+            document_id: The document ID.
+            bank_id: The memory bank ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with deletion counts.
+        """
+        ...
+
+    @abstractmethod
+    async def get_chunk(
+        self,
+        chunk_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any] | None:
+        """
+        Get a specific chunk.
+
+        Args:
+            chunk_id: The chunk ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            Chunk dict or None if not found.
+        """
+        ...
+
+    # =========================================================================
+    # Entities
+    # =========================================================================
+
+    @abstractmethod
+    async def list_entities(
+        self,
+        bank_id: str,
+        *,
+        limit: int = 100,
+        request_context: "RequestContext",
+    ) -> list[dict[str, Any]]:
+        """
+        List entities for a bank.
+
+        Args:
+            bank_id: The memory bank ID.
+            limit: Maximum results.
+            request_context: Request context for authentication.
+
+        Returns:
+            List of entity dicts.
+        """
+        ...
+
+    @abstractmethod
+    async def get_entity_observations(
+        self,
+        bank_id: str,
+        entity_id: str,
+        *,
+        limit: int = 10,
+        request_context: "RequestContext",
+    ) -> list[Any]:
+        """
+        Get observations for an entity.
+
+        Args:
+            bank_id: The memory bank ID.
+            entity_id: The entity ID.
+            limit: Maximum observations.
+            request_context: Request context for authentication.
+
+        Returns:
+            List of EntityObservation objects.
+        """
+        ...
+
+    @abstractmethod
+    async def regenerate_entity_observations(
+        self,
+        bank_id: str,
+        entity_id: str,
+        entity_name: str,
+        *,
+        request_context: "RequestContext",
+    ) -> None:
+        """
+        Regenerate observations for an entity.
+
+        Args:
+            bank_id: The memory bank ID.
+            entity_id: The entity ID.
+            entity_name: The entity's canonical name.
+            request_context: Request context for authentication.
+        """
+        ...
+
+    # =========================================================================
+    # Statistics & Operations
+    # =========================================================================
+
+    @abstractmethod
+    async def get_bank_stats(
+        self,
+        bank_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Get statistics about memory nodes and links for a bank.
+
+        Args:
+            bank_id: The memory bank ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with node_counts, link_counts, link_counts_by_fact_type,
+            link_breakdown, and operations stats.
+        """
+        ...
+
+    @abstractmethod
+    async def get_entity(
+        self,
+        bank_id: str,
+        entity_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any] | None:
+        """
+        Get entity details including metadata and observations.
+
+        Args:
+            bank_id: The memory bank ID.
+            entity_id: The entity ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            Entity dict with id, canonical_name, mention_count, first_seen,
+            last_seen, metadata, and observations. None if not found.
+        """
+        ...
+
+    @abstractmethod
+    async def list_operations(
+        self,
+        bank_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> list[dict[str, Any]]:
+        """
+        List async operations for a bank.
+
+        Args:
+            bank_id: The memory bank ID.
+            request_context: Request context for authentication.
+
+        Returns:
+            List of operation dicts with id, task_type, status, etc.
+        """
+        ...
+
+    @abstractmethod
+    async def cancel_operation(
+        self,
+        bank_id: str,
+        operation_id: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Cancel a pending async operation.
+
+        Args:
+            bank_id: The memory bank ID.
+            operation_id: The operation ID to cancel.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with success status and message.
+
+        Raises:
+            ValueError: If operation not found.
+        """
+        ...
+
+    @abstractmethod
+    async def update_bank(
+        self,
+        bank_id: str,
+        *,
+        name: str | None = None,
+        background: str | None = None,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Update bank name and/or background.
+
+        Args:
+            bank_id: The memory bank ID.
+            name: New bank name (optional).
+            background: New background text (optional, replaces existing).
+            request_context: Request context for authentication.
+
+        Returns:
+            Updated bank profile dict.
+        """
+        ...
+
+    @abstractmethod
+    async def submit_async_retain(
+        self,
+        bank_id: str,
+        contents: list[dict[str, Any]],
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Submit a batch retain operation to run asynchronously.
+
+        Args:
+            bank_id: The memory bank ID.
+            contents: List of content dicts to retain.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with operation_id and items_count.
+        """
+        ...

hindsight-api/hindsight_api/engine/llm_wrapper.py

@@ -1,15 +1,26 @@
 """
 LLM wrapper for unified configuration across providers.
 """
+
+import asyncio
+import json
+import logging
 import os
 import time
-import asyncio
-from typing import Optional, Any, Dict, List
-from openai import AsyncOpenAI, RateLimitError, APIError, APIStatusError, APIConnectionError, LengthFinishReasonError
+from typing import Any
+
+import httpx
 from google import genai
-from google.genai import types as genai_types
 from google.genai import errors as genai_errors
-import logging
+from google.genai import types as genai_types
+from openai import APIConnectionError, APIStatusError, AsyncOpenAI, LengthFinishReasonError
+
+from ..config import (
+    DEFAULT_LLM_MAX_CONCURRENT,
+    DEFAULT_LLM_TIMEOUT,
+    ENV_LLM_MAX_CONCURRENT,
+    ENV_LLM_TIMEOUT,
+)
 
 # Seed applied to every Groq request for deterministic behavior.
 DEFAULT_LLM_SEED = 4242
@@ -20,7 +31,9 @@ logger = logging.getLogger(__name__)
 logging.getLogger("httpx").setLevel(logging.WARNING)
 
 # Global semaphore to limit concurrent LLM requests across all instances
-_global_llm_semaphore = asyncio.Semaphore(32)
+# Set HINDSIGHT_API_LLM_MAX_CONCURRENT=1 for local LLMs (LM Studio, Ollama)
+_llm_max_concurrent = int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_CONCURRENT)))
+_global_llm_semaphore = asyncio.Semaphore(_llm_max_concurrent)
 
 
 class OutputTooLongError(Exception):
@@ -31,6 +44,7 @@ class OutputTooLongError(Exception):
     to allow callers to handle output length issues without depending on
     provider-specific implementations.
     """
+
     pass
 
 
@@ -53,7 +67,7 @@ class LLMProvider:
         Initialize LLM provider.
 
         Args:
-            provider: Provider name ("openai", "groq", "ollama", "gemini").
+            provider: Provider name ("openai", "groq", "ollama", "gemini", "anthropic", "lmstudio").
             api_key: API key.
             base_url: Base URL for the API.
             model: Model name.
@@ -66,11 +80,9 @@ class LLMProvider:
         self.reasoning_effort = reasoning_effort
 
         # Validate provider
-        valid_providers = ["openai", "groq", "ollama", "gemini"]
+        valid_providers = ["openai", "groq", "ollama", "gemini", "anthropic", "lmstudio"]
         if self.provider not in valid_providers:
-            raise ValueError(
-                f"Invalid LLM provider: {self.provider}. Must be one of: {', '.join(valid_providers)}"
-            )
+            raise ValueError(f"Invalid LLM provider: {self.provider}. Must be one of: {', '.join(valid_providers)}")
 
         # Set default base URLs
         if not self.base_url:
@@ -78,25 +90,48 @@ class LLMProvider:
                 self.base_url = "https://api.groq.com/openai/v1"
             elif self.provider == "ollama":
                 self.base_url = "http://localhost:11434/v1"
+            elif self.provider == "lmstudio":
+                self.base_url = "http://localhost:1234/v1"
 
-        # Validate API key (not needed for ollama)
-        if self.provider != "ollama" and not self.api_key:
+        # Validate API key (not needed for ollama or lmstudio)
+        if self.provider not in ("ollama", "lmstudio") and not self.api_key:
             raise ValueError(f"API key not found for {self.provider}")
 
+        # Get timeout config (set HINDSIGHT_API_LLM_TIMEOUT for local LLMs that need longer timeouts)
+        self.timeout = float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT)))
+
         # Create client based on provider
+        self._client = None
+        self._gemini_client = None
+        self._anthropic_client = None
+
         if self.provider == "gemini":
             self._gemini_client = genai.Client(api_key=self.api_key)
-            self._client = None
-        elif self.provider == "ollama":
-            self._client = AsyncOpenAI(api_key="ollama", base_url=self.base_url, max_retries=0)
-            self._gemini_client = None
+        elif self.provider == "anthropic":
+            from anthropic import AsyncAnthropic
+
+            # Only pass base_url if it's set (Anthropic uses default URL otherwise)
+            anthropic_kwargs = {"api_key": self.api_key}
+            if self.base_url:
+                anthropic_kwargs["base_url"] = self.base_url
+            if self.timeout:
+                anthropic_kwargs["timeout"] = self.timeout
+            self._anthropic_client = AsyncAnthropic(**anthropic_kwargs)
+        elif self.provider in ("ollama", "lmstudio"):
+            # Use dummy key if not provided for local
+            api_key = self.api_key or "local"
+            client_kwargs = {"api_key": api_key, "base_url": self.base_url, "max_retries": 0}
+            if self.timeout:
+                client_kwargs["timeout"] = self.timeout
+            self._client = AsyncOpenAI(**client_kwargs)
         else:
             # Only pass base_url if it's set (OpenAI uses default URL otherwise)
             client_kwargs = {"api_key": self.api_key, "max_retries": 0}
             if self.base_url:
                 client_kwargs["base_url"] = self.base_url
+            if self.timeout:
+                client_kwargs["timeout"] = self.timeout
             self._client = AsyncOpenAI(**client_kwargs)
-            self._gemini_client = None
 
     async def verify_connection(self) -> None:
         """
@@ -106,10 +141,12 @@ class LLMProvider:
             RuntimeError: If the connection test fails.
         """
         try:
-            logger.info(f"Verifying LLM: provider={self.provider}, model={self.model}, base_url={self.base_url or 'default'}...")
+            logger.info(
+                f"Verifying LLM: provider={self.provider}, model={self.model}, base_url={self.base_url or 'default'}..."
+            )
             await self.call(
                 messages=[{"role": "user", "content": "Say 'ok'"}],
-                max_completion_tokens=10,
+                max_completion_tokens=100,
                 max_retries=2,
                 initial_backoff=0.5,
                 max_backoff=2.0,
@@ -117,21 +154,20 @@ class LLMProvider:
             # If we get here without exception, the connection is working
             logger.info(f"LLM verified: {self.provider}/{self.model}")
         except Exception as e:
-            raise RuntimeError(
-                f"LLM connection verification failed for {self.provider}/{self.model}: {e}"
-            ) from e
+            raise RuntimeError(f"LLM connection verification failed for {self.provider}/{self.model}: {e}") from e
 
     async def call(
         self,
-        messages: List[Dict[str, str]],
-        response_format: Optional[Any] = None,
-        max_completion_tokens: Optional[int] = None,
-        temperature: Optional[float] = None,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
         scope: str = "memory",
         max_retries: int = 10,
         initial_backoff: float = 1.0,
         max_backoff: float = 60.0,
         skip_validation: bool = False,
+        strict_schema: bool = False,
     ) -> Any:
         """
         Make an LLM API call with retry logic.
@@ -146,6 +182,7 @@ class LLMProvider:
             initial_backoff: Initial backoff time in seconds.
             max_backoff: Maximum backoff time in seconds.
             skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (OpenAI only). Guarantees all required fields.
 
         Returns:
             Parsed response if response_format is provided, otherwise text content.
@@ -156,13 +193,38 @@ class LLMProvider:
         """
         async with _global_llm_semaphore:
             start_time = time.time()
-            import json
 
             # Handle Gemini provider separately
             if self.provider == "gemini":
                 return await self._call_gemini(
-                    messages, response_format, max_retries, initial_backoff,
-                    max_backoff, skip_validation, start_time
+                    messages, response_format, max_retries, initial_backoff, max_backoff, skip_validation, start_time
+                )
+
+            # Handle Anthropic provider separately
+            if self.provider == "anthropic":
+                return await self._call_anthropic(
+                    messages,
+                    response_format,
+                    max_completion_tokens,
+                    max_retries,
+                    initial_backoff,
+                    max_backoff,
+                    skip_validation,
+                    start_time,
+                )
+
+            # Handle Ollama with native API for structured output (better schema enforcement)
+            if self.provider == "ollama" and response_format is not None:
+                return await self._call_ollama_native(
+                    messages,
+                    response_format,
+                    max_completion_tokens,
+                    temperature,
+                    max_retries,
+                    initial_backoff,
+                    max_backoff,
+                    skip_validation,
+                    start_time,
                 )
 
             call_params = {
@@ -172,7 +234,7 @@ class LLMProvider:
 
             # Check if model supports reasoning parameter (o1, o3, gpt-5 families)
             model_lower = self.model.lower()
-            is_reasoning_model = any(x in model_lower for x in ["gpt-5", "o1", "o3"])
+            is_reasoning_model = any(x in model_lower for x in ["gpt-5", "o1", "o3", "deepseek"])
 
             # For GPT-4 and GPT-4.1 models, cap max_completion_tokens to 32000
             # For GPT-4o models, cap to 16384
@@ -194,7 +256,7 @@ class LLMProvider:
                 call_params["temperature"] = temperature
 
             # Set reasoning_effort for reasoning models (OpenAI gpt-5, o1, o3)
-            if is_reasoning_model and self.provider == "openai":
+            if is_reasoning_model:
                 call_params["reasoning_effort"] = self.reasoning_effort
 
             # Provider-specific parameters
@@ -203,7 +265,6 @@ class LLMProvider:
                 extra_body = {"service_tier": "auto"}
                 # Only add reasoning parameters for reasoning models
                 if is_reasoning_model:
-                    extra_body["reasoning_effort"] = self.reasoning_effort
                     extra_body["include_reasoning"] = False
                 call_params["extra_body"] = extra_body
 
@@ -212,21 +273,78 @@ class LLMProvider:
             for attempt in range(max_retries + 1):
                 try:
                     if response_format is not None:
-                        # Add schema to system message for JSON mode
-                        if hasattr(response_format, 'model_json_schema'):
+                        schema = None
+                        if hasattr(response_format, "model_json_schema"):
                             schema = response_format.model_json_schema()
-                            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
 
-                            if call_params['messages'] and call_params['messages'][0].get('role') == 'system':
-                                call_params['messages'][0]['content'] += schema_msg
-                            elif call_params['messages']:
-                                call_params['messages'][0]['content'] = schema_msg + "\n\n" + call_params['messages'][0]['content']
+                        if strict_schema and schema is not None:
+                            # Use OpenAI's strict JSON schema enforcement
+                            # This guarantees all required fields are returned
+                            call_params["response_format"] = {
+                                "type": "json_schema",
+                                "json_schema": {
+                                    "name": "response",
+                                    "strict": True,
+                                    "schema": schema,
+                                },
+                            }
+                        else:
+                            # Soft enforcement: add schema to prompt and use json_object mode
+                            if schema is not None:
+                                schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
 
-                        call_params['response_format'] = {"type": "json_object"}
+                                if call_params["messages"] and call_params["messages"][0].get("role") == "system":
+                                    call_params["messages"][0]["content"] += schema_msg
+                                elif call_params["messages"]:
+                                    call_params["messages"][0]["content"] = (
+                                        schema_msg + "\n\n" + call_params["messages"][0]["content"]
+                                    )
+                            if self.provider not in ("lmstudio", "ollama"):
+                                call_params["response_format"] = {"type": "json_object"}
+
+                        logger.debug(f"Sending request to {self.provider}/{self.model} (timeout={self.timeout})")
                         response = await self._client.chat.completions.create(**call_params)
+                        logger.debug(f"Received response from {self.provider}/{self.model}")
 
                         content = response.choices[0].message.content
-                        json_data = json.loads(content)
+
+                        # For local models, they may wrap JSON in markdown code blocks
+                        if self.provider in ("lmstudio", "ollama"):
+                            clean_content = content
+                            if "```json" in content:
+                                clean_content = content.split("```json")[1].split("```")[0].strip()
+                            elif "```" in content:
+                                clean_content = content.split("```")[1].split("```")[0].strip()
+                            try:
+                                json_data = json.loads(clean_content)
+                            except json.JSONDecodeError:
+                                # Fallback to parsing raw content
+                                json_data = json.loads(content)
+                        else:
+                            # Log raw LLM response for debugging JSON parse issues
+                            try:
+                                json_data = json.loads(content)
+                            except json.JSONDecodeError as json_err:
+                                # Truncate content for logging (first 500 and last 200 chars)
+                                content_preview = content[:500] if content else "<empty>"
+                                if content and len(content) > 700:
+                                    content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                                logger.warning(
+                                    f"JSON parse error from LLM response (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                                    f"  Model: {self.provider}/{self.model}\n"
+                                    f"  Content length: {len(content) if content else 0} chars\n"
+                                    f"  Content preview: {content_preview!r}\n"
+                                    f"  Finish reason: {response.choices[0].finish_reason if response.choices else 'unknown'}"
+                                )
+                                # Retry on JSON parse errors - LLM may return valid JSON on next attempt
+                                if attempt < max_retries:
+                                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                                    await asyncio.sleep(backoff)
+                                    last_exception = json_err
+                                    continue
+                                else:
+                                    logger.error(f"JSON parse error after {max_retries + 1} attempts, giving up")
+                                    raise
 
                         if skip_validation:
                             result = json_data
@@ -242,8 +360,8 @@ class LLMProvider:
                     if duration > 10.0:
                         ratio = max(1, usage.completion_tokens) / usage.prompt_tokens
                         cached_tokens = 0
-                        if hasattr(usage, 'prompt_tokens_details') and usage.prompt_tokens_details:
-                            cached_tokens = getattr(usage.prompt_tokens_details, 'cached_tokens', 0) or 0
+                        if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
+                            cached_tokens = getattr(usage.prompt_tokens_details, "cached_tokens", 0) or 0
                         cache_info = f", cached_tokens={cached_tokens}" if cached_tokens > 0 else ""
                         logger.info(
                             f"slow llm call: model={self.provider}/{self.model}, "
@@ -256,15 +374,19 @@ class LLMProvider:
                 except LengthFinishReasonError as e:
                     logger.warning(f"LLM output exceeded token limits: {str(e)}")
                     raise OutputTooLongError(
-                        f"LLM output exceeded token limits. Input may need to be split into smaller chunks."
+                        "LLM output exceeded token limits. Input may need to be split into smaller chunks."
                     ) from e
 
                 except APIConnectionError as e:
                     last_exception = e
                     if attempt < max_retries:
-                        status_code = getattr(e, 'status_code', None) or getattr(getattr(e, 'response', None), 'status_code', None)
-                        logger.warning(f"Connection error, retrying... (attempt {attempt + 1}/{max_retries + 1}) - status_code={status_code}, message={e}")
-                        backoff = min(initial_backoff * (2 ** attempt), max_backoff)
+                        status_code = getattr(e, "status_code", None) or getattr(
+                            getattr(e, "response", None), "status_code", None
+                        )
+                        logger.warning(
+                            f"Connection error, retrying... (attempt {attempt + 1}/{max_retries + 1}) - status_code={status_code}, message={e}"
+                        )
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
                         await asyncio.sleep(backoff)
                         continue
                     else:
@@ -279,7 +401,7 @@ class LLMProvider:
 
                     last_exception = e
                     if attempt < max_retries:
-                        backoff = min(initial_backoff * (2 ** attempt), max_backoff)
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
                         jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
                         sleep_time = backoff + jitter
                         await asyncio.sleep(sleep_time)
@@ -293,12 +415,271 @@ class LLMProvider:
 
             if last_exception:
                 raise last_exception
-            raise RuntimeError(f"LLM call failed after all retries with no exception captured")
+            raise RuntimeError("LLM call failed after all retries with no exception captured")
+
+    async def _call_anthropic(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None,
+        max_completion_tokens: int | None,
+        max_retries: int,
+        initial_backoff: float,
+        max_backoff: float,
+        skip_validation: bool,
+        start_time: float,
+    ) -> Any:
+        """Handle Anthropic-specific API calls."""
+        from anthropic import APIConnectionError, APIStatusError, RateLimitError
+
+        # Convert OpenAI-style messages to Anthropic format
+        system_prompt = None
+        anthropic_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                if system_prompt:
+                    system_prompt += "\n\n" + content
+                else:
+                    system_prompt = content
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            if system_prompt:
+                system_prompt += schema_msg
+            else:
+                system_prompt = schema_msg
+
+        # Prepare parameters
+        call_params = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "max_tokens": max_completion_tokens if max_completion_tokens is not None else 4096,
+        }
+
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._anthropic_client.messages.create(**call_params)
+
+                # Anthropic response content is a list of blocks
+                content = ""
+                for block in response.content:
+                    if block.type == "text":
+                        content += block.text
+
+                if response_format is not None:
+                    # Models may wrap JSON in markdown code blocks
+                    clean_content = content
+                    if "```json" in content:
+                        clean_content = content.split("```json")[1].split("```")[0].strip()
+                    elif "```" in content:
+                        clean_content = content.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_content)
+                    except json.JSONDecodeError:
+                        # Fallback to parsing raw content if markdown stripping failed
+                        json_data = json.loads(content)
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Log slow calls
+                duration = time.time() - start_time
+                if duration > 10.0:
+                    input_tokens = response.usage.input_tokens
+                    output_tokens = response.usage.output_tokens
+                    logger.info(
+                        f"slow llm call: model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"time={duration:.3f}s"
+                    )
+
+                return result
+
+            except json.JSONDecodeError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    logger.warning("Anthropic returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Anthropic returned invalid JSON after {max_retries + 1} attempts")
+                    raise
+
+            except (APIConnectionError, RateLimitError, APIStatusError) as e:
+                # Fast fail on 401/403
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    logger.error(f"Anthropic auth error (HTTP {e.status_code}), not retrying: {str(e)}")
+                    raise
+
+                last_exception = e
+                if attempt < max_retries:
+                    # Check if it's a rate limit or server error
+                    should_retry = isinstance(e, (APIConnectionError, RateLimitError)) or (
+                        isinstance(e, APIStatusError) and e.status_code >= 500
+                    )
+
+                    if should_retry:
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                        await asyncio.sleep(backoff + jitter)
+                        continue
+
+                logger.error(f"Anthropic API error after {max_retries + 1} attempts: {str(e)}")
+                raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Anthropic call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic call failed after all retries")
+
+    async def _call_ollama_native(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any,
+        max_completion_tokens: int | None,
+        temperature: float | None,
+        max_retries: int,
+        initial_backoff: float,
+        max_backoff: float,
+        skip_validation: bool,
+        start_time: float,
+    ) -> Any:
+        """
+        Call Ollama using native API with JSON schema enforcement.
+
+        Ollama's native API supports passing a full JSON schema in the 'format' parameter,
+        which provides better structured output control than the OpenAI-compatible API.
+        """
+        # Get the JSON schema from the Pydantic model
+        schema = response_format.model_json_schema() if hasattr(response_format, "model_json_schema") else None
+
+        # Build the base URL for Ollama's native API
+        # Default OpenAI-compatible URL is http://localhost:11434/v1
+        # Native API is at http://localhost:11434/api/chat
+        base_url = self.base_url or "http://localhost:11434/v1"
+        if base_url.endswith("/v1"):
+            native_url = base_url[:-3] + "/api/chat"
+        else:
+            native_url = base_url.rstrip("/") + "/api/chat"
+
+        # Build request payload
+        payload = {
+            "model": self.model,
+            "messages": messages,
+            "stream": False,
+        }
+
+        # Add schema as format parameter for structured output
+        if schema:
+            payload["format"] = schema
+
+        # Add optional parameters with optimized defaults for Ollama
+        # Benchmarking shows num_ctx=16384 + num_batch=512 is optimal
+        options = {
+            "num_ctx": 16384,  # 16k context window for larger prompts
+            "num_batch": 512,  # Optimal batch size for prompt processing
+        }
+        if max_completion_tokens:
+            options["num_predict"] = max_completion_tokens
+        if temperature is not None:
+            options["temperature"] = temperature
+        payload["options"] = options
+
+        last_exception = None
+
+        async with httpx.AsyncClient(timeout=300.0) as client:
+            for attempt in range(max_retries + 1):
+                try:
+                    response = await client.post(native_url, json=payload)
+                    response.raise_for_status()
+
+                    result = response.json()
+                    content = result.get("message", {}).get("content", "")
+
+                    # Parse JSON response
+                    try:
+                        json_data = json.loads(content)
+                    except json.JSONDecodeError as json_err:
+                        content_preview = content[:500] if content else "<empty>"
+                        if content and len(content) > 700:
+                            content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                        logger.warning(
+                            f"Ollama JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                            f"  Model: ollama/{self.model}\n"
+                            f"  Content length: {len(content) if content else 0} chars\n"
+                            f"  Content preview: {content_preview!r}"
+                        )
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = json_err
+                            continue
+                        else:
+                            raise
+
+                    # Validate against Pydantic model or return raw JSON
+                    if skip_validation:
+                        return json_data
+                    else:
+                        return response_format.model_validate(json_data)
+
+                except httpx.HTTPStatusError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(
+                            f"Ollama HTTP error (attempt {attempt + 1}/{max_retries + 1}): {e.response.status_code}"
+                        )
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama HTTP error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except httpx.RequestError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(f"Ollama connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama connection error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except Exception as e:
+                    logger.error(f"Unexpected error during Ollama call: {type(e).__name__}: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Ollama call failed after all retries")
 
     async def _call_gemini(
         self,
-        messages: List[Dict[str, str]],
-        response_format: Optional[Any],
+        messages: list[dict[str, str]],
+        response_format: Any | None,
         max_retries: int,
         initial_backoff: float,
         max_backoff: float,
@@ -306,34 +687,26 @@ class LLMProvider:
         start_time: float,
     ) -> Any:
         """Handle Gemini-specific API calls."""
-        import json
-
         # Convert OpenAI-style messages to Gemini format
         system_instruction = None
         gemini_contents = []
 
         for msg in messages:
-            role = msg.get('role', 'user')
-            content = msg.get('content', '')
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
 
-            if role == 'system':
+            if role == "system":
                 if system_instruction:
                     system_instruction += "\n\n" + content
                 else:
                     system_instruction = content
-            elif role == 'assistant':
-                gemini_contents.append(genai_types.Content(
-                    role="model",
-                    parts=[genai_types.Part(text=content)]
-                ))
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
             else:
-                gemini_contents.append(genai_types.Content(
-                    role="user",
-                    parts=[genai_types.Part(text=content)]
-                ))
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
 
         # Add JSON schema instruction if response_format is provided
-        if response_format is not None and hasattr(response_format, 'model_json_schema'):
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
             schema = response_format.model_json_schema()
             schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
             if system_instruction:
@@ -344,10 +717,10 @@ class LLMProvider:
         # Build generation config
         config_kwargs = {}
         if system_instruction:
-            config_kwargs['system_instruction'] = system_instruction
+            config_kwargs["system_instruction"] = system_instruction
         if response_format is not None:
-            config_kwargs['response_mime_type'] = 'application/json'
-            config_kwargs['response_schema'] = response_format
+            config_kwargs["response_mime_type"] = "application/json"
+            config_kwargs["response_schema"] = response_format
 
         generation_config = genai_types.GenerateContentConfig(**config_kwargs) if config_kwargs else None
 
@@ -366,14 +739,14 @@ class LLMProvider:
                 # Handle empty response
                 if content is None:
                     block_reason = None
-                    if hasattr(response, 'candidates') and response.candidates:
+                    if hasattr(response, "candidates") and response.candidates:
                         candidate = response.candidates[0]
-                        if hasattr(candidate, 'finish_reason'):
+                        if hasattr(candidate, "finish_reason"):
                             block_reason = candidate.finish_reason
 
                     if attempt < max_retries:
                         logger.warning(f"Gemini returned empty response (reason: {block_reason}), retrying...")
-                        backoff = min(initial_backoff * (2 ** attempt), max_backoff)
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
                         await asyncio.sleep(backoff)
                         continue
                     else:
@@ -390,7 +763,7 @@ class LLMProvider:
 
                 # Log slow calls
                 duration = time.time() - start_time
-                if duration > 10.0 and hasattr(response, 'usage_metadata') and response.usage_metadata:
+                if duration > 10.0 and hasattr(response, "usage_metadata") and response.usage_metadata:
                     usage = response.usage_metadata
                     logger.info(
                         f"slow llm call: model={self.provider}/{self.model}, "
@@ -403,8 +776,8 @@ class LLMProvider:
             except json.JSONDecodeError as e:
                 last_exception = e
                 if attempt < max_retries:
-                    logger.warning(f"Gemini returned invalid JSON, retrying...")
-                    backoff = min(initial_backoff * (2 ** attempt), max_backoff)
+                    logger.warning("Gemini returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
                     await asyncio.sleep(backoff)
                     continue
                 else:
@@ -421,7 +794,7 @@ class LLMProvider:
                 if e.code in (400, 429, 500, 502, 503, 504) or (e.code and e.code >= 500):
                     last_exception = e
                     if attempt < max_retries:
-                        backoff = min(initial_backoff * (2 ** attempt), max_backoff)
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
                         jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
                         await asyncio.sleep(backoff + jitter)
                     else:
@@ -437,55 +810,47 @@ class LLMProvider:
 
         if last_exception:
             raise last_exception
-        raise RuntimeError(f"Gemini call failed after all retries")
+        raise RuntimeError("Gemini call failed after all retries")
 
     @classmethod
     def for_memory(cls) -> "LLMProvider":
         """Create provider for memory operations from environment variables."""
         provider = os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq")
         api_key = os.getenv("HINDSIGHT_API_LLM_API_KEY")
+        if not api_key:
+            raise ValueError("HINDSIGHT_API_LLM_API_KEY environment variable is required")
         base_url = os.getenv("HINDSIGHT_API_LLM_BASE_URL", "")
         model = os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b")
 
-        return cls(
-            provider=provider,
-            api_key=api_key,
-            base_url=base_url,
-            model=model,
-            reasoning_effort="low"
-        )
+        return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="low")
 
     @classmethod
     def for_answer_generation(cls) -> "LLMProvider":
         """Create provider for answer generation. Falls back to memory config if not set."""
         provider = os.getenv("HINDSIGHT_API_ANSWER_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
         api_key = os.getenv("HINDSIGHT_API_ANSWER_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY"))
+        if not api_key:
+            raise ValueError(
+                "HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_ANSWER_LLM_API_KEY environment variable is required"
+            )
         base_url = os.getenv("HINDSIGHT_API_ANSWER_LLM_BASE_URL", os.getenv("HINDSIGHT_API_LLM_BASE_URL", ""))
         model = os.getenv("HINDSIGHT_API_ANSWER_LLM_MODEL", os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b"))
 
-        return cls(
-            provider=provider,
-            api_key=api_key,
-            base_url=base_url,
-            model=model,
-            reasoning_effort="high"
-        )
+        return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="high")
 
     @classmethod
     def for_judge(cls) -> "LLMProvider":
         """Create provider for judge/evaluator operations. Falls back to memory config if not set."""
         provider = os.getenv("HINDSIGHT_API_JUDGE_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
         api_key = os.getenv("HINDSIGHT_API_JUDGE_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY"))
+        if not api_key:
+            raise ValueError(
+                "HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_JUDGE_LLM_API_KEY environment variable is required"
+            )
         base_url = os.getenv("HINDSIGHT_API_JUDGE_LLM_BASE_URL", os.getenv("HINDSIGHT_API_LLM_BASE_URL", ""))
         model = os.getenv("HINDSIGHT_API_JUDGE_LLM_MODEL", os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b"))
 
-        return cls(
-            provider=provider,
-            api_key=api_key,
-            base_url=base_url,
-            model=model,
-            reasoning_effort="high"
-        )
+        return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="high")
 
 
 # Backwards compatibility alias

hindsight-api/hindsight_api/engine/query_analyzer.py

@@ -4,11 +4,12 @@ Query analysis abstraction for the memory system.
 Provides an interface for analyzing natural language queries to extract
 structured information like temporal constraints.
 """
-from abc import ABC, abstractmethod
-from typing import Optional
-from datetime import datetime, timedelta
+
 import logging
 import re
+from abc import ABC, abstractmethod
+from datetime import datetime, timedelta
+
 from pydantic import BaseModel, Field
 
 logger = logging.getLogger(__name__)
@@ -20,6 +21,7 @@ class TemporalConstraint(BaseModel):
 
     Represents a time range with start and end dates.
     """
+
     start_date: datetime = Field(description="Start of the time range (inclusive)")
     end_date: datetime = Field(description="End of the time range (inclusive)")
 
@@ -33,9 +35,9 @@ class QueryAnalysis(BaseModel):
 
     Contains extracted structured information like temporal constraints.
     """
-    temporal_constraint: Optional[TemporalConstraint] = Field(
-        default=None,
-        description="Extracted temporal constraint, if any"
+
+    temporal_constraint: TemporalConstraint | None = Field(
+        default=None, description="Extracted temporal constraint, if any"
     )
 
 
@@ -58,9 +60,7 @@ class QueryAnalyzer(ABC):
         pass
 
     @abstractmethod
-    def analyze(
-        self, query: str, reference_date: Optional[datetime] = None
-    ) -> QueryAnalysis:
+    def analyze(self, query: str, reference_date: datetime | None = None) -> QueryAnalysis:
         """
         Analyze a natural language query.
 
@@ -95,11 +95,10 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
         """Load dateparser (lazy import)."""
         if self._search_dates is None:
             from dateparser.search import search_dates
+
             self._search_dates = search_dates
 
-    def analyze(
-        self, query: str, reference_date: Optional[datetime] = None
-    ) -> QueryAnalysis:
+    def analyze(self, query: str, reference_date: datetime | None = None) -> QueryAnalysis:
         """
         Analyze query using dateparser.
 
@@ -126,9 +125,9 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
 
         # Use dateparser's search_dates to find temporal expressions
         settings = {
-            'RELATIVE_BASE': reference_date,
-            'PREFER_DATES_FROM': 'past',
-            'RETURN_AS_TIMEZONE_AWARE': False,
+            "RELATIVE_BASE": reference_date,
+            "PREFER_DATES_FROM": "past",
+            "RETURN_AS_TIMEZONE_AWARE": False,
         }
 
         results = self._search_dates(query, settings=settings)
@@ -137,11 +136,8 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
             return QueryAnalysis(temporal_constraint=None)
 
         # Filter out false positives (common words parsed as dates)
-        false_positives = {'do', 'may', 'march', 'will', 'can', 'sat', 'sun', 'mon', 'tue', 'wed', 'thu', 'fri'}
-        valid_results = [
-            (text, date) for text, date in results
-            if text.lower() not in false_positives or len(text) > 3
-        ]
+        false_positives = {"do", "may", "march", "will", "can", "sat", "sun", "mon", "tue", "wed", "thu", "fri"}
+        valid_results = [(text, date) for text, date in results if text.lower() not in false_positives or len(text) > 3]
 
         if not valid_results:
             return QueryAnalysis(temporal_constraint=None)
@@ -153,84 +149,94 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
         start_date = parsed_date.replace(hour=0, minute=0, second=0, microsecond=0)
         end_date = parsed_date.replace(hour=23, minute=59, second=59, microsecond=999999)
 
-        return QueryAnalysis(
-            temporal_constraint=TemporalConstraint(
-                start_date=start_date,
-                end_date=end_date
-            )
-        )
+        return QueryAnalysis(temporal_constraint=TemporalConstraint(start_date=start_date, end_date=end_date))
 
-    def _extract_period(
-        self, query: str, reference_date: datetime
-    ) -> Optional[TemporalConstraint]:
+    def _extract_period(self, query: str, reference_date: datetime) -> TemporalConstraint | None:
         """
         Extract period-based temporal expressions (week, month, year, weekend).
 
         These need special handling as they represent date ranges, not single dates.
         Supports multiple languages.
         """
+
         def constraint(start: datetime, end: datetime) -> TemporalConstraint:
             return TemporalConstraint(
                 start_date=start.replace(hour=0, minute=0, second=0, microsecond=0),
-                end_date=end.replace(hour=23, minute=59, second=59, microsecond=999999)
+                end_date=end.replace(hour=23, minute=59, second=59, microsecond=999999),
             )
 
         # Yesterday patterns (English, Spanish, Italian, French, German)
-        if re.search(r'\b(yesterday|ayer|ieri|hier|gestern)\b', query, re.IGNORECASE):
+        if re.search(r"\b(yesterday|ayer|ieri|hier|gestern)\b", query, re.IGNORECASE):
             d = reference_date - timedelta(days=1)
             return constraint(d, d)
 
         # Today patterns
-        if re.search(r'\b(today|hoy|oggi|aujourd\'?hui|heute)\b', query, re.IGNORECASE):
+        if re.search(r"\b(today|hoy|oggi|aujourd\'?hui|heute)\b", query, re.IGNORECASE):
             return constraint(reference_date, reference_date)
 
         # "a couple of days ago" / "a few days ago" patterns
         # These are imprecise so we create a range
-        if re.search(r'\b(a\s+)?couple\s+(of\s+)?days?\s+ago\b', query, re.IGNORECASE):
+        if re.search(r"\b(a\s+)?couple\s+(of\s+)?days?\s+ago\b", query, re.IGNORECASE):
             # "a couple of days" = approximately 2 days, give range of 1-3 days
             return constraint(reference_date - timedelta(days=3), reference_date - timedelta(days=1))
 
-        if re.search(r'\b(a\s+)?few\s+days?\s+ago\b', query, re.IGNORECASE):
+        if re.search(r"\b(a\s+)?few\s+days?\s+ago\b", query, re.IGNORECASE):
             # "a few days" = approximately 3-4 days, give range of 2-5 days
             return constraint(reference_date - timedelta(days=5), reference_date - timedelta(days=2))
 
         # "a couple of weeks ago" / "a few weeks ago" patterns
-        if re.search(r'\b(a\s+)?couple\s+(of\s+)?weeks?\s+ago\b', query, re.IGNORECASE):
+        if re.search(r"\b(a\s+)?couple\s+(of\s+)?weeks?\s+ago\b", query, re.IGNORECASE):
             # "a couple of weeks" = approximately 2 weeks, give range of 1-3 weeks
             return constraint(reference_date - timedelta(weeks=3), reference_date - timedelta(weeks=1))
 
-        if re.search(r'\b(a\s+)?few\s+weeks?\s+ago\b', query, re.IGNORECASE):
+        if re.search(r"\b(a\s+)?few\s+weeks?\s+ago\b", query, re.IGNORECASE):
             # "a few weeks" = approximately 3-4 weeks, give range of 2-5 weeks
             return constraint(reference_date - timedelta(weeks=5), reference_date - timedelta(weeks=2))
 
         # "a couple of months ago" / "a few months ago" patterns
-        if re.search(r'\b(a\s+)?couple\s+(of\s+)?months?\s+ago\b', query, re.IGNORECASE):
+        if re.search(r"\b(a\s+)?couple\s+(of\s+)?months?\s+ago\b", query, re.IGNORECASE):
             # "a couple of months" = approximately 2 months, give range of 1-3 months
             return constraint(reference_date - timedelta(days=90), reference_date - timedelta(days=30))
 
-        if re.search(r'\b(a\s+)?few\s+months?\s+ago\b', query, re.IGNORECASE):
+        if re.search(r"\b(a\s+)?few\s+months?\s+ago\b", query, re.IGNORECASE):
             # "a few months" = approximately 3-4 months, give range of 2-5 months
             return constraint(reference_date - timedelta(days=150), reference_date - timedelta(days=60))
 
         # Last week patterns (English, Spanish, Italian, French, German)
-        if re.search(r'\b(last\s+week|la\s+semana\s+pasada|la\s+settimana\s+scorsa|la\s+semaine\s+derni[eè]re|letzte\s+woche)\b', query, re.IGNORECASE):
+        if re.search(
+            r"\b(last\s+week|la\s+semana\s+pasada|la\s+settimana\s+scorsa|la\s+semaine\s+derni[eè]re|letzte\s+woche)\b",
+            query,
+            re.IGNORECASE,
+        ):
             start = reference_date - timedelta(days=reference_date.weekday() + 7)
             return constraint(start, start + timedelta(days=6))
 
         # Last month patterns
-        if re.search(r'\b(last\s+month|el\s+mes\s+pasado|il\s+mese\s+scorso|le\s+mois\s+dernier|letzten?\s+monat)\b', query, re.IGNORECASE):
+        if re.search(
+            r"\b(last\s+month|el\s+mes\s+pasado|il\s+mese\s+scorso|le\s+mois\s+dernier|letzten?\s+monat)\b",
+            query,
+            re.IGNORECASE,
+        ):
             first = reference_date.replace(day=1)
             end = first - timedelta(days=1)
             start = end.replace(day=1)
             return constraint(start, end)
 
         # Last year patterns
-        if re.search(r'\b(last\s+year|el\s+a[ñn]o\s+pasado|l\'anno\s+scorso|l\'ann[ée]e\s+derni[eè]re|letztes?\s+jahr)\b', query, re.IGNORECASE):
+        if re.search(
+            r"\b(last\s+year|el\s+a[ñn]o\s+pasado|l\'anno\s+scorso|l\'ann[ée]e\s+derni[eè]re|letztes?\s+jahr)\b",
+            query,
+            re.IGNORECASE,
+        ):
             year = reference_date.year - 1
             return constraint(datetime(year, 1, 1), datetime(year, 12, 31))
 
         # Last weekend patterns
-        if re.search(r'\b(last\s+weekend|el\s+fin\s+de\s+semana\s+pasado|lo\s+scorso\s+fine\s+settimana|le\s+week-?end\s+dernier|letztes?\s+wochenende)\b', query, re.IGNORECASE):
+        if re.search(
+            r"\b(last\s+weekend|el\s+fin\s+de\s+semana\s+pasado|lo\s+scorso\s+fine\s+settimana|le\s+week-?end\s+dernier|letztes?\s+wochenende)\b",
+            query,
+            re.IGNORECASE,
+        ):
             days_since_sat = (reference_date.weekday() + 2) % 7
             if days_since_sat == 0:
                 days_since_sat = 7
@@ -239,22 +245,22 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
 
         # Month + Year patterns (e.g., "June 2024", "junio 2024", "giugno 2024")
         month_patterns = {
-            'january|enero|gennaio|janvier|januar': 1,
-            'february|febrero|febbraio|f[ée]vrier|februar': 2,
-            'march|marzo|mars|m[äa]rz': 3,
-            'april|abril|aprile|avril': 4,
-            'may|mayo|maggio|mai': 5,
-            'june|junio|giugno|juin|juni': 6,
-            'july|julio|luglio|juillet|juli': 7,
-            'august|agosto|ao[uû]t': 8,
-            'september|septiembre|settembre|septembre': 9,
-            'october|octubre|ottobre|octobre|oktober': 10,
-            'november|noviembre|novembre': 11,
-            'december|diciembre|dicembre|d[ée]cembre|dezember': 12,
+            "january|enero|gennaio|janvier|januar": 1,
+            "february|febrero|febbraio|f[ée]vrier|februar": 2,
+            "march|marzo|mars|m[äa]rz": 3,
+            "april|abril|aprile|avril": 4,
+            "may|mayo|maggio|mai": 5,
+            "june|junio|giugno|juin|juni": 6,
+            "july|julio|luglio|juillet|juli": 7,
+            "august|agosto|ao[uû]t": 8,
+            "september|septiembre|settembre|septembre": 9,
+            "october|octubre|ottobre|octobre|oktober": 10,
+            "november|noviembre|novembre": 11,
+            "december|diciembre|dicembre|d[ée]cembre|dezember": 12,
         }
 
         for pattern, month_num in month_patterns.items():
-            match = re.search(rf'\b({pattern})\s+(\d{{4}})\b', query, re.IGNORECASE)
+            match = re.search(rf"\b({pattern})\s+(\d{{4}})\b", query, re.IGNORECASE)
             if match:
                 year = int(match.group(2))
                 start = datetime(year, month_num, 1)
@@ -279,11 +285,7 @@ class TransformerQueryAnalyzer(QueryAnalyzer):
     - Model size: ~80M params (~300MB download)
     """
 
-    def __init__(
-        self,
-        model_name: str = "google/flan-t5-small",
-        device: str = "cpu"
-    ):
+    def __init__(self, model_name: str = "google/flan-t5-small", device: str = "cpu"):
         """
         Initialize T5 query analyzer.
 
@@ -304,11 +306,10 @@ class TransformerQueryAnalyzer(QueryAnalyzer):
             return
 
         try:
-            from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+            from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
         except ImportError:
             raise ImportError(
-                "transformers is required for TransformerQueryAnalyzer. "
-                "Install it with: pip install transformers"
+                "transformers is required for TransformerQueryAnalyzer. Install it with: pip install transformers"
             )
 
         logger.info(f"Loading query analyzer model: {self.model_name}...")
@@ -322,9 +323,7 @@ class TransformerQueryAnalyzer(QueryAnalyzer):
         """Lazy load the T5 model for temporal extraction (calls load())."""
         self.load()
 
-    def _extract_with_rules(
-        self, query: str, reference_date: datetime
-    ) -> Optional[TemporalConstraint]:
+    def _extract_with_rules(self, query: str, reference_date: datetime) -> TemporalConstraint | None:
         """
         Extract temporal expressions using rule-based patterns.
 
@@ -332,6 +331,7 @@ class TransformerQueryAnalyzer(QueryAnalyzer):
         patterns that need model-based extraction.
         """
         import re
+
         query_lower = query.lower()
 
         def get_last_weekday(weekday: int) -> datetime:
@@ -343,50 +343,60 @@ class TransformerQueryAnalyzer(QueryAnalyzer):
         def constraint(start: datetime, end: datetime) -> TemporalConstraint:
             return TemporalConstraint(
                 start_date=start.replace(hour=0, minute=0, second=0, microsecond=0),
-                end_date=end.replace(hour=23, minute=59, second=59, microsecond=999999)
+                end_date=end.replace(hour=23, minute=59, second=59, microsecond=999999),
             )
 
         # Yesterday
-        if re.search(r'\byesterday\b', query_lower):
+        if re.search(r"\byesterday\b", query_lower):
             d = reference_date - timedelta(days=1)
             return constraint(d, d)
 
         # Last week
-        if re.search(r'\blast\s+week\b', query_lower):
+        if re.search(r"\blast\s+week\b", query_lower):
             start = reference_date - timedelta(days=reference_date.weekday() + 7)
             return constraint(start, start + timedelta(days=6))
 
         # Last month
-        if re.search(r'\blast\s+month\b', query_lower):
+        if re.search(r"\blast\s+month\b", query_lower):
             first = reference_date.replace(day=1)
             end = first - timedelta(days=1)
             start = end.replace(day=1)
             return constraint(start, end)
 
         # Last year
-        if re.search(r'\blast\s+year\b', query_lower):
+        if re.search(r"\blast\s+year\b", query_lower):
             y = reference_date.year - 1
             return constraint(datetime(y, 1, 1), datetime(y, 12, 31))
 
         # Last weekend
-        if re.search(r'\blast\s+weekend\b', query_lower):
+        if re.search(r"\blast\s+weekend\b", query_lower):
             sat = get_last_weekday(5)
             return constraint(sat, sat + timedelta(days=1))
 
         # Last <weekday>
-        weekdays = {'monday': 0, 'tuesday': 1, 'wednesday': 2, 'thursday': 3,
-                    'friday': 4, 'saturday': 5, 'sunday': 6}
+        weekdays = {"monday": 0, "tuesday": 1, "wednesday": 2, "thursday": 3, "friday": 4, "saturday": 5, "sunday": 6}
         for name, num in weekdays.items():
-            if re.search(rf'\blast\s+{name}\b', query_lower):
+            if re.search(rf"\blast\s+{name}\b", query_lower):
                 d = get_last_weekday(num)
                 return constraint(d, d)
 
         # Month + Year: "June 2024", "in March 2023"
-        months = {'january': 1, 'february': 2, 'march': 3, 'april': 4, 'may': 5,
-                  'june': 6, 'july': 7, 'august': 8, 'september': 9, 'october': 10,
-                  'november': 11, 'december': 12}
+        months = {
+            "january": 1,
+            "february": 2,
+            "march": 3,
+            "april": 4,
+            "may": 5,
+            "june": 6,
+            "july": 7,
+            "august": 8,
+            "september": 9,
+            "october": 10,
+            "november": 11,
+            "december": 12,
+        }
         for name, num in months.items():
-            match = re.search(rf'\b{name}\s+(\d{{4}})\b', query_lower)
+            match = re.search(rf"\b{name}\s+(\d{{4}})\b", query_lower)
             if match:
                 year = int(match.group(1))
                 if num == 12:
@@ -397,9 +407,7 @@ class TransformerQueryAnalyzer(QueryAnalyzer):
 
         return None
 
-    def analyze(
-        self, query: str, reference_date: Optional[datetime] = None
-    ) -> QueryAnalysis:
+    def analyze(self, query: str, reference_date: datetime | None = None) -> QueryAnalysis:
         """
         Analyze query for temporal expressions.
 
@@ -435,11 +443,11 @@ class TransformerQueryAnalyzer(QueryAnalyzer):
         last_saturday = get_last_weekday(5)
 
         # Build prompt for T5
-        prompt = f"""Today is {reference_date.strftime('%Y-%m-%d')}. Extract date range or "none".
+        prompt = f"""Today is {reference_date.strftime("%Y-%m-%d")}. Extract date range or "none".
 
 June 2024 = 2024-06-01 to 2024-06-30
-yesterday = {yesterday.strftime('%Y-%m-%d')} to {yesterday.strftime('%Y-%m-%d')}
-last Saturday = {last_saturday.strftime('%Y-%m-%d')} to {last_saturday.strftime('%Y-%m-%d')}
+yesterday = {yesterday.strftime("%Y-%m-%d")} to {yesterday.strftime("%Y-%m-%d")}
+last Saturday = {last_saturday.strftime("%Y-%m-%d")} to {last_saturday.strftime("%Y-%m-%d")}
 what is the weather = none
 {query} ="""
 
@@ -448,13 +456,7 @@ what is the weather = none
         inputs = {k: v.to(self.device) for k, v in inputs.items()}
 
         with self._no_grad():
-            outputs = self._model.generate(
-                **inputs,
-                max_new_tokens=30,
-                num_beams=3,
-                do_sample=False,
-                temperature=1.0
-            )
+            outputs = self._model.generate(**inputs, max_new_tokens=30, num_beams=3, do_sample=False, temperature=1.0)
 
         result = self._tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
 
@@ -466,14 +468,14 @@ what is the weather = none
         """Get torch.no_grad context manager."""
         try:
             import torch
+
             return torch.no_grad()
         except ImportError:
             from contextlib import nullcontext
+
             return nullcontext()
 
-    def _parse_generated_output(
-        self, result: str, reference_date: datetime
-    ) -> Optional[TemporalConstraint]:
+    def _parse_generated_output(self, result: str, reference_date: datetime) -> TemporalConstraint | None:
         """
         Parse T5 generated output into TemporalConstraint.
 
@@ -492,7 +494,8 @@ what is the weather = none
         try:
             # Parse "YYYY-MM-DD to YYYY-MM-DD"
             import re
-            pattern = r'(\d{4}-\d{2}-\d{2})\s+to\s+(\d{4}-\d{2}-\d{2})'
+
+            pattern = r"(\d{4}-\d{2}-\d{2})\s+to\s+(\d{4}-\d{2}-\d{2})"
             match = re.search(pattern, result, re.IGNORECASE)
 
             if match:
@@ -513,7 +516,7 @@ what is the weather = none
 
                 return TemporalConstraint(start_date=start_date, end_date=end_date)
 
-        except (ValueError, AttributeError) as e:
+        except (ValueError, AttributeError):
             return None
 
         return None

hindsight-api/hindsight_api/engine/response_models.py

@@ -6,9 +6,9 @@ API response models should be kept separate and convert from these core models t
 API stability even if internal models change.
 """
 
-from typing import Optional, List, Dict, Any
-from pydantic import BaseModel, Field, ConfigDict
+from typing import Any
 
+from pydantic import BaseModel, ConfigDict, Field
 
 # Valid fact types for recall operations (excludes 'observation' which is internal)
 VALID_RECALL_FACT_TYPES = frozenset(["world", "experience", "opinion"])
@@ -23,17 +23,12 @@ class DispositionTraits(BaseModel):
     - literalism: 1=flexible interpretation, 5=literal interpretation (how strictly to interpret information)
     - empathy: 1=detached, 5=empathetic (how much to consider emotional context)
     """
+
     skepticism: int = Field(ge=1, le=5, description="How skeptical vs trusting (1=trusting, 5=skeptical)")
     literalism: int = Field(ge=1, le=5, description="How literally to interpret information (1=flexible, 5=literal)")
     empathy: int = Field(ge=1, le=5, description="How much to consider emotional context (1=detached, 5=empathetic)")
 
-    model_config = ConfigDict(json_schema_extra={
-        "example": {
-            "skepticism": 3,
-            "literalism": 3,
-            "empathy": 3
-        }
-    })
+    model_config = ConfigDict(json_schema_extra={"example": {"skepticism": 3, "literalism": 3, "empathy": 3}})
 
 
 class MemoryFact(BaseModel):
@@ -43,38 +38,44 @@ class MemoryFact(BaseModel):
     This represents a unit of information stored in the memory system,
     including both the content and metadata.
     """
-    model_config = ConfigDict(json_schema_extra={
-        "example": {
-            "id": "123e4567-e89b-12d3-a456-426614174000",
-            "text": "Alice works at Google on the AI team",
-            "fact_type": "world",
-            "entities": ["Alice", "Google"],
-            "context": "work info",
-            "occurred_start": "2024-01-15T10:30:00Z",
-            "occurred_end": "2024-01-15T10:30:00Z",
-            "mentioned_at": "2024-01-15T10:30:00Z",
-            "document_id": "session_abc123",
-            "metadata": {"source": "slack"},
-            "chunk_id": "bank123_session_abc123_0",
-            "activation": 0.95
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "id": "123e4567-e89b-12d3-a456-426614174000",
+                "text": "Alice works at Google on the AI team",
+                "fact_type": "world",
+                "entities": ["Alice", "Google"],
+                "context": "work info",
+                "occurred_start": "2024-01-15T10:30:00Z",
+                "occurred_end": "2024-01-15T10:30:00Z",
+                "mentioned_at": "2024-01-15T10:30:00Z",
+                "document_id": "session_abc123",
+                "metadata": {"source": "slack"},
+                "chunk_id": "bank123_session_abc123_0",
+                "activation": 0.95,
+            }
         }
-    })
+    )
 
     id: str = Field(description="Unique identifier for the memory fact")
     text: str = Field(description="The actual text content of the memory")
     fact_type: str = Field(description="Type of fact: 'world', 'experience', 'opinion', or 'observation'")
-    entities: Optional[List[str]] = Field(None, description="Entity names mentioned in this fact")
-    context: Optional[str] = Field(None, description="Additional context for the memory")
-    occurred_start: Optional[str] = Field(None, description="ISO format date when the event started occurring")
-    occurred_end: Optional[str] = Field(None, description="ISO format date when the event ended occurring")
-    mentioned_at: Optional[str] = Field(None, description="ISO format date when the fact was mentioned/learned")
-    document_id: Optional[str] = Field(None, description="ID of the document this memory belongs to")
-    metadata: Optional[Dict[str, str]] = Field(None, description="User-defined metadata")
-    chunk_id: Optional[str] = Field(None, description="ID of the chunk this fact was extracted from (format: bank_id_document_id_chunk_index)")
+    entities: list[str] | None = Field(None, description="Entity names mentioned in this fact")
+    context: str | None = Field(None, description="Additional context for the memory")
+    occurred_start: str | None = Field(None, description="ISO format date when the event started occurring")
+    occurred_end: str | None = Field(None, description="ISO format date when the event ended occurring")
+    mentioned_at: str | None = Field(None, description="ISO format date when the fact was mentioned/learned")
+    document_id: str | None = Field(None, description="ID of the document this memory belongs to")
+    metadata: dict[str, str] | None = Field(None, description="User-defined metadata")
+    chunk_id: str | None = Field(
+        None, description="ID of the chunk this fact was extracted from (format: bank_id_document_id_chunk_index)"
+    )
 
 
 class ChunkInfo(BaseModel):
     """Information about a chunk."""
+
     chunk_text: str = Field(description="The raw chunk text")
     chunk_index: int = Field(description="Index of the chunk within the document")
     truncated: bool = Field(default=False, description="Whether the chunk was truncated due to token limits")
@@ -87,35 +88,33 @@ class RecallResult(BaseModel):
     Contains a list of matching memory facts and optional trace information
     for debugging and transparency.
     """
-    model_config = ConfigDict(json_schema_extra={
-        "example": {
-            "results": [
-                {
-                    "id": "123e4567-e89b-12d3-a456-426614174000",
-                    "text": "Alice works at Google on the AI team",
-                    "fact_type": "world",
-                    "context": "work info",
-                    "occurred_start": "2024-01-15T10:30:00Z",
-                    "occurred_end": "2024-01-15T10:30:00Z",
-                    "activation": 0.95
-                }
-            ],
-            "trace": {
-                "query": "What did Alice say about machine learning?",
-                "num_results": 1
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "results": [
+                    {
+                        "id": "123e4567-e89b-12d3-a456-426614174000",
+                        "text": "Alice works at Google on the AI team",
+                        "fact_type": "world",
+                        "context": "work info",
+                        "occurred_start": "2024-01-15T10:30:00Z",
+                        "occurred_end": "2024-01-15T10:30:00Z",
+                        "activation": 0.95,
+                    }
+                ],
+                "trace": {"query": "What did Alice say about machine learning?", "num_results": 1},
             }
         }
-    })
-
-    results: List[MemoryFact] = Field(description="List of memory facts matching the query")
-    trace: Optional[Dict[str, Any]] = Field(None, description="Trace information for debugging")
-    entities: Optional[Dict[str, "EntityState"]] = Field(
-        None,
-        description="Entity states for entities mentioned in results (keyed by canonical name)"
     )
-    chunks: Optional[Dict[str, ChunkInfo]] = Field(
-        None,
-        description="Chunks for facts, keyed by '{document_id}_{chunk_index}'"
+
+    results: list[MemoryFact] = Field(description="List of memory facts matching the query")
+    trace: dict[str, Any] | None = Field(None, description="Trace information for debugging")
+    entities: dict[str, "EntityState"] | None = Field(
+        None, description="Entity states for entities mentioned in results (keyed by canonical name)"
+    )
+    chunks: dict[str, ChunkInfo] | None = Field(
+        None, description="Chunks for facts, keyed by '{document_id}_{chunk_index}'"
     )
 
 
@@ -124,38 +123,42 @@ class ReflectResult(BaseModel):
     Result from a reflect operation.
 
     Contains the formulated answer, the facts it was based on (organized by type),
-    and any new opinions that were formed during the reflection process.
+    any new opinions that were formed during the reflection process, and optionally
+    structured output if a response schema was provided.
     """
-    model_config = ConfigDict(json_schema_extra={
-        "example": {
-            "text": "Based on my knowledge, machine learning is being actively used in healthcare...",
-            "based_on": {
-                "world": [
-                    {
-                        "id": "123e4567-e89b-12d3-a456-426614174000",
-                        "text": "Machine learning is used in medical diagnosis",
-                        "fact_type": "world",
-                        "context": "healthcare",
-                        "occurred_start": "2024-01-15T10:30:00Z",
-                        "occurred_end": "2024-01-15T10:30:00Z"
-                    }
-                ],
-                "experience": [],
-                "opinion": []
-            },
-            "new_opinions": [
-                "Machine learning has great potential in healthcare"
-            ]
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "text": "Based on my knowledge, machine learning is being actively used in healthcare...",
+                "based_on": {
+                    "world": [
+                        {
+                            "id": "123e4567-e89b-12d3-a456-426614174000",
+                            "text": "Machine learning is used in medical diagnosis",
+                            "fact_type": "world",
+                            "context": "healthcare",
+                            "occurred_start": "2024-01-15T10:30:00Z",
+                            "occurred_end": "2024-01-15T10:30:00Z",
+                        }
+                    ],
+                    "experience": [],
+                    "opinion": [],
+                },
+                "new_opinions": ["Machine learning has great potential in healthcare"],
+                "structured_output": {"summary": "ML in healthcare", "confidence": 0.9},
+            }
         }
-    })
+    )
 
     text: str = Field(description="The formulated answer text")
-    based_on: Dict[str, List[MemoryFact]] = Field(
+    based_on: dict[str, list[MemoryFact]] = Field(
         description="Facts used to formulate the answer, organized by type (world, experience, opinion)"
     )
-    new_opinions: List[str] = Field(
-        default_factory=list,
-        description="List of newly formed opinions during reflection"
+    new_opinions: list[str] = Field(default_factory=list, description="List of newly formed opinions during reflection")
+    structured_output: dict[str, Any] | None = Field(
+        default=None,
+        description="Structured output parsed according to the provided response schema. Only present when response_schema was provided.",
     )
 
 
@@ -166,12 +169,12 @@ class Opinion(BaseModel):
     Opinions represent the bank's formed perspectives on topics,
     with a confidence level indicating strength of belief.
     """
-    model_config = ConfigDict(json_schema_extra={
-        "example": {
-            "text": "Machine learning has great potential in healthcare",
-            "confidence": 0.85
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {"text": "Machine learning has great potential in healthcare", "confidence": 0.85}
         }
-    })
+    )
 
     text: str = Field(description="The opinion text")
     confidence: float = Field(description="Confidence score between 0.0 and 1.0")
@@ -184,15 +187,15 @@ class EntityObservation(BaseModel):
     Observations are objective facts synthesized from multiple memory facts
     about an entity, without personality influence.
     """
-    model_config = ConfigDict(json_schema_extra={
-        "example": {
-            "text": "John is detail-oriented and works at Google",
-            "mentioned_at": "2024-01-15T10:30:00Z"
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {"text": "John is detail-oriented and works at Google", "mentioned_at": "2024-01-15T10:30:00Z"}
         }
-    })
+    )
 
     text: str = Field(description="The observation text")
-    mentioned_at: Optional[str] = Field(None, description="ISO format date when this observation was created")
+    mentioned_at: str | None = Field(None, description="ISO format date when this observation was created")
 
 
 class EntityState(BaseModel):
@@ -201,20 +204,22 @@ class EntityState(BaseModel):
 
     Contains observations synthesized from facts about the entity.
     """
-    model_config = ConfigDict(json_schema_extra={
-        "example": {
-            "entity_id": "123e4567-e89b-12d3-a456-426614174000",
-            "canonical_name": "John",
-            "observations": [
-                {"text": "John is detail-oriented", "mentioned_at": "2024-01-15T10:30:00Z"},
-                {"text": "John works at Google on the AI team", "mentioned_at": "2024-01-14T09:00:00Z"}
-            ]
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "entity_id": "123e4567-e89b-12d3-a456-426614174000",
+                "canonical_name": "John",
+                "observations": [
+                    {"text": "John is detail-oriented", "mentioned_at": "2024-01-15T10:30:00Z"},
+                    {"text": "John works at Google on the AI team", "mentioned_at": "2024-01-14T09:00:00Z"},
+                ],
+            }
         }
-    })
+    )
 
     entity_id: str = Field(description="Unique identifier for the entity")
     canonical_name: str = Field(description="Canonical name of the entity")
-    observations: List[EntityObservation] = Field(
-        default_factory=list,
-        description="List of observations about this entity"
+    observations: list[EntityObservation] = Field(
+        default_factory=list, description="List of observations about this entity"
     )

hindsight-api/hindsight_api/engine/retain/__init__.py

@@ -12,23 +12,16 @@ This package contains modular components for the retain operation:
 - fact_storage: Handle fact insertion into database
 """
 
-from .types import (
-    RetainContent,
-    ExtractedFact,
-    ProcessedFact,
-    ChunkMetadata,
-    EntityRef,
-    CausalRelation,
-    RetainBatch
+from . import (
+    chunk_storage,
+    deduplication,
+    embedding_processing,
+    entity_processing,
+    fact_extraction,
+    fact_storage,
+    link_creation,
 )
-
-from . import fact_extraction
-from . import embedding_processing
-from . import deduplication
-from . import entity_processing
-from . import link_creation
-from . import chunk_storage
-from . import fact_storage
+from .types import CausalRelation, ChunkMetadata, EntityRef, ExtractedFact, ProcessedFact, RetainBatch, RetainContent
 
 __all__ = [
     # Types

hindsight-api/hindsight_api/engine/retain/bank_utils.py

@@ -5,9 +5,12 @@ bank profile utilities for disposition and background management.
 import json
 import logging
 import re
-from typing import Dict, Optional, TypedDict
+from typing import TypedDict
+
 from pydantic import BaseModel, Field
+
 from ..db_utils import acquire_with_retry
+from ..memory_engine import fq_table
 from ..response_models import DispositionTraits
 
 logger = logging.getLogger(__name__)
@@ -21,6 +24,7 @@ DEFAULT_DISPOSITION = {
 
 class BankProfile(TypedDict):
     """Type for bank profile data."""
+
     name: str
     disposition: DispositionTraits
     background: str
@@ -28,6 +32,7 @@ class BankProfile(TypedDict):
 
 class BackgroundMergeResponse(BaseModel):
     """LLM response for background merge with disposition inference."""
+
     background: str = Field(description="Merged background in first person perspective")
     disposition: DispositionTraits = Field(description="Inferred disposition traits (skepticism, literalism, empathy)")
 
@@ -47,11 +52,11 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
     async with acquire_with_retry(pool) as conn:
         # Try to get existing bank
         row = await conn.fetchrow(
-            """
+            f"""
             SELECT name, disposition, background
-            FROM banks WHERE bank_id = $1
+            FROM {fq_table("banks")} WHERE bank_id = $1
             """,
-            bank_id
+            bank_id,
         )
 
         if row:
@@ -61,36 +66,26 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
                 disposition_data = json.loads(disposition_data)
 
             return BankProfile(
-                name=row["name"],
-                disposition=DispositionTraits(**disposition_data),
-                background=row["background"]
+                name=row["name"], disposition=DispositionTraits(**disposition_data), background=row["background"]
             )
 
         # Bank doesn't exist, create with defaults
         await conn.execute(
-            """
-            INSERT INTO banks (bank_id, name, disposition, background)
+            f"""
+            INSERT INTO {fq_table("banks")} (bank_id, name, disposition, background)
             VALUES ($1, $2, $3::jsonb, $4)
             ON CONFLICT (bank_id) DO NOTHING
             """,
             bank_id,
             bank_id,  # Default name is the bank_id
             json.dumps(DEFAULT_DISPOSITION),
-            ""
+            "",
         )
 
-        return BankProfile(
-            name=bank_id,
-            disposition=DispositionTraits(**DEFAULT_DISPOSITION),
-            background=""
-        )
+        return BankProfile(name=bank_id, disposition=DispositionTraits(**DEFAULT_DISPOSITION), background="")
 
 
-async def update_bank_disposition(
-    pool,
-    bank_id: str,
-    disposition: Dict[str, int]
-) -> None:
+async def update_bank_disposition(pool, bank_id: str, disposition: dict[str, int]) -> None:
     """
     Update bank disposition traits.
 
@@ -104,24 +99,18 @@ async def update_bank_disposition(
 
     async with acquire_with_retry(pool) as conn:
         await conn.execute(
-            """
-            UPDATE banks
+            f"""
+            UPDATE {fq_table("banks")}
             SET disposition = $2::jsonb,
                 updated_at = NOW()
             WHERE bank_id = $1
             """,
             bank_id,
-            json.dumps(disposition)
+            json.dumps(disposition),
         )
 
 
-async def merge_bank_background(
-    pool,
-    llm_config,
-    bank_id: str,
-    new_info: str,
-    update_disposition: bool = True
-) -> dict:
+async def merge_bank_background(pool, llm_config, bank_id: str, new_info: str, update_disposition: bool = True) -> dict:
     """
     Merge new background information with existing background using LLM.
     Normalizes to first person ("I") and resolves conflicts.
@@ -142,12 +131,7 @@ async def merge_bank_background(
     current_background = profile["background"]
 
     # Use LLM to merge backgrounds and optionally infer disposition
-    result = await _llm_merge_background(
-        llm_config,
-        current_background,
-        new_info,
-        infer_disposition=update_disposition
-    )
+    result = await _llm_merge_background(llm_config, current_background, new_info, infer_disposition=update_disposition)
 
     merged_background = result["background"]
     inferred_disposition = result.get("disposition")
@@ -157,8 +141,8 @@ async def merge_bank_background(
         if inferred_disposition:
             # Update both background and disposition
             await conn.execute(
-                """
-                UPDATE banks
+                f"""
+                UPDATE {fq_table("banks")}
                 SET background = $2,
                     disposition = $3::jsonb,
                     updated_at = NOW()
@@ -166,19 +150,19 @@ async def merge_bank_background(
                 """,
                 bank_id,
                 merged_background,
-                json.dumps(inferred_disposition)
+                json.dumps(inferred_disposition),
             )
         else:
             # Update only background
             await conn.execute(
-                """
-                UPDATE banks
+                f"""
+                UPDATE {fq_table("banks")}
                 SET background = $2,
                     updated_at = NOW()
                 WHERE bank_id = $1
                 """,
                 bank_id,
-                merged_background
+                merged_background,
             )
 
     response = {"background": merged_background}
@@ -188,12 +172,7 @@ async def merge_bank_background(
     return response
 
 
-async def _llm_merge_background(
-    llm_config,
-    current: str,
-    new_info: str,
-    infer_disposition: bool = False
-) -> dict:
+async def _llm_merge_background(llm_config, current: str, new_info: str, infer_disposition: bool = False) -> dict:
     """
     Use LLM to intelligently merge background information.
     Optionally infer Big Five disposition traits from the merged background.
@@ -273,25 +252,19 @@ Merged background:"""
                     response_format=BackgroundMergeResponse,
                     scope="bank_background",
                     temperature=0.3,
-                    max_completion_tokens=8192
+                    max_completion_tokens=8192,
                 )
                 logger.info(f"Successfully got structured response: background={parsed.background[:100]}")
 
                 # Convert Pydantic model to dict format
-                return {
-                    "background": parsed.background,
-                    "disposition": parsed.disposition.model_dump()
-                }
+                return {"background": parsed.background, "disposition": parsed.disposition.model_dump()}
             except Exception as e:
                 logger.warning(f"Structured output failed, falling back to manual parsing: {e}")
                 # Fall through to manual parsing below
 
         # Manual parsing fallback or non-disposition merge
         content = await llm_config.call(
-            messages=messages,
-            scope="bank_background",
-            temperature=0.3,
-            max_completion_tokens=8192
+            messages=messages, scope="bank_background", temperature=0.3, max_completion_tokens=8192
         )
 
         logger.info(f"LLM response for background merge (first 500 chars): {content[:500]}")
@@ -310,7 +283,7 @@ Merged background:"""
             # Method 2: Extract from markdown code blocks
             if result is None:
                 # Remove markdown code blocks
-                code_block_match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', content, re.DOTALL)
+                code_block_match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", content, re.DOTALL)
                 if code_block_match:
                     try:
                         result = json.loads(code_block_match.group(1))
@@ -321,7 +294,9 @@ Merged background:"""
             # Method 3: Find nested JSON structure
             if result is None:
                 # Look for JSON object with nested structure
-                json_match = re.search(r'\{[^{}]*"background"[^{}]*"disposition"[^{}]*\{[^{}]*\}[^{}]*\}', content, re.DOTALL)
+                json_match = re.search(
+                    r'\{[^{}]*"background"[^{}]*"disposition"[^{}]*\{[^{}]*\}[^{}]*\}', content, re.DOTALL
+                )
                 if json_match:
                     try:
                         result = json.loads(json_match.group())
@@ -335,7 +310,7 @@ Merged background:"""
                 # Fallback: use new_info as background with default disposition
                 return {
                     "background": new_info if new_info else current if current else "",
-                    "disposition": DEFAULT_DISPOSITION.copy()
+                    "disposition": DEFAULT_DISPOSITION.copy(),
                 }
 
             # Validate disposition values
@@ -387,9 +362,9 @@ async def list_banks(pool) -> list:
     """
     async with acquire_with_retry(pool) as conn:
         rows = await conn.fetch(
-            """
+            f"""
             SELECT bank_id, name, disposition, background, created_at, updated_at
-            FROM banks
+            FROM {fq_table("banks")}
             ORDER BY updated_at DESC
             """
         )
@@ -401,13 +376,15 @@ async def list_banks(pool) -> list:
             if isinstance(disposition_data, str):
                 disposition_data = json.loads(disposition_data)
 
-            result.append({
-                "bank_id": row["bank_id"],
-                "name": row["name"],
-                "disposition": disposition_data,
-                "background": row["background"],
-                "created_at": row["created_at"].isoformat() if row["created_at"] else None,
-                "updated_at": row["updated_at"].isoformat() if row["updated_at"] else None,
-            })
+            result.append(
+                {
+                    "bank_id": row["bank_id"],
+                    "name": row["name"],
+                    "disposition": disposition_data,
+                    "background": row["background"],
+                    "created_at": row["created_at"].isoformat() if row["created_at"] else None,
+                    "updated_at": row["updated_at"].isoformat() if row["updated_at"] else None,
+                }
+            )
 
         return result

hindsight-api/hindsight_api/engine/retain/chunk_storage.py

@@ -3,20 +3,16 @@ Chunk storage for retain pipeline.
 
 Handles storage of document chunks in the database.
 """
-import logging
-from typing import List, Dict, Optional
 
+import logging
+
+from ..memory_engine import fq_table
 from .types import ChunkMetadata
 
 logger = logging.getLogger(__name__)
 
 
-async def store_chunks_batch(
-    conn,
-    bank_id: str,
-    document_id: str,
-    chunks: List[ChunkMetadata]
-) -> Dict[int, str]:
+async def store_chunks_batch(conn, bank_id: str, document_id: str, chunks: list[ChunkMetadata]) -> dict[int, str]:
     """
     Store document chunks in the database.
 
@@ -47,24 +43,21 @@ async def store_chunks_batch(
 
     # Batch insert all chunks
     await conn.execute(
-        """
-        INSERT INTO chunks (chunk_id, document_id, bank_id, chunk_text, chunk_index)
+        f"""
+        INSERT INTO {fq_table("chunks")} (chunk_id, document_id, bank_id, chunk_text, chunk_index)
         SELECT * FROM unnest($1::text[], $2::text[], $3::text[], $4::text[], $5::integer[])
         """,
         chunk_ids,
         [document_id] * len(chunk_texts),
         [bank_id] * len(chunk_texts),
         chunk_texts,
-        chunk_indices
+        chunk_indices,
     )
 
     return chunk_id_map
 
 
-def map_facts_to_chunks(
-    facts_chunk_indices: List[int],
-    chunk_id_map: Dict[int, str]
-) -> List[Optional[str]]:
+def map_facts_to_chunks(facts_chunk_indices: list[int], chunk_id_map: dict[int, str]) -> list[str | None]:
     """
     Map fact chunk indices to chunk IDs.
 

hindsight-api/hindsight_api/engine/retain/deduplication.py

@@ -3,22 +3,17 @@ Deduplication logic for retain pipeline.
 
 Checks for duplicate facts using semantic similarity and temporal proximity.
 """
+
 import logging
-from datetime import datetime
-from typing import List
 from collections import defaultdict
+from datetime import UTC
 
 from .types import ProcessedFact
 
 logger = logging.getLogger(__name__)
 
 
-async def check_duplicates_batch(
-    conn,
-    bank_id: str,
-    facts: List[ProcessedFact],
-    duplicate_checker_fn
-) -> List[bool]:
+async def check_duplicates_batch(conn, bank_id: str, facts: list[ProcessedFact], duplicate_checker_fn) -> list[bool]:
     """
     Check which facts are duplicates using batched time-window queries.
 
@@ -47,16 +42,12 @@ async def check_duplicates_batch(
 
         # Defensive: if both are None (shouldn't happen), use now()
         if fact_date is None:
-            from datetime import datetime, timezone
-            fact_date = datetime.now(timezone.utc)
+            from datetime import datetime
+
+            fact_date = datetime.now(UTC)
 
         # Round to 12-hour bucket to group similar times
-        bucket_key = fact_date.replace(
-            hour=(fact_date.hour // 12) * 12,
-            minute=0,
-            second=0,
-            microsecond=0
-        )
+        bucket_key = fact_date.replace(hour=(fact_date.hour // 12) * 12, minute=0, second=0, microsecond=0)
         time_buckets[bucket_key].append((idx, fact))
 
     # Process each bucket in batch
@@ -68,14 +59,7 @@ async def check_duplicates_batch(
         embeddings = [item[1].embedding for item in bucket_items]
 
         # Check duplicates for this time bucket
-        dup_flags = await duplicate_checker_fn(
-            conn,
-            bank_id,
-            texts,
-            embeddings,
-            bucket_date,
-            time_window_hours=24
-        )
+        dup_flags = await duplicate_checker_fn(conn, bank_id, texts, embeddings, bucket_date, time_window_hours=24)
 
         # Map results back to original indices
         for idx, is_dup in zip(indices, dup_flags):
@@ -84,10 +68,7 @@ async def check_duplicates_batch(
     return all_is_duplicate
 
 
-def filter_duplicates(
-    facts: List[ProcessedFact],
-    is_duplicate_flags: List[bool]
-) -> List[ProcessedFact]:
+def filter_duplicates(facts: list[ProcessedFact], is_duplicate_flags: list[bool]) -> list[ProcessedFact]:
     """
     Filter out duplicate facts based on duplicate flags.
 

hindsight-api/hindsight_api/engine/retain/embedding_processing.py

@@ -3,9 +3,8 @@ Embedding processing for retain pipeline.
 
 Handles augmenting fact texts with temporal information and generating embeddings.
 """
+
 import logging
-from typing import List
-from datetime import datetime
 
 from . import embedding_utils
 from .types import ExtractedFact
@@ -13,7 +12,7 @@ from .types import ExtractedFact
 logger = logging.getLogger(__name__)
 
 
-def augment_texts_with_dates(facts: List[ExtractedFact], format_date_fn) -> List[str]:
+def augment_texts_with_dates(facts: list[ExtractedFact], format_date_fn) -> list[str]:
     """
     Augment fact texts with readable dates for better temporal matching.
 
@@ -37,10 +36,7 @@ def augment_texts_with_dates(facts: List[ExtractedFact], format_date_fn) -> List
     return augmented_texts
 
 
-async def generate_embeddings_batch(
-    embeddings_model,
-    texts: List[str]
-) -> List[List[float]]:
+async def generate_embeddings_batch(embeddings_model, texts: list[str]) -> list[list[float]]:
     """
     Generate embeddings for a batch of texts.
 
@@ -54,9 +50,6 @@ async def generate_embeddings_batch(
     if not texts:
         return []
 
-    embeddings = await embedding_utils.generate_embeddings_batch(
-        embeddings_model,
-        texts
-    )
+    embeddings = await embedding_utils.generate_embeddings_batch(embeddings_model, texts)
 
     return embeddings

hindsight-api/hindsight_api/engine/retain/embedding_utils.py

@@ -4,12 +4,11 @@ Embedding generation utilities for memory units.
 
 import asyncio
 import logging
-from typing import List
 
 logger = logging.getLogger(__name__)
 
 
-def generate_embedding(embeddings_backend, text: str) -> List[float]:
+def generate_embedding(embeddings_backend, text: str) -> list[float]:
     """
     Generate embedding for text using the provided embeddings backend.
 
@@ -27,7 +26,7 @@ def generate_embedding(embeddings_backend, text: str) -> List[float]:
         raise Exception(f"Failed to generate embedding: {str(e)}")
 
 
-async def generate_embeddings_batch(embeddings_backend, texts: List[str]) -> List[List[float]]:
+async def generate_embeddings_batch(embeddings_backend, texts: list[str]) -> list[list[float]]:
     """
     Generate embeddings for multiple texts using the provided embeddings backend.
 
@@ -47,7 +46,7 @@ async def generate_embeddings_batch(embeddings_backend, texts: List[str]) -> Lis
         embeddings = await loop.run_in_executor(
             None,  # Use default thread pool
             embeddings_backend.encode,
-            texts
+            texts,
         )
         return embeddings
     except Exception as e:

hindsight-api/hindsight_api/engine/retain/entity_processing.py

@@ -3,24 +3,18 @@ Entity processing for retain pipeline.
 
 Handles entity extraction, resolution, and link creation for stored facts.
 """
-import logging
-from typing import List, Tuple, Dict, Any
-from uuid import UUID
 
-from .types import ProcessedFact, EntityRef, EntityLink
+import logging
+
 from . import link_utils
+from .types import EntityLink, ProcessedFact
 
 logger = logging.getLogger(__name__)
 
 
 async def process_entities_batch(
-    entity_resolver,
-    conn,
-    bank_id: str,
-    unit_ids: List[str],
-    facts: List[ProcessedFact],
-    log_buffer: List[str] = None
-) -> List[EntityLink]:
+    entity_resolver, conn, bank_id: str, unit_ids: list[str], facts: list[ProcessedFact], log_buffer: list[str] = None
+) -> list[EntityLink]:
     """
     Process entities for all facts and create entity links.
 
@@ -53,8 +47,7 @@ async def process_entities_batch(
     fact_dates = [fact.occurred_start if fact.occurred_start is not None else fact.mentioned_at for fact in facts]
     # Convert EntityRef objects to dict format expected by link_utils
     entities_per_fact = [
-        [{'text': entity.name, 'type': 'CONCEPT'} for entity in (fact.entities or [])]
-        for fact in facts
+        [{"text": entity.name, "type": "CONCEPT"} for entity in (fact.entities or [])] for fact in facts
     ]
 
     # Use existing link_utils function for entity processing
@@ -67,16 +60,13 @@ async def process_entities_batch(
         "",  # context (not used in current implementation)
         fact_dates,
         entities_per_fact,
-        log_buffer  # Pass log_buffer for detailed logging
+        log_buffer,  # Pass log_buffer for detailed logging
     )
 
     return entity_links
 
 
-async def insert_entity_links_batch(
-    conn,
-    entity_links: List[EntityLink]
-) -> None:
+async def insert_entity_links_batch(conn, entity_links: list[EntityLink]) -> None:
     """
     Insert entity links in batch.
 

hindsight-api/hindsight_api/engine/retain/fact_extraction.py

@@ -4,16 +4,17 @@ Fact extraction from text using LLM.
 Extracts semantic facts, entities, and temporal information from text.
 Uses the LLMConfig wrapper for all LLM calls.
 """
-import logging
-import os
-import json
-import re
+
 import asyncio
+import json
+import logging
+import re
 from datetime import datetime, timedelta
-from typing import List, Dict, Optional, Literal
-from openai import AsyncOpenAI
-from pydantic import BaseModel, Field, field_validator, ConfigDict
-from ..llm_wrapper import OutputTooLongError, LLMConfig
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict, Field, field_validator
+
+from ..llm_wrapper import LLMConfig, OutputTooLongError
 
 
 def _sanitize_text(text: str) -> str:
@@ -31,11 +32,12 @@ def _sanitize_text(text: str) -> str:
         return text
     # Remove surrogate characters (U+D800 to U+DFFF) using regex
     # These are invalid in UTF-8 and cause encoding errors
-    return re.sub(r'[\ud800-\udfff]', '', text)
+    return re.sub(r"[\ud800-\udfff]", "", text)
 
 
 class Entity(BaseModel):
     """An entity extracted from text."""
+
     text: str = Field(
         description="The specific, named entity as it appears in the fact. Must be a proper noun or specific identifier."
     )
@@ -48,42 +50,46 @@ class Fact(BaseModel):
     This is what fact_extraction returns and what the rest of the pipeline expects.
     Combined fact text format: "what | when | where | who | why"
     """
+
     # Required fields
     fact: str = Field(description="Combined fact text: what | when | where | who | why")
     fact_type: Literal["world", "experience", "opinion"] = Field(description="Perspective: world/experience/opinion")
 
     # Optional temporal fields
-    occurred_start: Optional[str] = None
-    occurred_end: Optional[str] = None
-    mentioned_at: Optional[str] = None
+    occurred_start: str | None = None
+    occurred_end: str | None = None
+    mentioned_at: str | None = None
 
     # Optional location field
-    where: Optional[str] = Field(None, description="WHERE the fact occurred or is about (specific location, place, or area)")
+    where: str | None = Field(
+        None, description="WHERE the fact occurred or is about (specific location, place, or area)"
+    )
 
     # Optional structured data
-    entities: Optional[List[Entity]] = None
-    causal_relations: Optional[List['CausalRelation']] = None
+    entities: list[Entity] | None = None
+    causal_relations: list["CausalRelation"] | None = None
 
 
 class CausalRelation(BaseModel):
     """Causal relationship between facts."""
+
     target_fact_index: int = Field(
         description="Index of the related fact in the facts array (0-based). "
-                   "This creates a directed causal link to another fact in the extraction."
+        "This creates a directed causal link to another fact in the extraction."
     )
     relation_type: Literal["causes", "caused_by", "enables", "prevents"] = Field(
         description="Type of causal relationship: "
-                   "'causes' = this fact directly causes the target fact, "
-                   "'caused_by' = this fact was caused by the target fact, "
-                   "'enables' = this fact enables/allows the target fact, "
-                   "'prevents' = this fact prevents/blocks the target fact"
+        "'causes' = this fact directly causes the target fact, "
+        "'caused_by' = this fact was caused by the target fact, "
+        "'enables' = this fact enables/allows the target fact, "
+        "'prevents' = this fact prevents/blocks the target fact"
     )
     strength: float = Field(
         description="Strength of causal relationship (0.0 to 1.0). "
-                   "1.0 = direct/strong causation, 0.5 = moderate, 0.3 = weak/indirect",
+        "1.0 = direct/strong causation, 0.5 = moderate, 0.3 = weak/indirect",
         ge=0.0,
         le=1.0,
-        default=1.0
+        default=1.0,
     )
 
 
@@ -92,9 +98,7 @@ class ExtractedFact(BaseModel):
 
     model_config = ConfigDict(
         json_schema_mode="validation",
-        json_schema_extra={
-            "required": ["what", "when", "where", "who", "why", "fact_type"]
-        }
+        json_schema_extra={"required": ["what", "when", "where", "who", "why", "fact_type"]},
     )
 
     # ==========================================================================
@@ -103,43 +107,43 @@ class ExtractedFact(BaseModel):
 
     what: str = Field(
         description="WHAT happened - COMPLETE, DETAILED description with ALL specifics. "
-                   "NEVER summarize or omit details. Include: exact actions, objects, quantities, specifics. "
-                   "BE VERBOSE - capture every detail that was mentioned. "
-                   "Example: 'Emily got married to Sarah at a rooftop garden ceremony with 50 guests attending and a live jazz band playing' "
-                   "NOT: 'A wedding happened' or 'Emily got married'"
+        "NEVER summarize or omit details. Include: exact actions, objects, quantities, specifics. "
+        "BE VERBOSE - capture every detail that was mentioned. "
+        "Example: 'Emily got married to Sarah at a rooftop garden ceremony with 50 guests attending and a live jazz band playing' "
+        "NOT: 'A wedding happened' or 'Emily got married'"
     )
 
     when: str = Field(
         description="WHEN it happened - ALWAYS include temporal information if mentioned. "
-                   "Include: specific dates, times, durations, relative time references. "
-                   "Examples: 'on June 15th, 2024 at 3pm', 'last weekend', 'for the past 3 years', 'every morning at 6am'. "
-                   "Write 'N/A' ONLY if absolutely no temporal context exists. Prefer converting to absolute dates when possible."
+        "Include: specific dates, times, durations, relative time references. "
+        "Examples: 'on June 15th, 2024 at 3pm', 'last weekend', 'for the past 3 years', 'every morning at 6am'. "
+        "Write 'N/A' ONLY if absolutely no temporal context exists. Prefer converting to absolute dates when possible."
     )
 
     where: str = Field(
         description="WHERE it happened or is about - SPECIFIC locations, places, areas, regions if applicable. "
-                   "Include: cities, neighborhoods, venues, buildings, countries, specific addresses when mentioned. "
-                   "Examples: 'downtown San Francisco at a rooftop garden venue', 'at the user's home in Brooklyn', 'online via Zoom', 'Paris, France'. "
-                   "Write 'N/A' ONLY if absolutely no location context exists or if the fact is completely location-agnostic."
+        "Include: cities, neighborhoods, venues, buildings, countries, specific addresses when mentioned. "
+        "Examples: 'downtown San Francisco at a rooftop garden venue', 'at the user's home in Brooklyn', 'online via Zoom', 'Paris, France'. "
+        "Write 'N/A' ONLY if absolutely no location context exists or if the fact is completely location-agnostic."
     )
 
     who: str = Field(
         description="WHO is involved - ALL people/entities with FULL context and relationships. "
-                   "Include: names, roles, relationships to user, background details. "
-                   "Resolve coreferences (if 'my roommate' is later named 'Emily', write 'Emily, the user's college roommate'). "
-                   "BE DETAILED about relationships and roles. "
-                   "Example: 'Emily (user's college roommate from Stanford, now works at Google), Sarah (Emily's partner of 5 years, software engineer)' "
-                   "NOT: 'my friend' or 'Emily and Sarah'"
+        "Include: names, roles, relationships to user, background details. "
+        "Resolve coreferences (if 'my roommate' is later named 'Emily', write 'Emily, the user's college roommate'). "
+        "BE DETAILED about relationships and roles. "
+        "Example: 'Emily (user's college roommate from Stanford, now works at Google), Sarah (Emily's partner of 5 years, software engineer)' "
+        "NOT: 'my friend' or 'Emily and Sarah'"
     )
 
     why: str = Field(
         description="WHY it matters - ALL emotional, contextual, and motivational details. "
-                   "Include EVERYTHING: feelings, preferences, motivations, observations, context, background, significance. "
-                   "BE VERBOSE - capture all the nuance and meaning. "
-                   "FOR ASSISTANT FACTS: MUST include what the user asked/requested that led to this interaction! "
-                   "Example (world): 'The user felt thrilled and inspired, has always dreamed of an outdoor ceremony, mentioned wanting a similar garden venue, was particularly moved by the intimate atmosphere and personal vows' "
-                   "Example (assistant): 'User asked how to fix slow API performance with 1000+ concurrent users, expected 70-80% reduction in database load' "
-                   "NOT: 'User liked it' or 'To help user'"
+        "Include EVERYTHING: feelings, preferences, motivations, observations, context, background, significance. "
+        "BE VERBOSE - capture all the nuance and meaning. "
+        "FOR ASSISTANT FACTS: MUST include what the user asked/requested that led to this interaction! "
+        "Example (world): 'The user felt thrilled and inspired, has always dreamed of an outdoor ceremony, mentioned wanting a similar garden venue, was particularly moved by the intimate atmosphere and personal vows' "
+        "Example (assistant): 'User asked how to fix slow API performance with 1000+ concurrent users, expected 70-80% reduction in database load' "
+        "NOT: 'User liked it' or 'To help user'"
     )
 
     # ==========================================================================
@@ -148,17 +152,17 @@ class ExtractedFact(BaseModel):
 
     fact_kind: str = Field(
         default="conversation",
-        description="'event' = specific datable occurrence (set occurred dates), 'conversation' = general info (no occurred dates)"
+        description="'event' = specific datable occurrence (set occurred dates), 'conversation' = general info (no occurred dates)",
     )
 
     # Temporal fields - optional
-    occurred_start: Optional[str] = Field(
+    occurred_start: str | None = Field(
         default=None,
-        description="WHEN the event happened (ISO timestamp). Only for fact_kind='event'. Leave null for conversations."
+        description="WHEN the event happened (ISO timestamp). Only for fact_kind='event'. Leave null for conversations.",
     )
-    occurred_end: Optional[str] = Field(
+    occurred_end: str | None = Field(
         default=None,
-        description="WHEN the event ended (ISO timestamp). Only for events with duration. Leave null for conversations."
+        description="WHEN the event ended (ISO timestamp). Only for events with duration. Leave null for conversations.",
     )
 
     # Classification (CRITICAL - required)
@@ -168,16 +172,15 @@ class ExtractedFact(BaseModel):
     )
 
     # Entities - extracted from fact content
-    entities: Optional[List[Entity]] = Field(
+    entities: list[Entity] | None = Field(
         default=None,
-        description="Named entities, objects, AND abstract concepts from the fact. Include: people names, organizations, places, significant objects (e.g., 'coffee maker', 'car'), AND abstract concepts/themes (e.g., 'friendship', 'career growth', 'loss', 'celebration'). Extract anything that could help link related facts together."
+        description="Named entities, objects, AND abstract concepts from the fact. Include: people names, organizations, places, significant objects (e.g., 'coffee maker', 'car'), AND abstract concepts/themes (e.g., 'friendship', 'career growth', 'loss', 'celebration'). Extract anything that could help link related facts together.",
     )
-    causal_relations: Optional[List[CausalRelation]] = Field(
-        default=None,
-        description="Causal links to other facts. Can be null."
+    causal_relations: list[CausalRelation] | None = Field(
+        default=None, description="Causal links to other facts. Can be null."
     )
 
-    @field_validator('entities', mode='before')
+    @field_validator("entities", mode="before")
     @classmethod
     def ensure_entities_list(cls, v):
         """Ensure entities is always a list (convert None to empty list)."""
@@ -185,7 +188,7 @@ class ExtractedFact(BaseModel):
             return []
         return v
 
-    @field_validator('causal_relations', mode='before')
+    @field_validator("causal_relations", mode="before")
     @classmethod
     def ensure_causal_relations_list(cls, v):
         """Ensure causal_relations is always a list (convert None to empty list)."""
@@ -198,11 +201,11 @@ class ExtractedFact(BaseModel):
         parts = [self.what]
 
         # Add 'who' if not N/A
-        if self.who and self.who.upper() != 'N/A':
+        if self.who and self.who.upper() != "N/A":
             parts.append(f"Involving: {self.who}")
 
         # Add 'why' if not N/A
-        if self.why and self.why.upper() != 'N/A':
+        if self.why and self.why.upper() != "N/A":
             parts.append(self.why)
 
         if len(parts) == 1:
@@ -213,12 +216,11 @@ class ExtractedFact(BaseModel):
 
 class FactExtractionResponse(BaseModel):
     """Response containing all extracted facts."""
-    facts: List[ExtractedFact] = Field(
-        description="List of extracted factual statements"
-    )
+
+    facts: list[ExtractedFact] = Field(description="List of extracted factual statements")
 
 
-def chunk_text(text: str, max_chars: int) -> List[str]:
+def chunk_text(text: str, max_chars: int) -> list[str]:
     """
     Split text into chunks, preserving conversation structure when possible.
 
@@ -232,7 +234,6 @@ def chunk_text(text: str, max_chars: int) -> List[str]:
     Returns:
         List of text chunks, roughly under max_chars
     """
-    import json
     from langchain_text_splitters import RecursiveCharacterTextSplitter
 
     # If text is small enough, return as-is
@@ -256,21 +257,21 @@ def chunk_text(text: str, max_chars: int) -> List[str]:
         is_separator_regex=False,
         separators=[
             "\n\n",  # Paragraph breaks
-            "\n",    # Line breaks
-            ". ",    # Sentence endings
-            "! ",    # Exclamations
-            "? ",    # Questions
-            "; ",    # Semicolons
-            ", ",    # Commas
-            " ",     # Words
-            "",      # Characters (last resort)
+            "\n",  # Line breaks
+            ". ",  # Sentence endings
+            "! ",  # Exclamations
+            "? ",  # Questions
+            "; ",  # Semicolons
+            ", ",  # Commas
+            " ",  # Words
+            "",  # Characters (last resort)
         ],
     )
 
     return splitter.split_text(text)
 
 
-def _chunk_conversation(turns: List[dict], max_chars: int) -> List[str]:
+def _chunk_conversation(turns: list[dict], max_chars: int) -> list[str]:
     """
     Chunk a conversation array at turn boundaries, preserving complete turns.
 
@@ -281,7 +282,6 @@ def _chunk_conversation(turns: List[dict], max_chars: int) -> List[str]:
     Returns:
         List of JSON-serialized chunks, each containing complete turns
     """
-    import json
 
     chunks = []
     current_chunk = []
@@ -315,10 +315,10 @@ async def _extract_facts_from_chunk(
     total_chunks: int,
     event_date: datetime,
     context: str,
-    llm_config: 'LLMConfig',
+    llm_config: "LLMConfig",
     agent_name: str = None,
-    extract_opinions: bool = False
-) -> List[Dict[str, str]]:
+    extract_opinions: bool = False,
+) -> list[dict[str, str]]:
     """
     Extract facts from a single chunk (internal helper for parallel processing).
 
@@ -333,7 +333,9 @@ async def _extract_facts_from_chunk(
         # Opinion extraction uses a separate prompt (not this one)
         fact_types_instruction = "Extract ONLY 'opinion' type facts (formed opinions, beliefs, and perspectives). DO NOT extract 'world' or 'assistant' facts."
     else:
-        fact_types_instruction = "Extract ONLY 'world' and 'assistant' type facts. DO NOT extract opinions - those are extracted separately."
+        fact_types_instruction = (
+            "Extract ONLY 'world' and 'assistant' type facts. DO NOT extract opinions - those are extracted separately."
+        )
 
     prompt = f"""Extract facts from text into structured format with FOUR required dimensions - BE EXTREMELY DETAILED.
 
@@ -534,10 +536,8 @@ WHAT TO EXTRACT vs SKIP
 ✅ EXTRACT: User preferences (ALWAYS as separate facts!), feelings, plans, events, relationships, achievements
 ❌ SKIP: Greetings, filler ("thanks", "cool"), purely structural statements"""
 
-
-
-
     import logging
+
     from openai import BadRequestError
 
     logger = logging.getLogger(__name__)
@@ -548,11 +548,11 @@ WHAT TO EXTRACT vs SKIP
 
     # Sanitize input text to prevent Unicode encoding errors (e.g., unpaired surrogates)
     sanitized_chunk = _sanitize_text(chunk)
-    sanitized_context = _sanitize_text(context) if context else 'none'
+    sanitized_context = _sanitize_text(context) if context else "none"
 
     # Build user message with metadata and chunk content in a clear format
     # Format event_date with day of week for better temporal reasoning
-    event_date_formatted = event_date.strftime('%A, %B %d, %Y')  # e.g., "Monday, June 10, 2024"
+    event_date_formatted = event_date.strftime("%A, %B %d, %Y")  # e.g., "Monday, June 10, 2024"
     user_message = f"""Extract facts from the following text chunk.
 {memory_bank_context}
 
@@ -566,16 +566,7 @@ Text:
     for attempt in range(max_retries):
         try:
             extraction_response_json = await llm_config.call(
-                messages=[
-                    {
-                        "role": "system",
-                        "content": prompt
-                    },
-                    {
-                        "role": "user",
-                        "content": user_message
-                    }
-                ],
+                messages=[{"role": "system", "content": prompt}, {"role": "user", "content": user_message}],
                 response_format=FactExtractionResponse,
                 scope="memory_extract_facts",
                 temperature=0.1,
@@ -601,7 +592,7 @@ Text:
                     )
                     return []
 
-            raw_facts = extraction_response_json.get('facts', [])
+            raw_facts = extraction_response_json.get("facts", [])
             if not raw_facts:
                 logger.debug(
                     f"LLM response missing 'facts' field or returned empty list. "
@@ -622,48 +613,48 @@ Text:
                 # Helper to get non-empty value
                 def get_value(field_name):
                     value = llm_fact.get(field_name)
-                    if value and value != '' and value != [] and value != {} and str(value).upper() != 'N/A':
+                    if value and value != "" and value != [] and value != {} and str(value).upper() != "N/A":
                         return value
                     return None
 
                 # NEW FORMAT: what, when, who, why (all required)
-                what = get_value('what')
-                when = get_value('when')
-                who = get_value('who')
-                why = get_value('why')
+                what = get_value("what")
+                when = get_value("when")
+                who = get_value("who")
+                why = get_value("why")
 
                 # Fallback to old format if new fields not present
                 if not what:
-                    what = get_value('factual_core')
+                    what = get_value("factual_core")
                 if not what:
                     logger.warning(f"Skipping fact {i}: missing 'what' field")
                     continue
 
                 # Critical field: fact_type
                 # LLM uses "assistant" but we convert to "experience" for storage
-                fact_type = llm_fact.get('fact_type')
+                fact_type = llm_fact.get("fact_type")
 
                 # Convert "assistant" → "experience" for storage
-                if fact_type == 'assistant':
-                    fact_type = 'experience'
+                if fact_type == "assistant":
+                    fact_type = "experience"
 
                 # Validate fact_type (after conversion)
-                if fact_type not in ['world', 'experience', 'opinion']:
+                if fact_type not in ["world", "experience", "opinion"]:
                     # Try to fix common mistakes - check if they swapped fact_type and fact_kind
-                    fact_kind = llm_fact.get('fact_kind')
-                    if fact_kind == 'assistant':
-                        fact_type = 'experience'
-                    elif fact_kind in ['world', 'experience', 'opinion']:
+                    fact_kind = llm_fact.get("fact_kind")
+                    if fact_kind == "assistant":
+                        fact_type = "experience"
+                    elif fact_kind in ["world", "experience", "opinion"]:
                         fact_type = fact_kind
                     else:
                         # Default to 'world' if we can't determine
-                        fact_type = 'world'
+                        fact_type = "world"
                         logger.warning(f"Fact {i}: defaulting to fact_type='world'")
 
                 # Get fact_kind for temporal handling (but don't store it)
-                fact_kind = llm_fact.get('fact_kind', 'conversation')
-                if fact_kind not in ['conversation', 'event', 'other']:
-                    fact_kind = 'conversation'
+                fact_kind = llm_fact.get("fact_kind", "conversation")
+                if fact_kind not in ["conversation", "event", "other"]:
+                    fact_kind = "conversation"
 
                 # Build combined fact text from the 4 dimensions: what | when | who | why
                 fact_data = {}
@@ -682,20 +673,20 @@ Text:
 
                 # Add temporal fields
                 # For events: occurred_start/occurred_end (when the event happened)
-                if fact_kind == 'event':
-                    occurred_start = get_value('occurred_start')
-                    occurred_end = get_value('occurred_end')
+                if fact_kind == "event":
+                    occurred_start = get_value("occurred_start")
+                    occurred_end = get_value("occurred_end")
                     if occurred_start:
-                        fact_data['occurred_start'] = occurred_start
+                        fact_data["occurred_start"] = occurred_start
                         # For point events: if occurred_end not set, default to occurred_start
                         if occurred_end:
-                            fact_data['occurred_end'] = occurred_end
+                            fact_data["occurred_end"] = occurred_end
                         else:
-                            fact_data['occurred_end'] = occurred_start
+                            fact_data["occurred_end"] = occurred_start
 
                 # Add entities if present (validate as Entity objects)
                 # LLM sometimes returns strings instead of {"text": "..."} format
-                entities = get_value('entities')
+                entities = get_value("entities")
                 if entities:
                     # Validate and normalize each entity
                     validated_entities = []
@@ -703,38 +694,34 @@ Text:
                         if isinstance(ent, str):
                             # Normalize string to Entity object
                             validated_entities.append(Entity(text=ent))
-                        elif isinstance(ent, dict) and 'text' in ent:
+                        elif isinstance(ent, dict) and "text" in ent:
                             try:
                                 validated_entities.append(Entity.model_validate(ent))
                             except Exception as e:
                                 logger.warning(f"Invalid entity {ent}: {e}")
                     if validated_entities:
-                        fact_data['entities'] = validated_entities
+                        fact_data["entities"] = validated_entities
 
                 # Add causal relations if present (validate as CausalRelation objects)
                 # Filter out invalid relations (missing required fields)
-                causal_relations = get_value('causal_relations')
+                causal_relations = get_value("causal_relations")
                 if causal_relations:
                     validated_relations = []
                     for rel in causal_relations:
-                        if isinstance(rel, dict) and 'target_fact_index' in rel and 'relation_type' in rel:
+                        if isinstance(rel, dict) and "target_fact_index" in rel and "relation_type" in rel:
                             try:
                                 validated_relations.append(CausalRelation.model_validate(rel))
                             except Exception as e:
                                 logger.warning(f"Invalid causal relation {rel}: {e}")
                     if validated_relations:
-                        fact_data['causal_relations'] = validated_relations
+                        fact_data["causal_relations"] = validated_relations
 
                 # Always set mentioned_at to the event_date (when the conversation/document occurred)
-                fact_data['mentioned_at'] = event_date.isoformat()
+                fact_data["mentioned_at"] = event_date.isoformat()
 
                 # Build Fact model instance
                 try:
-                    fact = Fact(
-                        fact=combined_text,
-                        fact_type=fact_type,
-                        **fact_data
-                    )
+                    fact = Fact(fact=combined_text, fact_type=fact_type, **fact_data)
                     chunk_facts.append(fact)
                 except Exception as e:
                     logger.error(f"Failed to create Fact model for fact {i}: {e}")
@@ -753,7 +740,9 @@ Text:
         except BadRequestError as e:
             last_error = e
             if "json_validate_failed" in str(e):
-                logger.warning(f"          [1.3.{chunk_index + 1}] Attempt {attempt + 1}/{max_retries} failed with JSON validation error: {e}")
+                logger.warning(
+                    f"          [1.3.{chunk_index + 1}] Attempt {attempt + 1}/{max_retries} failed with JSON validation error: {e}"
+                )
                 if attempt < max_retries - 1:
                     logger.info(f"          [1.3.{chunk_index + 1}] Retrying...")
                     continue
@@ -772,8 +761,8 @@ async def _extract_facts_with_auto_split(
     context: str,
     llm_config: LLMConfig,
     agent_name: str = None,
-    extract_opinions: bool = False
-) -> List[Dict[str, str]]:
+    extract_opinions: bool = False,
+) -> list[dict[str, str]]:
     """
     Extract facts from a chunk with automatic splitting if output exceeds token limits.
 
@@ -794,6 +783,7 @@ async def _extract_facts_with_auto_split(
         List of fact dictionaries extracted from the chunk (possibly from sub-chunks)
     """
     import logging
+
     logger = logging.getLogger(__name__)
 
     try:
@@ -806,9 +796,9 @@ async def _extract_facts_with_auto_split(
             context=context,
             llm_config=llm_config,
             agent_name=agent_name,
-            extract_opinions=extract_opinions
+            extract_opinions=extract_opinions,
         )
-    except OutputTooLongError as e:
+    except OutputTooLongError:
         # Output exceeded token limits - split the chunk in half and retry
         logger.warning(
             f"Output too long for chunk {chunk_index + 1}/{total_chunks} "
@@ -824,7 +814,7 @@ async def _extract_facts_with_auto_split(
         search_start = max(0, mid_point - search_range)
         search_end = min(len(chunk), mid_point + search_range)
 
-        sentence_endings = ['. ', '! ', '? ', '\n\n']
+        sentence_endings = [". ", "! ", "? ", "\n\n"]
         best_split = mid_point
 
         for ending in sentence_endings:
@@ -838,8 +828,7 @@ async def _extract_facts_with_auto_split(
         second_half = chunk[best_split:].strip()
 
         logger.info(
-            f"Split chunk {chunk_index + 1} into two sub-chunks: "
-            f"{len(first_half)} chars and {len(second_half)} chars"
+            f"Split chunk {chunk_index + 1} into two sub-chunks: {len(first_half)} chars and {len(second_half)} chars"
         )
 
         # Process both halves recursively (in parallel)
@@ -852,7 +841,7 @@ async def _extract_facts_with_auto_split(
                 context=context,
                 llm_config=llm_config,
                 agent_name=agent_name,
-                extract_opinions=extract_opinions
+                extract_opinions=extract_opinions,
             ),
             _extract_facts_with_auto_split(
                 chunk=second_half,
@@ -862,8 +851,8 @@ async def _extract_facts_with_auto_split(
                 context=context,
                 llm_config=llm_config,
                 agent_name=agent_name,
-                extract_opinions=extract_opinions
-            )
+                extract_opinions=extract_opinions,
+            ),
         ]
 
         sub_results = await asyncio.gather(*sub_tasks)
@@ -873,9 +862,7 @@ async def _extract_facts_with_auto_split(
         for sub_result in sub_results:
             all_facts.extend(sub_result)
 
-        logger.info(
-            f"Successfully extracted {len(all_facts)} facts from split chunk {chunk_index + 1}"
-        )
+        logger.info(f"Successfully extracted {len(all_facts)} facts from split chunk {chunk_index + 1}")
 
         return all_facts
 
@@ -887,7 +874,7 @@ async def extract_facts_from_text(
     agent_name: str,
     context: str = "",
     extract_opinions: bool = False,
-) -> tuple[List[Fact], List[tuple[str, int]]]:
+) -> tuple[list[Fact], list[tuple[str, int]]]:
     """
     Extract semantic facts from conversational or narrative text using LLM.
 
@@ -920,7 +907,7 @@ async def extract_facts_from_text(
             context=context,
             llm_config=llm_config,
             agent_name=agent_name,
-            extract_opinions=extract_opinions
+            extract_opinions=extract_opinions,
         )
         for i, chunk in enumerate(chunks)
     ]
@@ -938,8 +925,10 @@ async def extract_facts_from_text(
 # ============================================================================
 
 # Import types for the orchestration layer (note: ExtractedFact here is different from the Pydantic model above)
-from .types import RetainContent, ExtractedFact as ExtractedFactType, ChunkMetadata, CausalRelation as CausalRelationType
-from typing import Tuple
+
+from .types import CausalRelation as CausalRelationType
+from .types import ChunkMetadata, RetainContent
+from .types import ExtractedFact as ExtractedFactType
 
 logger = logging.getLogger(__name__)
 
@@ -948,11 +937,8 @@ SECONDS_PER_FACT = 10
 
 
 async def extract_facts_from_contents(
-    contents: List[RetainContent],
-    llm_config,
-    agent_name: str,
-    extract_opinions: bool = False
-) -> Tuple[List[ExtractedFactType], List[ChunkMetadata]]:
+    contents: list[RetainContent], llm_config, agent_name: str, extract_opinions: bool = False
+) -> tuple[list[ExtractedFactType], list[ChunkMetadata]]:
     """
     Extract facts from multiple content items in parallel.
 
@@ -985,7 +971,7 @@ async def extract_facts_from_contents(
             context=item.context,
             llm_config=llm_config,
             agent_name=agent_name,
-            extract_opinions=extract_opinions
+            extract_opinions=extract_opinions,
         )
         fact_extraction_tasks.append(task)
 
@@ -993,8 +979,8 @@ async def extract_facts_from_contents(
     all_fact_results = await asyncio.gather(*fact_extraction_tasks)
 
     # Step 3: Flatten and convert to typed objects
-    extracted_facts: List[ExtractedFactType] = []
-    chunks_metadata: List[ChunkMetadata] = []
+    extracted_facts: list[ExtractedFactType] = []
+    chunks_metadata: list[ChunkMetadata] = []
 
     global_chunk_idx = 0
     global_fact_idx = 0
@@ -1008,7 +994,7 @@ async def extract_facts_from_contents(
                 chunk_text=chunk_text,
                 fact_count=chunk_fact_count,
                 content_index=content_index,
-                chunk_index=global_chunk_idx
+                chunk_index=global_chunk_idx,
             )
             chunks_metadata.append(chunk_metadata)
             global_chunk_idx += 1
@@ -1029,18 +1015,21 @@ async def extract_facts_from_contents(
                         fact_type=fact_from_llm.fact_type,
                         entities=[e.text for e in (fact_from_llm.entities or [])],
                         # occurred_start/end: from LLM only, leave None if not provided
-                        occurred_start=_parse_datetime(fact_from_llm.occurred_start) if fact_from_llm.occurred_start else None,
-                        occurred_end=_parse_datetime(fact_from_llm.occurred_end) if fact_from_llm.occurred_end else None,
+                        occurred_start=_parse_datetime(fact_from_llm.occurred_start)
+                        if fact_from_llm.occurred_start
+                        else None,
+                        occurred_end=_parse_datetime(fact_from_llm.occurred_end)
+                        if fact_from_llm.occurred_end
+                        else None,
                         causal_relations=_convert_causal_relations(
-                            fact_from_llm.causal_relations or [],
-                            global_fact_idx
+                            fact_from_llm.causal_relations or [], global_fact_idx
                         ),
                         content_index=content_index,
                         chunk_index=chunk_global_idx,
                         context=content.context,
                         # mentioned_at: always the event_date (when the conversation/document occurred)
                         mentioned_at=content.event_date,
-                        metadata=content.metadata
+                        metadata=content.metadata,
                     )
 
                     extracted_facts.append(extracted_fact)
@@ -1056,13 +1045,14 @@ async def extract_facts_from_contents(
 def _parse_datetime(date_str: str):
     """Parse ISO datetime string."""
     from dateutil import parser as date_parser
+
     try:
         return date_parser.isoparse(date_str)
     except Exception:
         return None
 
 
-def _convert_causal_relations(relations_from_llm, fact_start_idx: int) -> List[CausalRelationType]:
+def _convert_causal_relations(relations_from_llm, fact_start_idx: int) -> list[CausalRelationType]:
     """
     Convert causal relations from LLM format to ExtractedFact format.
 
@@ -1073,13 +1063,13 @@ def _convert_causal_relations(relations_from_llm, fact_start_idx: int) -> List[C
         causal_relation = CausalRelationType(
             relation_type=rel.relation_type,
             target_fact_index=fact_start_idx + rel.target_fact_index,
-            strength=rel.strength
+            strength=rel.strength,
         )
         causal_relations.append(causal_relation)
     return causal_relations
 
 
-def _add_temporal_offsets(facts: List[ExtractedFactType], contents: List[RetainContent]) -> None:
+def _add_temporal_offsets(facts: list[ExtractedFactType], contents: list[RetainContent]) -> None:
     """
     Add time offsets to preserve fact ordering within each content.
 

hindsight-api/hindsight_api/engine/retain/fact_storage.py

@@ -3,22 +3,19 @@ Fact storage for retain pipeline.
 
 Handles insertion of facts into the database.
 """
-import logging
-import json
-from typing import List, Optional
-from uuid import UUID
 
+import json
+import logging
+
+from ..memory_engine import fq_table
 from .types import ProcessedFact
 
 logger = logging.getLogger(__name__)
 
 
 async def insert_facts_batch(
-    conn,
-    bank_id: str,
-    facts: List[ProcessedFact],
-    document_id: Optional[str] = None
-) -> List[str]:
+    conn, bank_id: str, facts: list[ProcessedFact], document_id: str | None = None
+) -> list[str]:
     """
     Insert facts into the database in batch.
 
@@ -62,7 +59,7 @@ async def insert_facts_batch(
         contexts.append(fact.context)
         fact_types.append(fact.fact_type)
         # confidence_score is only for opinion facts
-        confidence_scores.append(1.0 if fact.fact_type == 'opinion' else None)
+        confidence_scores.append(1.0 if fact.fact_type == "opinion" else None)
         access_counts.append(0)  # Initial access count
         metadata_jsons.append(json.dumps(fact.metadata))
         chunk_ids.append(fact.chunk_id)
@@ -71,8 +68,8 @@ async def insert_facts_batch(
 
     # Batch insert all facts
     results = await conn.fetch(
-        """
-        INSERT INTO memory_units (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+        f"""
+        INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
                                  context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id)
         SELECT $1, * FROM unnest(
             $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
@@ -93,10 +90,10 @@ async def insert_facts_batch(
         access_counts,
         metadata_jsons,
         chunk_ids,
-        document_ids
+        document_ids,
     )
 
-    unit_ids = [str(row['id']) for row in results]
+    unit_ids = [str(row["id"]) for row in results]
     return unit_ids
 
 
@@ -111,25 +108,20 @@ async def ensure_bank_exists(conn, bank_id: str) -> None:
         bank_id: Bank identifier
     """
     await conn.execute(
-        """
-        INSERT INTO banks (bank_id, disposition, background)
+        f"""
+        INSERT INTO {fq_table("banks")} (bank_id, disposition, background)
         VALUES ($1, $2::jsonb, $3)
         ON CONFLICT (bank_id) DO UPDATE
         SET updated_at = NOW()
         """,
         bank_id,
         '{"skepticism": 3, "literalism": 3, "empathy": 3}',
-        ""
+        "",
     )
 
 
 async def handle_document_tracking(
-    conn,
-    bank_id: str,
-    document_id: str,
-    combined_content: str,
-    is_first_batch: bool,
-    retain_params: Optional[dict] = None
+    conn, bank_id: str, document_id: str, combined_content: str, is_first_batch: bool, retain_params: dict | None = None
 ) -> None:
     """
     Handle document tracking in the database.
@@ -151,14 +143,13 @@ async def handle_document_tracking(
     # Only delete on the first batch to avoid deleting data we just inserted
     if is_first_batch:
         await conn.fetchval(
-            "DELETE FROM documents WHERE id = $1 AND bank_id = $2 RETURNING id",
-            document_id, bank_id
+            f"DELETE FROM {fq_table('documents')} WHERE id = $1 AND bank_id = $2 RETURNING id", document_id, bank_id
         )
 
     # Insert document (or update if exists from concurrent operations)
     await conn.execute(
-        """
-        INSERT INTO documents (id, bank_id, original_text, content_hash, metadata, retain_params)
+        f"""
+        INSERT INTO {fq_table("documents")} (id, bank_id, original_text, content_hash, metadata, retain_params)
         VALUES ($1, $2, $3, $4, $5, $6)
         ON CONFLICT (id, bank_id) DO UPDATE
         SET original_text = EXCLUDED.original_text,
@@ -172,5 +163,5 @@ async def handle_document_tracking(
         combined_content,
         content_hash,
         json.dumps({}),  # Empty metadata dict
-        json.dumps(retain_params) if retain_params else None
+        json.dumps(retain_params) if retain_params else None,
     )

hindsight-api/hindsight_api/engine/retain/link_creation.py

@@ -3,20 +3,16 @@ Link creation for retain pipeline.
 
 Handles creation of temporal, semantic, and causal links between facts.
 """
-import logging
-from typing import List
 
-from .types import ProcessedFact, CausalRelation
+import logging
+
 from . import link_utils
+from .types import ProcessedFact
 
 logger = logging.getLogger(__name__)
 
 
-async def create_temporal_links_batch(
-    conn,
-    bank_id: str,
-    unit_ids: List[str]
-) -> int:
+async def create_temporal_links_batch(conn, bank_id: str, unit_ids: list[str]) -> int:
     """
     Create temporal links between facts.
 
@@ -33,20 +29,10 @@ async def create_temporal_links_batch(
     if not unit_ids:
         return 0
 
-    return await link_utils.create_temporal_links_batch_per_fact(
-        conn,
-        bank_id,
-        unit_ids,
-        log_buffer=[]
-    )
+    return await link_utils.create_temporal_links_batch_per_fact(conn, bank_id, unit_ids, log_buffer=[])
 
 
-async def create_semantic_links_batch(
-    conn,
-    bank_id: str,
-    unit_ids: List[str],
-    embeddings: List[List[float]]
-) -> int:
+async def create_semantic_links_batch(conn, bank_id: str, unit_ids: list[str], embeddings: list[list[float]]) -> int:
     """
     Create semantic links between facts.
 
@@ -67,20 +53,10 @@ async def create_semantic_links_batch(
     if len(unit_ids) != len(embeddings):
         raise ValueError(f"Mismatch between unit_ids ({len(unit_ids)}) and embeddings ({len(embeddings)})")
 
-    return await link_utils.create_semantic_links_batch(
-        conn,
-        bank_id,
-        unit_ids,
-        embeddings,
-        log_buffer=[]
-    )
+    return await link_utils.create_semantic_links_batch(conn, bank_id, unit_ids, embeddings, log_buffer=[])
 
 
-async def create_causal_links_batch(
-    conn,
-    unit_ids: List[str],
-    facts: List[ProcessedFact]
-) -> int:
+async def create_causal_links_batch(conn, unit_ids: list[str], facts: list[ProcessedFact]) -> int:
     """
     Create causal links between facts.
 
@@ -108,9 +84,9 @@ async def create_causal_links_batch(
             # Convert CausalRelation objects to dicts
             relations_dicts = [
                 {
-                    'relation_type': rel.relation_type,
-                    'target_fact_index': rel.target_fact_index,
-                    'strength': rel.strength
+                    "relation_type": rel.relation_type,
+                    "target_fact_index": rel.target_fact_index,
+                    "strength": rel.strength,
                 }
                 for rel in fact.causal_relations
             ]
@@ -118,10 +94,6 @@ async def create_causal_links_batch(
         else:
             causal_relations_per_fact.append([])
 
-    link_count = await link_utils.create_causal_links_batch(
-        conn,
-        unit_ids,
-        causal_relations_per_fact
-    )
+    link_count = await link_utils.create_causal_links_batch(conn, unit_ids, causal_relations_per_fact)
 
     return link_count

hindsight-api/hindsight_api/engine/retain/link_utils.py

@@ -2,12 +2,12 @@
 Link creation utilities for temporal, semantic, and entity links.
 """
 
-import time
 import logging
-from typing import List
-from datetime import timedelta, datetime, timezone
+import time
+from datetime import UTC, datetime, timedelta
 from uuid import UUID
 
+from ..memory_engine import fq_table
 from .types import EntityLink
 
 logger = logging.getLogger(__name__)
@@ -19,7 +19,7 @@ def _normalize_datetime(dt):
         return None
     if dt.tzinfo is None:
         # Naive datetime - assume UTC
-        return dt.replace(tzinfo=timezone.utc)
+        return dt.replace(tzinfo=UTC)
     return dt
 
 
@@ -54,24 +54,26 @@ def compute_temporal_links(
         try:
             time_lower = unit_event_date_norm - timedelta(hours=time_window_hours)
         except OverflowError:
-            time_lower = datetime.min.replace(tzinfo=timezone.utc)
+            time_lower = datetime.min.replace(tzinfo=UTC)
         try:
             time_upper = unit_event_date_norm + timedelta(hours=time_window_hours)
         except OverflowError:
-            time_upper = datetime.max.replace(tzinfo=timezone.utc)
+            time_upper = datetime.max.replace(tzinfo=UTC)
 
         # Filter candidates within this unit's time window
         matching_neighbors = [
-            (row['id'], row['event_date'])
+            (row["id"], row["event_date"])
             for row in candidates
-            if time_lower <= _normalize_datetime(row['event_date']) <= time_upper
+            if time_lower <= _normalize_datetime(row["event_date"]) <= time_upper
         ][:10]  # Limit to top 10
 
         for recent_id, recent_event_date in matching_neighbors:
             # Calculate temporal proximity weight
-            time_diff_hours = abs((unit_event_date_norm - _normalize_datetime(recent_event_date)).total_seconds() / 3600)
+            time_diff_hours = abs(
+                (unit_event_date_norm - _normalize_datetime(recent_event_date)).total_seconds() / 3600
+            )
             weight = max(0.3, 1.0 - (time_diff_hours / time_window_hours))
-            links.append((unit_id, str(recent_id), 'temporal', weight, None))
+            links.append((unit_id, str(recent_id), "temporal", weight, None))
 
     return links
 
@@ -99,17 +101,17 @@ def compute_temporal_query_bounds(
     try:
         min_date = min(all_dates) - timedelta(hours=time_window_hours)
     except OverflowError:
-        min_date = datetime.min.replace(tzinfo=timezone.utc)
+        min_date = datetime.min.replace(tzinfo=UTC)
 
     try:
         max_date = max(all_dates) + timedelta(hours=time_window_hours)
     except OverflowError:
-        max_date = datetime.max.replace(tzinfo=timezone.utc)
+        max_date = datetime.max.replace(tzinfo=UTC)
 
     return min_date, max_date
 
 
-def _log(log_buffer, message, level='info'):
+def _log(log_buffer, message, level="info"):
     """Helper to log to buffer if available, otherwise use logger.
 
     Args:
@@ -117,7 +119,7 @@ def _log(log_buffer, message, level='info'):
         message: The log message
         level: 'info', 'debug', 'warning', or 'error'. Debug messages are not added to buffer.
     """
-    if level == 'debug':
+    if level == "debug":
         # Debug messages only go to logger, not to buffer
         logger.debug(message)
         return
@@ -125,23 +127,23 @@ def _log(log_buffer, message, level='info'):
     if log_buffer is not None:
         log_buffer.append(message)
     else:
-        if level == 'info':
+        if level == "info":
             logger.info(message)
         else:
-            logger.log(logging.WARNING if level == 'warning' else logging.ERROR, message)
+            logger.log(logging.WARNING if level == "warning" else logging.ERROR, message)
 
 
 async def extract_entities_batch_optimized(
     entity_resolver,
     conn,
     bank_id: str,
-    unit_ids: List[str],
-    sentences: List[str],
+    unit_ids: list[str],
+    sentences: list[str],
     context: str,
-    fact_dates: List,
-    llm_entities: List[List[dict]],
-    log_buffer: List[str] = None,
-) -> List[tuple]:
+    fact_dates: list,
+    llm_entities: list[list[dict]],
+    log_buffer: list[str] = None,
+) -> list[tuple]:
     """
     Process LLM-extracted entities for ALL facts in batch.
 
@@ -171,15 +173,19 @@ async def extract_entities_batch_optimized(
             formatted_entities = []
             for ent in entity_list:
                 # Handle both Entity objects and dicts
-                if hasattr(ent, 'text'):
+                if hasattr(ent, "text"):
                     # Entity objects only have 'text', default type to 'CONCEPT'
-                    formatted_entities.append({'text': ent.text, 'type': 'CONCEPT'})
+                    formatted_entities.append({"text": ent.text, "type": "CONCEPT"})
                 elif isinstance(ent, dict):
-                    formatted_entities.append({'text': ent.get('text', ''), 'type': ent.get('type', 'CONCEPT')})
+                    formatted_entities.append({"text": ent.get("text", ""), "type": ent.get("type", "CONCEPT")})
             all_entities.append(formatted_entities)
 
         total_entities = sum(len(ents) for ents in all_entities)
-        _log(log_buffer, f"  [6.1] Process LLM entities: {total_entities} entities from {len(sentences)} facts in {time.time() - substep_start:.3f}s", level='debug')
+        _log(
+            log_buffer,
+            f"  [6.1] Process LLM entities: {total_entities} entities from {len(sentences)} facts in {time.time() - substep_start:.3f}s",
+            level="debug",
+        )
 
         # Step 2: Resolve entities in BATCH (much faster!)
         substep_start = time.time()
@@ -195,13 +201,19 @@ async def extract_entities_batch_optimized(
                 continue
 
             for local_idx, entity in enumerate(entities):
-                all_entities_flat.append({
-                    'text': entity['text'],
-                    'type': entity['type'],
-                    'nearby_entities': entities,
-                })
+                all_entities_flat.append(
+                    {
+                        "text": entity["text"],
+                        "type": entity["type"],
+                        "nearby_entities": entities,
+                    }
+                )
                 entity_to_unit.append((unit_id, local_idx, fact_date))
-        _log(log_buffer, f"    [6.2.1] Prepare entities: {len(all_entities_flat)} entities in {time.time() - substep_6_2_1_start:.3f}s", level='debug')
+        _log(
+            log_buffer,
+            f"    [6.2.1] Prepare entities: {len(all_entities_flat)} entities in {time.time() - substep_6_2_1_start:.3f}s",
+            level="debug",
+        )
 
         # Resolve ALL entities in one batch call
         if all_entities_flat:
@@ -210,7 +222,7 @@ async def extract_entities_batch_optimized(
 
             # Add per-entity dates to entity data for batch resolution
             for idx, (unit_id, local_idx, fact_date) in enumerate(entity_to_unit):
-                all_entities_flat[idx]['event_date'] = fact_date
+                all_entities_flat[idx]["event_date"] = fact_date
 
             # Resolve ALL entities in ONE batch call (much faster than sequential buckets)
             # INSERT ... ON CONFLICT handles any race conditions at the DB level
@@ -219,10 +231,14 @@ async def extract_entities_batch_optimized(
                 entities_data=all_entities_flat,
                 context=context,
                 unit_event_date=None,  # Not used when per-entity dates provided
-                conn=conn  # Use main transaction connection
+                conn=conn,  # Use main transaction connection
             )
 
-            _log(log_buffer, f"    [6.2.2] Resolve entities: {len(all_entities_flat)} entities in single batch in {time.time() - substep_6_2_2_start:.3f}s", level='debug')
+            _log(
+                log_buffer,
+                f"    [6.2.2] Resolve entities: {len(all_entities_flat)} entities in single batch in {time.time() - substep_6_2_2_start:.3f}s",
+                level="debug",
+            )
 
             # [6.2.3] Create unit-entity links in BATCH
             substep_6_2_3_start = time.time()
@@ -239,12 +255,24 @@ async def extract_entities_batch_optimized(
 
             # Batch insert all unit-entity links (MUCH faster!)
             await entity_resolver.link_units_to_entities_batch(unit_entity_pairs, conn=conn)
-            _log(log_buffer, f"    [6.2.3] Create unit-entity links (batched): {len(unit_entity_pairs)} links in {time.time() - substep_6_2_3_start:.3f}s", level='debug')
+            _log(
+                log_buffer,
+                f"    [6.2.3] Create unit-entity links (batched): {len(unit_entity_pairs)} links in {time.time() - substep_6_2_3_start:.3f}s",
+                level="debug",
+            )
 
-            _log(log_buffer, f"  [6.2] Entity resolution (batched): {len(all_entities_flat)} entities resolved in {time.time() - step_6_2_start:.3f}s", level='debug')
+            _log(
+                log_buffer,
+                f"  [6.2] Entity resolution (batched): {len(all_entities_flat)} entities resolved in {time.time() - step_6_2_start:.3f}s",
+                level="debug",
+            )
         else:
             unit_to_entity_ids = {}
-            _log(log_buffer, f"  [6.2] Entity resolution (batched): 0 entities in {time.time() - step_6_2_start:.3f}s", level='debug')
+            _log(
+                log_buffer,
+                f"  [6.2] Entity resolution (batched): 0 entities in {time.time() - step_6_2_start:.3f}s",
+                level="debug",
+            )
 
         # Step 3: Create entity links between units that share entities
         substep_start = time.time()
@@ -253,39 +281,44 @@ async def extract_entities_batch_optimized(
         for entity_ids in unit_to_entity_ids.values():
             all_entity_ids.update(entity_ids)
 
-        _log(log_buffer, f"  [6.3] Creating entity links for {len(all_entity_ids)} unique entities...", level='debug')
+        _log(log_buffer, f"  [6.3] Creating entity links for {len(all_entity_ids)} unique entities...", level="debug")
 
         # Find all units that reference these entities (ONE batched query)
         entity_to_units = {}
         if all_entity_ids:
             query_start = time.time()
             import uuid
+
             entity_id_list = [uuid.UUID(eid) if isinstance(eid, str) else eid for eid in all_entity_ids]
             rows = await conn.fetch(
-                """
+                f"""
                 SELECT entity_id, unit_id
-                FROM unit_entities
+                FROM {fq_table("unit_entities")}
                 WHERE entity_id = ANY($1::uuid[])
                 """,
-                entity_id_list
+                entity_id_list,
+            )
+            _log(
+                log_buffer,
+                f"      [6.3.1] Query unit_entities: {len(rows)} rows in {time.time() - query_start:.3f}s",
+                level="debug",
             )
-            _log(log_buffer, f"      [6.3.1] Query unit_entities: {len(rows)} rows in {time.time() - query_start:.3f}s", level='debug')
 
             # Group by entity_id
             group_start = time.time()
             for row in rows:
-                entity_id = row['entity_id']
+                entity_id = row["entity_id"]
                 if entity_id not in entity_to_units:
                     entity_to_units[entity_id] = []
-                entity_to_units[entity_id].append(row['unit_id'])
-            _log(log_buffer, f"      [6.3.2] Group by entity_id: {time.time() - group_start:.3f}s", level='debug')
+                entity_to_units[entity_id].append(row["unit_id"])
+            _log(log_buffer, f"      [6.3.2] Group by entity_id: {time.time() - group_start:.3f}s", level="debug")
 
         # Create bidirectional links between units that share entities
         # OPTIMIZATION: Limit links per entity to avoid N² explosion
         # Only link each new unit to the most recent MAX_LINKS_PER_ENTITY units
         MAX_LINKS_PER_ENTITY = 50  # Limit to prevent explosion when entity appears in many facts
         link_gen_start = time.time()
-        links: List[EntityLink] = []
+        links: list[EntityLink] = []
         new_unit_set = set(unit_ids)  # Units from this batch
 
         def to_uuid(val) -> UUID:
@@ -299,27 +332,52 @@ async def extract_entities_batch_optimized(
 
             # Link new units to each other (within batch) - also limited
             # For very common entities, limit within-batch links too
-            new_units_to_link = new_units[-MAX_LINKS_PER_ENTITY:] if len(new_units) > MAX_LINKS_PER_ENTITY else new_units
+            new_units_to_link = (
+                new_units[-MAX_LINKS_PER_ENTITY:] if len(new_units) > MAX_LINKS_PER_ENTITY else new_units
+            )
             for i, unit_id_1 in enumerate(new_units_to_link):
-                for unit_id_2 in new_units_to_link[i+1:]:
-                    links.append(EntityLink(from_unit_id=to_uuid(unit_id_1), to_unit_id=to_uuid(unit_id_2), entity_id=entity_uuid))
-                    links.append(EntityLink(from_unit_id=to_uuid(unit_id_2), to_unit_id=to_uuid(unit_id_1), entity_id=entity_uuid))
+                for unit_id_2 in new_units_to_link[i + 1 :]:
+                    links.append(
+                        EntityLink(
+                            from_unit_id=to_uuid(unit_id_1), to_unit_id=to_uuid(unit_id_2), entity_id=entity_uuid
+                        )
+                    )
+                    links.append(
+                        EntityLink(
+                            from_unit_id=to_uuid(unit_id_2), to_unit_id=to_uuid(unit_id_1), entity_id=entity_uuid
+                        )
+                    )
 
             # Link new units to LIMITED existing units (most recent)
             existing_to_link = existing_units[-MAX_LINKS_PER_ENTITY:]  # Take most recent
             for new_unit in new_units:
                 for existing_unit in existing_to_link:
-                    links.append(EntityLink(from_unit_id=to_uuid(new_unit), to_unit_id=to_uuid(existing_unit), entity_id=entity_uuid))
-                    links.append(EntityLink(from_unit_id=to_uuid(existing_unit), to_unit_id=to_uuid(new_unit), entity_id=entity_uuid))
+                    links.append(
+                        EntityLink(
+                            from_unit_id=to_uuid(new_unit), to_unit_id=to_uuid(existing_unit), entity_id=entity_uuid
+                        )
+                    )
+                    links.append(
+                        EntityLink(
+                            from_unit_id=to_uuid(existing_unit), to_unit_id=to_uuid(new_unit), entity_id=entity_uuid
+                        )
+                    )
 
-        _log(log_buffer, f"      [6.3.3] Generate {len(links)} links: {time.time() - link_gen_start:.3f}s", level='debug')
-        _log(log_buffer, f"  [6.3] Entity link creation: {len(links)} links for {len(all_entity_ids)} unique entities in {time.time() - substep_start:.3f}s", level='debug')
+        _log(
+            log_buffer, f"      [6.3.3] Generate {len(links)} links: {time.time() - link_gen_start:.3f}s", level="debug"
+        )
+        _log(
+            log_buffer,
+            f"  [6.3] Entity link creation: {len(links)} links for {len(all_entity_ids)} unique entities in {time.time() - substep_start:.3f}s",
+            level="debug",
+        )
 
         return links
 
     except Exception as e:
         logger.error(f"Failed to extract entities in batch: {str(e)}")
         import traceback
+
         traceback.print_exc()
         raise
 
@@ -327,9 +385,9 @@ async def extract_entities_batch_optimized(
 async def create_temporal_links_batch_per_fact(
     conn,
     bank_id: str,
-    unit_ids: List[str],
+    unit_ids: list[str],
     time_window_hours: int = 24,
-    log_buffer: List[str] = None,
+    log_buffer: list[str] = None,
 ) -> int:
     """
     Create temporal links for multiple units, each with their own event_date.
@@ -356,15 +414,18 @@ async def create_temporal_links_batch_per_fact(
         # Get the event_date for each new unit
         fetch_dates_start = time_mod.time()
         rows = await conn.fetch(
-            """
+            f"""
             SELECT id, event_date
-            FROM memory_units
+            FROM {fq_table("memory_units")}
             WHERE id::text = ANY($1)
             """,
-            unit_ids
+            unit_ids,
+        )
+        new_units = {str(row["id"]): row["event_date"] for row in rows}
+        _log(
+            log_buffer,
+            f"      [7.1] Fetch event_dates for {len(unit_ids)} units: {time_mod.time() - fetch_dates_start:.3f}s",
         )
-        new_units = {str(row['id']): row['event_date'] for row in rows}
-        _log(log_buffer, f"      [7.1] Fetch event_dates for {len(unit_ids)} units: {time_mod.time() - fetch_dates_start:.3f}s")
 
         # Fetch ALL potential temporal neighbors in ONE query (much faster!)
         # Get time range across all units with overflow protection
@@ -372,9 +433,9 @@ async def create_temporal_links_batch_per_fact(
 
         fetch_neighbors_start = time_mod.time()
         all_candidates = await conn.fetch(
-            """
+            f"""
             SELECT id, event_date
-            FROM memory_units
+            FROM {fq_table("memory_units")}
             WHERE bank_id = $1
               AND event_date BETWEEN $2 AND $3
               AND id::text != ALL($4)
@@ -383,9 +444,12 @@ async def create_temporal_links_batch_per_fact(
             bank_id,
             min_date,
             max_date,
-            unit_ids
+            unit_ids,
+        )
+        _log(
+            log_buffer,
+            f"      [7.2] Fetch {len(all_candidates)} candidate neighbors (1 query): {time_mod.time() - fetch_neighbors_start:.3f}s",
         )
-        _log(log_buffer, f"      [7.2] Fetch {len(all_candidates)} candidate neighbors (1 query): {time_mod.time() - fetch_neighbors_start:.3f}s")
 
         # Filter and create links in memory (much faster than N queries)
         link_gen_start = time_mod.time()
@@ -408,20 +472,20 @@ async def create_temporal_links_batch_per_fact(
                     if time_diff_hours <= time_window_hours:
                         weight = max(0.3, 1.0 - (time_diff_hours / time_window_hours))
                         # Create bidirectional links
-                        links.append((unit_id, other_id, 'temporal', weight, None))
-                        links.append((other_id, unit_id, 'temporal', weight, None))
+                        links.append((unit_id, other_id, "temporal", weight, None))
+                        links.append((other_id, unit_id, "temporal", weight, None))
 
         _log(log_buffer, f"      [7.3] Generate {len(links)} temporal links: {time_mod.time() - link_gen_start:.3f}s")
 
         if links:
             insert_start = time_mod.time()
             await conn.executemany(
-                """
-                INSERT INTO memory_links (from_unit_id, to_unit_id, link_type, weight, entity_id)
+                f"""
+                INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
                 VALUES ($1, $2, $3, $4, $5)
                 ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
                 """,
-                links
+                links,
             )
             _log(log_buffer, f"      [7.4] Insert {len(links)} temporal links: {time_mod.time() - insert_start:.3f}s")
 
@@ -430,6 +494,7 @@ async def create_temporal_links_batch_per_fact(
     except Exception as e:
         logger.error(f"Failed to create temporal links: {str(e)}")
         import traceback
+
         traceback.print_exc()
         raise
 
@@ -437,11 +502,11 @@ async def create_temporal_links_batch_per_fact(
 async def create_semantic_links_batch(
     conn,
     bank_id: str,
-    unit_ids: List[str],
-    embeddings: List[List[float]],
+    unit_ids: list[str],
+    embeddings: list[list[float]],
     top_k: int = 5,
     threshold: float = 0.7,
-    log_buffer: List[str] = None,
+    log_buffer: list[str] = None,
 ) -> int:
     """
     Create semantic links for multiple units efficiently.
@@ -465,22 +530,26 @@ async def create_semantic_links_batch(
 
     try:
         import time as time_mod
+
         import numpy as np
 
         # Fetch ALL existing units with embeddings in ONE query
         fetch_start = time_mod.time()
         all_existing = await conn.fetch(
-            """
+            f"""
             SELECT id, embedding
-            FROM memory_units
+            FROM {fq_table("memory_units")}
             WHERE bank_id = $1
               AND embedding IS NOT NULL
               AND id::text != ALL($2)
             """,
             bank_id,
-            unit_ids
+            unit_ids,
+        )
+        _log(
+            log_buffer,
+            f"      [8.1] Fetch {len(all_existing)} existing embeddings (1 query): {time_mod.time() - fetch_start:.3f}s",
         )
-        _log(log_buffer, f"      [8.1] Fetch {len(all_existing)} existing embeddings (1 query): {time_mod.time() - fetch_start:.3f}s")
 
         # Convert to numpy for vectorized similarity computation
         compute_start = time_mod.time()
@@ -488,15 +557,16 @@ async def create_semantic_links_batch(
 
         if all_existing:
             # Convert existing embeddings to numpy array
-            existing_ids = [str(row['id']) for row in all_existing]
+            existing_ids = [str(row["id"]) for row in all_existing]
             # Stack embeddings as 2D array: (num_embeddings, embedding_dim)
             embedding_arrays = []
             for row in all_existing:
-                raw_emb = row['embedding']
+                raw_emb = row["embedding"]
                 # Handle different pgvector formats
                 if isinstance(raw_emb, str):
                     # Parse string format: "[1.0, 2.0, ...]"
                     import json
+
                     emb = np.array(json.loads(raw_emb), dtype=np.float32)
                 elif isinstance(raw_emb, (list, tuple)):
                     emb = np.array(raw_emb, dtype=np.float32)
@@ -537,7 +607,7 @@ async def create_semantic_links_batch(
                         similar_id = existing_ids[idx]
                         # Clamp to [0, 1] to handle floating point precision issues
                         similarity = float(min(1.0, max(0.0, similarities[idx])))
-                        all_links.append((unit_id, similar_id, 'semantic', similarity, None))
+                        all_links.append((unit_id, similar_id, "semantic", similarity, None))
 
         # Also compute similarities WITHIN the new batch (new units to each other)
         # Apply the same top_k limit per unit as we do for existing units
@@ -565,32 +635,38 @@ async def create_semantic_links_batch(
                         other_id = unit_ids[other_idx]
                         # Clamp to [0, 1] to handle floating point precision issues
                         similarity = float(min(1.0, max(0.0, similarities[local_idx])))
-                        all_links.append((unit_id, other_id, 'semantic', similarity, None))
+                        all_links.append((unit_id, other_id, "semantic", similarity, None))
 
-        _log(log_buffer, f"      [8.2] Compute similarities & generate {len(all_links)} semantic links: {time_mod.time() - compute_start:.3f}s")
+        _log(
+            log_buffer,
+            f"      [8.2] Compute similarities & generate {len(all_links)} semantic links: {time_mod.time() - compute_start:.3f}s",
+        )
 
         if all_links:
             insert_start = time_mod.time()
             await conn.executemany(
-                """
-                INSERT INTO memory_links (from_unit_id, to_unit_id, link_type, weight, entity_id)
+                f"""
+                INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
                 VALUES ($1, $2, $3, $4, $5)
                 ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
                 """,
-                all_links
+                all_links,
+            )
+            _log(
+                log_buffer, f"      [8.3] Insert {len(all_links)} semantic links: {time_mod.time() - insert_start:.3f}s"
             )
-            _log(log_buffer, f"      [8.3] Insert {len(all_links)} semantic links: {time_mod.time() - insert_start:.3f}s")
 
         return len(all_links)
 
     except Exception as e:
         logger.error(f"Failed to create semantic links: {str(e)}")
         import traceback
+
         traceback.print_exc()
         raise
 
 
-async def insert_entity_links_batch(conn, links: List[EntityLink], chunk_size: int = 50000):
+async def insert_entity_links_batch(conn, links: list[EntityLink], chunk_size: int = 50000):
     """
     Insert all entity links using COPY to temp table + INSERT for maximum speed.
 
@@ -606,7 +682,6 @@ async def insert_entity_links_batch(conn, links: List[EntityLink], chunk_size: i
     if not links:
         return
 
-    import uuid as uuid_mod
     import time as time_mod
 
     total_start = time_mod.time()
@@ -633,28 +708,22 @@ async def insert_entity_links_batch(conn, links: List[EntityLink], chunk_size: i
     convert_start = time_mod.time()
     records = []
     for link in links:
-        records.append((
-            link.from_unit_id,
-            link.to_unit_id,
-            link.link_type,
-            link.weight,
-            link.entity_id
-        ))
+        records.append((link.from_unit_id, link.to_unit_id, link.link_type, link.weight, link.entity_id))
     logger.debug(f"      [9.3] Convert {len(records)} records: {time_mod.time() - convert_start:.3f}s")
 
     # Bulk load using COPY (fastest method)
     copy_start = time_mod.time()
     await conn.copy_records_to_table(
-        '_temp_entity_links',
+        "_temp_entity_links",
         records=records,
-        columns=['from_unit_id', 'to_unit_id', 'link_type', 'weight', 'entity_id']
+        columns=["from_unit_id", "to_unit_id", "link_type", "weight", "entity_id"],
     )
     logger.debug(f"      [9.4] COPY {len(records)} records to temp table: {time_mod.time() - copy_start:.3f}s")
 
     # Insert from temp table with ON CONFLICT (single query for all rows)
     insert_start = time_mod.time()
-    await conn.execute("""
-        INSERT INTO memory_links (from_unit_id, to_unit_id, link_type, weight, entity_id)
+    await conn.execute(f"""
+        INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
         SELECT from_unit_id, to_unit_id, link_type, weight, entity_id
         FROM _temp_entity_links
         ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
@@ -665,8 +734,8 @@ async def insert_entity_links_batch(conn, links: List[EntityLink], chunk_size: i
 
 async def create_causal_links_batch(
     conn,
-    unit_ids: List[str],
-    causal_relations_per_fact: List[List[dict]],
+    unit_ids: list[str],
+    causal_relations_per_fact: list[list[dict]],
 ) -> int:
     """
     Create causal links between facts based on LLM-extracted causal relationships.
@@ -694,6 +763,7 @@ async def create_causal_links_batch(
 
     try:
         import time as time_mod
+
         create_start = time_mod.time()
 
         # Build links list
@@ -705,12 +775,12 @@ async def create_causal_links_batch(
             from_unit_id = unit_ids[fact_idx]
 
             for relation in causal_relations:
-                target_idx = relation['target_fact_index']
-                relation_type = relation['relation_type']
-                strength = relation.get('strength', 1.0)
+                target_idx = relation["target_fact_index"]
+                relation_type = relation["relation_type"]
+                strength = relation.get("strength", 1.0)
 
                 # Validate relation_type - must match database constraint
-                valid_types = {'causes', 'caused_by', 'enables', 'prevents'}
+                valid_types = {"causes", "caused_by", "enables", "prevents"}
                 if relation_type not in valid_types:
                     logger.error(
                         f"Invalid relation_type '{relation_type}' (type: {type(relation_type).__name__}) "
@@ -735,24 +805,25 @@ async def create_causal_links_batch(
                 # weight is the strength of the relationship
                 links.append((from_unit_id, to_unit_id, relation_type, strength, None))
 
-
         if links:
             insert_start = time_mod.time()
             try:
                 await conn.executemany(
-                    """
-                    INSERT INTO memory_links (from_unit_id, to_unit_id, link_type, weight, entity_id)
+                    f"""
+                    INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
                     VALUES ($1, $2, $3, $4, $5)
                     ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
                     """,
-                    links
+                    links,
                 )
             except Exception as db_error:
                 # Log the actual data being inserted for debugging
                 logger.error(f"Database insert failed for causal links. Error: {db_error}")
                 logger.error(f"Attempted to insert {len(links)} links. First few:")
                 for i, link in enumerate(links[:3]):
-                    logger.error(f"  Link {i}: from={link[0]}, to={link[1]}, type='{link[2]}' (repr={repr(link[2])}), weight={link[3]}, entity={link[4]}")
+                    logger.error(
+                        f"  Link {i}: from={link[0]}, to={link[1]}, type='{link[2]}' (repr={repr(link[2])}), weight={link[3]}, entity={link[4]}"
+                    )
                 raise
 
         return len(links)
@@ -760,5 +831,6 @@ async def create_causal_links_batch(
     except Exception as e:
         logger.error(f"Failed to create causal links: {str(e)}")
         import traceback
+
         traceback.print_exc()
         raise

hindsight-api/hindsight_api/engine/retain/observation_regeneration.py

@@ -3,15 +3,16 @@ Observation regeneration for retain pipeline.
 
 Regenerates entity observations as part of the retain transaction.
 """
+
 import logging
 import time
 import uuid
-from datetime import datetime, timezone
-from typing import List, Dict, Optional
+from datetime import UTC, datetime
 
+from ...config import get_config
+from ..memory_engine import fq_table
 from ..search import observation_utils
 from . import embedding_utils
-from ..db_utils import acquire_with_retry
 from .types import EntityLink
 
 logger = logging.getLogger(__name__)
@@ -19,12 +20,12 @@ logger = logging.getLogger(__name__)
 
 def utcnow():
     """Get current UTC time."""
-    return datetime.now(timezone.utc)
+    return datetime.now(UTC)
 
 
 # Simple dataclass-like container for facts (avoid importing from memory_engine)
 class MemoryFactForObservation:
-    def __init__(self, id: str, text: str, fact_type: str, context: str, occurred_start: Optional[str]):
+    def __init__(self, id: str, text: str, fact_type: str, context: str, occurred_start: str | None):
         self.id = id
         self.text = text
         self.fact_type = fact_type
@@ -33,12 +34,7 @@ class MemoryFactForObservation:
 
 
 async def regenerate_observations_batch(
-    conn,
-    embeddings_model,
-    llm_config,
-    bank_id: str,
-    entity_links: List[EntityLink],
-    log_buffer: List[str] = None
+    conn, embeddings_model, llm_config, bank_id: str, entity_links: list[EntityLink], log_buffer: list[str] = None
 ) -> None:
     """
     Regenerate observations for top entities in this batch.
@@ -54,14 +50,15 @@ async def regenerate_observations_batch(
         entity_links: Entity links from this batch
         log_buffer: Optional log buffer for timing
     """
-    TOP_N_ENTITIES = 5
-    MIN_FACTS_THRESHOLD = 5
+    config = get_config()
+    TOP_N_ENTITIES = config.observation_top_entities
+    MIN_FACTS_THRESHOLD = config.observation_min_facts
 
     if not entity_links:
         return
 
     # Count mentions per entity in this batch
-    entity_mention_counts: Dict[str, int] = {}
+    entity_mention_counts: dict[str, int] = {}
     for link in entity_links:
         if link.entity_id:
             entity_id = str(link.entity_id)
@@ -71,11 +68,7 @@ async def regenerate_observations_batch(
         return
 
     # Sort by mention count descending and take top N
-    sorted_entities = sorted(
-        entity_mention_counts.items(),
-        key=lambda x: x[1],
-        reverse=True
-    )
+    sorted_entities = sorted(entity_mention_counts.items(), key=lambda x: x[1], reverse=True)
     entities_to_process = [e[0] for e in sorted_entities[:TOP_N_ENTITIES]]
 
     obs_start = time.time()
@@ -85,26 +78,28 @@ async def regenerate_observations_batch(
 
     # Batch query for entity names
     entity_rows = await conn.fetch(
-        """
-        SELECT id, canonical_name FROM entities
+        f"""
+        SELECT id, canonical_name FROM {fq_table("entities")}
         WHERE id = ANY($1) AND bank_id = $2
         """,
-        entity_uuids, bank_id
+        entity_uuids,
+        bank_id,
     )
-    entity_names = {row['id']: row['canonical_name'] for row in entity_rows}
+    entity_names = {row["id"]: row["canonical_name"] for row in entity_rows}
 
     # Batch query for fact counts
     fact_counts = await conn.fetch(
-        """
+        f"""
         SELECT ue.entity_id, COUNT(*) as cnt
-        FROM unit_entities ue
-        JOIN memory_units mu ON ue.unit_id = mu.id
+        FROM {fq_table("unit_entities")} ue
+        JOIN {fq_table("memory_units")} mu ON ue.unit_id = mu.id
         WHERE ue.entity_id = ANY($1) AND mu.bank_id = $2
         GROUP BY ue.entity_id
         """,
-        entity_uuids, bank_id
+        entity_uuids,
+        bank_id,
     )
-    entity_fact_counts = {row['entity_id']: row['cnt'] for row in fact_counts}
+    entity_fact_counts = {row["entity_id"]: row["cnt"] for row in fact_counts}
 
     # Filter entities that meet the threshold
     entities_with_names = []
@@ -126,8 +121,7 @@ async def regenerate_observations_batch(
     for entity_id, entity_name in entities_with_names:
         try:
             obs_ids = await _regenerate_entity_observations(
-                conn, embeddings_model, llm_config,
-                bank_id, entity_id, entity_name
+                conn, embeddings_model, llm_config, bank_id, entity_id, entity_name
             )
             total_observations += len(obs_ids)
         except Exception as e:
@@ -135,17 +129,14 @@ async def regenerate_observations_batch(
 
     obs_time = time.time() - obs_start
     if log_buffer is not None:
-        log_buffer.append(f"[11] Observations: {total_observations} observations for {len(entities_with_names)} entities in {obs_time:.3f}s")
+        log_buffer.append(
+            f"[11] Observations: {total_observations} observations for {len(entities_with_names)} entities in {obs_time:.3f}s"
+        )
 
 
 async def _regenerate_entity_observations(
-    conn,
-    embeddings_model,
-    llm_config,
-    bank_id: str,
-    entity_id: str,
-    entity_name: str
-) -> List[str]:
+    conn, embeddings_model, llm_config, bank_id: str, entity_id: str, entity_name: str
+) -> list[str]:
     """
     Regenerate observations for a single entity.
 
@@ -166,17 +157,18 @@ async def _regenerate_entity_observations(
 
     # Get all facts mentioning this entity (exclude observations themselves)
     rows = await conn.fetch(
-        """
+        f"""
         SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.fact_type
-        FROM memory_units mu
-        JOIN unit_entities ue ON mu.id = ue.unit_id
+        FROM {fq_table("memory_units")} mu
+        JOIN {fq_table("unit_entities")} ue ON mu.id = ue.unit_id
         WHERE mu.bank_id = $1
           AND ue.entity_id = $2
           AND mu.fact_type IN ('world', 'experience')
         ORDER BY mu.occurred_start DESC
         LIMIT 50
         """,
-        bank_id, entity_uuid
+        bank_id,
+        entity_uuid,
     )
 
     if not rows:
@@ -185,45 +177,42 @@ async def _regenerate_entity_observations(
     # Convert to fact objects for observation extraction
     facts = []
     for row in rows:
-        occurred_start = row['occurred_start'].isoformat() if row['occurred_start'] else None
-        facts.append(MemoryFactForObservation(
-            id=str(row['id']),
-            text=row['text'],
-            fact_type=row['fact_type'],
-            context=row['context'],
-            occurred_start=occurred_start
-        ))
+        occurred_start = row["occurred_start"].isoformat() if row["occurred_start"] else None
+        facts.append(
+            MemoryFactForObservation(
+                id=str(row["id"]),
+                text=row["text"],
+                fact_type=row["fact_type"],
+                context=row["context"],
+                occurred_start=occurred_start,
+            )
+        )
 
     # Extract observations using LLM
-    observations = await observation_utils.extract_observations_from_facts(
-        llm_config,
-        entity_name,
-        facts
-    )
+    observations = await observation_utils.extract_observations_from_facts(llm_config, entity_name, facts)
 
     if not observations:
         return []
 
     # Delete old observations for this entity
     await conn.execute(
-        """
-        DELETE FROM memory_units
+        f"""
+        DELETE FROM {fq_table("memory_units")}
         WHERE id IN (
             SELECT mu.id
-            FROM memory_units mu
-            JOIN unit_entities ue ON mu.id = ue.unit_id
+            FROM {fq_table("memory_units")} mu
+            JOIN {fq_table("unit_entities")} ue ON mu.id = ue.unit_id
             WHERE mu.bank_id = $1
               AND mu.fact_type = 'observation'
               AND ue.entity_id = $2
         )
         """,
-        bank_id, entity_uuid
+        bank_id,
+        entity_uuid,
     )
 
     # Generate embeddings for new observations
-    embeddings = await embedding_utils.generate_embeddings_batch(
-        embeddings_model, observations
-    )
+    embeddings = await embedding_utils.generate_embeddings_batch(embeddings_model, observations)
 
     # Insert new observations
     current_time = utcnow()
@@ -231,8 +220,8 @@ async def _regenerate_entity_observations(
 
     for obs_text, embedding in zip(observations, embeddings):
         result = await conn.fetchrow(
-            """
-            INSERT INTO memory_units (
+            f"""
+            INSERT INTO {fq_table("memory_units")} (
                 bank_id, text, embedding, context, event_date,
                 occurred_start, occurred_end, mentioned_at,
                 fact_type, access_count
@@ -247,18 +236,19 @@ async def _regenerate_entity_observations(
             current_time,
             current_time,
             current_time,
-            current_time
+            current_time,
         )
-        obs_id = str(result['id'])
+        obs_id = str(result["id"])
         created_ids.append(obs_id)
 
         # Link observation to entity
         await conn.execute(
-            """
-            INSERT INTO unit_entities (unit_id, entity_id)
+            f"""
+            INSERT INTO {fq_table("unit_entities")} (unit_id, entity_id)
             VALUES ($1, $2)
             """,
-            uuid.UUID(obs_id), entity_uuid
+            uuid.UUID(obs_id),
+            entity_uuid,
         )
 
     return created_ids

hindsight-api/hindsight_api/engine/retain/orchestrator.py

@@ -3,31 +3,32 @@ Main orchestrator for the retain pipeline.
 
 Coordinates all retain pipeline modules to store memories efficiently.
 """
+
 import logging
 import time
 import uuid
-from datetime import datetime, timezone
-from typing import List, Dict, Any, Optional
+from datetime import UTC, datetime
 
-from . import bank_utils
 from ..db_utils import acquire_with_retry
+from . import bank_utils
 
 
 def utcnow():
     """Get current UTC time."""
-    return datetime.now(timezone.utc)
+    return datetime.now(UTC)
+
 
-from .types import RetainContent, ExtractedFact, ProcessedFact, EntityLink
 from . import (
-    fact_extraction,
-    embedding_processing,
-    deduplication,
     chunk_storage,
-    fact_storage,
+    deduplication,
+    embedding_processing,
     entity_processing,
+    fact_extraction,
+    fact_storage,
     link_creation,
-    observation_regeneration
+    observation_regeneration,
 )
+from .types import ExtractedFact, ProcessedFact, RetainContent, RetainContentDict
 
 logger = logging.getLogger(__name__)
 
@@ -41,12 +42,12 @@ async def retain_batch(
     format_date_fn,
     duplicate_checker_fn,
     bank_id: str,
-    contents_dicts: List[Dict[str, Any]],
-    document_id: Optional[str] = None,
+    contents_dicts: list[RetainContentDict],
+    document_id: str | None = None,
     is_first_batch: bool = True,
-    fact_type_override: Optional[str] = None,
-    confidence_score: Optional[float] = None,
-) -> List[List[str]]:
+    fact_type_override: str | None = None,
+    confidence_score: float | None = None,
+) -> list[list[str]]:
     """
     Process a batch of content through the retain pipeline.
 
@@ -73,10 +74,10 @@ async def retain_batch(
 
     # Buffer all logs
     log_buffer = []
-    log_buffer.append(f"{'='*60}")
+    log_buffer.append(f"{'=' * 60}")
     log_buffer.append(f"RETAIN_BATCH START: {bank_id}")
     log_buffer.append(f"Batch size: {len(contents_dicts)} content items, {total_chars:,} chars")
-    log_buffer.append(f"{'='*60}")
+    log_buffer.append(f"{'=' * 60}")
 
     # Get bank profile
     profile = await bank_utils.get_bank_profile(pool, bank_id)
@@ -89,23 +90,79 @@ async def retain_batch(
             content=item["content"],
             context=item.get("context", ""),
             event_date=item.get("event_date") or utcnow(),
-            metadata=item.get("metadata", {})
+            metadata=item.get("metadata", {}),
         )
         contents.append(content)
 
     # Step 1: Extract facts from all contents
     step_start = time.time()
-    extract_opinions = (fact_type_override == 'opinion')
+    extract_opinions = fact_type_override == "opinion"
 
     extracted_facts, chunks = await fact_extraction.extract_facts_from_contents(
-        contents,
-        llm_config,
-        agent_name,
-        extract_opinions
+        contents, llm_config, agent_name, extract_opinions
+    )
+    log_buffer.append(
+        f"[1] Extract facts: {len(extracted_facts)} facts, {len(chunks)} chunks from {len(contents)} contents in {time.time() - step_start:.3f}s"
     )
-    log_buffer.append(f"[1] Extract facts: {len(extracted_facts)} facts, {len(chunks)} chunks from {len(contents)} contents in {time.time() - step_start:.3f}s")
 
     if not extracted_facts:
+        # Still need to create document if document_id was provided
+        async with acquire_with_retry(pool) as conn:
+            async with conn.transaction():
+                await fact_storage.ensure_bank_exists(conn, bank_id)
+
+                # Handle document tracking even with no facts
+                if document_id:
+                    combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
+                    retain_params = {}
+                    if contents_dicts:
+                        first_item = contents_dicts[0]
+                        if first_item.get("context"):
+                            retain_params["context"] = first_item["context"]
+                        if first_item.get("event_date"):
+                            retain_params["event_date"] = (
+                                first_item["event_date"].isoformat()
+                                if hasattr(first_item["event_date"], "isoformat")
+                                else str(first_item["event_date"])
+                            )
+                        if first_item.get("metadata"):
+                            retain_params["metadata"] = first_item["metadata"]
+                    await fact_storage.handle_document_tracking(
+                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params
+                    )
+                else:
+                    # Check for per-item document_ids
+                    from collections import defaultdict
+
+                    contents_by_doc = defaultdict(list)
+                    for idx, content_dict in enumerate(contents_dicts):
+                        doc_id = content_dict.get("document_id")
+                        if doc_id:
+                            contents_by_doc[doc_id].append((idx, content_dict))
+
+                    for doc_id, doc_contents in contents_by_doc.items():
+                        combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
+                        retain_params = {}
+                        if doc_contents:
+                            first_item = doc_contents[0][1]
+                            if first_item.get("context"):
+                                retain_params["context"] = first_item["context"]
+                            if first_item.get("event_date"):
+                                retain_params["event_date"] = (
+                                    first_item["event_date"].isoformat()
+                                    if hasattr(first_item["event_date"], "isoformat")
+                                    else str(first_item["event_date"])
+                                )
+                            if first_item.get("metadata"):
+                                retain_params["metadata"] = first_item["metadata"]
+                        await fact_storage.handle_document_tracking(
+                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params
+                        )
+
+        total_time = time.time() - start_time
+        logger.info(
+            f"RETAIN_BATCH COMPLETE: 0 facts extracted from {len(contents)} contents in {total_time:.3f}s (document tracked, no facts)"
+        )
         return [[] for _ in contents]
 
     # Apply fact_type_override if provided
@@ -130,6 +187,7 @@ async def retain_batch(
 
     # Group contents by document_id for document tracking and chunk storage
     from collections import defaultdict
+
     contents_by_doc = defaultdict(list)
     for idx, content_dict in enumerate(contents_dicts):
         doc_id = content_dict.get("document_id")
@@ -155,7 +213,11 @@ async def retain_batch(
                     if first_item.get("context"):
                         retain_params["context"] = first_item["context"]
                     if first_item.get("event_date"):
-                        retain_params["event_date"] = first_item["event_date"].isoformat() if hasattr(first_item["event_date"], "isoformat") else str(first_item["event_date"])
+                        retain_params["event_date"] = (
+                            first_item["event_date"].isoformat()
+                            if hasattr(first_item["event_date"], "isoformat")
+                            else str(first_item["event_date"])
+                        )
                     if first_item.get("metadata"):
                         retain_params["metadata"] = first_item["metadata"]
 
@@ -195,7 +257,11 @@ async def retain_batch(
                                 if first_item.get("context"):
                                     retain_params["context"] = first_item["context"]
                                 if first_item.get("event_date"):
-                                    retain_params["event_date"] = first_item["event_date"].isoformat() if hasattr(first_item["event_date"], "isoformat") else str(first_item["event_date"])
+                                    retain_params["event_date"] = (
+                                        first_item["event_date"].isoformat()
+                                        if hasattr(first_item["event_date"], "isoformat")
+                                        else str(first_item["event_date"])
+                                    )
                                 if first_item.get("metadata"):
                                     retain_params["metadata"] = first_item["metadata"]
 
@@ -205,7 +271,9 @@ async def retain_batch(
                             document_ids_added.append(actual_doc_id)
 
             if document_ids_added:
-                log_buffer.append(f"[2.5] Document tracking: {len(document_ids_added)} documents in {time.time() - step_start:.3f}s")
+                log_buffer.append(
+                    f"[2.5] Document tracking: {len(document_ids_added)} documents in {time.time() - step_start:.3f}s"
+                )
 
             # Store chunks and map to facts for all documents
             step_start = time.time()
@@ -230,7 +298,9 @@ async def retain_batch(
                     for chunk_idx, chunk_id in chunk_id_map.items():
                         chunk_id_map_by_doc[(doc_id, chunk_idx)] = chunk_id
 
-                log_buffer.append(f"[3] Store chunks: {len(chunks)} chunks for {len(chunks_by_doc)} documents in {time.time() - step_start:.3f}s")
+                log_buffer.append(
+                    f"[3] Store chunks: {len(chunks)} chunks for {len(chunks_by_doc)} documents in {time.time() - step_start:.3f}s"
+                )
 
                 # Map chunk_ids and document_ids to facts
                 for fact, processed_fact in zip(extracted_facts, processed_facts):
@@ -265,7 +335,9 @@ async def retain_batch(
             is_duplicate_flags = await deduplication.check_duplicates_batch(
                 conn, bank_id, processed_facts, duplicate_checker_fn
             )
-            log_buffer.append(f"[4] Deduplication: {sum(is_duplicate_flags)} duplicates in {time.time() - step_start:.3f}s")
+            log_buffer.append(
+                f"[4] Deduplication: {sum(is_duplicate_flags)} duplicates in {time.time() - step_start:.3f}s"
+            )
 
             # Filter out duplicates
             non_duplicate_facts = deduplication.filter_duplicates(processed_facts, is_duplicate_flags)
@@ -293,14 +365,18 @@ async def retain_batch(
             # Create semantic links
             step_start = time.time()
             embeddings_for_links = [fact.embedding for fact in non_duplicate_facts]
-            semantic_link_count = await link_creation.create_semantic_links_batch(conn, bank_id, unit_ids, embeddings_for_links)
+            semantic_link_count = await link_creation.create_semantic_links_batch(
+                conn, bank_id, unit_ids, embeddings_for_links
+            )
             log_buffer.append(f"[8] Semantic links: {semantic_link_count} links in {time.time() - step_start:.3f}s")
 
             # Insert entity links
             step_start = time.time()
             if entity_links:
                 await entity_processing.insert_entity_links_batch(conn, entity_links)
-            log_buffer.append(f"[9] Entity links: {len(entity_links) if entity_links else 0} links in {time.time() - step_start:.3f}s")
+            log_buffer.append(
+                f"[9] Entity links: {len(entity_links) if entity_links else 0} links in {time.time() - step_start:.3f}s"
+            )
 
             # Create causal links
             step_start = time.time()
@@ -309,34 +385,22 @@ async def retain_batch(
 
             # Regenerate observations INSIDE transaction for atomicity
             await observation_regeneration.regenerate_observations_batch(
-                conn,
-                embeddings_model,
-                llm_config,
-                bank_id,
-                entity_links,
-                log_buffer
+                conn, embeddings_model, llm_config, bank_id, entity_links, log_buffer
             )
 
             # Map results back to original content items
-            result_unit_ids = _map_results_to_contents(
-                contents, extracted_facts, is_duplicate_flags, unit_ids
-            )
+            result_unit_ids = _map_results_to_contents(contents, extracted_facts, is_duplicate_flags, unit_ids)
 
         # Trigger background tasks AFTER transaction commits (opinion reinforcement only)
-        await _trigger_background_tasks(
-            task_backend,
-            bank_id,
-            unit_ids,
-            non_duplicate_facts
-        )
+        await _trigger_background_tasks(task_backend, bank_id, unit_ids, non_duplicate_facts)
 
         # Log final summary
         total_time = time.time() - start_time
-        log_buffer.append(f"{'='*60}")
+        log_buffer.append(f"{'=' * 60}")
         log_buffer.append(f"RETAIN_BATCH COMPLETE: {len(unit_ids)} units in {total_time:.3f}s")
         if document_ids_added:
             log_buffer.append(f"Documents: {', '.join(document_ids_added)}")
-        log_buffer.append(f"{'='*60}")
+        log_buffer.append(f"{'=' * 60}")
 
         logger.info("\n" + "\n".join(log_buffer) + "\n")
 
@@ -344,11 +408,11 @@ async def retain_batch(
 
 
 def _map_results_to_contents(
-    contents: List[RetainContent],
-    extracted_facts: List[ExtractedFact],
-    is_duplicate_flags: List[bool],
-    unit_ids: List[str]
-) -> List[List[str]]:
+    contents: list[RetainContent],
+    extracted_facts: list[ExtractedFact],
+    is_duplicate_flags: list[bool],
+    unit_ids: list[str],
+) -> list[list[str]]:
     """
     Map created unit IDs back to original content items.
 
@@ -376,17 +440,19 @@ def _map_results_to_contents(
 async def _trigger_background_tasks(
     task_backend,
     bank_id: str,
-    unit_ids: List[str],
-    facts: List[ProcessedFact],
+    unit_ids: list[str],
+    facts: list[ProcessedFact],
 ) -> None:
     """Trigger opinion reinforcement as background task (after transaction commits)."""
     # Trigger opinion reinforcement if there are entities
     fact_entities = [[e.name for e in fact.entities] for fact in facts]
     if any(fact_entities):
-        await task_backend.submit_task({
-            'type': 'reinforce_opinion',
-            'bank_id': bank_id,
-            'created_unit_ids': unit_ids,
-            'unit_texts': [fact.fact_text for fact in facts],
-            'unit_entities': fact_entities
-        })
+        await task_backend.submit_task(
+            {
+                "type": "reinforce_opinion",
+                "bank_id": bank_id,
+                "created_unit_ids": unit_ids,
+                "unit_texts": [fact.fact_text for fact in facts],
+                "unit_entities": fact_entities,
+            }
+        )

hindsight-api/hindsight_api/engine/retain/types.py

@@ -6,11 +6,34 @@ from content input to fact storage.
 """
 
 from dataclasses import dataclass, field
-from typing import List, Optional, Dict, Any
-from datetime import datetime
+from datetime import UTC, datetime
+from typing import TypedDict
 from uuid import UUID
 
 
+class RetainContentDict(TypedDict, total=False):
+    """Type definition for content items in retain_batch_async.
+
+    Fields:
+        content: Text content to store (required)
+        context: Context about the content (optional)
+        event_date: When the content occurred (optional, defaults to now)
+        metadata: Custom key-value metadata (optional)
+        document_id: Document ID for this content item (optional)
+    """
+
+    content: str  # Required
+    context: str
+    event_date: datetime
+    metadata: dict[str, str]
+    document_id: str
+
+
+def _now_utc() -> datetime:
+    """Factory function for default event_date."""
+    return datetime.now(UTC)
+
+
 @dataclass
 class RetainContent:
     """
@@ -18,16 +41,11 @@ class RetainContent:
 
     Represents a single piece of content to extract facts from.
     """
+
     content: str
     context: str = ""
-    event_date: Optional[datetime] = None
-    metadata: Dict[str, str] = field(default_factory=dict)
-
-    def __post_init__(self):
-        """Ensure event_date is set."""
-        if self.event_date is None:
-            from datetime import datetime, timezone
-            self.event_date = datetime.now(timezone.utc)
+    event_date: datetime = field(default_factory=_now_utc)
+    metadata: dict[str, str] = field(default_factory=dict)
 
 
 @dataclass
@@ -37,6 +55,7 @@ class ChunkMetadata:
 
     Used to track which facts were extracted from which chunks.
     """
+
     chunk_text: str
     fact_count: int
     content_index: int  # Index of the source content
@@ -50,9 +69,10 @@ class EntityRef:
 
     Entities are extracted by the LLM during fact extraction.
     """
+
     name: str
-    canonical_name: Optional[str] = None  # Resolved canonical name
-    entity_id: Optional[UUID] = None  # Resolved entity ID
+    canonical_name: str | None = None  # Resolved canonical name
+    entity_id: UUID | None = None  # Resolved entity ID
 
 
 @dataclass
@@ -62,6 +82,7 @@ class CausalRelation:
 
     Represents how one fact causes, enables, or prevents another.
     """
+
     relation_type: str  # "causes", "enables", "prevents", "caused_by"
     target_fact_index: int  # Index of the target fact in the batch
     strength: float = 1.0  # Strength of the causal relationship
@@ -74,20 +95,21 @@ class ExtractedFact:
 
     This is the raw output from fact extraction before processing.
     """
+
     fact_text: str
     fact_type: str  # "world", "experience", "opinion", "observation"
-    entities: List[str] = field(default_factory=list)
-    occurred_start: Optional[datetime] = None
-    occurred_end: Optional[datetime] = None
-    where: Optional[str] = None  # WHERE the fact occurred or is about
-    causal_relations: List[CausalRelation] = field(default_factory=list)
+    entities: list[str] = field(default_factory=list)
+    occurred_start: datetime | None = None
+    occurred_end: datetime | None = None
+    where: str | None = None  # WHERE the fact occurred or is about
+    causal_relations: list[CausalRelation] = field(default_factory=list)
 
     # Context from the content item
     content_index: int = 0  # Which content this fact came from
     chunk_index: int = 0  # Which chunk this fact came from
     context: str = ""
-    mentioned_at: Optional[datetime] = None
-    metadata: Dict[str, str] = field(default_factory=dict)
+    mentioned_at: datetime | None = None
+    metadata: dict[str, str] = field(default_factory=dict)
 
 
 @dataclass
@@ -97,37 +119,38 @@ class ProcessedFact:
 
     Includes resolved entities, embeddings, and all necessary fields.
     """
+
     # Core fact data
     fact_text: str
     fact_type: str
-    embedding: List[float]
+    embedding: list[float]
 
     # Temporal data
-    occurred_start: Optional[datetime]
-    occurred_end: Optional[datetime]
+    occurred_start: datetime | None
+    occurred_end: datetime | None
     mentioned_at: datetime
 
     # Context and metadata
     context: str
-    metadata: Dict[str, str]
+    metadata: dict[str, str]
 
     # Location data
-    where: Optional[str] = None
+    where: str | None = None
 
     # Entities
-    entities: List[EntityRef] = field(default_factory=list)
+    entities: list[EntityRef] = field(default_factory=list)
 
     # Causal relations
-    causal_relations: List[CausalRelation] = field(default_factory=list)
+    causal_relations: list[CausalRelation] = field(default_factory=list)
 
     # Chunk reference
-    chunk_id: Optional[str] = None
+    chunk_id: str | None = None
 
     # Document reference (denormalized for query performance)
-    document_id: Optional[str] = None
+    document_id: str | None = None
 
     # DB fields (set after insertion)
-    unit_id: Optional[UUID] = None
+    unit_id: UUID | None = None
 
     @property
     def is_duplicate(self) -> bool:
@@ -136,10 +159,8 @@ class ProcessedFact:
 
     @staticmethod
     def from_extracted_fact(
-        extracted_fact: 'ExtractedFact',
-        embedding: List[float],
-        chunk_id: Optional[str] = None
-    ) -> 'ProcessedFact':
+        extracted_fact: "ExtractedFact", embedding: list[float], chunk_id: str | None = None
+    ) -> "ProcessedFact":
         """
         Create ProcessedFact from ExtractedFact.
 
@@ -151,12 +172,12 @@ class ProcessedFact:
         Returns:
             ProcessedFact ready for storage
         """
-        from datetime import datetime, timezone
+        from datetime import datetime
 
         # Use occurred dates only if explicitly provided by LLM
         occurred_start = extracted_fact.occurred_start
         occurred_end = extracted_fact.occurred_end
-        mentioned_at = extracted_fact.mentioned_at or datetime.now(timezone.utc)
+        mentioned_at = extracted_fact.mentioned_at or datetime.now(UTC)
 
         # Convert entity strings to EntityRef objects
         entities = [EntityRef(name=name) for name in extracted_fact.entities]
@@ -172,7 +193,7 @@ class ProcessedFact:
             metadata=extracted_fact.metadata,
             entities=entities,
             causal_relations=extracted_fact.causal_relations,
-            chunk_id=chunk_id
+            chunk_id=chunk_id,
         )
 
 
@@ -183,10 +204,11 @@ class EntityLink:
 
     Used for entity-based graph connections in the memory graph.
     """
+
     from_unit_id: UUID
     to_unit_id: UUID
     entity_id: UUID
-    link_type: str = 'entity'
+    link_type: str = "entity"
     weight: float = 1.0
 
 
@@ -197,24 +219,25 @@ class RetainBatch:
 
     Tracks all facts, chunks, and metadata for a batch operation.
     """
+
     bank_id: str
-    contents: List[RetainContent]
-    document_id: Optional[str] = None
-    fact_type_override: Optional[str] = None
-    confidence_score: Optional[float] = None
+    contents: list[RetainContent]
+    document_id: str | None = None
+    fact_type_override: str | None = None
+    confidence_score: float | None = None
 
     # Extracted data (populated during processing)
-    extracted_facts: List[ExtractedFact] = field(default_factory=list)
-    processed_facts: List[ProcessedFact] = field(default_factory=list)
-    chunks: List[ChunkMetadata] = field(default_factory=list)
+    extracted_facts: list[ExtractedFact] = field(default_factory=list)
+    processed_facts: list[ProcessedFact] = field(default_factory=list)
+    chunks: list[ChunkMetadata] = field(default_factory=list)
 
     # Results (populated after storage)
-    unit_ids_by_content: List[List[str]] = field(default_factory=list)
+    unit_ids_by_content: list[list[str]] = field(default_factory=list)
 
-    def get_facts_for_content(self, content_index: int) -> List[ExtractedFact]:
+    def get_facts_for_content(self, content_index: int) -> list[ExtractedFact]:
         """Get all extracted facts for a specific content item."""
         return [f for f in self.extracted_facts if f.content_index == content_index]
 
-    def get_chunks_for_content(self, content_index: int) -> List[ChunkMetadata]:
+    def get_chunks_for_content(self, content_index: int) -> list[ChunkMetadata]:
         """Get all chunks for a specific content item."""
         return [c for c in self.chunks if c.content_index == content_index]

hindsight-api/hindsight_api/engine/search/__init__.py

@@ -7,15 +7,15 @@ Provides modular search architecture:
 - Reranking: Pluggable strategies (heuristic, cross-encoder)
 """
 
-from .retrieval import (
-    retrieve_parallel,
-    get_default_graph_retriever,
-    set_default_graph_retriever,
-    ParallelRetrievalResult,
-)
-from .graph_retrieval import GraphRetriever, BFSGraphRetriever
+from .graph_retrieval import BFSGraphRetriever, GraphRetriever
 from .mpfp_retrieval import MPFPGraphRetriever
 from .reranking import CrossEncoderReranker
+from .retrieval import (
+    ParallelRetrievalResult,
+    get_default_graph_retriever,
+    retrieve_parallel,
+    set_default_graph_retriever,
+)
 
 __all__ = [
     "retrieve_parallel",

hindsight-api/hindsight_api/engine/search/fusion.py

@@ -2,15 +2,12 @@
 Helper functions for hybrid search (semantic + BM25 + graph).
 """
 
-from typing import List, Dict, Any, Tuple
-import asyncio
-from .types import RetrievalResult, MergedCandidate
+from typing import Any
+
+from .types import MergedCandidate, RetrievalResult
 
 
-def reciprocal_rank_fusion(
-    result_lists: List[List[RetrievalResult]],
-    k: int = 60
-) -> List[MergedCandidate]:
+def reciprocal_rank_fusion(result_lists: list[list[RetrievalResult]], k: int = 60) -> list[MergedCandidate]:
     """
     Merge multiple ranked result lists using Reciprocal Rank Fusion.
 
@@ -73,20 +70,14 @@ def reciprocal_rank_fusion(
         sorted(rrf_scores.items(), key=lambda x: x[1], reverse=True), start=1
     ):
         merged_candidate = MergedCandidate(
-            retrieval=all_retrievals[doc_id],
-            rrf_score=rrf_score,
-            rrf_rank=rrf_rank,
-            source_ranks=source_ranks[doc_id]
+            retrieval=all_retrievals[doc_id], rrf_score=rrf_score, rrf_rank=rrf_rank, source_ranks=source_ranks[doc_id]
         )
         merged_results.append(merged_candidate)
 
     return merged_results
 
 
-def normalize_scores_on_deltas(
-    results: List[Dict[str, Any]],
-    score_keys: List[str]
-) -> List[Dict[str, Any]]:
+def normalize_scores_on_deltas(results: list[dict[str, Any]], score_keys: list[str]) -> list[dict[str, Any]]:
     """
     Normalize scores based on deltas (min-max normalization within result set).
 

hindsight-api/hindsight_api/engine/search/graph_retrieval.py

@@ -6,13 +6,12 @@ allowing different algorithms (BFS spreading activation, PPR, etc.) to be
 swapped without changing the rest of the recall pipeline.
 """
 
-from abc import ABC, abstractmethod
-from typing import List, Optional
-from datetime import datetime
 import logging
+from abc import ABC, abstractmethod
 
-from .types import RetrievalResult
 from ..db_utils import acquire_with_retry
+from ..memory_engine import fq_table
+from .types import RetrievalResult
 
 logger = logging.getLogger(__name__)
 
@@ -40,10 +39,10 @@ class GraphRetriever(ABC):
         bank_id: str,
         fact_type: str,
         budget: int,
-        query_text: Optional[str] = None,
-        semantic_seeds: Optional[List[RetrievalResult]] = None,
-        temporal_seeds: Optional[List[RetrievalResult]] = None,
-    ) -> List[RetrievalResult]:
+        query_text: str | None = None,
+        semantic_seeds: list[RetrievalResult] | None = None,
+        temporal_seeds: list[RetrievalResult] | None = None,
+    ) -> list[RetrievalResult]:
         """
         Retrieve relevant facts via graph traversal.
 
@@ -109,10 +108,10 @@ class BFSGraphRetriever(GraphRetriever):
         bank_id: str,
         fact_type: str,
         budget: int,
-        query_text: Optional[str] = None,
-        semantic_seeds: Optional[List[RetrievalResult]] = None,
-        temporal_seeds: Optional[List[RetrievalResult]] = None,
-    ) -> List[RetrievalResult]:
+        query_text: str | None = None,
+        semantic_seeds: list[RetrievalResult] | None = None,
+        temporal_seeds: list[RetrievalResult] | None = None,
+    ) -> list[RetrievalResult]:
         """
         Retrieve facts using BFS spreading activation.
 
@@ -127,9 +126,7 @@ class BFSGraphRetriever(GraphRetriever):
         for interface compatibility but not used.
         """
         async with acquire_with_retry(pool) as conn:
-            return await self._retrieve_with_conn(
-                conn, query_embedding_str, bank_id, fact_type, budget
-            )
+            return await self._retrieve_with_conn(conn, query_embedding_str, bank_id, fact_type, budget)
 
     async def _retrieve_with_conn(
         self,
@@ -138,16 +135,16 @@ class BFSGraphRetriever(GraphRetriever):
         bank_id: str,
         fact_type: str,
         budget: int,
-    ) -> List[RetrievalResult]:
+    ) -> list[RetrievalResult]:
         """Internal implementation with connection."""
 
         # Step 1: Find entry points
         entry_points = await conn.fetch(
-            """
+            f"""
             SELECT id, text, context, event_date, occurred_start, occurred_end,
                    mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
                    1 - (embedding <=> $1::vector) AS similarity
-            FROM memory_units
+            FROM {fq_table("memory_units")}
             WHERE bank_id = $2
               AND embedding IS NOT NULL
               AND fact_type = $3
@@ -155,8 +152,11 @@ class BFSGraphRetriever(GraphRetriever):
             ORDER BY embedding <=> $1::vector
             LIMIT $5
             """,
-            query_embedding_str, bank_id, fact_type,
-            self.entry_point_threshold, self.entry_point_limit
+            query_embedding_str,
+            bank_id,
+            fact_type,
+            self.entry_point_threshold,
+            self.entry_point_limit,
         )
 
         if not entry_points:
@@ -165,10 +165,7 @@ class BFSGraphRetriever(GraphRetriever):
         # Step 2: BFS spreading activation
         visited = set()
         results = []
-        queue = [
-            (RetrievalResult.from_db_row(dict(r)), r["similarity"])
-            for r in entry_points
-        ]
+        queue = [(RetrievalResult.from_db_row(dict(r)), r["similarity"]) for r in entry_points]
         budget_remaining = budget
 
         while queue and budget_remaining > 0:
@@ -192,20 +189,23 @@ class BFSGraphRetriever(GraphRetriever):
             if batch_nodes and budget_remaining > 0:
                 max_neighbors = len(batch_nodes) * 20
                 neighbors = await conn.fetch(
-                    """
+                    f"""
                     SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.occurred_end,
                            mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type,
                            mu.document_id, mu.chunk_id,
                            ml.weight, ml.link_type, ml.from_unit_id
-                    FROM memory_links ml
-                    JOIN memory_units mu ON ml.to_unit_id = mu.id
+                    FROM {fq_table("memory_links")} ml
+                    JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
                     WHERE ml.from_unit_id = ANY($1::uuid[])
                       AND ml.weight >= $2
                       AND mu.fact_type = $3
                     ORDER BY ml.weight DESC
                     LIMIT $4
                     """,
-                    batch_nodes, self.min_activation, fact_type, max_neighbors
+                    batch_nodes,
+                    self.min_activation,
+                    fact_type,
+                    max_neighbors,
                 )
 
                 for n in neighbors:

hindsight-api/hindsight_api/engine/search/mpfp_retrieval.py

@@ -16,13 +16,13 @@ Key properties:
 
 import asyncio
 import logging
-from dataclasses import dataclass, field
-from typing import List, Dict, Optional, Tuple
 from collections import defaultdict
+from dataclasses import dataclass, field
 
-from .types import RetrievalResult
-from .graph_retrieval import GraphRetriever
 from ..db_utils import acquire_with_retry
+from ..memory_engine import fq_table
+from .graph_retrieval import GraphRetriever
+from .types import RetrievalResult
 
 logger = logging.getLogger(__name__)
 
@@ -31,9 +31,11 @@ logger = logging.getLogger(__name__)
 # Data Classes
 # -----------------------------------------------------------------------------
 
+
 @dataclass
 class EdgeTarget:
     """A neighbor node with its edge weight."""
+
     node_id: str
     weight: float
 
@@ -41,19 +43,15 @@ class EdgeTarget:
 @dataclass
 class TypedAdjacency:
     """Adjacency lists split by edge type."""
-    # edge_type -> from_node_id -> list of (to_node_id, weight)
-    graphs: Dict[str, Dict[str, List[EdgeTarget]]] = field(default_factory=dict)
 
-    def get_neighbors(self, edge_type: str, node_id: str) -> List[EdgeTarget]:
+    # edge_type -> from_node_id -> list of (to_node_id, weight)
+    graphs: dict[str, dict[str, list[EdgeTarget]]] = field(default_factory=dict)
+
+    def get_neighbors(self, edge_type: str, node_id: str) -> list[EdgeTarget]:
         """Get neighbors for a node via a specific edge type."""
         return self.graphs.get(edge_type, {}).get(node_id, [])
 
-    def get_normalized_neighbors(
-        self,
-        edge_type: str,
-        node_id: str,
-        top_k: int
-    ) -> List[EdgeTarget]:
+    def get_normalized_neighbors(self, edge_type: str, node_id: str, top_k: int) -> list[EdgeTarget]:
         """Get top-k neighbors with weights normalized to sum to 1."""
         neighbors = self.get_neighbors(edge_type, node_id)[:top_k]
         if not neighbors:
@@ -63,45 +61,49 @@ class TypedAdjacency:
         if total == 0:
             return []
 
-        return [
-            EdgeTarget(node_id=n.node_id, weight=n.weight / total)
-            for n in neighbors
-        ]
+        return [EdgeTarget(node_id=n.node_id, weight=n.weight / total) for n in neighbors]
 
 
 @dataclass
 class PatternResult:
     """Result from a single pattern traversal."""
-    pattern: List[str]
-    scores: Dict[str, float]  # node_id -> accumulated mass
+
+    pattern: list[str]
+    scores: dict[str, float]  # node_id -> accumulated mass
 
 
 @dataclass
 class MPFPConfig:
     """Configuration for MPFP algorithm."""
-    alpha: float = 0.15           # teleport/keep probability
-    threshold: float = 1e-6       # mass pruning threshold (lower = explore more)
-    top_k_neighbors: int = 20     # fan-out limit per node
+
+    alpha: float = 0.15  # teleport/keep probability
+    threshold: float = 1e-6  # mass pruning threshold (lower = explore more)
+    top_k_neighbors: int = 20  # fan-out limit per node
 
     # Patterns from semantic seeds
-    patterns_semantic: List[List[str]] = field(default_factory=lambda: [
-        ['semantic', 'semantic'],   # topic expansion
-        ['entity', 'temporal'],     # entity timeline
-        ['semantic', 'causes'],     # reasoning chains (forward)
-        ['semantic', 'caused_by'],  # reasoning chains (backward)
-        ['entity', 'semantic'],     # entity context
-    ])
+    patterns_semantic: list[list[str]] = field(
+        default_factory=lambda: [
+            ["semantic", "semantic"],  # topic expansion
+            ["entity", "temporal"],  # entity timeline
+            ["semantic", "causes"],  # reasoning chains (forward)
+            ["semantic", "caused_by"],  # reasoning chains (backward)
+            ["entity", "semantic"],  # entity context
+        ]
+    )
 
     # Patterns from temporal seeds
-    patterns_temporal: List[List[str]] = field(default_factory=lambda: [
-        ['temporal', 'semantic'],   # what was happening then
-        ['temporal', 'entity'],     # who was involved then
-    ])
+    patterns_temporal: list[list[str]] = field(
+        default_factory=lambda: [
+            ["temporal", "semantic"],  # what was happening then
+            ["temporal", "entity"],  # who was involved then
+        ]
+    )
 
 
 @dataclass
 class SeedNode:
     """An entry point node with its initial score."""
+
     node_id: str
     score: float  # initial mass (e.g., similarity score)
 
@@ -110,9 +112,10 @@ class SeedNode:
 # Core Algorithm
 # -----------------------------------------------------------------------------
 
+
 def mpfp_traverse(
-    seeds: List[SeedNode],
-    pattern: List[str],
+    seeds: list[SeedNode],
+    pattern: list[str],
     adjacency: TypedAdjacency,
     config: MPFPConfig,
 ) -> PatternResult:
@@ -131,20 +134,18 @@ def mpfp_traverse(
     if not seeds:
         return PatternResult(pattern=pattern, scores={})
 
-    scores: Dict[str, float] = {}
+    scores: dict[str, float] = {}
 
     # Initialize frontier with seed masses (normalized)
     total_seed_score = sum(s.score for s in seeds)
     if total_seed_score == 0:
         total_seed_score = len(seeds)  # fallback to uniform
 
-    frontier: Dict[str, float] = {
-        s.node_id: s.score / total_seed_score for s in seeds
-    }
+    frontier: dict[str, float] = {s.node_id: s.score / total_seed_score for s in seeds}
 
     # Follow pattern hop by hop
     for edge_type in pattern:
-        next_frontier: Dict[str, float] = {}
+        next_frontier: dict[str, float] = {}
 
         for node_id, mass in frontier.items():
             if mass < config.threshold:
@@ -155,15 +156,10 @@ def mpfp_traverse(
 
             # Push (1-α) to neighbors
             push_mass = (1 - config.alpha) * mass
-            neighbors = adjacency.get_normalized_neighbors(
-                edge_type, node_id, config.top_k_neighbors
-            )
+            neighbors = adjacency.get_normalized_neighbors(edge_type, node_id, config.top_k_neighbors)
 
             for neighbor in neighbors:
-                next_frontier[neighbor.node_id] = (
-                    next_frontier.get(neighbor.node_id, 0) +
-                    push_mass * neighbor.weight
-                )
+                next_frontier[neighbor.node_id] = next_frontier.get(neighbor.node_id, 0) + push_mass * neighbor.weight
 
         frontier = next_frontier
 
@@ -176,10 +172,10 @@ def mpfp_traverse(
 
 
 def rrf_fusion(
-    results: List[PatternResult],
+    results: list[PatternResult],
     k: int = 60,
     top_k: int = 50,
-) -> List[Tuple[str, float]]:
+) -> list[tuple[str, float]]:
     """
     Reciprocal Rank Fusion to combine pattern results.
 
@@ -191,28 +187,20 @@ def rrf_fusion(
     Returns:
         List of (node_id, fused_score) tuples, sorted by score descending
     """
-    fused: Dict[str, float] = {}
+    fused: dict[str, float] = {}
 
     for result in results:
         if not result.scores:
             continue
 
         # Rank nodes by their score in this pattern
-        ranked = sorted(
-            result.scores.keys(),
-            key=lambda n: result.scores[n],
-            reverse=True
-        )
+        ranked = sorted(result.scores.keys(), key=lambda n: result.scores[n], reverse=True)
 
         for rank, node_id in enumerate(ranked):
             fused[node_id] = fused.get(node_id, 0) + 1.0 / (k + rank + 1)
 
     # Sort by fused score and return top-k
-    sorted_results = sorted(
-        fused.items(),
-        key=lambda x: x[1],
-        reverse=True
-    )
+    sorted_results = sorted(fused.items(), key=lambda x: x[1], reverse=True)
 
     return sorted_results[:top_k]
 
@@ -221,6 +209,7 @@ def rrf_fusion(
 # Database Loading
 # -----------------------------------------------------------------------------
 
+
 async def load_typed_adjacency(pool, bank_id: str) -> TypedAdjacency:
     """
     Load all edges for a bank, split by edge type.
@@ -229,54 +218,50 @@ async def load_typed_adjacency(pool, bank_id: str) -> TypedAdjacency:
     """
     async with acquire_with_retry(pool) as conn:
         rows = await conn.fetch(
-            """
+            f"""
             SELECT ml.from_unit_id, ml.to_unit_id, ml.link_type, ml.weight
-            FROM memory_links ml
-            JOIN memory_units mu ON ml.from_unit_id = mu.id
+            FROM {fq_table("memory_links")} ml
+            JOIN {fq_table("memory_units")} mu ON ml.from_unit_id = mu.id
             WHERE mu.bank_id = $1
               AND ml.weight >= 0.1
             ORDER BY ml.from_unit_id, ml.weight DESC
             """,
-            bank_id
+            bank_id,
         )
 
-    graphs: Dict[str, Dict[str, List[EdgeTarget]]] = defaultdict(
-        lambda: defaultdict(list)
-    )
+    graphs: dict[str, dict[str, list[EdgeTarget]]] = defaultdict(lambda: defaultdict(list))
 
     for row in rows:
-        from_id = str(row['from_unit_id'])
-        to_id = str(row['to_unit_id'])
-        link_type = row['link_type']
-        weight = row['weight']
+        from_id = str(row["from_unit_id"])
+        to_id = str(row["to_unit_id"])
+        link_type = row["link_type"]
+        weight = row["weight"]
 
-        graphs[link_type][from_id].append(
-            EdgeTarget(node_id=to_id, weight=weight)
-        )
+        graphs[link_type][from_id].append(EdgeTarget(node_id=to_id, weight=weight))
 
     return TypedAdjacency(graphs=dict(graphs))
 
 
 async def fetch_memory_units_by_ids(
     pool,
-    node_ids: List[str],
+    node_ids: list[str],
     fact_type: str,
-) -> List[RetrievalResult]:
+) -> list[RetrievalResult]:
     """Fetch full memory unit details for a list of node IDs."""
     if not node_ids:
         return []
 
     async with acquire_with_retry(pool) as conn:
         rows = await conn.fetch(
-            """
+            f"""
             SELECT id, text, context, event_date, occurred_start, occurred_end,
                    mentioned_at, access_count, embedding, fact_type, document_id, chunk_id
-            FROM memory_units
+            FROM {fq_table("memory_units")}
             WHERE id = ANY($1::uuid[])
               AND fact_type = $2
             """,
             node_ids,
-            fact_type
+            fact_type,
         )
 
     return [RetrievalResult.from_db_row(dict(r)) for r in rows]
@@ -286,6 +271,7 @@ async def fetch_memory_units_by_ids(
 # Graph Retriever Implementation
 # -----------------------------------------------------------------------------
 
+
 class MPFPGraphRetriever(GraphRetriever):
     """
     Graph retrieval using Meta-Path Forward Push.
@@ -294,7 +280,7 @@ class MPFPGraphRetriever(GraphRetriever):
     then fuses results via RRF.
     """
 
-    def __init__(self, config: Optional[MPFPConfig] = None):
+    def __init__(self, config: MPFPConfig | None = None):
         """
         Initialize MPFP retriever.
 
@@ -302,7 +288,7 @@ class MPFPGraphRetriever(GraphRetriever):
             config: Algorithm configuration (uses defaults if None)
         """
         self.config = config or MPFPConfig()
-        self._adjacency_cache: Dict[str, TypedAdjacency] = {}
+        self._adjacency_cache: dict[str, TypedAdjacency] = {}
 
     @property
     def name(self) -> str:
@@ -315,10 +301,10 @@ class MPFPGraphRetriever(GraphRetriever):
         bank_id: str,
         fact_type: str,
         budget: int,
-        query_text: Optional[str] = None,
-        semantic_seeds: Optional[List[RetrievalResult]] = None,
-        temporal_seeds: Optional[List[RetrievalResult]] = None,
-    ) -> List[RetrievalResult]:
+        query_text: str | None = None,
+        semantic_seeds: list[RetrievalResult] | None = None,
+        temporal_seeds: list[RetrievalResult] | None = None,
+    ) -> list[RetrievalResult]:
         """
         Retrieve facts using MPFP algorithm.
 
@@ -339,14 +325,12 @@ class MPFPGraphRetriever(GraphRetriever):
         adjacency = await load_typed_adjacency(pool, bank_id)
 
         # Convert seeds to SeedNode format
-        semantic_seed_nodes = self._convert_seeds(semantic_seeds, 'similarity')
-        temporal_seed_nodes = self._convert_seeds(temporal_seeds, 'temporal_score')
+        semantic_seed_nodes = self._convert_seeds(semantic_seeds, "similarity")
+        temporal_seed_nodes = self._convert_seeds(temporal_seeds, "temporal_score")
 
         # If no semantic seeds provided, fall back to finding our own
         if not semantic_seed_nodes:
-            semantic_seed_nodes = await self._find_semantic_seeds(
-                pool, query_embedding_str, bank_id, fact_type
-            )
+            semantic_seed_nodes = await self._find_semantic_seeds(pool, query_embedding_str, bank_id, fact_type)
 
         # Run all patterns in parallel
         tasks = []
@@ -407,9 +391,9 @@ class MPFPGraphRetriever(GraphRetriever):
 
     def _convert_seeds(
         self,
-        seeds: Optional[List[RetrievalResult]],
+        seeds: list[RetrievalResult] | None,
         score_attr: str,
-    ) -> List[SeedNode]:
+    ) -> list[SeedNode]:
         """Convert RetrievalResult seeds to SeedNode format."""
         if not seeds:
             return []
@@ -431,13 +415,13 @@ class MPFPGraphRetriever(GraphRetriever):
         fact_type: str,
         limit: int = 20,
         threshold: float = 0.3,
-    ) -> List[SeedNode]:
+    ) -> list[SeedNode]:
         """Fallback: find semantic seeds via embedding search."""
         async with acquire_with_retry(pool) as conn:
             rows = await conn.fetch(
-                """
+                f"""
                 SELECT id, 1 - (embedding <=> $1::vector) AS similarity
-                FROM memory_units
+                FROM {fq_table("memory_units")}
                 WHERE bank_id = $2
                   AND embedding IS NOT NULL
                   AND fact_type = $3
@@ -445,10 +429,11 @@ class MPFPGraphRetriever(GraphRetriever):
                 ORDER BY embedding <=> $1::vector
                 LIMIT $5
                 """,
-                query_embedding_str, bank_id, fact_type, threshold, limit
+                query_embedding_str,
+                bank_id,
+                fact_type,
+                threshold,
+                limit,
             )
 
-        return [
-            SeedNode(node_id=str(r['id']), score=r['similarity'])
-            for r in rows
-        ]
+        return [SeedNode(node_id=str(r["id"]), score=r["similarity"]) for r in rows]

hindsight-api/hindsight_api/engine/search/observation_utils.py

@@ -6,7 +6,7 @@ about an entity, without personality influence.
 """
 
 import logging
-from typing import List, Dict, Any
+
 from pydantic import BaseModel, Field
 
 from ..response_models import MemoryFact
@@ -16,18 +16,17 @@ logger = logging.getLogger(__name__)
 
 class Observation(BaseModel):
     """An observation about an entity."""
+
     observation: str = Field(description="The observation text - a factual statement about the entity")
 
 
 class ObservationExtractionResponse(BaseModel):
     """Response containing extracted observations."""
-    observations: List[Observation] = Field(
-        default_factory=list,
-        description="List of observations about the entity"
-    )
+
+    observations: list[Observation] = Field(default_factory=list, description="List of observations about the entity")
 
 
-def format_facts_for_observation_prompt(facts: List[MemoryFact]) -> str:
+def format_facts_for_observation_prompt(facts: list[MemoryFact]) -> str:
     """Format facts as text for observation extraction prompt."""
     import json
 
@@ -35,9 +34,7 @@ def format_facts_for_observation_prompt(facts: List[MemoryFact]) -> str:
         return "[]"
     formatted = []
     for fact in facts:
-        fact_obj = {
-            "text": fact.text
-        }
+        fact_obj = {"text": fact.text}
 
         # Add context if available
         if fact.context:
@@ -92,11 +89,7 @@ def get_observation_system_message() -> str:
     return "You are an objective observer synthesizing facts about an entity. Generate clear, factual observations without opinions or personality influence. Be concise and accurate."
 
 
-async def extract_observations_from_facts(
-    llm_config,
-    entity_name: str,
-    facts: List[MemoryFact]
-) -> List[str]:
+async def extract_observations_from_facts(llm_config, entity_name: str, facts: list[MemoryFact]) -> list[str]:
     """
     Extract observations from facts about an entity using LLM.
 
@@ -118,10 +111,10 @@ async def extract_observations_from_facts(
         result = await llm_config.call(
             messages=[
                 {"role": "system", "content": get_observation_system_message()},
-                {"role": "user", "content": prompt}
+                {"role": "user", "content": prompt},
             ],
             response_format=ObservationExtractionResponse,
-            scope="memory_extract_observation"
+            scope="memory_extract_observation",
         )
 
         observations = [op.observation for op in result.observations]

hindsight-api/hindsight_api/engine/search/reranking.py

@@ -2,7 +2,6 @@
 Cross-encoder neural reranking for search results.
 """
 
-from typing import List
 from .types import MergedCandidate, ScoredResult
 
 
@@ -24,14 +23,28 @@ class CrossEncoderReranker:
         """
         if cross_encoder is None:
             from hindsight_api.engine.cross_encoder import create_cross_encoder_from_env
+
             cross_encoder = create_cross_encoder_from_env()
         self.cross_encoder = cross_encoder
+        self._initialized = False
 
-    def rerank(
-        self,
-        query: str,
-        candidates: List[MergedCandidate]
-    ) -> List[ScoredResult]:
+    async def ensure_initialized(self):
+        """Ensure the cross-encoder model is initialized (for lazy initialization)."""
+        if self._initialized:
+            return
+
+        import asyncio
+
+        cross_encoder = self.cross_encoder
+        # For local providers, run in thread pool to avoid blocking event loop
+        if cross_encoder.provider_name == "local":
+            loop = asyncio.get_event_loop()
+            await loop.run_in_executor(None, lambda: asyncio.run(cross_encoder.initialize()))
+        else:
+            await cross_encoder.initialize()
+        self._initialized = True
+
+    def rerank(self, query: str, candidates: list[MergedCandidate]) -> list[ScoredResult]:
         """
         Rerank candidates using cross-encoder scores.
 
@@ -77,6 +90,7 @@ class CrossEncoderReranker:
         # Normalize scores using sigmoid to [0, 1] range
         # Cross-encoder returns logits which can be negative
         import numpy as np
+
         def sigmoid(x):
             return 1 / (1 + np.exp(-x))
 
@@ -89,7 +103,7 @@ class CrossEncoderReranker:
                 candidate=candidate,
                 cross_encoder_score=float(raw_score),
                 cross_encoder_score_normalized=float(norm_score),
-                weight=float(norm_score)  # Initial weight is just cross-encoder score
+                weight=float(norm_score),  # Initial weight is just cross-encoder score
             )
             scored_results.append(scored_result)
 

hindsight-api/hindsight_api/engine/search/retrieval.py

@@ -8,16 +8,18 @@ Implements:
 4. Temporal retrieval (time-aware search with spreading)
 """
 
-from typing import List, Dict, Optional
-from dataclasses import dataclass, field
-from datetime import datetime
 import asyncio
 import logging
-from ..db_utils import acquire_with_retry
-from .types import RetrievalResult
-from .graph_retrieval import GraphRetriever, BFSGraphRetriever
-from .mpfp_retrieval import MPFPGraphRetriever
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+from typing import Optional
+
 from ...config import get_config
+from ..db_utils import acquire_with_retry
+from ..memory_engine import fq_table
+from .graph_retrieval import BFSGraphRetriever, GraphRetriever
+from .mpfp_retrieval import MPFPGraphRetriever
+from .types import RetrievalResult
 
 logger = logging.getLogger(__name__)
 
@@ -25,16 +27,17 @@ logger = logging.getLogger(__name__)
 @dataclass
 class ParallelRetrievalResult:
     """Result from parallel retrieval across all methods."""
-    semantic: List[RetrievalResult]
-    bm25: List[RetrievalResult]
-    graph: List[RetrievalResult]
-    temporal: Optional[List[RetrievalResult]]
-    timings: Dict[str, float] = field(default_factory=dict)
-    temporal_constraint: Optional[tuple] = None  # (start_date, end_date)
+
+    semantic: list[RetrievalResult]
+    bm25: list[RetrievalResult]
+    graph: list[RetrievalResult]
+    temporal: list[RetrievalResult] | None
+    timings: dict[str, float] = field(default_factory=dict)
+    temporal_constraint: tuple | None = None  # (start_date, end_date)
 
 
 # Default graph retriever instance (can be overridden)
-_default_graph_retriever: Optional[GraphRetriever] = None
+_default_graph_retriever: GraphRetriever | None = None
 
 
 def get_default_graph_retriever() -> GraphRetriever:
@@ -62,12 +65,8 @@ def set_default_graph_retriever(retriever: GraphRetriever) -> None:
 
 
 async def retrieve_semantic(
-    conn,
-    query_emb_str: str,
-    bank_id: str,
-    fact_type: str,
-    limit: int
-) -> List[RetrievalResult]:
+    conn, query_emb_str: str, bank_id: str, fact_type: str, limit: int
+) -> list[RetrievalResult]:
     """
     Semantic retrieval via vector similarity.
 
@@ -82,10 +81,10 @@ async def retrieve_semantic(
         List of RetrievalResult objects
     """
     results = await conn.fetch(
-        """
+        f"""
         SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
                1 - (embedding <=> $1::vector) AS similarity
-        FROM memory_units
+        FROM {fq_table("memory_units")}
         WHERE bank_id = $2
           AND embedding IS NOT NULL
           AND fact_type = $3
@@ -93,18 +92,15 @@ async def retrieve_semantic(
         ORDER BY embedding <=> $1::vector
         LIMIT $4
         """,
-        query_emb_str, bank_id, fact_type, limit
+        query_emb_str,
+        bank_id,
+        fact_type,
+        limit,
     )
     return [RetrievalResult.from_db_row(dict(r)) for r in results]
 
 
-async def retrieve_bm25(
-    conn,
-    query_text: str,
-    bank_id: str,
-    fact_type: str,
-    limit: int
-) -> List[RetrievalResult]:
+async def retrieve_bm25(conn, query_text: str, bank_id: str, fact_type: str, limit: int) -> list[RetrievalResult]:
     """
     BM25 keyword retrieval via full-text search.
 
@@ -122,7 +118,7 @@ async def retrieve_bm25(
 
     # Sanitize query text: remove special characters that have meaning in tsquery
     # Keep only alphanumeric characters and spaces
-    sanitized_text = re.sub(r'[^\w\s]', ' ', query_text.lower())
+    sanitized_text = re.sub(r"[^\w\s]", " ", query_text.lower())
 
     # Split and filter empty strings
     tokens = [token for token in sanitized_text.split() if token]
@@ -136,17 +132,20 @@ async def retrieve_bm25(
     query_tsquery = " | ".join(tokens)
 
     results = await conn.fetch(
-        """
+        f"""
         SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
                ts_rank_cd(search_vector, to_tsquery('english', $1)) AS bm25_score
-        FROM memory_units
+        FROM {fq_table("memory_units")}
         WHERE bank_id = $2
           AND fact_type = $3
           AND search_vector @@ to_tsquery('english', $1)
         ORDER BY bm25_score DESC
         LIMIT $4
         """,
-        query_tsquery, bank_id, fact_type, limit
+        query_tsquery,
+        bank_id,
+        fact_type,
+        limit,
     )
     return [RetrievalResult.from_db_row(dict(r)) for r in results]
 
@@ -159,8 +158,8 @@ async def retrieve_temporal(
     start_date: datetime,
     end_date: datetime,
     budget: int,
-    semantic_threshold: float = 0.1
-) -> List[RetrievalResult]:
+    semantic_threshold: float = 0.1,
+) -> list[RetrievalResult]:
     """
     Temporal retrieval with spreading activation.
 
@@ -182,19 +181,18 @@ async def retrieve_temporal(
     Returns:
         List of RetrievalResult objects with temporal scores
     """
-    from datetime import timezone
 
     # Ensure start_date and end_date are timezone-aware (UTC) to match database datetimes
     if start_date.tzinfo is None:
-        start_date = start_date.replace(tzinfo=timezone.utc)
+        start_date = start_date.replace(tzinfo=UTC)
     if end_date.tzinfo is None:
-        end_date = end_date.replace(tzinfo=timezone.utc)
+        end_date = end_date.replace(tzinfo=UTC)
 
     entry_points = await conn.fetch(
-        """
+        f"""
         SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
                1 - (embedding <=> $1::vector) AS similarity
-        FROM memory_units
+        FROM {fq_table("memory_units")}
         WHERE bank_id = $2
           AND fact_type = $3
           AND embedding IS NOT NULL
@@ -215,7 +213,12 @@ async def retrieve_temporal(
         ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC, (embedding <=> $1::vector) ASC
         LIMIT 10
         """,
-        query_emb_str, bank_id, fact_type, start_date, end_date, semantic_threshold
+        query_emb_str,
+        bank_id,
+        fact_type,
+        start_date,
+        end_date,
+        semantic_threshold,
     )
 
     if not entry_points:
@@ -258,7 +261,9 @@ async def retrieve_temporal(
         results.append(ep_result)
 
     # Spread through temporal links
-    queue = [(RetrievalResult.from_db_row(dict(ep)), ep["similarity"], 1.0) for ep in entry_points]  # (unit, semantic_sim, temporal_score)
+    queue = [
+        (RetrievalResult.from_db_row(dict(ep)), ep["similarity"], 1.0) for ep in entry_points
+    ]  # (unit, semantic_sim, temporal_score)
     budget_remaining = budget - len(entry_points)
 
     while queue and budget_remaining > 0:
@@ -268,12 +273,12 @@ async def retrieve_temporal(
         # Get neighbors via temporal and causal links
         if budget_remaining > 0:
             neighbors = await conn.fetch(
-                """
+                f"""
                 SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id,
                        ml.weight, ml.link_type,
                        1 - (mu.embedding <=> $1::vector) AS similarity
-                FROM memory_links ml
-                JOIN memory_units mu ON ml.to_unit_id = mu.id
+                FROM {fq_table("memory_links")} ml
+                JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
                 WHERE ml.from_unit_id = $2
                   AND ml.link_type IN ('temporal', 'causes', 'caused_by', 'enables', 'prevents')
                   AND ml.weight >= 0.1
@@ -283,7 +288,10 @@ async def retrieve_temporal(
                 ORDER BY ml.weight DESC
                 LIMIT 10
                 """,
-                query_emb_str, current.id, fact_type, semantic_threshold
+                query_emb_str,
+                current.id,
+                fact_type,
+                semantic_threshold,
             )
 
             for n in neighbors:
@@ -307,7 +315,9 @@ async def retrieve_temporal(
 
                 if neighbor_best_date:
                     days_from_mid = abs((neighbor_best_date - mid_date).total_seconds() / 86400)
-                    neighbor_temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
+                    neighbor_temporal_proximity = (
+                        1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
+                    )
                 else:
                     neighbor_temporal_proximity = 0.3  # Lower score if no temporal data
 
@@ -349,9 +359,9 @@ async def retrieve_parallel(
     bank_id: str,
     fact_type: str,
     thinking_budget: int,
-    question_date: Optional[datetime] = None,
+    question_date: datetime | None = None,
     query_analyzer: Optional["QueryAnalyzer"] = None,
-    graph_retriever: Optional[GraphRetriever] = None,
+    graph_retriever: GraphRetriever | None = None,
 ) -> ParallelRetrievalResult:
     """
     Run 3-way or 4-way parallel retrieval (adds temporal if detected).
@@ -372,29 +382,26 @@ async def retrieve_parallel(
     """
     from .temporal_extraction import extract_temporal_constraint
 
-    temporal_constraint = extract_temporal_constraint(
-        query_text, reference_date=question_date, analyzer=query_analyzer
-    )
+    temporal_constraint = extract_temporal_constraint(query_text, reference_date=question_date, analyzer=query_analyzer)
 
     retriever = graph_retriever or get_default_graph_retriever()
 
     if retriever.name == "mpfp":
         return await _retrieve_parallel_mpfp(
-            pool, query_text, query_embedding_str, bank_id, fact_type,
-            thinking_budget, temporal_constraint, retriever
+            pool, query_text, query_embedding_str, bank_id, fact_type, thinking_budget, temporal_constraint, retriever
         )
     else:
         return await _retrieve_parallel_bfs(
-            pool, query_text, query_embedding_str, bank_id, fact_type,
-            thinking_budget, temporal_constraint, retriever
+            pool, query_text, query_embedding_str, bank_id, fact_type, thinking_budget, temporal_constraint, retriever
         )
 
 
 @dataclass
 class _SemanticGraphResult:
     """Internal result from semantic→graph chain."""
-    semantic: List[RetrievalResult]
-    graph: List[RetrievalResult]
+
+    semantic: list[RetrievalResult]
+    graph: list[RetrievalResult]
     semantic_time: float
     graph_time: float
 
@@ -402,7 +409,8 @@ class _SemanticGraphResult:
 @dataclass
 class _TimedResult:
     """Internal result with timing."""
-    results: List[RetrievalResult]
+
+    results: list[RetrievalResult]
     time: float
 
 
@@ -413,7 +421,7 @@ async def _retrieve_parallel_mpfp(
     bank_id: str,
     fact_type: str,
     thinking_budget: int,
-    temporal_constraint: Optional[tuple],
+    temporal_constraint: tuple | None,
     retriever: GraphRetriever,
 ) -> ParallelRetrievalResult:
     """
@@ -430,9 +438,7 @@ async def _retrieve_parallel_mpfp(
         """Chain: semantic retrieval → graph retrieval (using semantic as seeds)."""
         start = time.time()
         async with acquire_with_retry(pool) as conn:
-            semantic = await retrieve_semantic(
-                conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget
-            )
+            semantic = await retrieve_semantic(conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget)
         semantic_time = time.time() - start
 
         # Get temporal seeds if needed (quick query, part of this chain)
@@ -441,8 +447,7 @@ async def _retrieve_parallel_mpfp(
             tc_start, tc_end = temporal_constraint
             async with acquire_with_retry(pool) as conn:
                 temporal_seeds = await _get_temporal_entry_points(
-                    conn, query_embedding_str, bank_id, fact_type,
-                    tc_start, tc_end, limit=20
+                    conn, query_embedding_str, bank_id, fact_type, tc_start, tc_end, limit=20
                 )
 
         # Run graph with seeds
@@ -473,8 +478,14 @@ async def _retrieve_parallel_mpfp(
         start = time.time()
         async with acquire_with_retry(pool) as conn:
             results = await retrieve_temporal(
-                conn, query_embedding_str, bank_id, fact_type,
-                tc_start, tc_end, budget=thinking_budget, semantic_threshold=0.1
+                conn,
+                query_embedding_str,
+                bank_id,
+                fact_type,
+                tc_start,
+                tc_end,
+                budget=thinking_budget,
+                semantic_threshold=0.1,
             )
         return _TimedResult(results, time.time() - start)
 
@@ -527,21 +538,20 @@ async def _get_temporal_entry_points(
     end_date: datetime,
     limit: int = 20,
     semantic_threshold: float = 0.1,
-) -> List[RetrievalResult]:
+) -> list[RetrievalResult]:
     """Get temporal entry points (facts in date range with semantic relevance)."""
-    from datetime import timezone
 
     if start_date.tzinfo is None:
-        start_date = start_date.replace(tzinfo=timezone.utc)
+        start_date = start_date.replace(tzinfo=UTC)
     if end_date.tzinfo is None:
-        end_date = end_date.replace(tzinfo=timezone.utc)
+        end_date = end_date.replace(tzinfo=UTC)
 
     rows = await conn.fetch(
-        """
+        f"""
         SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at,
                access_count, embedding, fact_type, document_id, chunk_id,
                1 - (embedding <=> $1::vector) AS similarity
-        FROM memory_units
+        FROM {fq_table("memory_units")}
         WHERE bank_id = $2
           AND fact_type = $3
           AND embedding IS NOT NULL
@@ -557,7 +567,13 @@ async def _get_temporal_entry_points(
                  (embedding <=> $1::vector) ASC
         LIMIT $7
         """,
-        query_embedding_str, bank_id, fact_type, start_date, end_date, semantic_threshold, limit
+        query_embedding_str,
+        bank_id,
+        fact_type,
+        start_date,
+        end_date,
+        semantic_threshold,
+        limit,
     )
 
     results = []
@@ -597,7 +613,7 @@ async def _retrieve_parallel_bfs(
     bank_id: str,
     fact_type: str,
     thinking_budget: int,
-    temporal_constraint: Optional[tuple],
+    temporal_constraint: tuple | None,
     retriever: GraphRetriever,
 ) -> ParallelRetrievalResult:
     """BFS retrieval: all methods run in parallel (original behavior)."""
@@ -631,8 +647,14 @@ async def _retrieve_parallel_bfs(
         start = time.time()
         async with acquire_with_retry(pool) as conn:
             results = await retrieve_temporal(
-                conn, query_embedding_str, bank_id, fact_type,
-                tc_start, tc_end, budget=thinking_budget, semantic_threshold=0.1
+                conn,
+                query_embedding_str,
+                bank_id,
+                fact_type,
+                tc_start,
+                tc_end,
+                budget=thinking_budget,
+                semantic_threshold=0.1,
             )
         return _TimedResult(results, time.time() - start)
 

hindsight-api/hindsight_api/engine/search/scoring.py

@@ -4,11 +4,11 @@ Scoring functions for memory search and retrieval.
 Includes recency weighting, frequency weighting, temporal proximity,
 and similarity calculations used in memory activation and ranking.
 """
+
 from datetime import datetime
-from typing import List
 
 
-def cosine_similarity(vec1: List[float], vec2: List[float]) -> float:
+def cosine_similarity(vec1: list[float], vec2: list[float]) -> float:
     """
     Calculate cosine similarity between two vectors.
 
@@ -58,6 +58,7 @@ def calculate_recency_weight(days_since: float, half_life_days: float = 365.0) -
         Weight between 0 and 1
     """
     import math
+
     # Logarithmic decay: 1 / (1 + log(1 + days_since/half_life))
     # This decays much slower than exponential, giving better long-term differentiation
     normalized_age = days_since / half_life_days
@@ -79,6 +80,7 @@ def calculate_frequency_weight(access_count: int, max_boost: float = 2.0) -> flo
         Weight between 1.0 and max_boost
     """
     import math
+
     if access_count <= 0:
         return 1.0
 
@@ -116,11 +118,7 @@ def calculate_temporal_anchor(occurred_start: datetime, occurred_end: datetime)
     return midpoint
 
 
-def calculate_temporal_proximity(
-    anchor_a: datetime,
-    anchor_b: datetime,
-    half_life_days: float = 30.0
-) -> float:
+def calculate_temporal_proximity(anchor_a: datetime, anchor_b: datetime, half_life_days: float = 30.0) -> float:
     """
     Calculate temporal proximity between two temporal anchors.
 

hindsight-api/hindsight_api/engine/search/temporal_extraction.py

@@ -4,16 +4,16 @@ Temporal extraction for time-aware search queries.
 Handles natural language temporal expressions using transformer-based query analysis.
 """
 
-from typing import Optional, Tuple
-from datetime import datetime
 import logging
-from hindsight_api.engine.query_analyzer import QueryAnalyzer, DateparserQueryAnalyzer
+from datetime import datetime
+
+from hindsight_api.engine.query_analyzer import DateparserQueryAnalyzer, QueryAnalyzer
 
 logger = logging.getLogger(__name__)
 
 # Global default analyzer instance
 # Can be overridden by passing a custom analyzer to extract_temporal_constraint
-_default_analyzer: Optional[QueryAnalyzer] = None
+_default_analyzer: QueryAnalyzer | None = None
 
 
 def get_default_analyzer() -> QueryAnalyzer:
@@ -33,9 +33,9 @@ def get_default_analyzer() -> QueryAnalyzer:
 
 def extract_temporal_constraint(
     query: str,
-    reference_date: Optional[datetime] = None,
-    analyzer: Optional[QueryAnalyzer] = None,
-) -> Optional[Tuple[datetime, datetime]]:
+    reference_date: datetime | None = None,
+    analyzer: QueryAnalyzer | None = None,
+) -> tuple[datetime, datetime] | None:
     """
     Extract temporal constraint from query.
 
@@ -55,10 +55,7 @@ def extract_temporal_constraint(
     analysis = analyzer.analyze(query, reference_date)
 
     if analysis.temporal_constraint:
-        result = (
-            analysis.temporal_constraint.start_date,
-            analysis.temporal_constraint.end_date
-        )
+        result = (analysis.temporal_constraint.start_date, analysis.temporal_constraint.end_date)
         return result
 
     return None

hindsight-api/hindsight_api/engine/search/think_utils.py

@@ -2,41 +2,35 @@
 Think operation utilities for formulating answers based on agent and world facts.
 """
 
-import asyncio
 import logging
 import re
-from datetime import datetime, timezone
-from typing import Dict, List, Any
+from datetime import datetime
+
 from pydantic import BaseModel, Field
 
-from ..response_models import ReflectResult, MemoryFact, DispositionTraits
+from ..response_models import DispositionTraits, MemoryFact
 
 logger = logging.getLogger(__name__)
 
 
 class Opinion(BaseModel):
     """An opinion formed by the bank."""
+
     opinion: str = Field(description="The opinion or perspective with reasoning included")
     confidence: float = Field(description="Confidence score for this opinion (0.0 to 1.0, where 1.0 is very confident)")
 
 
 class OpinionExtractionResponse(BaseModel):
     """Response containing extracted opinions."""
-    opinions: List[Opinion] = Field(
-        default_factory=list,
-        description="List of opinions formed with their supporting reasons and confidence scores"
+
+    opinions: list[Opinion] = Field(
+        default_factory=list, description="List of opinions formed with their supporting reasons and confidence scores"
     )
 
 
 def describe_trait_level(value: int) -> str:
     """Convert trait value (1-5) to descriptive text."""
-    levels = {
-        1: "very low",
-        2: "low",
-        3: "moderate",
-        4: "high",
-        5: "very high"
-    }
+    levels = {1: "very low", 2: "low", 3: "moderate", 4: "high", 5: "very high"}
     return levels.get(value, "moderate")
 
 
@@ -47,7 +41,7 @@ def build_disposition_description(disposition: DispositionTraits) -> str:
         2: "You tend to trust information but may question obvious inconsistencies.",
         3: "You have a balanced approach to information, neither too trusting nor too skeptical.",
         4: "You are somewhat skeptical and often question the reliability of information.",
-        5: "You are highly skeptical and critically examine all information for accuracy and hidden motives."
+        5: "You are highly skeptical and critically examine all information for accuracy and hidden motives.",
     }
 
     literalism_desc = {
@@ -55,7 +49,7 @@ def build_disposition_description(disposition: DispositionTraits) -> str:
         2: "You tend to consider context and implied meaning alongside literal statements.",
         3: "You balance literal interpretation with contextual understanding.",
         4: "You prefer to interpret information more literally and precisely.",
-        5: "You interpret information very literally and focus on exact wording and commitments."
+        5: "You interpret information very literally and focus on exact wording and commitments.",
     }
 
     empathy_desc = {
@@ -63,7 +57,7 @@ def build_disposition_description(disposition: DispositionTraits) -> str:
         2: "You consider facts first but acknowledge emotional factors exist.",
         3: "You balance factual analysis with emotional understanding.",
         4: "You give significant weight to emotional context and human factors.",
-        5: "You strongly consider the emotional state and circumstances of others when forming memories."
+        5: "You strongly consider the emotional state and circumstances of others when forming memories.",
     }
 
     return f"""Your disposition traits:
@@ -72,7 +66,7 @@ def build_disposition_description(disposition: DispositionTraits) -> str:
 - Empathy ({describe_trait_level(disposition.empathy)}): {empathy_desc.get(disposition.empathy, empathy_desc[3])}"""
 
 
-def format_facts_for_prompt(facts: List[MemoryFact]) -> str:
+def format_facts_for_prompt(facts: list[MemoryFact]) -> str:
     """Format facts as JSON for LLM prompt."""
     import json
 
@@ -80,9 +74,7 @@ def format_facts_for_prompt(facts: List[MemoryFact]) -> str:
         return "[]"
     formatted = []
     for fact in facts:
-        fact_obj = {
-            "text": fact.text
-        }
+        fact_obj = {"text": fact.text}
 
         # Add context if available
         if fact.context:
@@ -94,7 +86,7 @@ def format_facts_for_prompt(facts: List[MemoryFact]) -> str:
             if isinstance(occurred_start, str):
                 fact_obj["occurred_start"] = occurred_start
             elif isinstance(occurred_start, datetime):
-                fact_obj["occurred_start"] = occurred_start.strftime('%Y-%m-%d %H:%M:%S')
+                fact_obj["occurred_start"] = occurred_start.strftime("%Y-%m-%d %H:%M:%S")
 
         formatted.append(fact_obj)
 
@@ -109,7 +101,7 @@ def build_think_prompt(
     name: str,
     disposition: DispositionTraits,
     background: str,
-    context: str = None,
+    context: str | None = None,
 ) -> str:
     """Build the think prompt for the LLM."""
     disposition_desc = build_disposition_description(disposition)
@@ -176,16 +168,14 @@ def get_system_message(disposition: DispositionTraits) -> str:
     elif disposition.empathy <= 2:
         instructions.append("Focus on facts and outcomes rather than emotional context.")
 
-    disposition_instruction = " ".join(instructions) if instructions else "Balance your disposition traits when interpreting information."
+    disposition_instruction = (
+        " ".join(instructions) if instructions else "Balance your disposition traits when interpreting information."
+    )
 
     return f"You are a person with your own thoughts, experiences, opinions, and disposition. When asked a question, respond naturally like a human would - share what you genuinely think based on what you know and have experienced. {disposition_instruction} Be direct, express your views confidently, and use 'I think', 'I believe', 'in my view', etc. Respond in plain text without markdown formatting."
 
 
-async def extract_opinions_from_text(
-    llm_config,
-    text: str,
-    query: str
-) -> List[Opinion]:
+async def extract_opinions_from_text(llm_config, text: str, query: str) -> list[Opinion]:
     """
     Extract opinions with reasons and confidence from text using LLM.
 
@@ -238,11 +228,14 @@ If no genuine opinions are expressed (e.g., the response just says "I don't know
     try:
         result = await llm_config.call(
             messages=[
-                {"role": "system", "content": "You are converting opinions from text into first-person statements. Always use 'I think', 'I believe', 'I feel', etc. NEVER use third-person like 'The speaker' or 'They'."},
-                {"role": "user", "content": extraction_prompt}
+                {
+                    "role": "system",
+                    "content": "You are converting opinions from text into first-person statements. Always use 'I think', 'I believe', 'I feel', etc. NEVER use third-person like 'The speaker' or 'They'.",
+                },
+                {"role": "user", "content": extraction_prompt},
             ],
             response_format=OpinionExtractionResponse,
-            scope="memory_extract_opinion"
+            scope="memory_extract_opinion",
         )
 
         # Format opinions with confidence score and convert to first-person
@@ -253,14 +246,18 @@ If no genuine opinions are expressed (e.g., the response just says "I don't know
 
             # Replace common third-person patterns with first-person
             def singularize_verb(verb):
-                if verb.endswith('es'):
+                if verb.endswith("es"):
                     return verb[:-1]  # believes -> believe
-                elif verb.endswith('s'):
+                elif verb.endswith("s"):
                     return verb[:-1]  # thinks -> think
                 return verb
 
             # Pattern: "The speaker/user [verb]..." -> "I [verb]..."
-            match = re.match(r'^(The speaker|The user|They|It is believed) (believes?|thinks?|feels?|says|asserts?|considers?)(\s+that)?(.*)$', opinion_text, re.IGNORECASE)
+            match = re.match(
+                r"^(The speaker|The user|They|It is believed) (believes?|thinks?|feels?|says|asserts?|considers?)(\s+that)?(.*)$",
+                opinion_text,
+                re.IGNORECASE,
+            )
             if match:
                 verb = singularize_verb(match.group(2))
                 that_part = match.group(3) or ""  # Keep " that" if present
@@ -268,17 +265,96 @@ If no genuine opinions are expressed (e.g., the response just says "I don't know
                 opinion_text = f"I {verb}{that_part}{rest}"
 
             # If still doesn't start with first-person, prepend "I believe that "
-            first_person_starters = ["I think", "I believe", "I feel", "In my view", "I've come to believe", "Previously I"]
+            first_person_starters = [
+                "I think",
+                "I believe",
+                "I feel",
+                "In my view",
+                "I've come to believe",
+                "Previously I",
+            ]
             if not any(opinion_text.startswith(starter) for starter in first_person_starters):
                 opinion_text = "I believe that " + opinion_text[0].lower() + opinion_text[1:]
 
-            formatted_opinions.append(Opinion(
-                opinion=opinion_text,
-                confidence=op.confidence
-            ))
+            formatted_opinions.append(Opinion(opinion=opinion_text, confidence=op.confidence))
 
         return formatted_opinions
 
     except Exception as e:
         logger.warning(f"Failed to extract opinions: {str(e)}")
         return []
+
+
+async def reflect(
+    llm_config,
+    query: str,
+    experience_facts: list[str] = None,
+    world_facts: list[str] = None,
+    opinion_facts: list[str] = None,
+    name: str = "Assistant",
+    disposition: DispositionTraits = None,
+    background: str = "",
+    context: str = None,
+) -> str:
+    """
+    Standalone reflect function for generating answers based on facts.
+
+    This is a static version of the reflect operation that can be called
+    without a MemoryEngine instance, useful for testing.
+
+    Args:
+        llm_config: LLM provider instance
+        query: Question to answer
+        experience_facts: List of experience/agent fact strings
+        world_facts: List of world fact strings
+        opinion_facts: List of opinion fact strings
+        name: Name of the agent/persona
+        disposition: Disposition traits (defaults to neutral)
+        background: Background information
+        context: Additional context for the prompt
+
+    Returns:
+        Generated answer text
+    """
+    # Default disposition if not provided
+    if disposition is None:
+        disposition = DispositionTraits(skepticism=3, literalism=3, empathy=3)
+
+    # Convert string lists to MemoryFact format for formatting
+    def to_memory_facts(facts: list[str], fact_type: str) -> list[MemoryFact]:
+        if not facts:
+            return []
+        return [MemoryFact(id=f"test-{i}", text=f, fact_type=fact_type) for i, f in enumerate(facts)]
+
+    agent_results = to_memory_facts(experience_facts or [], "experience")
+    world_results = to_memory_facts(world_facts or [], "world")
+    opinion_results = to_memory_facts(opinion_facts or [], "opinion")
+
+    # Format facts for prompt
+    agent_facts_text = format_facts_for_prompt(agent_results)
+    world_facts_text = format_facts_for_prompt(world_results)
+    opinion_facts_text = format_facts_for_prompt(opinion_results)
+
+    # Build prompt
+    prompt = build_think_prompt(
+        agent_facts_text=agent_facts_text,
+        world_facts_text=world_facts_text,
+        opinion_facts_text=opinion_facts_text,
+        query=query,
+        name=name,
+        disposition=disposition,
+        background=background,
+        context=context,
+    )
+
+    system_message = get_system_message(disposition)
+
+    # Call LLM
+    answer_text = await llm_config.call(
+        messages=[{"role": "system", "content": system_message}, {"role": "user", "content": prompt}],
+        scope="memory_think",
+        temperature=0.9,
+        max_completion_tokens=1000,
+    )
+
+    return answer_text.strip()

hindsight-api/hindsight_api/engine/search/trace.py

@@ -4,15 +4,18 @@ Search trace models for debugging and visualization.
 These Pydantic models define the structure of search traces, capturing
 every step of the spreading activation search process for analysis.
 """
+
 from datetime import datetime
-from typing import List, Optional, Dict, Any, Literal
+from typing import Any, Literal
+
 from pydantic import BaseModel, Field
 
 
 class QueryInfo(BaseModel):
     """Information about the search query."""
+
     query_text: str = Field(description="Original query text")
-    query_embedding: List[float] = Field(description="Generated query embedding vector")
+    query_embedding: list[float] = Field(description="Generated query embedding vector")
     timestamp: datetime = Field(description="When the query was executed")
     budget: int = Field(description="Maximum nodes to explore")
     max_tokens: int = Field(description="Maximum tokens to return in results")
@@ -20,6 +23,7 @@ class QueryInfo(BaseModel):
 
 class EntryPoint(BaseModel):
     """An entry point node selected for search."""
+
     node_id: str = Field(description="Memory unit ID")
     text: str = Field(description="Memory unit text content")
     similarity_score: float = Field(description="Cosine similarity to query", ge=0.0, le=1.0)
@@ -28,6 +32,7 @@ class EntryPoint(BaseModel):
 
 class WeightComponents(BaseModel):
     """Breakdown of weight calculation components."""
+
     activation: float = Field(description="Activation from spreading (can exceed 1.0 through accumulation)", ge=0.0)
     semantic_similarity: float = Field(description="Semantic similarity to query", ge=0.0, le=1.0)
     recency: float = Field(description="Recency weight", ge=0.0, le=1.0)
@@ -43,99 +48,120 @@ class WeightComponents(BaseModel):
 
 class LinkInfo(BaseModel):
     """Information about a link to a neighbor."""
+
     to_node_id: str = Field(description="Target node ID")
     link_type: Literal["temporal", "semantic", "entity"] = Field(description="Type of link")
-    link_weight: float = Field(description="Weight of the link (can exceed 1.0 when aggregating multiple connections)", ge=0.0)
-    entity_id: Optional[str] = Field(default=None, description="Entity ID if link_type is 'entity'")
-    new_activation: Optional[float] = Field(default=None, description="Activation that would be passed to neighbor (None for supplementary links)")
+    link_weight: float = Field(
+        description="Weight of the link (can exceed 1.0 when aggregating multiple connections)", ge=0.0
+    )
+    entity_id: str | None = Field(default=None, description="Entity ID if link_type is 'entity'")
+    new_activation: float | None = Field(
+        default=None, description="Activation that would be passed to neighbor (None for supplementary links)"
+    )
     followed: bool = Field(description="Whether this link was followed (or pruned)")
-    prune_reason: Optional[str] = Field(default=None, description="Why link was not followed (if not followed)")
-    is_supplementary: bool = Field(default=False, description="Whether this is a supplementary link (multiple connections to same node)")
+    prune_reason: str | None = Field(default=None, description="Why link was not followed (if not followed)")
+    is_supplementary: bool = Field(
+        default=False, description="Whether this is a supplementary link (multiple connections to same node)"
+    )
 
 
 class NodeVisit(BaseModel):
     """Information about visiting a node during search."""
+
     step: int = Field(description="Step number in search (1-based)")
     node_id: str = Field(description="Memory unit ID")
     text: str = Field(description="Memory unit text content")
     context: str = Field(description="Memory unit context")
-    event_date: Optional[datetime] = Field(default=None, description="When the memory occurred")
+    event_date: datetime | None = Field(default=None, description="When the memory occurred")
     access_count: int = Field(description="Number of times accessed before this search")
 
     # How this node was reached
     is_entry_point: bool = Field(description="Whether this is an entry point")
-    parent_node_id: Optional[str] = Field(default=None, description="Node that led to this one")
-    link_type: Optional[Literal["temporal", "semantic", "entity"]] = Field(default=None, description="Type of link from parent")
-    link_weight: Optional[float] = Field(default=None, description="Weight of link from parent")
+    parent_node_id: str | None = Field(default=None, description="Node that led to this one")
+    link_type: Literal["temporal", "semantic", "entity"] | None = Field(
+        default=None, description="Type of link from parent"
+    )
+    link_weight: float | None = Field(default=None, description="Weight of link from parent")
 
     # Weights
     weights: WeightComponents = Field(description="Weight calculation breakdown")
 
     # Neighbors discovered from this node
-    neighbors_explored: List[LinkInfo] = Field(default_factory=list, description="Links explored from this node")
+    neighbors_explored: list[LinkInfo] = Field(default_factory=list, description="Links explored from this node")
 
     # Ranking
-    final_rank: Optional[int] = Field(default=None, description="Final rank in results (1-based, None if not in top-k)")
+    final_rank: int | None = Field(default=None, description="Final rank in results (1-based, None if not in top-k)")
 
 
 class PruningDecision(BaseModel):
     """Records when a node was considered but not visited."""
+
     node_id: str = Field(description="Node that was pruned")
-    reason: Literal["already_visited", "activation_too_low", "budget_exhausted"] = Field(description="Why it was pruned")
+    reason: Literal["already_visited", "activation_too_low", "budget_exhausted"] = Field(
+        description="Why it was pruned"
+    )
     activation: float = Field(description="Activation value when pruned")
     would_have_been_step: int = Field(description="What step it would have been if visited")
 
 
 class SearchPhaseMetrics(BaseModel):
     """Performance metrics for a search phase."""
+
     phase_name: str = Field(description="Name of the phase")
     duration_seconds: float = Field(description="Time taken in seconds")
-    details: Dict[str, Any] = Field(default_factory=dict, description="Additional phase-specific metrics")
+    details: dict[str, Any] = Field(default_factory=dict, description="Additional phase-specific metrics")
 
 
 class RetrievalResult(BaseModel):
     """A single result from a retrieval method."""
+
     rank: int = Field(description="Rank in this retrieval method (1-based)")
     node_id: str = Field(description="Memory unit ID")
     text: str = Field(description="Memory unit text content")
     context: str = Field(default="", description="Memory unit context")
-    event_date: Optional[datetime] = Field(default=None, description="When the memory occurred")
-    fact_type: Optional[str] = Field(default=None, description="Fact type (world, experience, opinion)")
+    event_date: datetime | None = Field(default=None, description="When the memory occurred")
+    fact_type: str | None = Field(default=None, description="Fact type (world, experience, opinion)")
     score: float = Field(description="Score from this retrieval method")
     score_name: str = Field(description="Name of the score (e.g., 'similarity', 'bm25_score', 'activation')")
 
 
 class RetrievalMethodResults(BaseModel):
     """Results from a single retrieval method."""
+
     method_name: Literal["semantic", "bm25", "graph", "temporal"] = Field(description="Name of retrieval method")
-    fact_type: Optional[str] = Field(default=None, description="Fact type this retrieval was for (world, experience, opinion)")
-    results: List[RetrievalResult] = Field(description="Retrieved results with ranks")
+    fact_type: str | None = Field(
+        default=None, description="Fact type this retrieval was for (world, experience, opinion)"
+    )
+    results: list[RetrievalResult] = Field(description="Retrieved results with ranks")
     duration_seconds: float = Field(description="Time taken for this retrieval")
-    metadata: Dict[str, Any] = Field(default_factory=dict, description="Method-specific metadata")
+    metadata: dict[str, Any] = Field(default_factory=dict, description="Method-specific metadata")
 
 
 class RRFMergeResult(BaseModel):
     """A result after RRF merging."""
+
     node_id: str = Field(description="Memory unit ID")
     text: str = Field(description="Memory unit text content")
     rrf_score: float = Field(description="Reciprocal Rank Fusion score")
-    source_ranks: Dict[str, int] = Field(description="Rank in each source that contributed (method_name -> rank)")
+    source_ranks: dict[str, int] = Field(description="Rank in each source that contributed (method_name -> rank)")
     final_rrf_rank: int = Field(description="Rank after RRF merge (1-based)")
 
 
 class RerankedResult(BaseModel):
     """A result after reranking."""
+
     node_id: str = Field(description="Memory unit ID")
     text: str = Field(description="Memory unit text content")
     rerank_score: float = Field(description="Final reranking score")
     rerank_rank: int = Field(description="Rank after reranking (1-based)")
     rrf_rank: int = Field(description="Original RRF rank before reranking")
     rank_change: int = Field(description="Change in rank (positive = moved up)")
-    score_components: Dict[str, float] = Field(default_factory=dict, description="Score breakdown")
+    score_components: dict[str, float] = Field(default_factory=dict, description="Score breakdown")
 
 
 class SearchSummary(BaseModel):
     """Summary statistics about the search."""
+
     total_nodes_visited: int = Field(description="Total nodes visited")
     total_nodes_pruned: int = Field(description="Total nodes pruned")
     entry_points_found: int = Field(description="Number of entry points")
@@ -150,33 +176,36 @@ class SearchSummary(BaseModel):
     entity_links_followed: int = Field(default=0, description="Entity links followed")
 
     # Phase timings
-    phase_metrics: List[SearchPhaseMetrics] = Field(default_factory=list, description="Metrics for each phase")
+    phase_metrics: list[SearchPhaseMetrics] = Field(default_factory=list, description="Metrics for each phase")
 
 
 class SearchTrace(BaseModel):
     """Complete trace of a search operation."""
+
     query: QueryInfo = Field(description="Query information")
 
     # New 4-way retrieval architecture
-    retrieval_results: List[RetrievalMethodResults] = Field(default_factory=list, description="Results from each retrieval method")
-    rrf_merged: List[RRFMergeResult] = Field(default_factory=list, description="Results after RRF merging")
-    reranked: List[RerankedResult] = Field(default_factory=list, description="Results after reranking")
+    retrieval_results: list[RetrievalMethodResults] = Field(
+        default_factory=list, description="Results from each retrieval method"
+    )
+    rrf_merged: list[RRFMergeResult] = Field(default_factory=list, description="Results after RRF merging")
+    reranked: list[RerankedResult] = Field(default_factory=list, description="Results after reranking")
 
     # Legacy fields (kept for backward compatibility with graph/temporal visualizations)
-    entry_points: List[EntryPoint] = Field(default_factory=list, description="Entry points selected for search (legacy)")
-    visits: List[NodeVisit] = Field(default_factory=list, description="All nodes visited during search (legacy, for graph viz)")
-    pruned: List[PruningDecision] = Field(default_factory=list, description="Nodes that were pruned (legacy)")
+    entry_points: list[EntryPoint] = Field(
+        default_factory=list, description="Entry points selected for search (legacy)"
+    )
+    visits: list[NodeVisit] = Field(
+        default_factory=list, description="All nodes visited during search (legacy, for graph viz)"
+    )
+    pruned: list[PruningDecision] = Field(default_factory=list, description="Nodes that were pruned (legacy)")
 
     summary: SearchSummary = Field(description="Summary statistics")
 
     # Final results (for comparison with visits)
-    final_results: List[Dict[str, Any]] = Field(description="Final ranked results returned to user")
+    final_results: list[dict[str, Any]] = Field(description="Final ranked results returned to user")
 
-    model_config = {
-        "json_encoders": {
-            datetime: lambda v: v.isoformat()
-        }
-    }
+    model_config = {"json_encoders": {datetime: lambda v: v.isoformat()}}
 
     def to_json(self, **kwargs) -> str:
         """Export trace as JSON string."""
@@ -186,14 +215,14 @@ class SearchTrace(BaseModel):
         """Export trace as dictionary."""
         return self.model_dump()
 
-    def get_visit_by_node_id(self, node_id: str) -> Optional[NodeVisit]:
+    def get_visit_by_node_id(self, node_id: str) -> NodeVisit | None:
         """Find a visit by node ID."""
         for visit in self.visits:
             if visit.node_id == node_id:
                 return visit
         return None
 
-    def get_search_path_to_node(self, node_id: str) -> List[NodeVisit]:
+    def get_search_path_to_node(self, node_id: str) -> list[NodeVisit]:
         """Get the path from entry point to a specific node."""
         path = []
         current_visit = self.get_visit_by_node_id(node_id)
@@ -207,10 +236,10 @@ class SearchTrace(BaseModel):
 
         return path
 
-    def get_nodes_by_link_type(self, link_type: Literal["temporal", "semantic", "entity"]) -> List[NodeVisit]:
+    def get_nodes_by_link_type(self, link_type: Literal["temporal", "semantic", "entity"]) -> list[NodeVisit]:
         """Get all nodes reached via a specific link type."""
         return [v for v in self.visits if v.link_type == link_type]
 
-    def get_entry_point_nodes(self) -> List[NodeVisit]:
+    def get_entry_point_nodes(self) -> list[NodeVisit]:
         """Get all entry point visits."""
         return [v for v in self.visits if v.is_entry_point]

hindsight-api/hindsight_api/engine/search/tracer.py

@@ -4,24 +4,25 @@ Search tracer for collecting detailed search execution traces.
 The SearchTracer collects comprehensive information about each step
 of the spreading activation search process for debugging and visualization.
 """
+
 import time
-from datetime import datetime, timezone
-from typing import List, Optional, Dict, Any, Literal
+from datetime import UTC, datetime
+from typing import Any, Literal
 
 from .trace import (
-    SearchTrace,
-    QueryInfo,
     EntryPoint,
-    NodeVisit,
-    WeightComponents,
     LinkInfo,
+    NodeVisit,
     PruningDecision,
-    SearchSummary,
-    SearchPhaseMetrics,
-    RetrievalResult,
-    RetrievalMethodResults,
-    RRFMergeResult,
+    QueryInfo,
     RerankedResult,
+    RetrievalMethodResults,
+    RetrievalResult,
+    RRFMergeResult,
+    SearchPhaseMetrics,
+    SearchSummary,
+    SearchTrace,
+    WeightComponents,
 )
 
 
@@ -58,17 +59,17 @@ class SearchTracer:
         self.max_tokens = max_tokens
 
         # Trace data
-        self.query_embedding: Optional[List[float]] = None
-        self.start_time: Optional[float] = None
-        self.entry_points: List[EntryPoint] = []
-        self.visits: List[NodeVisit] = []
-        self.pruned: List[PruningDecision] = []
-        self.phase_metrics: List[SearchPhaseMetrics] = []
+        self.query_embedding: list[float] | None = None
+        self.start_time: float | None = None
+        self.entry_points: list[EntryPoint] = []
+        self.visits: list[NodeVisit] = []
+        self.pruned: list[PruningDecision] = []
+        self.phase_metrics: list[SearchPhaseMetrics] = []
 
         # New 4-way retrieval tracking
-        self.retrieval_results: List[RetrievalMethodResults] = []
-        self.rrf_merged: List[RRFMergeResult] = []
-        self.reranked: List[RerankedResult] = []
+        self.retrieval_results: list[RetrievalMethodResults] = []
+        self.rrf_merged: list[RRFMergeResult] = []
+        self.reranked: list[RerankedResult] = []
 
         # Tracking state
         self.current_step = 0
@@ -83,7 +84,7 @@ class SearchTracer:
         """Start timing the search."""
         self.start_time = time.time()
 
-    def record_query_embedding(self, embedding: List[float]):
+    def record_query_embedding(self, embedding: list[float]):
         """Record the query embedding."""
         self.query_embedding = embedding
 
@@ -114,12 +115,12 @@ class SearchTracer:
         node_id: str,
         text: str,
         context: str,
-        event_date: datetime,
+        event_date: datetime | None,
         access_count: int,
         is_entry_point: bool,
-        parent_node_id: Optional[str],
-        link_type: Optional[Literal["temporal", "semantic", "entity"]],
-        link_weight: Optional[float],
+        parent_node_id: str | None,
+        link_type: Literal["temporal", "semantic", "entity"] | None,
+        link_weight: float | None,
         activation: float,
         semantic_similarity: float,
         recency: float,
@@ -199,10 +200,10 @@ class SearchTracer:
         to_node_id: str,
         link_type: Literal["temporal", "semantic", "entity"],
         link_weight: float,
-        entity_id: Optional[str],
-        new_activation: Optional[float],
+        entity_id: str | None,
+        new_activation: float | None,
         followed: bool,
-        prune_reason: Optional[str] = None,
+        prune_reason: str | None = None,
         is_supplementary: bool = False,
     ):
         """
@@ -266,7 +267,7 @@ class SearchTracer:
             )
         )
 
-    def add_phase_metric(self, phase_name: str, duration_seconds: float, details: Optional[Dict[str, Any]] = None):
+    def add_phase_metric(self, phase_name: str, duration_seconds: float, details: dict[str, Any] | None = None):
         """
         Record metrics for a search phase.
 
@@ -286,11 +287,11 @@ class SearchTracer:
     def add_retrieval_results(
         self,
         method_name: Literal["semantic", "bm25", "graph", "temporal"],
-        results: List[tuple],  # List of (doc_id, data) tuples
+        results: list[tuple],  # List of (doc_id, data) tuples
         duration_seconds: float,
         score_field: str,  # e.g., "similarity", "bm25_score"
-        metadata: Optional[Dict[str, Any]] = None,
-        fact_type: Optional[str] = None
+        metadata: dict[str, Any] | None = None,
+        fact_type: str | None = None,
     ):
         """
         Record results from a single retrieval method.
@@ -331,7 +332,7 @@ class SearchTracer:
             )
         )
 
-    def add_rrf_merged(self, merged_results: List[tuple]):
+    def add_rrf_merged(self, merged_results: list[tuple]):
         """
         Record RRF merged results.
 
@@ -350,7 +351,7 @@ class SearchTracer:
                 )
             )
 
-    def add_reranked(self, reranked_results: List[Dict[str, Any]], rrf_merged: List):
+    def add_reranked(self, reranked_results: list[dict[str, Any]], rrf_merged: list):
         """
         Record reranked results.
 
@@ -373,7 +374,15 @@ class SearchTracer:
             # Keys from ScoredResult.to_dict(): cross_encoder_score, cross_encoder_score_normalized,
             # rrf_normalized, temporal, recency, combined_score, weight
             score_components = {}
-            for key in ["cross_encoder_score", "cross_encoder_score_normalized", "rrf_score", "rrf_normalized", "temporal", "recency", "combined_score"]:
+            for key in [
+                "cross_encoder_score",
+                "cross_encoder_score_normalized",
+                "rrf_score",
+                "rrf_normalized",
+                "temporal",
+                "recency",
+                "combined_score",
+            ]:
                 if key in result and result[key] is not None:
                     score_components[key] = result[key]
 
@@ -389,7 +398,7 @@ class SearchTracer:
                 )
             )
 
-    def finalize(self, final_results: List[Dict[str, Any]]) -> SearchTrace:
+    def finalize(self, final_results: list[dict[str, Any]]) -> SearchTrace:
         """
         Finalize the trace and return the complete SearchTrace object.
 
@@ -416,7 +425,7 @@ class SearchTracer:
         query_info = QueryInfo(
             query_text=self.query_text,
             query_embedding=self.query_embedding or [],
-            timestamp=datetime.now(timezone.utc),
+            timestamp=datetime.now(UTC),
             budget=self.budget,
             max_tokens=self.max_tokens,
         )

hindsight-api/hindsight_api/engine/search/types.py

@@ -6,8 +6,8 @@ providing type safety and making data flow explicit.
 """
 
 from dataclasses import dataclass, field
-from typing import Optional, List, Dict, Any
 from datetime import datetime
+from typing import Any
 
 
 @dataclass
@@ -17,28 +17,29 @@ class RetrievalResult:
 
     This represents a raw result from the database query, before merging or reranking.
     """
+
     id: str
     text: str
     fact_type: str
-    context: Optional[str] = None
-    event_date: Optional[datetime] = None
-    occurred_start: Optional[datetime] = None
-    occurred_end: Optional[datetime] = None
-    mentioned_at: Optional[datetime] = None
-    document_id: Optional[str] = None
-    chunk_id: Optional[str] = None
+    context: str | None = None
+    event_date: datetime | None = None
+    occurred_start: datetime | None = None
+    occurred_end: datetime | None = None
+    mentioned_at: datetime | None = None
+    document_id: str | None = None
+    chunk_id: str | None = None
     access_count: int = 0
-    embedding: Optional[List[float]] = None
+    embedding: list[float] | None = None
 
     # Retrieval-specific scores (only one will be set depending on retrieval method)
-    similarity: Optional[float] = None  # Semantic retrieval
-    bm25_score: Optional[float] = None  # BM25 retrieval
-    activation: Optional[float] = None  # Graph retrieval (spreading activation)
-    temporal_score: Optional[float] = None  # Temporal retrieval
-    temporal_proximity: Optional[float] = None  # Temporal retrieval
+    similarity: float | None = None  # Semantic retrieval
+    bm25_score: float | None = None  # BM25 retrieval
+    activation: float | None = None  # Graph retrieval (spreading activation)
+    temporal_score: float | None = None  # Temporal retrieval
+    temporal_proximity: float | None = None  # Temporal retrieval
 
     @classmethod
-    def from_db_row(cls, row: Dict[str, Any]) -> "RetrievalResult":
+    def from_db_row(cls, row: dict[str, Any]) -> "RetrievalResult":
         """Create from a database row (asyncpg Record converted to dict)."""
         return cls(
             id=str(row["id"]),
@@ -68,13 +69,14 @@ class MergedCandidate:
 
     Contains the original retrieval data plus RRF metadata.
     """
+
     # Original retrieval data
     retrieval: RetrievalResult
 
     # RRF metadata
     rrf_score: float
     rrf_rank: int = 0
-    source_ranks: Dict[str, int] = field(default_factory=dict)  # method_name -> rank
+    source_ranks: dict[str, int] = field(default_factory=dict)  # method_name -> rank
 
     @property
     def id(self) -> str:
@@ -89,6 +91,7 @@ class ScoredResult:
 
     Contains all retrieval/merge data plus reranking scores and combined score.
     """
+
     # Original merged candidate
     candidate: MergedCandidate
 
@@ -115,7 +118,7 @@ class ScoredResult:
         """Convenience property to access retrieval data."""
         return self.candidate.retrieval
 
-    def to_dict(self) -> Dict[str, Any]:
+    def to_dict(self) -> dict[str, Any]:
         """
         Convert to dict for backwards compatibility.
 

hindsight-api/hindsight_api/engine/task_backend.py

@@ -6,10 +6,12 @@ This provides an abstraction that can be adapted to different execution models:
 - Pub/Sub architectures (future)
 - Message brokers (future)
 """
-from abc import ABC, abstractmethod
-from typing import Any, Dict, Optional, Callable, Awaitable
+
 import asyncio
 import logging
+from abc import ABC, abstractmethod
+from collections.abc import Awaitable, Callable
+from typing import Any
 
 logger = logging.getLogger(__name__)
 
@@ -29,10 +31,10 @@ class TaskBackend(ABC):
 
     def __init__(self):
         """Initialize the task backend."""
-        self._executor: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None
+        self._executor: Callable[[dict[str, Any]], Awaitable[None]] | None = None
         self._initialized = False
 
-    def set_executor(self, executor: Callable[[Dict[str, Any]], Awaitable[None]]):
+    def set_executor(self, executor: Callable[[dict[str, Any]], Awaitable[None]]):
         """
         Set the executor callback for processing tasks.
 
@@ -49,7 +51,7 @@ class TaskBackend(ABC):
         pass
 
     @abstractmethod
-    async def submit_task(self, task_dict: Dict[str, Any]):
+    async def submit_task(self, task_dict: dict[str, Any]):
         """
         Submit a task for execution.
 
@@ -65,7 +67,7 @@ class TaskBackend(ABC):
         """
         pass
 
-    async def _execute_task(self, task_dict: Dict[str, Any]):
+    async def _execute_task(self, task_dict: dict[str, Any]):
         """
         Execute a task through the registered executor.
 
@@ -73,19 +75,52 @@ class TaskBackend(ABC):
             task_dict: Task dictionary to execute
         """
         if self._executor is None:
-            task_type = task_dict.get('type', 'unknown')
+            task_type = task_dict.get("type", "unknown")
             logger.warning(f"No executor registered, skipping task {task_type}")
             return
 
         try:
             await self._executor(task_dict)
         except Exception as e:
-            task_type = task_dict.get('type', 'unknown')
+            task_type = task_dict.get("type", "unknown")
             logger.error(f"Error executing task {task_type}: {e}")
             import traceback
+
             traceback.print_exc()
 
 
+class SyncTaskBackend(TaskBackend):
+    """
+    Synchronous task backend that executes tasks immediately.
+
+    This is useful for embedded/CLI usage where we don't want background
+    workers that prevent clean exit. Tasks are executed inline rather than
+    being queued.
+    """
+
+    async def initialize(self):
+        """No-op for sync backend."""
+        self._initialized = True
+        logger.debug("SyncTaskBackend initialized")
+
+    async def submit_task(self, task_dict: dict[str, Any]):
+        """
+        Execute the task immediately (synchronously).
+
+        Args:
+            task_dict: Task dictionary to execute
+        """
+        if not self._initialized:
+            await self.initialize()
+
+        await self._execute_task(task_dict)
+
+    async def shutdown(self):
+        """No-op for sync backend."""
+        self._initialized = False
+        logger.debug("SyncTaskBackend shutdown")
+
+
 class AsyncIOQueueBackend(TaskBackend):
     """
     Task backend implementation using asyncio queues.
@@ -94,11 +129,7 @@ class AsyncIOQueueBackend(TaskBackend):
     and a periodic consumer worker.
     """
 
-    def __init__(
-        self,
-        batch_size: int = 100,
-        batch_interval: float = 1.0
-    ):
+    def __init__(self, batch_size: int = 100, batch_interval: float = 1.0):
         """
         Initialize AsyncIO queue backend.
 
@@ -107,9 +138,9 @@ class AsyncIOQueueBackend(TaskBackend):
             batch_interval: Maximum time (seconds) to wait before processing batch
         """
         super().__init__()
-        self._queue: Optional[asyncio.Queue] = None
-        self._worker_task: Optional[asyncio.Task] = None
-        self._shutdown_event: Optional[asyncio.Event] = None
+        self._queue: asyncio.Queue | None = None
+        self._worker_task: asyncio.Task | None = None
+        self._shutdown_event: asyncio.Event | None = None
         self._batch_size = batch_size
         self._batch_interval = batch_interval
 
@@ -124,7 +155,7 @@ class AsyncIOQueueBackend(TaskBackend):
         self._initialized = True
         logger.info("AsyncIOQueueBackend initialized")
 
-    async def submit_task(self, task_dict: Dict[str, Any]):
+    async def submit_task(self, task_dict: dict[str, Any]):
         """
         Submit a task by putting it in the queue.
 
@@ -135,8 +166,8 @@ class AsyncIOQueueBackend(TaskBackend):
             await self.initialize()
 
         await self._queue.put(task_dict)
-        task_type = task_dict.get('type', 'unknown')
-        task_id = task_dict.get('id')
+        task_type = task_dict.get("type", "unknown")
+        task_id = task_dict.get("id")
 
     async def wait_for_pending_tasks(self, timeout: float = 5.0):
         """
@@ -200,20 +231,16 @@ class AsyncIOQueueBackend(TaskBackend):
                 while len(tasks) < self._batch_size and asyncio.get_event_loop().time() < deadline:
                     try:
                         remaining_time = max(0.1, deadline - asyncio.get_event_loop().time())
-                        task_dict = await asyncio.wait_for(
-                            self._queue.get(),
-                            timeout=remaining_time
-                        )
+                        task_dict = await asyncio.wait_for(self._queue.get(), timeout=remaining_time)
                         tasks.append(task_dict)
-                    except asyncio.TimeoutError:
+                    except TimeoutError:
                         break
 
                 # Process batch
                 if tasks:
                     # Execute tasks concurrently
                     await asyncio.gather(
-                        *[self._execute_task(task_dict) for task_dict in tasks],
-                        return_exceptions=True
+                        *[self._execute_task(task_dict) for task_dict in tasks], return_exceptions=True
                     )
 
             except asyncio.CancelledError:

hindsight-api/hindsight_api/engine/utils.py

@@ -1,9 +1,10 @@
 """
 Utility functions for memory system.
 """
+
 import logging
 from datetime import datetime
-from typing import List, Dict, TYPE_CHECKING
+from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     from .llm_wrapper import LLMConfig
@@ -12,7 +13,14 @@ if TYPE_CHECKING:
 from .retain.fact_extraction import extract_facts_from_text
 
 
-async def extract_facts(text: str, event_date: datetime, context: str = "", llm_config: 'LLMConfig' = None, agent_name: str = None, extract_opinions: bool = False) -> tuple[List['Fact'], List[tuple[str, int]]]:
+async def extract_facts(
+    text: str,
+    event_date: datetime,
+    context: str = "",
+    llm_config: "LLMConfig" = None,
+    agent_name: str = None,
+    extract_opinions: bool = False,
+) -> tuple[list["Fact"], list[tuple[str, int]]]:
     """
     Extract semantic facts from text using LLM.
 
@@ -41,16 +49,25 @@ async def extract_facts(text: str, event_date: datetime, context: str = "", llm_
     if not text or not text.strip():
         return [], []
 
-    facts, chunks = await extract_facts_from_text(text, event_date, context=context, llm_config=llm_config, agent_name=agent_name, extract_opinions=extract_opinions)
+    facts, chunks = await extract_facts_from_text(
+        text,
+        event_date,
+        context=context,
+        llm_config=llm_config,
+        agent_name=agent_name,
+        extract_opinions=extract_opinions,
+    )
 
     if not facts:
-        logging.warning(f"LLM extracted 0 facts from text of length {len(text)}. This may indicate the text contains no meaningful information, or the LLM failed to extract facts. Full text: {text}")
+        logging.warning(
+            f"LLM extracted 0 facts from text of length {len(text)}. This may indicate the text contains no meaningful information, or the LLM failed to extract facts. Full text: {text}"
+        )
         return [], chunks
 
     return facts, chunks
 
 
-def cosine_similarity(vec1: List[float], vec2: List[float]) -> float:
+def cosine_similarity(vec1: list[float], vec2: list[float]) -> float:
     """
     Calculate cosine similarity between two vectors.
 
@@ -100,6 +117,7 @@ def calculate_recency_weight(days_since: float, half_life_days: float = 365.0) -
         Weight between 0 and 1
     """
     import math
+
     # Logarithmic decay: 1 / (1 + log(1 + days_since/half_life))
     # This decays much slower than exponential, giving better long-term differentiation
     normalized_age = days_since / half_life_days
@@ -121,6 +139,7 @@ def calculate_frequency_weight(access_count: int, max_boost: float = 2.0) -> flo
         Weight between 1.0 and max_boost
     """
     import math
+
     if access_count <= 0:
         return 1.0
 
@@ -158,11 +177,7 @@ def calculate_temporal_anchor(occurred_start: datetime, occurred_end: datetime)
     return midpoint
 
 
-def calculate_temporal_proximity(
-    anchor_a: datetime,
-    anchor_b: datetime,
-    half_life_days: float = 30.0
-) -> float:
+def calculate_temporal_proximity(anchor_a: datetime, anchor_b: datetime, half_life_days: float = 30.0) -> float:
     """
     Calculate temporal proximity between two temporal anchors.
 

hindsight-api/hindsight_api/extensions/__init__.py

@@ -0,0 +1,66 @@
+"""
+Hindsight Extensions System.
+
+Extensions allow customizing and extending Hindsight behavior without modifying core code.
+Extensions are loaded via environment variables pointing to implementation classes.
+
+Example:
+    HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION=mypackage.validators:MyValidator
+    HINDSIGHT_API_OPERATION_VALIDATOR_MAX_RETRIES=3
+
+    HINDSIGHT_API_HTTP_EXTENSION=mypackage.http:MyHttpExtension
+    HINDSIGHT_API_HTTP_SOME_CONFIG=value
+
+Extensions receive an ExtensionContext that provides a controlled API for interacting
+with the system (e.g., running migrations for tenant schemas).
+"""
+
+from hindsight_api.extensions.base import Extension
+from hindsight_api.extensions.builtin import ApiKeyTenantExtension
+from hindsight_api.extensions.context import DefaultExtensionContext, ExtensionContext
+from hindsight_api.extensions.http import HttpExtension
+from hindsight_api.extensions.loader import load_extension
+from hindsight_api.extensions.operation_validator import (
+    OperationValidationError,
+    OperationValidatorExtension,
+    RecallContext,
+    RecallResult,
+    ReflectContext,
+    ReflectResultContext,
+    RetainContext,
+    RetainResult,
+    ValidationResult,
+)
+from hindsight_api.extensions.tenant import (
+    AuthenticationError,
+    TenantContext,
+    TenantExtension,
+)
+from hindsight_api.models import RequestContext
+
+__all__ = [
+    # Base
+    "Extension",
+    "load_extension",
+    # Context
+    "ExtensionContext",
+    "DefaultExtensionContext",
+    # HTTP Extension
+    "HttpExtension",
+    # Operation Validator
+    "OperationValidationError",
+    "OperationValidatorExtension",
+    "RecallContext",
+    "RecallResult",
+    "ReflectContext",
+    "ReflectResultContext",
+    "RetainContext",
+    "RetainResult",
+    "ValidationResult",
+    # Tenant/Auth
+    "ApiKeyTenantExtension",
+    "AuthenticationError",
+    "RequestContext",
+    "TenantContext",
+    "TenantExtension",
+]

hindsight-api/hindsight_api/extensions/base.py

@@ -0,0 +1,81 @@
+"""Base Extension class for all Hindsight extensions."""
+
+from abc import ABC
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from hindsight_api.extensions.context import ExtensionContext
+
+
+class Extension(ABC):
+    """
+    Base class for all Hindsight extensions.
+
+    Extensions are loaded via environment variables and receive configuration
+    from prefixed environment variables.
+
+    Example:
+        HINDSIGHT_API_MY_EXTENSION=mypackage.ext:MyExtension
+        HINDSIGHT_API_MY_SOME_CONFIG=value
+
+        The extension receives: {"some_config": "value"}
+
+    Extensions also receive an ExtensionContext that provides a controlled API
+    for interacting with the system (e.g., running migrations for tenant schemas).
+    """
+
+    def __init__(self, config: dict[str, str]):
+        """
+        Initialize the extension with configuration.
+
+        Args:
+            config: Dictionary of configuration values from environment variables.
+                    Keys are lowercased with the prefix stripped.
+        """
+        self.config = config
+        self._context: "ExtensionContext | None" = None
+
+    def set_context(self, context: "ExtensionContext") -> None:
+        """
+        Set the extension context.
+
+        Called by the extension loader after instantiation.
+        Extensions should not call this directly.
+
+        Args:
+            context: The ExtensionContext providing system APIs.
+        """
+        self._context = context
+
+    @property
+    def context(self) -> "ExtensionContext":
+        """
+        Get the extension context.
+
+        Returns:
+            The ExtensionContext providing system APIs.
+
+        Raises:
+            RuntimeError: If context has not been set yet.
+        """
+        if self._context is None:
+            raise RuntimeError(
+                "Extension context not set. Context is available after the extension is loaded by the system."
+            )
+        return self._context
+
+    async def on_startup(self) -> None:
+        """
+        Called when the application starts.
+
+        Override to perform initialization tasks like connecting to external services.
+        """
+        pass
+
+    async def on_shutdown(self) -> None:
+        """
+        Called when the application shuts down.
+
+        Override to perform cleanup tasks like closing connections.
+        """
+        pass

hindsight-api/hindsight_api/extensions/builtin/__init__.py

@@ -0,0 +1,18 @@
+"""
+Built-in extension implementations.
+
+These are ready-to-use implementations of the extension interfaces.
+They can be used directly or serve as examples for custom implementations.
+
+Available built-in extensions:
+    - ApiKeyTenantExtension: Simple API key validation with public schema
+
+Example usage:
+    HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
+"""
+
+from hindsight_api.extensions.builtin.tenant import ApiKeyTenantExtension
+
+__all__ = [
+    "ApiKeyTenantExtension",
+]

hindsight-api/hindsight_api/extensions/builtin/tenant.py

@@ -0,0 +1,33 @@
+"""Built-in tenant extension implementations."""
+
+from hindsight_api.extensions.tenant import AuthenticationError, TenantContext, TenantExtension
+from hindsight_api.models import RequestContext
+
+
+class ApiKeyTenantExtension(TenantExtension):
+    """
+    Built-in tenant extension that validates API key against an environment variable.
+
+    This is a simple implementation that:
+    1. Validates the API key matches HINDSIGHT_API_TENANT_API_KEY
+    2. Returns 'public' as the schema for all authenticated requests
+
+    Configuration:
+        HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
+        HINDSIGHT_API_TENANT_API_KEY=your-secret-key
+
+    For multi-tenant setups with separate schemas per tenant, implement a custom
+    TenantExtension that looks up the schema based on the API key or token claims.
+    """
+
+    def __init__(self, config: dict[str, str]):
+        super().__init__(config)
+        self.expected_api_key = config.get("api_key")
+        if not self.expected_api_key:
+            raise ValueError("HINDSIGHT_API_TENANT_API_KEY is required when using ApiKeyTenantExtension")
+
+    async def authenticate(self, context: RequestContext) -> TenantContext:
+        """Validate API key and return public schema context."""
+        if context.api_key != self.expected_api_key:
+            raise AuthenticationError("Invalid API key")
+        return TenantContext(schema_name="public")

hindsight-api/hindsight_api/extensions/context.py

@@ -0,0 +1,117 @@
+"""Extension context providing a controlled API for extensions to interact with the system."""
+
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from hindsight_api.engine.interface import MemoryEngineInterface
+
+
+class ExtensionContext(ABC):
+    """
+    Abstract context providing a controlled API for extensions.
+
+    Extensions receive this context instead of direct access to internal
+    components like MemoryEngine or database connections. This provides:
+    - A stable API that won't break when internals change
+    - Security by limiting what extensions can access
+    - Clear documentation of what extensions can do
+
+    Built-in implementation:
+        hindsight_api.extensions.builtin.context.DefaultExtensionContext
+
+    Example usage in an extension:
+        class MyTenantExtension(TenantExtension):
+            async def on_startup(self) -> None:
+                # Run migrations for a new tenant schema
+                await self.context.run_migration("tenant_acme")
+
+        class MyHttpExtension(HttpExtension):
+            def get_router(self, memory):
+                # Use memory engine for custom endpoints
+                engine = self.context.get_memory_engine()
+                ...
+    """
+
+    @abstractmethod
+    async def run_migration(self, schema: str) -> None:
+        """
+        Run database migrations for a specific schema.
+
+        This creates the schema if it doesn't exist and runs all pending
+        migrations. Uses advisory locks to coordinate between distributed workers.
+
+        Args:
+            schema: PostgreSQL schema name (e.g., "tenant_acme").
+                    The schema will be created if it doesn't exist.
+
+        Raises:
+            RuntimeError: If migrations fail to complete.
+
+        Example:
+            # Provision a new tenant schema
+            await context.run_migration("tenant_acme")
+        """
+        ...
+
+    @abstractmethod
+    def get_memory_engine(self) -> "MemoryEngineInterface":
+        """
+        Get the memory engine interface.
+
+        Returns the MemoryEngineInterface for performing memory operations
+        like retain, recall, reflect, and entity/document management.
+
+        Returns:
+            MemoryEngineInterface instance.
+
+        Example:
+            engine = context.get_memory_engine()
+            result = await engine.recall_async(bank_id, query)
+        """
+        ...
+
+
+class DefaultExtensionContext(ExtensionContext):
+    """
+    Default implementation of ExtensionContext.
+
+    Uses the system's database URL and migration infrastructure.
+    """
+
+    def __init__(
+        self,
+        database_url: str,
+        memory_engine: "MemoryEngineInterface | None" = None,
+    ):
+        """
+        Initialize the context.
+
+        Args:
+            database_url: SQLAlchemy database URL for migrations.
+            memory_engine: Optional MemoryEngine instance for memory operations.
+        """
+        self._database_url = database_url
+        self._memory_engine = memory_engine
+
+    async def run_migration(self, schema: str) -> None:
+        """Run migrations for a specific schema."""
+        from hindsight_api.migrations import run_migrations
+
+        # Prefer getting URL from memory engine (handles pg0 case where URL is set after init)
+        db_url = self._database_url
+        if self._memory_engine is not None:
+            engine_url = getattr(self._memory_engine, "db_url", None)
+            if engine_url:
+                db_url = engine_url
+
+        run_migrations(db_url, schema=schema)
+
+    def get_memory_engine(self) -> "MemoryEngineInterface":
+        """Get the memory engine interface."""
+        if self._memory_engine is None:
+            raise RuntimeError(
+                "Memory engine not configured in ExtensionContext. "
+                "Ensure the context was created with a memory_engine parameter."
+            )
+        return self._memory_engine

hindsight-api/hindsight_api/extensions/http.py

@@ -0,0 +1,89 @@
+"""
+HTTP Extension for adding custom endpoints to the Hindsight API.
+
+This extension allows adding custom HTTP endpoints under the /ext/ path prefix.
+The extension provides a FastAPI router that is mounted on the main application.
+"""
+
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING
+
+from fastapi import APIRouter
+
+from hindsight_api.extensions.base import Extension
+
+if TYPE_CHECKING:
+    from hindsight_api import MemoryEngine
+
+
+class HttpExtension(Extension, ABC):
+    """
+    Base class for HTTP extensions that add custom API endpoints.
+
+    HTTP extensions provide a FastAPI router that gets mounted under /ext/.
+    The extension has full control over the routes, request/response models, and handlers.
+
+    Example:
+        ```python
+        from fastapi import APIRouter
+        from hindsight_api.extensions import HttpExtension
+
+        class MyHttpExtension(HttpExtension):
+            def get_router(self, memory: MemoryEngine) -> APIRouter:
+                router = APIRouter()
+
+                @router.get("/hello")
+                async def hello():
+                    return {"message": "Hello from extension!"}
+
+                @router.post("/custom/{bank_id}/action")
+                async def custom_action(bank_id: str):
+                    # Access memory engine for database operations
+                    pool = await memory._get_pool()
+                    # ... custom logic
+                    return {"status": "ok"}
+
+                return router
+        ```
+
+    The routes will be available at:
+        - GET /ext/hello
+        - POST /ext/custom/{bank_id}/action
+
+    Configuration via environment variables:
+        HINDSIGHT_API_HTTP_EXTENSION=mypackage.ext:MyHttpExtension
+        HINDSIGHT_API_HTTP_SOME_CONFIG=value
+
+    The extension receives config: {"some_config": "value"}
+    """
+
+    @abstractmethod
+    def get_router(self, memory: "MemoryEngine") -> APIRouter:
+        """
+        Return a FastAPI router with custom endpoints.
+
+        The router will be mounted at /ext/ on the main application.
+        All routes defined in the router will be prefixed with /ext/.
+
+        Args:
+            memory: The MemoryEngine instance for database access and core operations.
+                   Use this to access the connection pool, run queries, or call
+                   memory operations like retain, recall, etc.
+
+        Returns:
+            A FastAPI APIRouter with the custom endpoints defined.
+
+        Example:
+            ```python
+            def get_router(self, memory: MemoryEngine) -> APIRouter:
+                router = APIRouter(tags=["My Extension"])
+
+                @router.get("/status")
+                async def status():
+                    health = await memory.health_check()
+                    return {"extension": "healthy", "memory": health}
+
+                return router
+            ```
+        """
+        pass

hindsight-api/hindsight_api/extensions/loader.py

@@ -0,0 +1,125 @@
+"""Extension loader utilities."""
+
+import importlib
+import logging
+import os
+from typing import TYPE_CHECKING, TypeVar
+
+from hindsight_api.extensions.base import Extension
+
+if TYPE_CHECKING:
+    from hindsight_api.extensions.context import ExtensionContext
+
+logger = logging.getLogger(__name__)
+
+T = TypeVar("T", bound=Extension)
+
+
+class ExtensionLoadError(Exception):
+    """Raised when an extension fails to load."""
+
+    pass
+
+
+def load_extension(
+    prefix: str,
+    base_class: type[T],
+    env_prefix: str = "HINDSIGHT_API",
+    context: "ExtensionContext | None" = None,
+) -> T | None:
+    """
+    Load an extension from environment variable configuration.
+
+    The extension class is specified via {env_prefix}_{prefix}_EXTENSION environment
+    variable in the format "module.path:ClassName".
+
+    Configuration for the extension is collected from all environment variables
+    matching {env_prefix}_{prefix}_* (excluding the EXTENSION variable itself).
+
+    Args:
+        prefix: The extension prefix (e.g., "OPERATION_VALIDATOR").
+        base_class: The base class that the extension must inherit from.
+        env_prefix: The environment variable prefix (default: "HINDSIGHT_API").
+        context: Optional ExtensionContext to provide system APIs to the extension.
+
+    Returns:
+        An instance of the extension, or None if not configured.
+
+    Raises:
+        ExtensionLoadError: If the extension fails to load or validate.
+
+    Example:
+        HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION=mypackage.validators:MyValidator
+        HINDSIGHT_API_OPERATION_VALIDATOR_MAX_REQUESTS=100
+
+        ext = load_extension("OPERATION_VALIDATOR", OperationValidatorExtension)
+        # ext.config == {"max_requests": "100"}
+    """
+    env_var = f"{env_prefix}_{prefix}_EXTENSION"
+    ext_path = os.getenv(env_var)
+
+    if not ext_path:
+        logger.debug(f"No extension configured for {env_var}")
+        return None
+
+    logger.info(f"Loading extension from {env_var}={ext_path}")
+
+    # Parse "module.path:ClassName"
+    if ":" not in ext_path:
+        raise ExtensionLoadError(f"Invalid extension path '{ext_path}'. Expected format: 'module.path:ClassName'")
+
+    module_path, class_name = ext_path.rsplit(":", 1)
+
+    # Import the module
+    try:
+        module = importlib.import_module(module_path)
+    except ImportError as e:
+        raise ExtensionLoadError(f"Failed to import extension module '{module_path}': {e}") from e
+
+    # Get the class
+    try:
+        ext_class = getattr(module, class_name)
+    except AttributeError as e:
+        raise ExtensionLoadError(f"Extension class '{class_name}' not found in module '{module_path}'") from e
+
+    # Validate inheritance
+    if not isinstance(ext_class, type) or not issubclass(ext_class, base_class):
+        raise ExtensionLoadError(f"Extension class '{ext_class.__name__}' must inherit from '{base_class.__name__}'")
+
+    # Collect configuration from environment variables
+    config = _collect_config(env_prefix, prefix)
+
+    logger.info(f"Loaded extension {ext_class.__name__} with config keys: {list(config.keys())}")
+
+    # Instantiate the extension
+    try:
+        extension = ext_class(config)
+    except Exception as e:
+        raise ExtensionLoadError(f"Failed to instantiate extension '{ext_class.__name__}': {e}") from e
+
+    # Set the context if provided
+    if context is not None:
+        extension.set_context(context)
+        logger.debug(f"Set context on extension {ext_class.__name__}")
+
+    return extension
+
+
+def _collect_config(env_prefix: str, prefix: str) -> dict[str, str]:
+    """
+    Collect configuration from environment variables.
+
+    Collects all variables matching {env_prefix}_{prefix}_* except for
+    {env_prefix}_{prefix}_EXTENSION, strips the prefix, and lowercases keys.
+    """
+    config = {}
+    full_prefix = f"{env_prefix}_{prefix}_"
+    extension_var = f"{full_prefix}EXTENSION"
+
+    for key, value in os.environ.items():
+        if key.startswith(full_prefix) and key != extension_var:
+            # Strip prefix and lowercase the key
+            config_key = key[len(full_prefix) :].lower()
+            config[config_key] = value
+
+    return config

hindsight-api/hindsight_api/extensions/operation_validator.py

@@ -0,0 +1,327 @@
+"""Operation Validator Extension for validating retain/recall/reflect operations."""
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from datetime import datetime
+from typing import TYPE_CHECKING, Any
+
+from hindsight_api.extensions.base import Extension
+
+if TYPE_CHECKING:
+    from hindsight_api.engine.memory_engine import Budget
+    from hindsight_api.engine.response_models import RecallResult as RecallResultModel
+    from hindsight_api.engine.response_models import ReflectResult
+    from hindsight_api.models import RequestContext
+
+
+class OperationValidationError(Exception):
+    """Raised when an operation fails validation."""
+
+    def __init__(self, reason: str, status_code: int = 403):
+        self.reason = reason
+        self.status_code = status_code
+        super().__init__(f"Operation validation failed: {reason}")
+
+
+@dataclass
+class ValidationResult:
+    """Result of an operation validation."""
+
+    allowed: bool
+    reason: str | None = None
+    status_code: int = 403  # Default to Forbidden
+
+    @classmethod
+    def accept(cls) -> "ValidationResult":
+        """Create an accepted validation result."""
+        return cls(allowed=True)
+
+    @classmethod
+    def reject(cls, reason: str, status_code: int = 403) -> "ValidationResult":
+        """Create a rejected validation result with a reason and HTTP status code."""
+        return cls(allowed=False, reason=reason, status_code=status_code)
+
+
+# =============================================================================
+# Pre-operation Contexts (all user-provided parameters)
+# =============================================================================
+
+
+@dataclass
+class RetainContext:
+    """Context for a retain operation validation (pre-operation).
+
+    Contains ALL user-provided parameters for the retain operation.
+    """
+
+    bank_id: str
+    contents: list[dict]  # List of {content, context, event_date, document_id}
+    request_context: "RequestContext"
+    document_id: str | None = None
+    fact_type_override: str | None = None
+    confidence_score: float | None = None
+
+
+@dataclass
+class RecallContext:
+    """Context for a recall operation validation (pre-operation).
+
+    Contains ALL user-provided parameters for the recall operation.
+    """
+
+    bank_id: str
+    query: str
+    request_context: "RequestContext"
+    budget: "Budget | None" = None
+    max_tokens: int = 4096
+    enable_trace: bool = False
+    fact_types: list[str] = field(default_factory=list)
+    question_date: datetime | None = None
+    include_entities: bool = False
+    max_entity_tokens: int = 500
+    include_chunks: bool = False
+    max_chunk_tokens: int = 8192
+
+
+@dataclass
+class ReflectContext:
+    """Context for a reflect operation validation (pre-operation).
+
+    Contains ALL user-provided parameters for the reflect operation.
+    """
+
+    bank_id: str
+    query: str
+    request_context: "RequestContext"
+    budget: "Budget | None" = None
+    context: str | None = None
+
+
+# =============================================================================
+# Post-operation Contexts (includes results)
+# =============================================================================
+
+
+@dataclass
+class RetainResult:
+    """Result context for post-retain hook.
+
+    Contains the operation parameters and the result.
+    """
+
+    bank_id: str
+    contents: list[dict]
+    request_context: "RequestContext"
+    document_id: str | None
+    fact_type_override: str | None
+    confidence_score: float | None
+    # Result
+    unit_ids: list[list[str]]  # List of unit IDs per content item
+    success: bool = True
+    error: str | None = None
+
+
+@dataclass
+class RecallResult:
+    """Result context for post-recall hook.
+
+    Contains the operation parameters and the result.
+    """
+
+    bank_id: str
+    query: str
+    request_context: "RequestContext"
+    budget: "Budget | None"
+    max_tokens: int
+    enable_trace: bool
+    fact_types: list[str]
+    question_date: datetime | None
+    include_entities: bool
+    max_entity_tokens: int
+    include_chunks: bool
+    max_chunk_tokens: int
+    # Result
+    result: "RecallResultModel | None" = None
+    success: bool = True
+    error: str | None = None
+
+
+@dataclass
+class ReflectResultContext:
+    """Result context for post-reflect hook.
+
+    Contains the operation parameters and the result.
+    """
+
+    bank_id: str
+    query: str
+    request_context: "RequestContext"
+    budget: "Budget | None"
+    context: str | None
+    # Result
+    result: "ReflectResult | None" = None
+    success: bool = True
+    error: str | None = None
+
+
+class OperationValidatorExtension(Extension, ABC):
+    """
+    Validates and hooks into retain/recall/reflect operations.
+
+    This extension allows implementing custom logic such as:
+    - Rate limiting (pre-operation)
+    - Quota enforcement (pre-operation)
+    - Permission checks (pre-operation)
+    - Content filtering (pre-operation)
+    - Usage tracking (post-operation)
+    - Audit logging (post-operation)
+    - Metrics collection (post-operation)
+
+    Enable via environment variable:
+        HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION=mypackage.validators:MyValidator
+
+    Configuration is passed from prefixed environment variables:
+        HINDSIGHT_API_OPERATION_VALIDATOR_MAX_REQUESTS=100
+        -> config = {"max_requests": "100"}
+
+    Hook execution order:
+        1. validate_retain/validate_recall/validate_reflect (pre-operation)
+        2. [operation executes]
+        3. on_retain_complete/on_recall_complete/on_reflect_complete (post-operation)
+    """
+
+    # =========================================================================
+    # Pre-operation validation hooks (abstract - must be implemented)
+    # =========================================================================
+
+    @abstractmethod
+    async def validate_retain(self, ctx: RetainContext) -> ValidationResult:
+        """
+        Validate a retain operation before execution.
+
+        Called before the retain operation is processed. Return ValidationResult.reject()
+        to prevent the operation from executing.
+
+        Args:
+            ctx: Context containing all user-provided parameters:
+                - bank_id: Bank identifier
+                - contents: List of content dicts
+                - request_context: Request context with auth info
+                - document_id: Optional document ID
+                - fact_type_override: Optional fact type override
+                - confidence_score: Optional confidence score
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        ...
+
+    @abstractmethod
+    async def validate_recall(self, ctx: RecallContext) -> ValidationResult:
+        """
+        Validate a recall operation before execution.
+
+        Called before the recall operation is processed. Return ValidationResult.reject()
+        to prevent the operation from executing.
+
+        Args:
+            ctx: Context containing all user-provided parameters:
+                - bank_id: Bank identifier
+                - query: Search query
+                - request_context: Request context with auth info
+                - budget: Budget level
+                - max_tokens: Maximum tokens to return
+                - enable_trace: Whether to include trace info
+                - fact_types: List of fact types to search
+                - question_date: Optional date context for query
+                - include_entities: Whether to include entity data
+                - max_entity_tokens: Max tokens for entities
+                - include_chunks: Whether to include chunks
+                - max_chunk_tokens: Max tokens for chunks
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        ...
+
+    @abstractmethod
+    async def validate_reflect(self, ctx: ReflectContext) -> ValidationResult:
+        """
+        Validate a reflect operation before execution.
+
+        Called before the reflect operation is processed. Return ValidationResult.reject()
+        to prevent the operation from executing.
+
+        Args:
+            ctx: Context containing all user-provided parameters:
+                - bank_id: Bank identifier
+                - query: Question to answer
+                - request_context: Request context with auth info
+                - budget: Budget level
+                - context: Optional additional context
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        ...
+
+    # =========================================================================
+    # Post-operation hooks (optional - override to implement)
+    # =========================================================================
+
+    async def on_retain_complete(self, result: RetainResult) -> None:
+        """
+        Called after a retain operation completes (success or failure).
+
+        Override this method to implement post-operation logic such as:
+        - Usage tracking
+        - Audit logging
+        - Metrics collection
+        - Notifications
+
+        Args:
+            result: Result context containing:
+                - All original operation parameters
+                - unit_ids: List of created unit IDs (if success)
+                - success: Whether the operation succeeded
+                - error: Error message (if failed)
+        """
+        pass
+
+    async def on_recall_complete(self, result: RecallResult) -> None:
+        """
+        Called after a recall operation completes (success or failure).
+
+        Override this method to implement post-operation logic such as:
+        - Usage tracking
+        - Audit logging
+        - Metrics collection
+        - Query analytics
+
+        Args:
+            result: Result context containing:
+                - All original operation parameters
+                - result: RecallResultModel (if success)
+                - success: Whether the operation succeeded
+                - error: Error message (if failed)
+        """
+        pass
+
+    async def on_reflect_complete(self, result: ReflectResultContext) -> None:
+        """
+        Called after a reflect operation completes (success or failure).
+
+        Override this method to implement post-operation logic such as:
+        - Usage tracking
+        - Audit logging
+        - Metrics collection
+        - Response analytics
+
+        Args:
+            result: Result context containing:
+                - All original operation parameters
+                - result: ReflectResult (if success)
+                - success: Whether the operation succeeded
+                - error: Error message (if failed)
+        """
+        pass

hindsight-api/hindsight_api/extensions/tenant.py

@@ -0,0 +1,63 @@
+"""Tenant Extension for multi-tenancy and API key authentication."""
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+
+from hindsight_api.extensions.base import Extension
+from hindsight_api.models import RequestContext
+
+
+class AuthenticationError(Exception):
+    """Raised when authentication fails."""
+
+    def __init__(self, reason: str):
+        self.reason = reason
+        super().__init__(f"Authentication failed: {reason}")
+
+
+@dataclass
+class TenantContext:
+    """
+    Tenant context returned by authentication.
+
+    Contains the PostgreSQL schema name for tenant isolation.
+    All database queries will use fully-qualified table names
+    with this schema (e.g., schema_name.memory_units).
+    """
+
+    schema_name: str
+
+
+class TenantExtension(Extension, ABC):
+    """
+    Extension for multi-tenancy and API key authentication.
+
+    This extension validates incoming requests and returns the tenant context
+    including the PostgreSQL schema to use for database operations.
+
+    Built-in implementation:
+        hindsight_api.extensions.builtin.tenant.ApiKeyTenantExtension
+
+    Enable via environment variable:
+        HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
+        HINDSIGHT_API_TENANT_API_KEY=your-secret-key
+
+    The returned schema_name is used for fully-qualified table names in queries,
+    enabling tenant isolation at the database level.
+    """
+
+    @abstractmethod
+    async def authenticate(self, context: RequestContext) -> TenantContext:
+        """
+        Authenticate the action context and return tenant context.
+
+        Args:
+            context: The action context containing API key and other auth data.
+
+        Returns:
+            TenantContext with the schema_name for database operations.
+
+        Raises:
+            AuthenticationError: If authentication fails.
+        """
+        ...

hindsight-api/hindsight_api/main.py

@@ -4,8 +4,12 @@ Command-line interface for Hindsight API.
 Run the server with:
     hindsight-api
 
+Run as background daemon:
+    hindsight-api --daemon
+
 Stop with Ctrl+C.
 """
+
 import argparse
 import asyncio
 import atexit
@@ -13,17 +17,21 @@ import os
 import signal
 import sys
 import warnings
-from typing import Optional
 
 import uvicorn
 
 from . import MemoryEngine
 from .api import create_app
-from .config import get_config, HindsightConfig
-
 from .banner import print_banner
-print()
-print_banner()
+from .config import HindsightConfig, get_config
+from .daemon import (
+    DEFAULT_DAEMON_PORT,
+    DEFAULT_IDLE_TIMEOUT,
+    DaemonLock,
+    IdleTimeoutMiddleware,
+    daemonize,
+)
+from .extensions import DefaultExtensionContext, OperationValidatorExtension, TenantExtension, load_extension
 
 # Filter deprecation warnings from third-party libraries
 warnings.filterwarnings("ignore", message="websockets.legacy is deprecated")
@@ -33,7 +41,7 @@ warnings.filterwarnings("ignore", message="websockets.server.WebSocketServerProt
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
 # Global reference for cleanup
-_memory: Optional[MemoryEngine] = None
+_memory: MemoryEngine | None = None
 
 
 def _cleanup():
@@ -70,62 +78,88 @@ def main():
 
     # Server options
     parser.add_argument(
-        "--host", default=config.host,
-        help=f"Host to bind to (default: {config.host}, env: HINDSIGHT_API_HOST)"
+        "--host", default=config.host, help=f"Host to bind to (default: {config.host}, env: HINDSIGHT_API_HOST)"
     )
     parser.add_argument(
-        "--port", type=int, default=config.port,
-        help=f"Port to bind to (default: {config.port}, env: HINDSIGHT_API_PORT)"
+        "--port",
+        type=int,
+        default=config.port,
+        help=f"Port to bind to (default: {config.port}, env: HINDSIGHT_API_PORT)",
     )
     parser.add_argument(
-        "--log-level", default=config.log_level,
+        "--log-level",
+        default=config.log_level,
         choices=["critical", "error", "warning", "info", "debug", "trace"],
-        help=f"Log level (default: {config.log_level}, env: HINDSIGHT_API_LOG_LEVEL)"
+        help=f"Log level (default: {config.log_level}, env: HINDSIGHT_API_LOG_LEVEL)",
     )
 
     # Development options
-    parser.add_argument(
-        "--reload", action="store_true",
-        help="Enable auto-reload on code changes (development only)"
-    )
-    parser.add_argument(
-        "--workers", type=int, default=1,
-        help="Number of worker processes (default: 1)"
-    )
+    parser.add_argument("--reload", action="store_true", help="Enable auto-reload on code changes (development only)")
+    parser.add_argument("--workers", type=int, default=1, help="Number of worker processes (default: 1)")
 
     # Access log options
-    parser.add_argument(
-        "--access-log", action="store_true",
-        help="Enable access log"
-    )
-    parser.add_argument(
-        "--no-access-log", dest="access_log", action="store_false",
-        help="Disable access log (default)"
-    )
+    parser.add_argument("--access-log", action="store_true", help="Enable access log")
+    parser.add_argument("--no-access-log", dest="access_log", action="store_false", help="Disable access log (default)")
     parser.set_defaults(access_log=False)
 
     # Proxy options
     parser.add_argument(
-        "--proxy-headers", action="store_true",
-        help="Enable X-Forwarded-Proto, X-Forwarded-For headers"
+        "--proxy-headers", action="store_true", help="Enable X-Forwarded-Proto, X-Forwarded-For headers"
     )
     parser.add_argument(
-        "--forwarded-allow-ips", default=None,
-        help="Comma separated list of IPs to trust with proxy headers"
+        "--forwarded-allow-ips", default=None, help="Comma separated list of IPs to trust with proxy headers"
     )
 
     # SSL options
+    parser.add_argument("--ssl-keyfile", default=None, help="SSL key file")
+    parser.add_argument("--ssl-certfile", default=None, help="SSL certificate file")
+
+    # Daemon mode options
     parser.add_argument(
-        "--ssl-keyfile", default=None,
-        help="SSL key file"
+        "--daemon",
+        action="store_true",
+        help=f"Run as background daemon (uses port {DEFAULT_DAEMON_PORT}, auto-exits after idle)",
     )
     parser.add_argument(
-        "--ssl-certfile", default=None,
-        help="SSL certificate file"
+        "--idle-timeout",
+        type=int,
+        default=DEFAULT_IDLE_TIMEOUT,
+        help=f"Idle timeout in seconds before auto-exit in daemon mode (default: {DEFAULT_IDLE_TIMEOUT})",
     )
 
     args = parser.parse_args()
 
+    # Daemon mode handling
+    if args.daemon:
+        # Use fixed daemon port
+        args.port = DEFAULT_DAEMON_PORT
+        args.host = "127.0.0.1"  # Only bind to localhost for security
+
+        # Check if another daemon is already running
+        daemon_lock = DaemonLock()
+        if not daemon_lock.acquire():
+            print(f"Daemon already running (PID: {daemon_lock.get_pid()})", file=sys.stderr)
+            sys.exit(1)
+
+        # Fork into background
+        daemonize()
+
+        # Re-acquire lock in child process
+        daemon_lock = DaemonLock()
+        if not daemon_lock.acquire():
+            sys.exit(1)
+
+        # Register cleanup to release lock
+        def release_lock():
+            daemon_lock.release()
+
+        atexit.register(release_lock)
+
+    # Print banner (not in daemon mode)
+    if not args.daemon:
+        print()
+        print_banner()
+
     # Configure Python logging based on log level
     # Update config with CLI override if provided
     if args.log_level != config.log_level:
@@ -135,6 +169,8 @@ def main():
             llm_api_key=config.llm_api_key,
             llm_model=config.llm_model,
             llm_base_url=config.llm_base_url,
+            llm_max_concurrent=config.llm_max_concurrent,
+            llm_timeout=config.llm_timeout,
             embeddings_provider=config.embeddings_provider,
             embeddings_local_model=config.embeddings_local_model,
             embeddings_tei_url=config.embeddings_tei_url,
@@ -145,16 +181,46 @@ def main():
             port=args.port,
             log_level=args.log_level,
             mcp_enabled=config.mcp_enabled,
+            graph_retriever=config.graph_retriever,
+            observation_min_facts=config.observation_min_facts,
+            observation_top_entities=config.observation_top_entities,
+            skip_llm_verification=config.skip_llm_verification,
+            lazy_reranker=config.lazy_reranker,
         )
     config.configure_logging()
+    if not args.daemon:
+        config.log_config()
 
     # Register cleanup handlers
     atexit.register(_cleanup)
     signal.signal(signal.SIGINT, _signal_handler)
     signal.signal(signal.SIGTERM, _signal_handler)
 
+    # Load operation validator extension if configured
+    operation_validator = load_extension("OPERATION_VALIDATOR", OperationValidatorExtension)
+    if operation_validator:
+        import logging
+
+        logging.info(f"Loaded operation validator: {operation_validator.__class__.__name__}")
+
+    # Load tenant extension if configured
+    tenant_extension = load_extension("TENANT", TenantExtension)
+    if tenant_extension:
+        import logging
+
+        logging.info(f"Loaded tenant extension: {tenant_extension.__class__.__name__}")
+
     # Create MemoryEngine (reads configuration from environment)
-    _memory = MemoryEngine()
+    _memory = MemoryEngine(operation_validator=operation_validator, tenant_extension=tenant_extension)
+
+    # Set extension context on tenant extension (needed for schema provisioning)
+    if tenant_extension:
+        extension_context = DefaultExtensionContext(
+            database_url=config.database_url,
+            memory_engine=_memory,
+        )
+        tenant_extension.set_context(extension_context)
+        logging.info("Extension context set on tenant extension")
 
     # Create FastAPI app
     app = create_app(
@@ -165,6 +231,12 @@ def main():
         initialize_memory=True,
     )
 
+    # Wrap with idle timeout middleware in daemon mode
+    idle_middleware = None
+    if args.daemon:
+        idle_middleware = IdleTimeoutMiddleware(app, idle_timeout=args.idle_timeout)
+        app = idle_middleware
+
     # Prepare uvicorn config
     uvicorn_config = {
         "app": app,
@@ -188,21 +260,40 @@ def main():
     if args.ssl_certfile:
         uvicorn_config["ssl_certfile"] = args.ssl_certfile
 
+    # Print startup info (not in daemon mode)
+    if not args.daemon:
+        from .banner import print_startup_info
 
+        print_startup_info(
+            host=args.host,
+            port=args.port,
+            database_url=config.database_url,
+            llm_provider=config.llm_provider,
+            llm_model=config.llm_model,
+            embeddings_provider=config.embeddings_provider,
+            reranker_provider=config.reranker_provider,
+            mcp_enabled=config.mcp_enabled,
+        )
 
-    from .banner import print_startup_info
-    print_startup_info(
-        host=args.host,
-        port=args.port,
-        database_url=config.database_url,
-        llm_provider=config.llm_provider,
-        llm_model=config.llm_model,
-        embeddings_provider=config.embeddings_provider,
-        reranker_provider=config.reranker_provider,
-        mcp_enabled=config.mcp_enabled,
-    )
+    # Start idle checker in daemon mode
+    if idle_middleware is not None:
+        # Start the idle checker in a background thread with its own event loop
+        import threading
 
-    uvicorn.run(**uvicorn_config)
+        def run_idle_checker():
+            import time
+
+            time.sleep(2)  # Wait for uvicorn to start
+            try:
+                loop = asyncio.new_event_loop()
+                asyncio.set_event_loop(loop)
+                loop.run_until_complete(idle_middleware._check_idle())
+            except Exception:
+                pass
+
+        threading.Thread(target=run_idle_checker, daemon=True).start()
+
+    uvicorn.run(**uvicorn_config)  # type: ignore[invalid-argument-type] - dict kwargs
 
 
 if __name__ == "__main__":

hindsight-api/hindsight_api/mcp_local.py

@@ -0,0 +1,199 @@
+"""
+Local MCP server for use with Claude Code (stdio transport).
+
+This runs a fully local Hindsight instance with embedded PostgreSQL (pg0).
+No external database or server required.
+
+Run with:
+    hindsight-local-mcp
+
+Or with uvx:
+    uvx hindsight-api@latest hindsight-local-mcp
+
+Configure in Claude Code's MCP settings:
+    {
+        "mcpServers": {
+            "hindsight": {
+                "command": "uvx",
+                "args": ["hindsight-api@latest", "hindsight-local-mcp"],
+                "env": {
+                    "HINDSIGHT_API_LLM_API_KEY": "your-openai-key"
+                }
+            }
+        }
+    }
+
+Environment variables:
+    HINDSIGHT_API_LLM_API_KEY: Required. API key for LLM provider.
+    HINDSIGHT_API_LLM_PROVIDER: Optional. LLM provider (default: "openai").
+    HINDSIGHT_API_LLM_MODEL: Optional. LLM model (default: "gpt-4o-mini").
+    HINDSIGHT_API_MCP_LOCAL_BANK_ID: Optional. Memory bank ID (default: "mcp").
+    HINDSIGHT_API_LOG_LEVEL: Optional. Log level (default: "warning").
+    HINDSIGHT_API_MCP_INSTRUCTIONS: Optional. Additional instructions appended to both retain and recall tools.
+
+Example custom instructions (these are ADDED to the default behavior):
+    To also store assistant actions:
+        HINDSIGHT_API_MCP_INSTRUCTIONS="Also store every action you take, including tool calls, code written, and decisions made."
+
+    To also store conversation summaries:
+        HINDSIGHT_API_MCP_INSTRUCTIONS="Also store summaries of important conversations and their outcomes."
+"""
+
+import logging
+import os
+import sys
+
+from mcp.server.fastmcp import FastMCP
+from mcp.types import Icon
+
+from hindsight_api.config import (
+    DEFAULT_MCP_LOCAL_BANK_ID,
+    DEFAULT_MCP_RECALL_DESCRIPTION,
+    DEFAULT_MCP_RETAIN_DESCRIPTION,
+    ENV_MCP_INSTRUCTIONS,
+    ENV_MCP_LOCAL_BANK_ID,
+)
+
+# Configure logging - default to warning to avoid polluting stderr during MCP init
+# MCP clients interpret stderr output as errors, so we suppress INFO logs by default
+_log_level_str = os.environ.get("HINDSIGHT_API_LOG_LEVEL", "warning").lower()
+_log_level_map = {
+    "critical": logging.CRITICAL,
+    "error": logging.ERROR,
+    "warning": logging.WARNING,
+    "info": logging.INFO,
+    "debug": logging.DEBUG,
+}
+logging.basicConfig(
+    level=_log_level_map.get(_log_level_str, logging.WARNING),
+    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+    stream=sys.stderr,  # MCP uses stdout for protocol, logs go to stderr
+)
+logger = logging.getLogger(__name__)
+
+
+def create_local_mcp_server(bank_id: str, memory=None) -> FastMCP:
+    """
+    Create a stdio MCP server with retain/recall tools.
+
+    Args:
+        bank_id: The memory bank ID to use for all operations.
+        memory: Optional MemoryEngine instance. If not provided, creates one with pg0.
+
+    Returns:
+        Configured FastMCP server instance.
+    """
+    # Import here to avoid slow startup if just checking --help
+    from hindsight_api import MemoryEngine
+    from hindsight_api.engine.memory_engine import Budget
+    from hindsight_api.engine.response_models import VALID_RECALL_FACT_TYPES
+    from hindsight_api.models import RequestContext
+
+    # Create memory engine with pg0 embedded database if not provided
+    if memory is None:
+        memory = MemoryEngine(db_url="pg0://hindsight-mcp")
+
+    # Get custom instructions from environment variable (appended to both tools)
+    extra_instructions = os.environ.get(ENV_MCP_INSTRUCTIONS, "")
+
+    retain_description = DEFAULT_MCP_RETAIN_DESCRIPTION
+    recall_description = DEFAULT_MCP_RECALL_DESCRIPTION
+
+    if extra_instructions:
+        retain_description = f"{DEFAULT_MCP_RETAIN_DESCRIPTION}\n\nAdditional instructions: {extra_instructions}"
+        recall_description = f"{DEFAULT_MCP_RECALL_DESCRIPTION}\n\nAdditional instructions: {extra_instructions}"
+
+    mcp = FastMCP("hindsight")
+
+    @mcp.tool(description=retain_description)
+    async def retain(content: str, context: str = "general") -> dict:
+        """
+        Args:
+            content: The fact/memory to store (be specific and include relevant details)
+            context: Category for the memory (e.g., 'preferences', 'work', 'hobbies', 'family'). Default: 'general'
+        """
+        import asyncio
+
+        async def _retain():
+            try:
+                await memory.retain_batch_async(
+                    bank_id=bank_id,
+                    contents=[{"content": content, "context": context}],
+                    request_context=RequestContext(),
+                )
+            except Exception as e:
+                logger.error(f"Error storing memory: {e}", exc_info=True)
+
+        # Fire and forget - don't block on memory storage
+        asyncio.create_task(_retain())
+        return {"status": "accepted", "message": "Memory storage initiated"}
+
+    @mcp.tool(description=recall_description)
+    async def recall(query: str, max_tokens: int = 4096, budget: str = "low") -> dict:
+        """
+        Args:
+            query: Natural language search query (e.g., "user's food preferences", "what projects is user working on")
+            max_tokens: Maximum tokens to return in results (default: 4096)
+            budget: Search budget level - "low", "mid", or "high" (default: "low")
+        """
+        try:
+            # Map string budget to enum
+            budget_map = {"low": Budget.LOW, "mid": Budget.MID, "high": Budget.HIGH}
+            budget_enum = budget_map.get(budget.lower(), Budget.LOW)
+
+            search_result = await memory.recall_async(
+                bank_id=bank_id,
+                query=query,
+                fact_type=list(VALID_RECALL_FACT_TYPES),
+                budget=budget_enum,
+                max_tokens=max_tokens,
+                request_context=RequestContext(),
+            )
+
+            return search_result.model_dump()
+        except Exception as e:
+            logger.error(f"Error searching: {e}", exc_info=True)
+            return {"error": str(e), "results": []}
+
+    return mcp
+
+
+async def _initialize_and_run(bank_id: str):
+    """Initialize memory and run the MCP server."""
+    from hindsight_api import MemoryEngine
+
+    # Create and initialize memory engine with pg0 embedded database
+    # Note: We avoid printing to stderr during init as MCP clients show it as "errors"
+    memory = MemoryEngine(db_url="pg0://hindsight-mcp")
+    await memory.initialize()
+
+    # Create and run the server
+    mcp = create_local_mcp_server(bank_id, memory=memory)
+    await mcp.run_stdio_async()
+
+
+def main():
+    """Main entry point for the stdio MCP server."""
+    import asyncio
+
+    from hindsight_api.config import ENV_LLM_API_KEY, get_config
+
+    # Check for required environment variables
+    config = get_config()
+    if not config.llm_api_key:
+        print(f"Error: {ENV_LLM_API_KEY} environment variable is required", file=sys.stderr)
+        print("Set it in your MCP configuration or shell environment", file=sys.stderr)
+        sys.exit(1)
+
+    # Get bank ID from environment, default to "mcp"
+    bank_id = os.environ.get(ENV_MCP_LOCAL_BANK_ID, DEFAULT_MCP_LOCAL_BANK_ID)
+
+    # Note: We don't print to stderr as MCP clients display it as "error output"
+    # Use HINDSIGHT_API_LOG_LEVEL=debug for verbose startup logging
+
+    # Run the async initialization and server
+    asyncio.run(_initialize_and_run(bank_id))
+
+
+if __name__ == "__main__":
+    main()

hindsight-api/hindsight_api/metrics.py

@@ -6,16 +6,15 @@ This module provides metrics for:
 - Token usage (input/output) per operation
 - Per-bank granularity via labels
 """
+
 import logging
-from typing import Dict, Any, Optional
-from contextlib import contextmanager
 import time
+from contextlib import contextmanager
 
 from opentelemetry import metrics
+from opentelemetry.exporter.prometheus import PrometheusMetricReader
 from opentelemetry.sdk.metrics import MeterProvider
 from opentelemetry.sdk.resources import Resource
-from opentelemetry.exporter.prometheus import PrometheusMetricReader
-from prometheus_client import REGISTRY
 
 logger = logging.getLogger(__name__)
 
@@ -39,19 +38,18 @@ def initialize_metrics(service_name: str = "hindsight-api", service_version: str
     global _meter
 
     # Create resource with service information
-    resource = Resource.create({
-        "service.name": service_name,
-        "service.version": service_version,
-    })
+    resource = Resource.create(
+        {
+            "service.name": service_name,
+            "service.version": service_version,
+        }
+    )
 
     # Create Prometheus metric reader
     prometheus_reader = PrometheusMetricReader()
 
     # Create meter provider with Prometheus exporter
-    provider = MeterProvider(
-        resource=resource,
-        metric_readers=[prometheus_reader]
-    )
+    provider = MeterProvider(resource=resource, metric_readers=[prometheus_reader])
 
     # Set the global meter provider
     metrics.set_meter_provider(provider)
@@ -73,11 +71,19 @@ class MetricsCollectorBase:
     """Base class for metrics collectors."""
 
     @contextmanager
-    def record_operation(self, operation: str, bank_id: str, budget: Optional[str] = None, max_tokens: Optional[int] = None):
+    def record_operation(self, operation: str, bank_id: str, budget: str | None = None, max_tokens: int | None = None):
         """Context manager to record operation duration and status."""
         raise NotImplementedError
 
-    def record_tokens(self, operation: str, bank_id: str, input_tokens: int = 0, output_tokens: int = 0, budget: Optional[str] = None, max_tokens: Optional[int] = None):
+    def record_tokens(
+        self,
+        operation: str,
+        bank_id: str,
+        input_tokens: int = 0,
+        output_tokens: int = 0,
+        budget: str | None = None,
+        max_tokens: int | None = None,
+    ):
         """Record token usage for an operation."""
         raise NotImplementedError
 
@@ -86,11 +92,19 @@ class NoOpMetricsCollector(MetricsCollectorBase):
     """No-op metrics collector that does nothing. Used when metrics are disabled."""
 
     @contextmanager
-    def record_operation(self, operation: str, bank_id: str, budget: Optional[str] = None, max_tokens: Optional[int] = None):
+    def record_operation(self, operation: str, bank_id: str, budget: str | None = None, max_tokens: int | None = None):
         """No-op context manager."""
         yield
 
-    def record_tokens(self, operation: str, bank_id: str, input_tokens: int = 0, output_tokens: int = 0, budget: Optional[str] = None, max_tokens: Optional[int] = None):
+    def record_tokens(
+        self,
+        operation: str,
+        bank_id: str,
+        input_tokens: int = 0,
+        output_tokens: int = 0,
+        budget: str | None = None,
+        max_tokens: int | None = None,
+    ):
         """No-op token recording."""
         pass
 
@@ -108,33 +122,25 @@ class MetricsCollector(MetricsCollectorBase):
         # Operation latency histogram (in seconds)
         # Records duration of retain, recall, reflect operations
         self.operation_duration = self.meter.create_histogram(
-            name="hindsight.operation.duration",
-            description="Duration of Hindsight operations in seconds",
-            unit="s"
+            name="hindsight.operation.duration", description="Duration of Hindsight operations in seconds", unit="s"
         )
 
         # Token usage counters
         self.tokens_input = self.meter.create_counter(
-            name="hindsight.tokens.input",
-            description="Number of input tokens consumed",
-            unit="tokens"
+            name="hindsight.tokens.input", description="Number of input tokens consumed", unit="tokens"
         )
 
         self.tokens_output = self.meter.create_counter(
-            name="hindsight.tokens.output",
-            description="Number of output tokens generated",
-            unit="tokens"
+            name="hindsight.tokens.output", description="Number of output tokens generated", unit="tokens"
         )
 
         # Operation counter (success/failure)
         self.operation_total = self.meter.create_counter(
-            name="hindsight.operation.total",
-            description="Total number of operations executed",
-            unit="operations"
+            name="hindsight.operation.total", description="Total number of operations executed", unit="operations"
         )
 
     @contextmanager
-    def record_operation(self, operation: str, bank_id: str, budget: Optional[str] = None, max_tokens: Optional[int] = None):
+    def record_operation(self, operation: str, bank_id: str, budget: str | None = None, max_tokens: int | None = None):
         """
         Context manager to record operation duration and status.
 
@@ -175,7 +181,15 @@ class MetricsCollector(MetricsCollectorBase):
             # Record operation count
             self.operation_total.add(1, attributes)
 
-    def record_tokens(self, operation: str, bank_id: str, input_tokens: int = 0, output_tokens: int = 0, budget: Optional[str] = None, max_tokens: Optional[int] = None):
+    def record_tokens(
+        self,
+        operation: str,
+        bank_id: str,
+        input_tokens: int = 0,
+        output_tokens: int = 0,
+        budget: str | None = None,
+        max_tokens: int | None = None,
+    ):
         """
         Record token usage for an operation.
 

hindsight-api/hindsight_api/migrations.py

@@ -6,16 +6,19 @@ on application startup. It is designed to be safe for concurrent
 execution using PostgreSQL advisory locks to coordinate between
 distributed workers.
 
+Supports multi-tenant schema isolation: migrations can target a specific
+PostgreSQL schema, allowing each tenant to have isolated tables.
+
 Important: All migrations must be backward-compatible to allow
 safe rolling deployments.
 
 No alembic.ini required - all configuration is done programmatically.
 """
+
+import hashlib
 import logging
 import os
-import shutil
 from pathlib import Path
-from typing import Optional
 
 from alembic import command
 from alembic.config import Config
@@ -27,11 +30,29 @@ logger = logging.getLogger(__name__)
 MIGRATION_LOCK_ID = 123456789
 
 
-def _run_migrations_internal(database_url: str, script_location: str) -> None:
+def _get_schema_lock_id(schema: str) -> int:
+    """
+    Generate a unique advisory lock ID for a schema.
+
+    Uses hash of schema name to create a deterministic lock ID.
+    """
+    # Use hash to create a unique lock ID per schema
+    # Keep within PostgreSQL's bigint range
+    hash_bytes = hashlib.sha256(schema.encode()).digest()[:8]
+    return int.from_bytes(hash_bytes, byteorder="big") % (2**31)
+
+
+def _run_migrations_internal(database_url: str, script_location: str, schema: str | None = None) -> None:
     """
     Internal function to run migrations without locking.
+
+    Args:
+        database_url: SQLAlchemy database URL
+        script_location: Path to alembic scripts
+        schema: Target schema (None for default/public)
     """
-    logger.info(f"Running database migrations to head...")
+    schema_name = schema or "public"
+    logger.info(f"Running database migrations to head for schema '{schema_name}'...")
     logger.info(f"Database URL: {database_url}")
     logger.info(f"Script location: {script_location}")
 
@@ -51,13 +72,22 @@ def _run_migrations_internal(database_url: str, script_location: str) -> None:
     # Set path_separator to avoid deprecation warning
     alembic_cfg.set_main_option("path_separator", "os")
 
-    # Run migrations to head (latest version)
+    # If targeting a specific schema, pass it to env.py via config
+    # env.py will handle setting search_path and version_table_schema
+    if schema:
+        alembic_cfg.set_main_option("target_schema", schema)
+
+    # Run migrations
     command.upgrade(alembic_cfg, "head")
 
-    logger.info("Database migrations completed successfully")
+    logger.info(f"Database migrations completed successfully for schema '{schema_name}'")
 
 
-def run_migrations(database_url: str, script_location: Optional[str] = None) -> None:
+def run_migrations(
+    database_url: str,
+    script_location: str | None = None,
+    schema: str | None = None,
+) -> None:
     """
     Run database migrations to the latest version using programmatic Alembic configuration.
 
@@ -66,19 +96,28 @@ def run_migrations(database_url: str, script_location: Optional[str] = None) ->
     - Other workers wait for the lock, then verify migrations are complete
     - If schema is already up-to-date, this is a fast no-op
 
+    Supports multi-tenant schema isolation: when a schema is specified, migrations
+    run in that schema instead of public. This allows tenant extensions to provision
+    new tenant schemas with their own isolated tables.
+
     Args:
         database_url: SQLAlchemy database URL (e.g., "postgresql://user:pass@host/db")
         script_location: Path to alembic migrations directory (e.g., "/path/to/alembic").
                         If None, defaults to hindsight-api/alembic directory.
+        schema: Target PostgreSQL schema name. If None, uses default (public).
+                When specified, creates the schema if needed and runs migrations there.
 
     Raises:
         RuntimeError: If migrations fail to complete
         FileNotFoundError: If script_location doesn't exist
 
     Example:
-        # Using default location (hindsight_api package)
+        # Using default location and public schema
         run_migrations("postgresql://user:pass@host/db")
 
+        # Run migrations for a specific tenant schema
+        run_migrations("postgresql://user:pass@host/db", schema="tenant_acme")
+
         # Using custom location (when importing from another project)
         run_migrations(
             "postgresql://user:pass@host/db",
@@ -97,25 +136,28 @@ def run_migrations(database_url: str, script_location: Optional[str] = None) ->
         script_path = Path(script_location)
         if not script_path.exists():
             raise FileNotFoundError(
-                f"Alembic script location not found at {script_location}. "
-                "Database migrations cannot be run."
+                f"Alembic script location not found at {script_location}. Database migrations cannot be run."
             )
 
+        # Use schema-specific lock ID for multi-tenant isolation
+        lock_id = _get_schema_lock_id(schema) if schema else MIGRATION_LOCK_ID
+        schema_name = schema or "public"
+
         # Use PostgreSQL advisory lock to coordinate between distributed workers
         engine = create_engine(database_url)
         with engine.connect() as conn:
             # pg_advisory_lock blocks until the lock is acquired
             # The lock is automatically released when the connection closes
-            logger.debug(f"Acquiring migration advisory lock (id={MIGRATION_LOCK_ID})...")
-            conn.execute(text(f"SELECT pg_advisory_lock({MIGRATION_LOCK_ID})"))
+            logger.debug(f"Acquiring migration advisory lock for schema '{schema_name}' (id={lock_id})...")
+            conn.execute(text(f"SELECT pg_advisory_lock({lock_id})"))
             logger.debug("Migration advisory lock acquired")
 
             try:
                 # Run migrations while holding the lock
-                _run_migrations_internal(database_url, script_location)
+                _run_migrations_internal(database_url, script_location, schema=schema)
             finally:
                 # Explicitly release the lock (also released on connection close)
-                conn.execute(text(f"SELECT pg_advisory_unlock({MIGRATION_LOCK_ID})"))
+                conn.execute(text(f"SELECT pg_advisory_unlock({lock_id})"))
                 logger.debug("Migration advisory lock released")
 
     except FileNotFoundError:
@@ -130,7 +172,9 @@ def run_migrations(database_url: str, script_location: Optional[str] = None) ->
         raise RuntimeError("Database migration failed") from e
 
 
-def check_migration_status(database_url: Optional[str] = None, script_location: Optional[str] = None) -> tuple[str | None, str | None]:
+def check_migration_status(
+    database_url: str | None = None, script_location: str | None = None
+) -> tuple[str | None, str | None]:
     """
     Check current database schema version and latest available version.
 
@@ -151,7 +195,9 @@ def check_migration_status(database_url: Optional[str] = None, script_location:
         if database_url is None:
             database_url = os.getenv("HINDSIGHT_API_DATABASE_URL")
         if not database_url:
-            logger.warning("Database URL not provided and HINDSIGHT_API_DATABASE_URL not set, cannot check migration status")
+            logger.warning(
+                "Database URL not provided and HINDSIGHT_API_DATABASE_URL not set, cannot check migration status"
+            )
             return None, None
 
         # Get current revision from database

hindsight-api/hindsight_api/models.py

@@ -1,49 +1,65 @@
 """
 SQLAlchemy models for the memory system.
 """
-from datetime import datetime
-from typing import Optional
-from uuid import UUID as PyUUID, uuid4
 
+from dataclasses import dataclass
+from datetime import datetime
+from uuid import UUID as PyUUID
+
+
+@dataclass
+class RequestContext:
+    """
+    Context for request authentication and authorization.
+
+    This dataclass carries authentication data from HTTP requests to the
+    memory engine operations. It can be extended to include additional
+    context like headers, tokens, user info, etc.
+    """
+
+    api_key: str | None = None
+    api_key_id: str | None = None  # UUID of the API key used for authentication
+    tenant_id: str | None = None  # Tenant identifier (set by extension after auth)
+    internal: bool = False  # True for background/internal operations (not user-visible)
+
+
+from pgvector.sqlalchemy import Vector
 from sqlalchemy import (
     CheckConstraint,
-    Column,
     Float,
     ForeignKey,
     ForeignKeyConstraint,
     Index,
     Integer,
-    PrimaryKeyConstraint,
     Text,
     func,
+)
+from sqlalchemy import (
     text as sql_text,
 )
 from sqlalchemy.dialects.postgresql import JSONB, TIMESTAMP, UUID
 from sqlalchemy.ext.asyncio import AsyncAttrs
 from sqlalchemy.orm import DeclarativeBase, Mapped, mapped_column, relationship
-from pgvector.sqlalchemy import Vector
 
 
 class Base(AsyncAttrs, DeclarativeBase):
     """Base class for all models."""
+
     pass
 
 
 class Document(Base):
     """Source documents for memory units."""
+
     __tablename__ = "documents"
 
     id: Mapped[str] = mapped_column(Text, primary_key=True)
     bank_id: Mapped[str] = mapped_column(Text, primary_key=True)
-    original_text: Mapped[Optional[str]] = mapped_column(Text)
-    content_hash: Mapped[Optional[str]] = mapped_column(Text)
+    original_text: Mapped[str | None] = mapped_column(Text)
+    content_hash: Mapped[str | None] = mapped_column(Text)
     doc_metadata: Mapped[dict] = mapped_column("metadata", JSONB, server_default=sql_text("'{}'::jsonb"))
-    created_at: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
-    updated_at: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
+    created_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
+    updated_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
 
     # Relationships
     memory_units = relationship("MemoryUnit", back_populates="document", cascade="all, delete-orphan")
@@ -56,45 +72,42 @@ class Document(Base):
 
 class MemoryUnit(Base):
     """Individual sentence-level memories."""
+
     __tablename__ = "memory_units"
 
     id: Mapped[PyUUID] = mapped_column(
         UUID(as_uuid=True), primary_key=True, server_default=sql_text("gen_random_uuid()")
     )
     bank_id: Mapped[str] = mapped_column(Text, nullable=False)
-    document_id: Mapped[Optional[str]] = mapped_column(Text)
+    document_id: Mapped[str | None] = mapped_column(Text)
     text: Mapped[str] = mapped_column(Text, nullable=False)
     embedding = mapped_column(Vector(384))  # pgvector type
-    context: Mapped[Optional[str]] = mapped_column(Text)
-    event_date: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), nullable=False)  # Kept for backward compatibility
-    occurred_start: Mapped[Optional[datetime]] = mapped_column(TIMESTAMP(timezone=True))  # When fact occurred (range start)
-    occurred_end: Mapped[Optional[datetime]] = mapped_column(TIMESTAMP(timezone=True))    # When fact occurred (range end)
-    mentioned_at: Mapped[Optional[datetime]] = mapped_column(TIMESTAMP(timezone=True))    # When fact was mentioned
+    context: Mapped[str | None] = mapped_column(Text)
+    event_date: Mapped[datetime] = mapped_column(
+        TIMESTAMP(timezone=True), nullable=False
+    )  # Kept for backward compatibility
+    occurred_start: Mapped[datetime | None] = mapped_column(
+        TIMESTAMP(timezone=True)
+    )  # When fact occurred (range start)
+    occurred_end: Mapped[datetime | None] = mapped_column(TIMESTAMP(timezone=True))  # When fact occurred (range end)
+    mentioned_at: Mapped[datetime | None] = mapped_column(TIMESTAMP(timezone=True))  # When fact was mentioned
     fact_type: Mapped[str] = mapped_column(Text, nullable=False, server_default="world")
-    confidence_score: Mapped[Optional[float]] = mapped_column(Float)
+    confidence_score: Mapped[float | None] = mapped_column(Float)
     access_count: Mapped[int] = mapped_column(Integer, server_default="0")
-    unit_metadata: Mapped[dict] = mapped_column("metadata", JSONB, server_default=sql_text("'{}'::jsonb"))  # User-defined metadata (str->str)
-    created_at: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
-    updated_at: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
+    unit_metadata: Mapped[dict] = mapped_column(
+        "metadata", JSONB, server_default=sql_text("'{}'::jsonb")
+    )  # User-defined metadata (str->str)
+    created_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
+    updated_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
 
     # Relationships
     document = relationship("Document", back_populates="memory_units")
     unit_entities = relationship("UnitEntity", back_populates="memory_unit", cascade="all, delete-orphan")
     outgoing_links = relationship(
-        "MemoryLink",
-        foreign_keys="MemoryLink.from_unit_id",
-        back_populates="from_unit",
-        cascade="all, delete-orphan"
+        "MemoryLink", foreign_keys="MemoryLink.from_unit_id", back_populates="from_unit", cascade="all, delete-orphan"
     )
     incoming_links = relationship(
-        "MemoryLink",
-        foreign_keys="MemoryLink.to_unit_id",
-        back_populates="to_unit",
-        cascade="all, delete-orphan"
+        "MemoryLink", foreign_keys="MemoryLink.to_unit_id", back_populates="to_unit", cascade="all, delete-orphan"
     )
 
     __table_args__ = (
@@ -110,7 +123,7 @@ class MemoryUnit(Base):
             "(fact_type = 'opinion' AND confidence_score IS NOT NULL) OR "
             "(fact_type = 'observation') OR "
             "(fact_type NOT IN ('opinion', 'observation') AND confidence_score IS NULL)",
-            name="confidence_score_fact_type_check"
+            name="confidence_score_fact_type_check",
         ),
         Index("idx_memory_units_bank_id", "bank_id"),
         Index("idx_memory_units_document_id", "document_id"),
@@ -119,39 +132,46 @@ class MemoryUnit(Base):
         Index("idx_memory_units_access_count", "access_count", postgresql_ops={"access_count": "DESC"}),
         Index("idx_memory_units_fact_type", "fact_type"),
         Index("idx_memory_units_bank_fact_type", "bank_id", "fact_type"),
-        Index("idx_memory_units_bank_type_date", "bank_id", "fact_type", "event_date", postgresql_ops={"event_date": "DESC"}),
+        Index(
+            "idx_memory_units_bank_type_date",
+            "bank_id",
+            "fact_type",
+            "event_date",
+            postgresql_ops={"event_date": "DESC"},
+        ),
         Index(
             "idx_memory_units_opinion_confidence",
             "bank_id",
             "confidence_score",
             postgresql_where=sql_text("fact_type = 'opinion'"),
-            postgresql_ops={"confidence_score": "DESC"}
+            postgresql_ops={"confidence_score": "DESC"},
         ),
         Index(
             "idx_memory_units_opinion_date",
             "bank_id",
             "event_date",
             postgresql_where=sql_text("fact_type = 'opinion'"),
-            postgresql_ops={"event_date": "DESC"}
+            postgresql_ops={"event_date": "DESC"},
         ),
         Index(
             "idx_memory_units_observation_date",
             "bank_id",
             "event_date",
             postgresql_where=sql_text("fact_type = 'observation'"),
-            postgresql_ops={"event_date": "DESC"}
+            postgresql_ops={"event_date": "DESC"},
         ),
         Index(
             "idx_memory_units_embedding",
             "embedding",
             postgresql_using="hnsw",
-            postgresql_ops={"embedding": "vector_cosine_ops"}
+            postgresql_ops={"embedding": "vector_cosine_ops"},
         ),
     )
 
 
 class Entity(Base):
     """Resolved entities (people, organizations, locations, etc.)."""
+
     __tablename__ = "entities"
 
     id: Mapped[PyUUID] = mapped_column(
@@ -160,12 +180,8 @@ class Entity(Base):
     canonical_name: Mapped[str] = mapped_column(Text, nullable=False)
     bank_id: Mapped[str] = mapped_column(Text, nullable=False)
     entity_metadata: Mapped[dict] = mapped_column("metadata", JSONB, server_default=sql_text("'{}'::jsonb"))
-    first_seen: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
-    last_seen: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
+    first_seen: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
+    last_seen: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
     mention_count: Mapped[int] = mapped_column(Integer, server_default="1")
 
     # Relationships
@@ -175,13 +191,13 @@ class Entity(Base):
         "EntityCooccurrence",
         foreign_keys="EntityCooccurrence.entity_id_1",
         back_populates="entity_1",
-        cascade="all, delete-orphan"
+        cascade="all, delete-orphan",
     )
     cooccurrences_2 = relationship(
         "EntityCooccurrence",
         foreign_keys="EntityCooccurrence.entity_id_2",
         back_populates="entity_2",
-        cascade="all, delete-orphan"
+        cascade="all, delete-orphan",
     )
 
     __table_args__ = (
@@ -193,6 +209,7 @@ class Entity(Base):
 
 class UnitEntity(Base):
     """Association between memory units and entities."""
+
     __tablename__ = "unit_entities"
 
     unit_id: Mapped[PyUUID] = mapped_column(
@@ -214,6 +231,7 @@ class UnitEntity(Base):
 
 class EntityCooccurrence(Base):
     """Materialized cache of entity co-occurrences."""
+
     __tablename__ = "entity_cooccurrences"
 
     entity_id_1: Mapped[PyUUID] = mapped_column(
@@ -223,9 +241,7 @@ class EntityCooccurrence(Base):
         UUID(as_uuid=True), ForeignKey("entities.id", ondelete="CASCADE"), primary_key=True
     )
     cooccurrence_count: Mapped[int] = mapped_column(Integer, server_default="1")
-    last_cooccurred: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
+    last_cooccurred: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
 
     # Relationships
     entity_1 = relationship("Entity", foreign_keys=[entity_id_1], back_populates="cooccurrences_1")
@@ -241,6 +257,7 @@ class EntityCooccurrence(Base):
 
 class MemoryLink(Base):
     """Links between memory units (temporal, semantic, entity)."""
+
     __tablename__ = "memory_links"
 
     from_unit_id: Mapped[PyUUID] = mapped_column(
@@ -250,13 +267,11 @@ class MemoryLink(Base):
         UUID(as_uuid=True), ForeignKey("memory_units.id", ondelete="CASCADE"), primary_key=True
     )
     link_type: Mapped[str] = mapped_column(Text, primary_key=True)
-    entity_id: Mapped[Optional[PyUUID]] = mapped_column(
+    entity_id: Mapped[PyUUID | None] = mapped_column(
         UUID(as_uuid=True), ForeignKey("entities.id", ondelete="CASCADE"), primary_key=True
     )
     weight: Mapped[float] = mapped_column(Float, nullable=False, server_default="1.0")
-    created_at: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
+    created_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
 
     # Relationships
     from_unit = relationship("MemoryUnit", foreign_keys=[from_unit_id], back_populates="outgoing_links")
@@ -266,7 +281,7 @@ class MemoryLink(Base):
     __table_args__ = (
         CheckConstraint(
             "link_type IN ('temporal', 'semantic', 'entity', 'causes', 'caused_by', 'enables', 'prevents')",
-            name="memory_links_link_type_check"
+            name="memory_links_link_type_check",
         ),
         CheckConstraint("weight >= 0.0 AND weight <= 1.0", name="memory_links_weight_check"),
         Index("idx_memory_links_from", "from_unit_id"),
@@ -278,31 +293,22 @@ class MemoryLink(Base):
             "from_unit_id",
             "weight",
             postgresql_where=sql_text("weight >= 0.1"),
-            postgresql_ops={"weight": "DESC"}
+            postgresql_ops={"weight": "DESC"},
         ),
     )
 
 
 class Bank(Base):
     """Memory bank profiles with disposition traits and background."""
+
     __tablename__ = "banks"
 
     bank_id: Mapped[str] = mapped_column(Text, primary_key=True)
     disposition: Mapped[dict] = mapped_column(
-        JSONB,
-        nullable=False,
-        server_default=sql_text(
-            '\'{"skepticism": 3, "literalism": 3, "empathy": 3}\'::jsonb'
-        )
+        JSONB, nullable=False, server_default=sql_text('\'{"skepticism": 3, "literalism": 3, "empathy": 3}\'::jsonb')
     )
     background: Mapped[str] = mapped_column(Text, nullable=False, server_default="")
-    created_at: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
-    updated_at: Mapped[datetime] = mapped_column(
-        TIMESTAMP(timezone=True), server_default=func.now()
-    )
+    created_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
+    updated_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
 
-    __table_args__ = (
-        Index("idx_banks_bank_id", "bank_id"),
-    )
+    __table_args__ = (Index("idx_banks_bank_id", "bank_id"),)

hindsight-api/hindsight_api/pg0.py

@@ -1,12 +1,10 @@
 import asyncio
 import logging
-from typing import Optional
 
 from pg0 import Pg0
 
 logger = logging.getLogger(__name__)
 
-DEFAULT_PORT = 5555
 DEFAULT_USERNAME = "hindsight"
 DEFAULT_PASSWORD = "hindsight"
 DEFAULT_DATABASE = "hindsight"
@@ -17,34 +15,38 @@ class EmbeddedPostgres:
 
     def __init__(
         self,
-        port: int = DEFAULT_PORT,
+        port: int | None = None,
         username: str = DEFAULT_USERNAME,
         password: str = DEFAULT_PASSWORD,
         database: str = DEFAULT_DATABASE,
         name: str = "hindsight",
         **kwargs,
     ):
-        self.port = port
+        self.port = port  # None means pg0 will auto-assign
         self.username = username
         self.password = password
         self.database = database
         self.name = name
-        self._pg0: Optional[Pg0] = None
+        self._pg0: Pg0 | None = None
 
     def _get_pg0(self) -> Pg0:
         if self._pg0 is None:
-            self._pg0 = Pg0(
-                name=self.name,
-                port=self.port,
-                username=self.username,
-                password=self.password,
-                database=self.database,
-            )
+            kwargs = {
+                "name": self.name,
+                "username": self.username,
+                "password": self.password,
+                "database": self.database,
+            }
+            # Only set port if explicitly specified
+            if self.port is not None:
+                kwargs["port"] = self.port
+            self._pg0 = Pg0(**kwargs)  # type: ignore[invalid-argument-type] - dict kwargs
         return self._pg0
 
-    async def start(self, max_retries: int = 3, retry_delay: float = 2.0) -> str:
+    async def start(self, max_retries: int = 5, retry_delay: float = 4.0) -> str:
         """Start the PostgreSQL server with retry logic."""
-        logger.info(f"Starting embedded PostgreSQL (name: {self.name}, port: {self.port})...")
+        port_info = f"port={self.port}" if self.port else "port=auto"
+        logger.info(f"Starting embedded PostgreSQL (name={self.name}, {port_info})...")
 
         pg0 = self._get_pg0()
         last_error = None
@@ -53,9 +55,9 @@ class EmbeddedPostgres:
             try:
                 loop = asyncio.get_event_loop()
                 info = await loop.run_in_executor(None, pg0.start)
-                logger.info(f"PostgreSQL started on port {self.port}")
-                # Construct URI manually since pg0-embedded may return None
-                uri = info.uri if info and info.uri else f"postgresql://{self.username}:{self.password}@localhost:{self.port}/{self.database}"
+                # Get URI from pg0 (includes auto-assigned port)
+                uri = info.uri
+                logger.info(f"PostgreSQL started: {uri}")
                 return uri
             except Exception as e:
                 last_error = str(e)
@@ -68,8 +70,7 @@ class EmbeddedPostgres:
                     logger.debug(f"pg0 start attempt {attempt}/{max_retries} failed: {last_error}")
 
         raise RuntimeError(
-            f"Failed to start embedded PostgreSQL after {max_retries} attempts. "
-            f"Last error: {last_error}"
+            f"Failed to start embedded PostgreSQL after {max_retries} attempts. Last error: {last_error}"
         )
 
     async def stop(self) -> None:
@@ -91,9 +92,7 @@ class EmbeddedPostgres:
         pg0 = self._get_pg0()
         loop = asyncio.get_event_loop()
         info = await loop.run_in_executor(None, pg0.info)
-        # Construct URI manually since pg0-embedded may return None
-        uri = info.uri if info and info.uri else f"postgresql://{self.username}:{self.password}@localhost:{self.port}/{self.database}"
-        return uri
+        return info.uri
 
     async def is_running(self) -> bool:
         """Check if the PostgreSQL server is currently running."""
@@ -112,7 +111,7 @@ class EmbeddedPostgres:
         return await self.start()
 
 
-_default_instance: Optional[EmbeddedPostgres] = None
+_default_instance: EmbeddedPostgres | None = None
 
 
 def get_embedded_postgres() -> EmbeddedPostgres:

hindsight-api/hindsight_api/server.py

@@ -6,6 +6,7 @@ This module provides the ASGI app for uvicorn import string usage:
 
 For CLI usage, use the hindsight-api command instead.
 """
+
 import os
 import warnings
 
@@ -29,15 +30,11 @@ config.configure_logging()
 _memory = MemoryEngine()
 
 # Create unified app with both HTTP and optionally MCP
-app = create_app(
-    memory=_memory,
-    http_api_enabled=True,
-    mcp_api_enabled=config.mcp_enabled,
-    mcp_mount_path="/mcp"
-)
+app = create_app(memory=_memory, http_api_enabled=True, mcp_api_enabled=config.mcp_enabled, mcp_mount_path="/mcp")
 
 
 if __name__ == "__main__":
     # When run directly, delegate to the CLI
     from hindsight_api.main import main
+
     main()

hindsight-api/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "hatchling.build"
 
 [project]
 name = "hindsight-api"
-version = "0.1.5"
-description = "Temporal + Semantic + Entity Memory System for AI agents using PostgreSQL"
+version = "0.1.16"
+description = "Hindsight: Agent Memory That Works Like Human Memory"
 readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
@@ -25,11 +25,11 @@ dependencies = [
     "greenlet>=3.2.4",
     "psycopg2-binary>=2.9.11",
     "transformers>=4.30.0,<4.46.0",
-    "torch>=2.0.0,<2.6.0",
+    "torch>=2.0.0",
     "tiktoken>=0.12.0",
     "httpx>=0.27.0",
     "fastmcp>=2.3.0",
-    "pg0-embedded>=0.1.0",
+    "pg0-embedded>=0.11.0",
     "python-dateutil>=2.8.0",
     "opentelemetry-api>=1.20.0",
     "opentelemetry-sdk>=1.20.0",
@@ -37,6 +37,7 @@ dependencies = [
     "opentelemetry-exporter-prometheus>=0.41b0",
     "dateparser>=1.2.2",
     "google-genai>=1.0.0",
+    "anthropic>=0.40.0",
 ]
 
 [project.optional-dependencies]
@@ -50,6 +51,7 @@ test = [
 
 [project.scripts]
 hindsight-api = "hindsight_api.main:main"
+hindsight-local-mcp = "hindsight_api.mcp_local:main"
 
 [tool.hatch.build.targets.wheel]
 packages = ["hindsight_api"]
@@ -90,4 +92,59 @@ dev = [
     "pytest-xdist>=3.8.0",
     "python-dotenv>=1.2.1",
     "filelock>=3.0.0",
+    "ruff>=0.8.0",
+    "ty>=0.0.1",
 ]
+
+[tool.ruff]
+line-length = 120
+target-version = "py311"
+exclude = [
+    "tests/",
+    "**/tests/",
+]
+
+[tool.ruff.lint]
+select = [
+    "E",      # pycodestyle errors
+    "W",      # pycodestyle warnings
+    "F",      # Pyflakes
+    "I",      # isort
+]
+ignore = [
+    "E501",   # line too long (handled by formatter)
+    "E402",   # module import not at top of file
+    "F401",   # unused import (too noisy during development)
+    "F841",   # unused variable (too noisy during development)
+    "F811",   # redefined while unused
+    "F821",   # undefined name (forward references in type hints)
+]
+
+[tool.ruff.format]
+quote-style = "double"
+indent-style = "space"
+
+[tool.ty]
+# Type checking configuration
+# ty is an extremely fast Python type checker from Astral (same team as ruff/uv)
+
+[tool.ty.environment]
+python-version = "3.11"
+
+[tool.ty.src]
+exclude = [
+    "tests/",
+    "hindsight_api/alembic/",
+]
+
+[tool.ty.rules]
+# Disable noisy rules while keeping important ones
+invalid-argument-type = "ignore"    # False positives with **kwargs patterns
+invalid-return-type = "ignore"      # Often intentional in async code
+invalid-parameter-default = "ignore" # Optional params with None default
+possibly-missing-attribute = "ignore" # Common with Optional types
+invalid-raise = "ignore"            # False positives with exception tracking
+call-non-callable = "ignore"        # False positives with Optional types
+invalid-key = "ignore"              # Pydantic ConfigDict not understood
+invalid-method-override = "ignore"  # Intentional signature differences
+unresolved-reference = "ignore"     # Forward references not always resolved

hindsight-api/tests/conftest.py

@@ -8,7 +8,7 @@ import os
 import filelock
 from pathlib import Path
 from dotenv import load_dotenv
-from hindsight_api import MemoryEngine, LLMConfig, LocalSTEmbeddings
+from hindsight_api import MemoryEngine, LLMConfig, LocalSTEmbeddings, RequestContext
 
 from hindsight_api.engine.cross_encoder import LocalSTCrossEncoder
 from hindsight_api.engine.query_analyzer import DateparserQueryAnalyzer
@@ -99,6 +99,12 @@ def pg0_db_url(db_url, tmp_path_factory, worker_id):
     return url
 
 
+@pytest.fixture(scope="function")
+def request_context():
+    """Provide a default RequestContext for tests."""
+    return RequestContext()
+
+
 @pytest.fixture(scope="session")
 def llm_config():
     """

hindsight-api/tests/test_agents_api.py

@@ -3,7 +3,7 @@ Tests for agent management API (profile, disposition, background).
 """
 import pytest
 import uuid
-from hindsight_api import MemoryEngine
+from hindsight_api import MemoryEngine, RequestContext
 from hindsight_api.api import CreateBankRequest, DispositionTraits
 from hindsight_api.engine.memory_engine import Budget
 
@@ -17,11 +17,11 @@ class TestAgentProfile:
     """Tests for agent profile management."""
 
     @pytest.mark.asyncio
-    async def test_get_agent_profile_creates_default(self, memory: MemoryEngine):
+    async def test_get_agent_profile_creates_default(self, memory: MemoryEngine, request_context):
         """Test that getting a profile for a new agent creates default disposition."""
         bank_id = unique_agent_id("test_profile_default")
 
-        profile = await memory.get_bank_profile(bank_id)
+        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
 
         assert profile is not None
         assert "disposition" in profile
@@ -35,11 +35,11 @@ class TestAgentProfile:
         assert profile["background"] == ""
 
     @pytest.mark.asyncio
-    async def test_update_agent_disposition(self, memory: MemoryEngine):
+    async def test_update_agent_disposition(self, memory: MemoryEngine, request_context):
         """Test updating agent disposition traits."""
         bank_id = unique_agent_id("test_profile_update")
 
-        profile = await memory.get_bank_profile(bank_id)
+        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
         assert profile["disposition"].skepticism == 3
 
         new_disposition = {
@@ -47,26 +47,26 @@ class TestAgentProfile:
             "literalism": 4,
             "empathy": 2,
         }
-        await memory.update_bank_disposition(bank_id, new_disposition)
+        await memory.update_bank_disposition(bank_id, new_disposition, request_context=request_context)
 
-        updated_profile = await memory.get_bank_profile(bank_id)
+        updated_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
         disposition = updated_profile["disposition"]
         assert disposition.skepticism == new_disposition["skepticism"]
         assert disposition.literalism == new_disposition["literalism"]
         assert disposition.empathy == new_disposition["empathy"]
 
     @pytest.mark.asyncio
-    async def test_list_agents(self, memory: MemoryEngine):
+    async def test_list_agents(self, memory: MemoryEngine, request_context):
         """Test listing all agents."""
         agent_id_1 = unique_agent_id("test_list")
         agent_id_2 = unique_agent_id("test_list")
         agent_id_3 = unique_agent_id("test_list")
 
-        await memory.get_bank_profile(agent_id_1)
-        await memory.get_bank_profile(agent_id_2)
-        await memory.get_bank_profile(agent_id_3)
+        await memory.get_bank_profile(agent_id_1, request_context=request_context)
+        await memory.get_bank_profile(agent_id_2, request_context=request_context)
+        await memory.get_bank_profile(agent_id_3, request_context=request_context)
 
-        agents = await memory.list_banks()
+        agents = await memory.list_banks(request_context=request_context)
 
         agent_ids = [a["bank_id"] for a in agents]
         assert agent_id_1 in agent_ids
@@ -85,46 +85,50 @@ class TestAgentBackground:
     """Tests for agent background management."""
 
     @pytest.mark.asyncio
-    async def test_merge_agent_background(self, memory: MemoryEngine):
+    async def test_merge_agent_background(self, memory: MemoryEngine, request_context):
         """Test merging agent background information."""
         bank_id = unique_agent_id("test_profile_merge")
 
-        profile = await memory.get_bank_profile(bank_id)
+        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
         assert profile["background"] == ""
 
         result1 = await memory.merge_bank_background(
             bank_id,
             "I was born in Texas",
-            update_disposition=False
+            update_disposition=False,
+            request_context=request_context,
         )
         assert "Texas" in result1["background"]
 
         result2 = await memory.merge_bank_background(
             bank_id,
             "I have 10 years of startup experience",
-            update_disposition=False
+            update_disposition=False,
+            request_context=request_context,
         )
         assert "Texas" in result2["background"] or "startup" in result2["background"]
 
-        final_profile = await memory.get_bank_profile(bank_id)
+        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
         assert final_profile["background"] != ""
 
     @pytest.mark.asyncio
-    async def test_merge_background_handles_conflicts(self, memory: MemoryEngine):
+    async def test_merge_background_handles_conflicts(self, memory: MemoryEngine, request_context):
         """Test that merging background handles conflicts (new overwrites old)."""
         bank_id = unique_agent_id("test_profile_conflict")
 
         result1 = await memory.merge_bank_background(
             bank_id,
             "I was born in Colorado",
-            update_disposition=False
+            update_disposition=False,
+            request_context=request_context,
         )
         assert "Colorado" in result1["background"]
 
         result2 = await memory.merge_bank_background(
             bank_id,
             "You were born in Texas",
-            update_disposition=False
+            update_disposition=False,
+            request_context=request_context,
         )
         assert "Texas" in result2["background"]
 
@@ -133,7 +137,7 @@ class TestAgentEndpoint:
     """Tests for agent PUT endpoint logic."""
 
     @pytest.mark.asyncio
-    async def test_put_agent_create(self, memory: MemoryEngine):
+    async def test_put_agent_create(self, memory: MemoryEngine, request_context):
         """Test creating an agent via PUT endpoint."""
         bank_id = unique_agent_id("test_put_create")
 
@@ -146,12 +150,13 @@ class TestAgentEndpoint:
             background="I am a creative software engineer"
         )
 
-        profile = await memory.get_bank_profile(bank_id)
+        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
 
         if request.disposition is not None:
             await memory.update_bank_disposition(
                 bank_id,
-                request.disposition.model_dump()
+                request.disposition.model_dump(),
+                request_context=request_context,
             )
 
         if request.background is not None:
@@ -168,14 +173,14 @@ class TestAgentEndpoint:
                     request.background
                 )
 
-        final_profile = await memory.get_bank_profile(bank_id)
+        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
 
         assert final_profile["disposition"].skepticism == 4
         assert final_profile["disposition"].literalism == 5
         assert final_profile["background"] == "I am a creative software engineer"
 
     @pytest.mark.asyncio
-    async def test_put_agent_partial_update(self, memory: MemoryEngine):
+    async def test_put_agent_partial_update(self, memory: MemoryEngine, request_context):
         """Test updating only background."""
         bank_id = unique_agent_id("test_put_partial")
 
@@ -183,7 +188,7 @@ class TestAgentEndpoint:
             background="I am a data scientist"
         )
 
-        profile = await memory.get_bank_profile(bank_id)
+        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
 
         if request.background is not None:
             pool = await memory._get_pool()
@@ -199,7 +204,7 @@ class TestAgentEndpoint:
                     request.background
                 )
 
-        final_profile = await memory.get_bank_profile(bank_id)
+        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
 
         assert final_profile["disposition"].skepticism == 3  # Default
         assert final_profile["background"] == "I am a data scientist"
@@ -209,7 +214,7 @@ class TestAgentDispositionIntegration:
     """Tests for disposition integration with other features."""
 
     @pytest.mark.asyncio
-    async def test_think_uses_disposition(self, memory: MemoryEngine):
+    async def test_think_uses_disposition(self, memory: MemoryEngine, request_context):
         """Test that THINK operation uses agent disposition."""
         bank_id = unique_agent_id("test_think")
 
@@ -218,12 +223,13 @@ class TestAgentDispositionIntegration:
             "literalism": 4,  # High literalism
             "empathy": 2,     # Low empathy
         }
-        await memory.update_bank_disposition(bank_id, disposition)
+        await memory.update_bank_disposition(bank_id, disposition, request_context=request_context)
 
         await memory.merge_bank_background(
             bank_id,
             "I am a creative artist who values innovation over tradition",
-            update_disposition=False
+            update_disposition=False,
+            request_context=request_context,
         )
 
         await memory.retain_batch_async(
@@ -232,13 +238,14 @@ class TestAgentDispositionIntegration:
                 {"content": "Traditional painting techniques have been used for centuries"},
                 {"content": "Modern digital art is changing the art world"}
             ],
-            document_id="art_facts"
+            request_context=request_context,
         )
 
         result = await memory.reflect_async(
             bank_id=bank_id,
             query="What do you think about traditional vs modern art?",
-            budget=Budget.LOW
+            budget=Budget.LOW,
+            request_context=request_context,
         )
 
         assert result.text is not None

hindsight-api/tests/test_batch_chunking.py

@@ -6,7 +6,7 @@ import os
 
 
 @pytest.mark.asyncio
-async def test_large_batch_auto_chunks(memory):
+async def test_large_batch_auto_chunks(memory, request_context):
     bank_id = "test_chunking_agent"
     # Create a large batch that should trigger chunking
     # Each item is ~2000 chars, so 30 items = 60k chars (exceeds 50k threshold)
@@ -24,7 +24,8 @@ async def test_large_batch_auto_chunks(memory):
     # Ingest the large batch (should auto-chunk)
     result = await memory.retain_batch_async(
         bank_id=bank_id,
-        contents=contents
+        contents=contents,
+        request_context=request_context,
     )
 
     # Verify we got results back
@@ -33,7 +34,7 @@ async def test_large_batch_auto_chunks(memory):
 
 
 @pytest.mark.asyncio
-async def test_small_batch_no_chunking(memory):
+async def test_small_batch_no_chunking(memory, request_context):
     bank_id = "test_no_chunking_agent"
 
     # Create a small batch that should NOT trigger chunking
@@ -50,7 +51,8 @@ async def test_small_batch_no_chunking(memory):
     # Ingest the small batch (should NOT auto-chunk)
     result = await memory.retain_batch_async(
         bank_id=bank_id,
-        contents=contents
+        contents=contents,
+        request_context=request_context,
     )
 
     # Verify we got results back

hindsight-api/tests/test_combined_scoring.py

@@ -10,6 +10,7 @@ import pytest
 from datetime import datetime, timezone
 from hindsight_api.engine.search.types import RetrievalResult, MergedCandidate, ScoredResult
 from hindsight_api.engine.memory_engine import Budget
+from hindsight_api import RequestContext
 
 
 class TestRRFNormalization:
@@ -125,7 +126,7 @@ class TestCombinedScoringFormula:
 
 
 @pytest.mark.asyncio
-async def test_trace_has_normalized_rrf(memory):
+async def test_trace_has_normalized_rrf(memory, request_context):
     """Integration test: verify trace contains normalized RRF values, not raw."""
     bank_id = f"test_scoring_{datetime.now(timezone.utc).timestamp()}"
 
@@ -135,21 +136,25 @@ async def test_trace_has_normalized_rrf(memory):
             bank_id=bank_id,
             content="Python is a programming language created by Guido van Rossum",
             context="tech facts",
+            request_context=request_context,
         )
         await memory.retain_async(
             bank_id=bank_id,
             content="JavaScript was created by Brendan Eich at Netscape",
             context="tech facts",
+            request_context=request_context,
         )
         await memory.retain_async(
             bank_id=bank_id,
             content="The Eiffel Tower is located in Paris, France",
             context="geography facts",
+            request_context=request_context,
         )
         await memory.retain_async(
             bank_id=bank_id,
             content="Mount Everest is the tallest mountain on Earth",
             context="geography facts",
+            request_context=request_context,
         )
 
         # Search with tracing
@@ -160,6 +165,7 @@ async def test_trace_has_normalized_rrf(memory):
             budget=Budget.LOW,
             max_tokens=1024,
             enable_trace=True,
+            request_context=request_context,
         )
 
         assert result.trace is not None, "Trace should be present"
@@ -210,11 +216,11 @@ async def test_trace_has_normalized_rrf(memory):
             print(f"  - First result score components: {sc}")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_rrf_normalized_not_raw_in_trace(memory):
+async def test_rrf_normalized_not_raw_in_trace(memory, request_context):
     """Verify that raw RRF scores (0.04-0.06 range) don't appear as normalized values."""
     bank_id = f"test_rrf_raw_{datetime.now(timezone.utc).timestamp()}"
 
@@ -225,6 +231,7 @@ async def test_rrf_normalized_not_raw_in_trace(memory):
                 bank_id=bank_id,
                 content=f"Test fact number {i} about various topics",
                 context="test context",
+                request_context=request_context,
             )
 
         result = await memory.recall_async(
@@ -234,6 +241,7 @@ async def test_rrf_normalized_not_raw_in_trace(memory):
             budget=Budget.LOW,
             max_tokens=512,
             enable_trace=True,
+            request_context=request_context,
         )
 
         trace = result.trace
@@ -268,11 +276,11 @@ async def test_rrf_normalized_not_raw_in_trace(memory):
         print("\n✓ RRF raw vs normalized test passed!")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_combined_score_matches_components(memory):
+async def test_combined_score_matches_components(memory, request_context):
     """Verify the final score actually equals the weighted sum of components."""
     bank_id = f"test_combined_{datetime.now(timezone.utc).timestamp()}"
 
@@ -281,11 +289,13 @@ async def test_combined_score_matches_components(memory):
             bank_id=bank_id,
             content="The quick brown fox jumps over the lazy dog",
             context="test",
+            request_context=request_context,
         )
         await memory.retain_async(
             bank_id=bank_id,
             content="A quick test of the emergency broadcast system",
             context="test",
+            request_context=request_context,
         )
 
         result = await memory.recall_async(
@@ -295,6 +305,7 @@ async def test_combined_score_matches_components(memory):
             budget=Budget.LOW,
             max_tokens=512,
             enable_trace=True,
+            request_context=request_context,
         )
 
         trace = result.trace
@@ -320,4 +331,4 @@ async def test_combined_score_matches_components(memory):
         print("\n✓ Combined score verification test passed!")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)

hindsight-api/tests/test_document_tracking.py

@@ -4,10 +4,11 @@ Tests for document tracking and upsert functionality.
 import logging
 import pytest
 from datetime import datetime, timezone
+from hindsight_api import RequestContext
 
 
 @pytest.mark.asyncio
-async def test_document_creation_and_retrieval(memory):
+async def test_document_creation_and_retrieval(memory, request_context):
     """Test that documents are created and can be retrieved."""
     bank_id = f"test_doc_{datetime.now(timezone.utc).timestamp()}"
 
@@ -19,11 +20,12 @@ async def test_document_creation_and_retrieval(memory):
             bank_id=bank_id,
             content="Alice works at Google. Bob works at Microsoft.",
             context="Team meeting",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         # Retrieve document
-        doc = await memory.get_document(document_id, bank_id)
+        doc = await memory.get_document(document_id, bank_id, request_context=request_context)
 
         assert doc is not None
         assert doc["id"] == document_id
@@ -32,11 +34,11 @@ async def test_document_creation_and_retrieval(memory):
         assert doc["memory_unit_count"] > 0
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_document_upsert(memory):
+async def test_document_upsert(memory, request_context):
     """Test that providing the same document_id automatically upserts (deletes old units and creates new ones)."""
     bank_id = f"test_upsert_{datetime.now(timezone.utc).timestamp()}"
 
@@ -48,11 +50,12 @@ async def test_document_upsert(memory):
             bank_id=bank_id,
             content="Alice works at Google.",
             context="Initial",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         # Get document stats
-        doc_v1 = await memory.get_document(document_id, bank_id)
+        doc_v1 = await memory.get_document(document_id, bank_id, request_context=request_context)
         count_v1 = doc_v1["memory_unit_count"]
 
         # Update with different content (automatic upsert when same document_id is provided)
@@ -60,11 +63,12 @@ async def test_document_upsert(memory):
             bank_id=bank_id,
             content="Alice works at Microsoft. Bob works at Apple.",
             context="Updated",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         # Get updated document stats
-        doc_v2 = await memory.get_document(document_id, bank_id)
+        doc_v2 = await memory.get_document(document_id, bank_id, request_context=request_context)
         count_v2 = doc_v2["memory_unit_count"]
 
         # Verify old units were replaced
@@ -75,11 +79,11 @@ async def test_document_upsert(memory):
         assert set(units_v1).isdisjoint(set(units_v2))
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_document_deletion(memory):
+async def test_document_deletion(memory, request_context):
     """Test that deleting a document cascades to memory units."""
     bank_id = f"test_delete_{datetime.now(timezone.utc).timestamp()}"
 
@@ -91,29 +95,30 @@ async def test_document_deletion(memory):
             bank_id=bank_id,
             content="Alice works at Google.",
             context="Test",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         # Verify it exists
-        doc = await memory.get_document(document_id, bank_id)
+        doc = await memory.get_document(document_id, bank_id, request_context=request_context)
         assert doc is not None
         assert doc["memory_unit_count"] > 0
 
         # Delete document
-        result = await memory.delete_document(document_id, bank_id)
+        result = await memory.delete_document(document_id, bank_id, request_context=request_context)
         assert result["document_deleted"] == 1
         assert result["memory_units_deleted"] > 0
 
         # Verify it's gone
-        doc_after = await memory.get_document(document_id, bank_id)
+        doc_after = await memory.get_document(document_id, bank_id, request_context=request_context)
         assert doc_after is None
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_memory_without_document(memory):
+async def test_memory_without_document(memory, request_context):
     """Test that memories can still be created without document tracking."""
     bank_id = f"test_no_doc_{datetime.now(timezone.utc).timestamp()}"
 
@@ -122,10 +127,11 @@ async def test_memory_without_document(memory):
         units = await memory.retain_async(
             bank_id=bank_id,
             content="Alice works at Google.",
-            context="Test"
+            context="Test",
+            request_context=request_context,
         )
 
         assert len(units) > 0
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)

hindsight-api/tests/test_extensions.py

@@ -0,0 +1,796 @@
+"""Tests for the Hindsight extensions system."""
+
+from collections import defaultdict
+
+import pytest
+from fastapi import APIRouter
+from fastapi.testclient import TestClient
+
+from hindsight_api.extensions import (
+    ApiKeyTenantExtension,
+    AuthenticationError,
+    Extension,
+    HttpExtension,
+    OperationValidationError,
+    OperationValidatorExtension,
+    RecallContext,
+    RecallResult,
+    ReflectContext,
+    ReflectResultContext,
+    RequestContext,
+    RetainContext,
+    RetainResult,
+    TenantContext,
+    TenantExtension,
+    ValidationResult,
+    load_extension,
+)
+
+
+class TestExtensionLoader:
+    """Tests for extension loading and lifecycle."""
+
+    def test_load_extension_with_config(self, monkeypatch):
+        """Extension receives config from prefixed env vars and supports lifecycle."""
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TEST_EXTENSION",
+            "tests.test_extensions:LifecycleTestExtension",
+        )
+        monkeypatch.setenv("HINDSIGHT_API_TEST_API_URL", "https://example.com")
+        monkeypatch.setenv("HINDSIGHT_API_TEST_MAX_RETRIES", "5")
+
+        ext = load_extension("TEST", Extension)
+
+        assert ext is not None
+        assert ext.config["api_url"] == "https://example.com"
+        assert ext.config["max_retries"] == "5"
+
+    @pytest.mark.asyncio
+    async def test_extension_lifecycle(self, monkeypatch):
+        """Extension on_startup and on_shutdown are called."""
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TEST_EXTENSION",
+            "tests.test_extensions:LifecycleTestExtension",
+        )
+
+        ext = load_extension("TEST", Extension)
+
+        assert not ext.started
+        assert not ext.stopped
+
+        await ext.on_startup()
+        assert ext.started
+
+        await ext.on_shutdown()
+        assert ext.stopped
+
+
+class LifecycleTestExtension(Extension):
+    """Test extension for config and lifecycle tests."""
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.started = False
+        self.stopped = False
+
+    async def on_startup(self):
+        self.started = True
+
+    async def on_shutdown(self):
+        self.stopped = True
+
+
+class RateLimitingValidator(OperationValidatorExtension):
+    """
+    Mock validator that blocks after N attempts per bank_id.
+
+    Used for testing the extension integration with MemoryEngine.
+    """
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+        self.max_attempts = int(config.get("max_attempts", "2"))
+        self.retain_counts: dict[str, int] = defaultdict(int)
+        self.recall_counts: dict[str, int] = defaultdict(int)
+        self.reflect_counts: dict[str, int] = defaultdict(int)
+
+    async def validate_retain(self, ctx: RetainContext) -> ValidationResult:
+        self.retain_counts[ctx.bank_id] += 1
+        if self.retain_counts[ctx.bank_id] > self.max_attempts:
+            return ValidationResult.reject(
+                f"Retain limit exceeded for bank {ctx.bank_id}"
+            )
+        return ValidationResult.accept()
+
+    async def validate_recall(self, ctx: RecallContext) -> ValidationResult:
+        self.recall_counts[ctx.bank_id] += 1
+        if self.recall_counts[ctx.bank_id] > self.max_attempts:
+            return ValidationResult.reject(
+                f"Recall limit exceeded for bank {ctx.bank_id}"
+            )
+        return ValidationResult.accept()
+
+    async def validate_reflect(self, ctx: ReflectContext) -> ValidationResult:
+        self.reflect_counts[ctx.bank_id] += 1
+        if self.reflect_counts[ctx.bank_id] > self.max_attempts:
+            return ValidationResult.reject(
+                f"Reflect limit exceeded for bank {ctx.bank_id}"
+            )
+        return ValidationResult.accept()
+
+
+class TrackingValidator(OperationValidatorExtension):
+    """
+    Mock validator that tracks all pre and post hook calls with full parameters.
+
+    Used for testing that hooks receive all user-provided parameters.
+    """
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+        # Pre-hook tracking
+        self.pre_retain_calls: list[RetainContext] = []
+        self.pre_recall_calls: list[RecallContext] = []
+        self.pre_reflect_calls: list[ReflectContext] = []
+        # Post-hook tracking
+        self.post_retain_calls: list[RetainResult] = []
+        self.post_recall_calls: list[RecallResult] = []
+        self.post_reflect_calls: list[ReflectResultContext] = []
+
+    async def validate_retain(self, ctx: RetainContext) -> ValidationResult:
+        self.pre_retain_calls.append(ctx)
+        return ValidationResult.accept()
+
+    async def validate_recall(self, ctx: RecallContext) -> ValidationResult:
+        self.pre_recall_calls.append(ctx)
+        return ValidationResult.accept()
+
+    async def validate_reflect(self, ctx: ReflectContext) -> ValidationResult:
+        self.pre_reflect_calls.append(ctx)
+        return ValidationResult.accept()
+
+    async def on_retain_complete(self, result: RetainResult) -> None:
+        self.post_retain_calls.append(result)
+
+    async def on_recall_complete(self, result: RecallResult) -> None:
+        self.post_recall_calls.append(result)
+
+    async def on_reflect_complete(self, result: ReflectResultContext) -> None:
+        self.post_reflect_calls.append(result)
+
+
+class TestMemoryEngineValidation:
+    """Tests for validation integration with MemoryEngine.
+
+    The OperationValidatorExtension is integrated at the MemoryEngine level,
+    so all interfaces (HTTP API, MCP, SDK) get the same validation behavior.
+
+    For retain, the batch is validated as a whole (all or nothing) using
+    retain_batch_async which is the public method used by the HTTP API.
+    """
+
+    @pytest.mark.asyncio
+    async def test_retain_batch_validation(self, memory_with_validator):
+        """Retain batch is validated as a whole - accepts or rejects entire batch."""
+        memory = memory_with_validator
+        bank_id = "test-retain-batch"
+        ctx = RequestContext()
+
+        # First batch should succeed
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[
+                {"content": "First item"},
+                {"content": "Second item"},
+            ],
+            request_context=ctx,
+        )
+
+        # Second batch should succeed (2nd attempt)
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[{"content": "Third item"}],
+            request_context=ctx,
+        )
+
+        # Third batch should be blocked entirely (exceeds limit)
+        with pytest.raises(OperationValidationError) as exc_info:
+            await memory.retain_batch_async(
+                bank_id=bank_id,
+                contents=[
+                    {"content": "Should not be stored"},
+                    {"content": "Neither should this"},
+                ],
+                request_context=ctx,
+            )
+
+        assert "limit exceeded" in str(exc_info.value).lower()
+
+    @pytest.mark.asyncio
+    async def test_recall_validation(self, memory_with_validator):
+        """Recall is validated before execution."""
+        memory = memory_with_validator
+        bank_id = "test-recall-validation"
+        ctx = RequestContext()
+
+        # First recall should pass validation
+        await memory.recall_async(bank_id, "test query", fact_type=["world"], request_context=ctx)
+
+        # Second recall should pass validation
+        await memory.recall_async(bank_id, "another query", fact_type=["world"], request_context=ctx)
+
+        # Third recall should be blocked by validator
+        with pytest.raises(OperationValidationError) as exc_info:
+            await memory.recall_async(bank_id, "blocked query", fact_type=["world"], request_context=ctx)
+
+        assert "limit exceeded" in str(exc_info.value).lower()
+
+    @pytest.mark.asyncio
+    async def test_reflect_validation(self, memory_with_validator):
+        """Reflect is validated before execution."""
+        memory = memory_with_validator
+        bank_id = "test-reflect-validation"
+        ctx = RequestContext()
+
+        # First reflect should pass validation (may fail internally but validation passes)
+        try:
+            await memory.reflect_async(bank_id, "test question", request_context=ctx)
+        except OperationValidationError:
+            raise  # Re-raise validation errors
+        except Exception:
+            pass  # Other errors are fine (e.g., no data)
+
+        # Second reflect should pass validation
+        try:
+            await memory.reflect_async(bank_id, "another question", request_context=ctx)
+        except OperationValidationError:
+            raise
+        except Exception:
+            pass
+
+        # Third reflect should be blocked by validator
+        with pytest.raises(OperationValidationError) as exc_info:
+            await memory.reflect_async(bank_id, "blocked question", request_context=ctx)
+
+        assert "limit exceeded" in str(exc_info.value).lower()
+
+
+@pytest.fixture
+def memory_with_validator(memory):
+    """Memory engine with a rate-limiting validator (max 2 attempts per bank)."""
+    validator = RateLimitingValidator({"max_attempts": "2"})
+    memory._operation_validator = validator
+    return memory
+
+
+@pytest.fixture
+def memory_with_tracking_validator(memory):
+    """Memory engine with a tracking validator that records all hook calls."""
+    validator = TrackingValidator({})
+    memory._operation_validator = validator
+    return memory, validator
+
+
+class TestOperationHooksParameters:
+    """Tests for pre and post operation hooks receiving all user-provided parameters."""
+
+    @pytest.mark.asyncio
+    async def test_retain_pre_hook_receives_all_parameters(self, memory_with_tracking_validator):
+        """Pre-retain hook receives all user-provided parameters."""
+        memory, validator = memory_with_tracking_validator
+        bank_id = "test-retain-params"
+        ctx = RequestContext(api_key="test-key")
+        contents = [{"content": "Test content", "context": "test context"}]
+        document_id = "doc-123"
+
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=contents,
+            document_id=document_id,
+            fact_type_override="world",
+            confidence_score=0.9,
+            request_context=ctx,
+        )
+
+        assert len(validator.pre_retain_calls) == 1
+        pre_ctx = validator.pre_retain_calls[0]
+
+        # Verify all parameters are present
+        assert pre_ctx.bank_id == bank_id
+        # Note: contents is copied before document_id is applied to individual items
+        assert len(pre_ctx.contents) == len(contents)
+        assert pre_ctx.contents[0]["content"] == contents[0]["content"]
+        assert pre_ctx.document_id == document_id
+        assert pre_ctx.fact_type_override == "world"
+        assert pre_ctx.confidence_score == 0.9
+        assert pre_ctx.request_context == ctx
+
+    @pytest.mark.asyncio
+    async def test_retain_post_hook_receives_all_parameters_and_result(self, memory_with_tracking_validator):
+        """Post-retain hook receives all parameters plus the result."""
+        memory, validator = memory_with_tracking_validator
+        bank_id = "test-retain-post"
+        ctx = RequestContext(api_key="test-key")
+        contents = [{"content": "Test content for post hook"}]
+        document_id = "doc-456"
+
+        result = await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=contents,
+            document_id=document_id,
+            fact_type_override="experience",
+            confidence_score=0.8,
+            request_context=ctx,
+        )
+
+        assert len(validator.post_retain_calls) == 1
+        post_result = validator.post_retain_calls[0]
+
+        # Verify all parameters are present
+        assert post_result.bank_id == bank_id
+        assert post_result.document_id == document_id
+        assert post_result.fact_type_override == "experience"
+        assert post_result.confidence_score == 0.8
+        assert post_result.request_context == ctx
+
+        # Verify result data
+        assert post_result.success is True
+        assert post_result.error is None
+        assert post_result.unit_ids == result  # Should match the return value
+
+    @pytest.mark.asyncio
+    async def test_recall_pre_hook_receives_all_parameters(self, memory_with_tracking_validator):
+        """Pre-recall hook receives all user-provided parameters."""
+        from datetime import datetime, timezone
+        from hindsight_api.engine.memory_engine import Budget
+
+        memory, validator = memory_with_tracking_validator
+        bank_id = "test-recall-params"
+        ctx = RequestContext(api_key="test-key")
+        query = "test query"
+        question_date = datetime(2024, 1, 15, tzinfo=timezone.utc)
+
+        await memory.recall_async(
+            bank_id=bank_id,
+            query=query,
+            budget=Budget.HIGH,
+            max_tokens=2048,
+            enable_trace=True,
+            fact_type=["world", "experience"],
+            question_date=question_date,
+            include_entities=True,
+            max_entity_tokens=300,
+            include_chunks=True,
+            max_chunk_tokens=4096,
+            request_context=ctx,
+        )
+
+        assert len(validator.pre_recall_calls) == 1
+        pre_ctx = validator.pre_recall_calls[0]
+
+        # Verify all parameters are present
+        assert pre_ctx.bank_id == bank_id
+        assert pre_ctx.query == query
+        assert pre_ctx.budget == Budget.HIGH
+        assert pre_ctx.max_tokens == 2048
+        assert pre_ctx.enable_trace is True
+        assert pre_ctx.fact_types == ["world", "experience"]
+        assert pre_ctx.question_date == question_date
+        assert pre_ctx.include_entities is True
+        assert pre_ctx.max_entity_tokens == 300
+        assert pre_ctx.include_chunks is True
+        assert pre_ctx.max_chunk_tokens == 4096
+        assert pre_ctx.request_context == ctx
+
+    @pytest.mark.asyncio
+    async def test_recall_post_hook_receives_all_parameters_and_result(self, memory_with_tracking_validator):
+        """Post-recall hook receives all parameters plus the result."""
+        from hindsight_api.engine.memory_engine import Budget
+
+        memory, validator = memory_with_tracking_validator
+        bank_id = "test-recall-post"
+        ctx = RequestContext(api_key="test-key")
+
+        result = await memory.recall_async(
+            bank_id=bank_id,
+            query="test query for post",
+            budget=Budget.LOW,
+            max_tokens=1024,
+            fact_type=["world"],
+            request_context=ctx,
+        )
+
+        assert len(validator.post_recall_calls) == 1
+        post_result = validator.post_recall_calls[0]
+
+        # Verify all parameters are present
+        assert post_result.bank_id == bank_id
+        assert post_result.query == "test query for post"
+        assert post_result.budget == Budget.LOW
+        assert post_result.max_tokens == 1024
+        assert post_result.fact_types == ["world"]
+        assert post_result.request_context == ctx
+
+        # Verify result data
+        assert post_result.success is True
+        assert post_result.error is None
+        assert post_result.result == result  # Should match the return value
+
+    @pytest.mark.asyncio
+    async def test_reflect_pre_hook_receives_all_parameters(self, memory_with_tracking_validator):
+        """Pre-reflect hook receives all user-provided parameters."""
+        from hindsight_api.engine.memory_engine import Budget
+
+        memory, validator = memory_with_tracking_validator
+        bank_id = "test-reflect-params"
+        ctx = RequestContext(api_key="test-key")
+
+        try:
+            await memory.reflect_async(
+                bank_id=bank_id,
+                query="test question",
+                budget=Budget.MID,
+                context="additional context",
+                request_context=ctx,
+            )
+        except Exception:
+            pass  # May fail if no data, but pre-hook should still be called
+
+        assert len(validator.pre_reflect_calls) == 1
+        pre_ctx = validator.pre_reflect_calls[0]
+
+        # Verify all parameters are present
+        assert pre_ctx.bank_id == bank_id
+        assert pre_ctx.query == "test question"
+        assert pre_ctx.budget == Budget.MID
+        assert pre_ctx.context == "additional context"
+        assert pre_ctx.request_context == ctx
+
+    @pytest.mark.asyncio
+    async def test_reflect_post_hook_receives_all_parameters_and_result(self, memory_with_tracking_validator):
+        """Post-reflect hook receives all parameters plus the result on success."""
+        from hindsight_api.engine.memory_engine import Budget
+
+        memory, validator = memory_with_tracking_validator
+        bank_id = "test-reflect-post"
+        ctx = RequestContext(api_key="test-key")
+
+        # Store some content first so reflect has something to work with
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[{"content": "Alice is a software engineer at Google."}],
+            request_context=ctx,
+        )
+
+        result = await memory.reflect_async(
+            bank_id=bank_id,
+            query="What does Alice do?",
+            budget=Budget.LOW,
+            context="work context",
+            request_context=ctx,
+        )
+
+        assert len(validator.post_reflect_calls) == 1
+        post_result = validator.post_reflect_calls[0]
+
+        # Verify all parameters are present
+        assert post_result.bank_id == bank_id
+        assert post_result.query == "What does Alice do?"
+        assert post_result.budget == Budget.LOW
+        assert post_result.context == "work context"
+        assert post_result.request_context == ctx
+
+        # Verify result data
+        assert post_result.success is True
+        assert post_result.error is None
+        assert post_result.result == result  # Should match the return value
+        assert post_result.result.text is not None
+
+    @pytest.mark.asyncio
+    async def test_post_hooks_called_in_order_after_pre_hooks(self, memory_with_tracking_validator):
+        """Post hooks are called after pre hooks and after operation completes."""
+        memory, validator = memory_with_tracking_validator
+        bank_id = "test-hook-order"
+        ctx = RequestContext()
+
+        # Retain operation
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[{"content": "Test content"}],
+            request_context=ctx,
+        )
+
+        # Pre-hook should be called before post-hook
+        assert len(validator.pre_retain_calls) == 1
+        assert len(validator.post_retain_calls) == 1
+
+        # Recall operation
+        await memory.recall_async(
+            bank_id=bank_id,
+            query="test",
+            fact_type=["world"],
+            request_context=ctx,
+        )
+
+        assert len(validator.pre_recall_calls) == 1
+        assert len(validator.post_recall_calls) == 1
+
+
+class TestTenantExtension:
+    """Tests for TenantExtension and ApiKeyTenantExtension."""
+
+    @pytest.mark.asyncio
+    async def test_api_key_tenant_extension_valid_key(self):
+        """ApiKeyTenantExtension accepts valid API key."""
+        ext = ApiKeyTenantExtension({"api_key": "secret-key-123"})
+
+        result = await ext.authenticate(RequestContext(api_key="secret-key-123"))
+
+        assert result.schema_name == "public"
+
+    @pytest.mark.asyncio
+    async def test_api_key_tenant_extension_invalid_key(self):
+        """ApiKeyTenantExtension rejects invalid API key."""
+        ext = ApiKeyTenantExtension({"api_key": "secret-key-123"})
+
+        with pytest.raises(AuthenticationError) as exc_info:
+            await ext.authenticate(RequestContext(api_key="wrong-key"))
+
+        assert "Invalid API key" in str(exc_info.value)
+
+    @pytest.mark.asyncio
+    async def test_api_key_tenant_extension_missing_key(self):
+        """ApiKeyTenantExtension rejects missing API key."""
+        ext = ApiKeyTenantExtension({"api_key": "secret-key-123"})
+
+        with pytest.raises(AuthenticationError):
+            await ext.authenticate(RequestContext(api_key=None))
+
+    def test_api_key_tenant_extension_requires_config(self):
+        """ApiKeyTenantExtension requires api_key in config."""
+        with pytest.raises(ValueError) as exc_info:
+            ApiKeyTenantExtension({})
+
+        assert "HINDSIGHT_API_TENANT_API_KEY is required" in str(exc_info.value)
+
+
+class TestMemoryEngineTenantAuth:
+    """Tests for tenant authentication in MemoryEngine."""
+
+    @pytest.mark.asyncio
+    async def test_retain_requires_tenant_request_when_extension_configured(
+        self, memory_with_tenant
+    ):
+        """Retain fails without RequestContext when tenant extension is configured."""
+        memory = memory_with_tenant
+
+        with pytest.raises(AuthenticationError) as exc_info:
+            await memory.retain_batch_async(
+                bank_id="test-bank",
+                contents=[{"content": "test"}],
+                request_context=None,  # Missing!
+            )
+
+        assert "RequestContext is required" in str(exc_info.value)
+
+    @pytest.mark.asyncio
+    async def test_retain_succeeds_with_valid_tenant_request(self, memory_with_tenant):
+        """Retain succeeds with valid RequestContext."""
+        memory = memory_with_tenant
+
+        # Should not raise
+        await memory.retain_batch_async(
+            bank_id="test-bank-tenant",
+            contents=[{"content": "test content"}],
+            request_context=RequestContext(api_key="test-api-key"),
+        )
+
+    @pytest.mark.asyncio
+    async def test_retain_fails_with_invalid_api_key(self, memory_with_tenant):
+        """Retain fails with invalid API key."""
+        memory = memory_with_tenant
+
+        with pytest.raises(AuthenticationError) as exc_info:
+            await memory.retain_batch_async(
+                bank_id="test-bank",
+                contents=[{"content": "test"}],
+                request_context=RequestContext(api_key="wrong-key"),
+            )
+
+        assert "Invalid API key" in str(exc_info.value)
+
+    @pytest.mark.asyncio
+    async def test_recall_requires_tenant_request_when_extension_configured(
+        self, memory_with_tenant
+    ):
+        """Recall fails without RequestContext when tenant extension is configured."""
+        memory = memory_with_tenant
+
+        with pytest.raises(AuthenticationError):
+            await memory.recall_async(
+                bank_id="test-bank",
+                query="test query",
+                fact_type=["world"],
+                request_context=None,
+            )
+
+    @pytest.mark.asyncio
+    async def test_no_tenant_request_needed_without_extension(self, memory):
+        """Operations work with empty RequestContext when no tenant extension configured."""
+        # Should not raise - no tenant extension configured, just pass empty RequestContext
+        await memory.retain_batch_async(
+            bank_id="test-bank-no-tenant",
+            contents=[{"content": "test content"}],
+            request_context=RequestContext(),
+        )
+
+
+@pytest.fixture
+def memory_with_tenant(memory):
+    """Memory engine with a tenant extension (API key auth)."""
+    tenant_ext = ApiKeyTenantExtension({"api_key": "test-api-key"})
+    memory._tenant_extension = tenant_ext
+    return memory
+
+
+class SampleHttpExtension(HttpExtension):
+    """Sample HTTP extension for testing that provides custom endpoints."""
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+        self.started = False
+        self.stopped = False
+        self.request_count = 0
+
+    async def on_startup(self):
+        self.started = True
+
+    async def on_shutdown(self):
+        self.stopped = True
+
+    def get_router(self, memory) -> APIRouter:
+        router = APIRouter()
+
+        @router.get("/hello")
+        async def hello():
+            self.request_count += 1
+            return {"message": "Hello from extension!"}
+
+        @router.get("/config")
+        async def get_config():
+            return {"config": self.config}
+
+        @router.get("/health-check")
+        async def extension_health():
+            health = await memory.health_check()
+            return {"extension": "healthy", "memory": health}
+
+        @router.post("/echo")
+        async def echo(data: dict):
+            return {"echoed": data}
+
+        return router
+
+
+class TestHttpExtensionIntegration:
+    """Tests for HTTP extension integration."""
+
+    def test_load_http_extension(self, monkeypatch):
+        """HttpExtension can be loaded from environment variable."""
+        monkeypatch.setenv(
+            "HINDSIGHT_API_HTTP_EXTENSION",
+            "tests.test_extensions:SampleHttpExtension",
+        )
+        monkeypatch.setenv("HINDSIGHT_API_HTTP_CUSTOM_PARAM", "custom_value")
+
+        ext = load_extension("HTTP", HttpExtension)
+
+        assert ext is not None
+        assert isinstance(ext, SampleHttpExtension)
+        assert ext.config["custom_param"] == "custom_value"
+
+    def test_http_extension_router_mounted_at_ext(self, memory):
+        """HTTP extension router is mounted at /ext/."""
+        from hindsight_api.api.http import create_app
+
+        ext = SampleHttpExtension({"test_key": "test_value"})
+        app = create_app(memory, initialize_memory=False, http_extension=ext)
+
+        client = TestClient(app)
+
+        # Extension endpoint should be accessible at /ext/
+        response = client.get("/ext/hello")
+        assert response.status_code == 200
+        assert response.json() == {"message": "Hello from extension!"}
+
+        # Should track request count
+        assert ext.request_count == 1
+
+        # Old path should NOT work
+        response = client.get("/extension/hello")
+        assert response.status_code == 404
+
+    def test_http_extension_config_endpoint(self, memory):
+        """Extension can expose its config via custom endpoint."""
+        from hindsight_api.api.http import create_app
+
+        ext = SampleHttpExtension({"api_key": "secret", "limit": "100"})
+        app = create_app(memory, initialize_memory=False, http_extension=ext)
+
+        client = TestClient(app)
+
+        response = client.get("/ext/config")
+        assert response.status_code == 200
+        assert response.json()["config"]["api_key"] == "secret"
+        assert response.json()["config"]["limit"] == "100"
+
+    def test_http_extension_can_access_memory(self, memory):
+        """Extension endpoints can access memory engine."""
+        from hindsight_api.api.http import create_app
+
+        ext = SampleHttpExtension({})
+        app = create_app(memory, initialize_memory=False, http_extension=ext)
+
+        client = TestClient(app)
+
+        response = client.get("/ext/health-check")
+        assert response.status_code == 200
+        data = response.json()
+        assert data["extension"] == "healthy"
+        assert "memory" in data
+
+    def test_http_extension_post_endpoint(self, memory):
+        """Extension can handle POST requests with JSON body."""
+        from hindsight_api.api.http import create_app
+
+        ext = SampleHttpExtension({})
+        app = create_app(memory, initialize_memory=False, http_extension=ext)
+
+        client = TestClient(app)
+
+        response = client.post("/ext/echo", json={"key": "value", "number": 42})
+        assert response.status_code == 200
+        assert response.json() == {"echoed": {"key": "value", "number": 42}}
+
+    def test_http_extension_not_mounted_when_none(self, memory):
+        """No extension routes when http_extension is None."""
+        from hindsight_api.api.http import create_app
+
+        app = create_app(memory, initialize_memory=False, http_extension=None)
+
+        client = TestClient(app)
+
+        # Extension endpoint should not exist
+        response = client.get("/ext/hello")
+        assert response.status_code == 404
+
+    @pytest.mark.asyncio
+    async def test_http_extension_lifecycle(self):
+        """HTTP extension on_startup and on_shutdown are called."""
+        ext = SampleHttpExtension({})
+
+        assert not ext.started
+        assert not ext.stopped
+
+        await ext.on_startup()
+        assert ext.started
+
+        await ext.on_shutdown()
+        assert ext.stopped
+
+    def test_core_routes_still_work_with_extension(self, memory):
+        """Core API routes still work when extension is mounted."""
+        from hindsight_api.api.http import create_app
+
+        ext = SampleHttpExtension({})
+        app = create_app(memory, initialize_memory=False, http_extension=ext)
+
+        client = TestClient(app)
+
+        # Health endpoint should work
+        response = client.get("/health")
+        assert response.status_code in (200, 503)  # May be unhealthy if DB not connected
+
+        # Banks list endpoint should work
+        response = client.get("/v1/default/banks")
+        assert response.status_code in (200, 500)  # May fail if DB not ready

hindsight-api/tests/test_fact_extraction_quality.py

@@ -12,12 +12,12 @@ This comprehensive test suite validates that the fact extraction system:
 These are quality/accuracy tests that verify the LLM-based extraction
 produces semantically correct and complete facts.
 """
-import pytest
-import re
-from datetime import datetime, timezone
-from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text
-from hindsight_api import LLMConfig
+from datetime import UTC, datetime
 
+import pytest
+
+from hindsight_api import LLMConfig
+from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text
 
 # =============================================================================
 # DIMENSION PRESERVATION TESTS
@@ -432,6 +432,7 @@ with a concert surrounded by music, joy and the warm summer breeze.
         assert birthday_fact is not None, "Should extract fact about birthday celebration"
 
         fact_date_str = birthday_fact.occurred_start
+        assert fact_date_str is not None, "occurred_start should not be None for temporal events"
 
         if 'T' in fact_date_str:
             fact_date = datetime.fromisoformat(fact_date_str.replace('Z', '+00:00'))
@@ -497,7 +498,7 @@ Yesterday I went for a morning jog for the first time in a nearby park.
     async def test_extract_facts_with_relative_dates(self):
         """Test that relative dates are converted to absolute dates."""
 
-        reference_date = datetime(2024, 3, 20, 14, 0, 0, tzinfo=timezone.utc)
+        reference_date = datetime(2024, 3, 20, 14, 0, 0, tzinfo=UTC)
         llm_config = LLMConfig.for_memory()
 
         text = """
@@ -531,7 +532,7 @@ Yesterday I went for a morning jog for the first time in a nearby park.
     async def test_extract_facts_with_no_temporal_info(self):
         """Test that facts without temporal info are still extracted."""
 
-        reference_date = datetime(2024, 3, 20, 14, 0, 0, tzinfo=timezone.utc)
+        reference_date = datetime(2024, 3, 20, 14, 0, 0, tzinfo=UTC)
         llm_config = LLMConfig.for_memory()
 
         text = "Alice works at Google. She loves Python programming."
@@ -555,7 +556,7 @@ Yesterday I went for a morning jog for the first time in a nearby park.
     async def test_extract_facts_with_absolute_dates(self):
         """Test that absolute dates in text are preserved."""
 
-        reference_date = datetime(2024, 3, 20, 14, 0, 0, tzinfo=timezone.utc)
+        reference_date = datetime(2024, 3, 20, 14, 0, 0, tzinfo=UTC)
         llm_config = LLMConfig.for_memory()
 
         text = """
@@ -896,7 +897,7 @@ class TestDispositionInference:
     """Tests for LLM-based disposition trait inference from background."""
 
     @pytest.mark.asyncio
-    async def test_background_merge_with_disposition_inference(self, memory):
+    async def test_background_merge_with_disposition_inference(self, memory, request_context):
         """Test that background merge infers disposition traits by default."""
         import uuid
         bank_id = f"test_infer_{uuid.uuid4().hex[:8]}"
@@ -904,7 +905,8 @@ class TestDispositionInference:
         result = await memory.merge_bank_background(
             bank_id,
             "I am a creative software engineer who loves innovation and trying new technologies",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
 
         assert "background" in result
@@ -922,30 +924,31 @@ class TestDispositionInference:
             assert 1 <= disposition[trait] <= 5
 
     @pytest.mark.asyncio
-    async def test_background_merge_without_disposition_inference(self, memory):
+    async def test_background_merge_without_disposition_inference(self, memory, request_context):
         """Test that background merge skips disposition inference when disabled."""
         import uuid
         bank_id = f"test_no_infer_{uuid.uuid4().hex[:8]}"
 
-        initial_profile = await memory.get_bank_profile(bank_id)
+        initial_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
         initial_disposition = initial_profile["disposition"]
 
         result = await memory.merge_bank_background(
             bank_id,
             "I am a data scientist",
-            update_disposition=False
+            update_disposition=False,
+            request_context=request_context,
         )
 
         assert "background" in result
         assert "disposition" not in result
 
-        final_profile = await memory.get_bank_profile(bank_id)
+        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
         final_disposition = final_profile["disposition"]
 
         assert initial_disposition == final_disposition
 
     @pytest.mark.asyncio
-    async def test_disposition_inference_for_lawyer(self, memory):
+    async def test_disposition_inference_for_lawyer(self, memory, request_context):
         """Test disposition inference for lawyer profile (high skepticism, high literalism)."""
         import uuid
         bank_id = f"test_lawyer_{uuid.uuid4().hex[:8]}"
@@ -953,7 +956,8 @@ class TestDispositionInference:
         result = await memory.merge_bank_background(
             bank_id,
             "I am a lawyer who focuses on contract details and never takes claims at face value",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
 
         disposition = result["disposition"]
@@ -963,7 +967,7 @@ class TestDispositionInference:
         assert disposition["literalism"] >= 3
 
     @pytest.mark.asyncio
-    async def test_disposition_inference_for_therapist(self, memory):
+    async def test_disposition_inference_for_therapist(self, memory, request_context):
         """Test disposition inference for therapist profile (high empathy)."""
         import uuid
         bank_id = f"test_therapist_{uuid.uuid4().hex[:8]}"
@@ -971,7 +975,8 @@ class TestDispositionInference:
         result = await memory.merge_bank_background(
             bank_id,
             "I am a therapist who deeply understands and connects with people's emotional struggles",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
 
         disposition = result["disposition"]
@@ -980,7 +985,7 @@ class TestDispositionInference:
         assert disposition["empathy"] >= 3
 
     @pytest.mark.asyncio
-    async def test_disposition_updates_in_database(self, memory):
+    async def test_disposition_updates_in_database(self, memory, request_context):
         """Test that inferred disposition is actually stored in database."""
         import uuid
         bank_id = f"test_db_update_{uuid.uuid4().hex[:8]}"
@@ -988,12 +993,13 @@ class TestDispositionInference:
         result = await memory.merge_bank_background(
             bank_id,
             "I am an innovative designer",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
 
         inferred_disposition = result["disposition"]
 
-        profile = await memory.get_bank_profile(bank_id)
+        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
         db_disposition = profile["disposition"]
 
         # Compare values (db_disposition is a Pydantic model)
@@ -1002,7 +1008,7 @@ class TestDispositionInference:
         assert db_disposition.empathy == inferred_disposition["empathy"]
 
     @pytest.mark.asyncio
-    async def test_multiple_background_merges_update_disposition(self, memory):
+    async def test_multiple_background_merges_update_disposition(self, memory, request_context):
         """Test that each background merge can update disposition."""
         import uuid
         bank_id = f"test_multi_merge_{uuid.uuid4().hex[:8]}"
@@ -1010,14 +1016,16 @@ class TestDispositionInference:
         result1 = await memory.merge_bank_background(
             bank_id,
             "I am a software engineer",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
         disposition1 = result1["disposition"]
 
         result2 = await memory.merge_bank_background(
             bank_id,
             "I love creative problem solving and innovation",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
         disposition2 = result2["disposition"]
 
@@ -1025,7 +1033,7 @@ class TestDispositionInference:
         assert "creative" in result2["background"].lower() or "innovation" in result2["background"].lower()
 
     @pytest.mark.asyncio
-    async def test_background_merge_conflict_resolution_with_disposition(self, memory):
+    async def test_background_merge_conflict_resolution_with_disposition(self, memory, request_context):
         """Test that conflicts are resolved and disposition reflects final background."""
         import uuid
         bank_id = f"test_conflict_{uuid.uuid4().hex[:8]}"
@@ -1033,13 +1041,15 @@ class TestDispositionInference:
         await memory.merge_bank_background(
             bank_id,
             "I was born in Colorado and prefer stability",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
 
         result = await memory.merge_bank_background(
             bank_id,
             "You were born in Texas and are very skeptical of people",
-            update_disposition=True
+            update_disposition=True,
+            request_context=request_context,
         )
 
         background = result["background"]
@@ -1047,4 +1057,4 @@ class TestDispositionInference:
 
         assert "texas" in background.lower()
         # Higher skepticism expected from "very skeptical of people"
-        assert disposition["skepticism"] >= 3
+        assert disposition["skepticism"] >= 3

hindsight-api/tests/test_fact_ordering.py

@@ -7,24 +7,24 @@ distinguish between things said earlier vs later.
 """
 import pytest
 from datetime import datetime, timezone
-from hindsight_api import MemoryEngine
+from hindsight_api import MemoryEngine, RequestContext
 from hindsight_api.engine.memory_engine import Budget
 import os
 
 
 @pytest.mark.asyncio
-async def test_fact_ordering_within_conversation(memory):
+async def test_fact_ordering_within_conversation(memory, request_context):
     bank_id = "test_ordering_agent"
 
     # Get/create agent (auto-creates with defaults)
-    await memory.get_bank_profile(bank_id)
+    await memory.get_bank_profile(bank_id, request_context=request_context)
 
     # Update disposition to match Marcus
     await memory.update_bank_disposition(bank_id, {
         "skepticism": 3,
         "literalism": 3,
         "empathy": 3
-    })
+    }, request_context=request_context)
 
     # A conversation where Marcus changes his position
     conversation = """
@@ -43,7 +43,8 @@ Marcus: Yeah, I realized I was being too optimistic about their defense.
         content=conversation,
         context="podcast discussion about NFL game",
         event_date=base_event_date,
-        document_id="test_conv_1"
+        document_id="test_conv_1",
+        request_context=request_context,
     )
 
     # Search for all facts about Marcus's predictions
@@ -52,7 +53,8 @@ Marcus: Yeah, I realized I was being too optimistic about their defense.
         query="Marcus prediction Rams",
         fact_type=['opinion', 'experience', 'world'],
         budget=Budget.LOW,
-        max_tokens=8192
+        max_tokens=8192,
+        request_context=request_context,
     )
 
     print(f"\n=== Retrieved {len(results.results)} facts ===")
@@ -113,17 +115,17 @@ Marcus: Yeah, I realized I was being too optimistic about their defense.
         print(f"\n✅ Temporal ordering preserved: First prediction came before changed prediction")
 
     # Cleanup
-    await memory.delete_bank(bank_id)
+    await memory.delete_bank(bank_id, request_context=request_context)
 
     print(f"\n✅ Test passed: Fact ordering within conversation is preserved")
 
 
 @pytest.mark.asyncio
-async def test_multiple_documents_ordering(memory):
+async def test_multiple_documents_ordering(memory, request_context):
 
     bank_id = "test_multi_doc_agent"
 
-    await memory.get_bank_profile(bank_id)  # Auto-creates with defaults
+    await memory.get_bank_profile(bank_id, request_context=request_context)  # Auto-creates with defaults
 
     # Two separate conversations with same base time
     base_time = datetime(2024, 11, 14, 10, 0, 0, tzinfo=timezone.utc)
@@ -146,7 +148,8 @@ Alice: I reconsidered the team's experience level.
         contents=[
             {"content": conv1, "context": "project discussion 1", "event_date": base_time},
             {"content": conv2, "context": "project discussion 2", "event_date": base_time}
-        ]
+        ],
+        request_context=request_context,
     )
 
     # Search for Alice's preferences
@@ -155,7 +158,8 @@ Alice: I reconsidered the team's experience level.
         query="Alice preference React Vue",
         fact_type=['opinion', 'experience'],
         budget=Budget.LOW,
-        max_tokens=8192
+        max_tokens=8192,
+        request_context=request_context,
     )
 
     print(f"\n=== Retrieved {len(results.results)} agent facts ===")
@@ -175,6 +179,6 @@ Alice: I reconsidered the team's experience level.
         print(f"\n✅ Facts from {len(agent_facts)} statements have {len(unique_timestamps)} unique timestamps")
 
     # Cleanup
-    await memory.delete_bank(bank_id)
+    await memory.delete_bank(bank_id, request_context=request_context)
 
     print(f"\n✅ Test passed: Multiple documents maintain separate ordering")

hindsight-api/tests/test_http_api_integration.py

@@ -426,3 +426,349 @@ async def test_document_deletion(api_client):
         f"/v1/default/banks/{test_bank_id}/documents/sales-report-q1-2024"
     )
     assert response.status_code == 404
+
+
+@pytest.mark.asyncio
+async def test_async_retain(api_client):
+    """Test asynchronous retain functionality.
+
+    When async=true is passed, the retain endpoint should:
+    1. Return immediately with success and async_=true
+    2. Process the content in the background
+    3. Eventually store the memories
+    """
+    import asyncio
+
+    test_bank_id = f"async_retain_test_{datetime.now().timestamp()}"
+
+    # Store memory with async=true
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "async": True,
+            "items": [
+                {
+                    "content": "Alice is a senior engineer at TechCorp. She has been working on the authentication system for 5 years.",
+                    "context": "team introduction"
+                }
+            ]
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+    assert result["success"] is True
+    assert result["async"] is True, "Response should indicate async processing"
+    assert result["items_count"] == 1
+
+    # Check operations endpoint to see the pending operation
+    response = await api_client.get(f"/v1/default/banks/{test_bank_id}/operations")
+    assert response.status_code == 200
+    ops_result = response.json()
+    assert "operations" in ops_result
+
+    # Wait for async processing to complete (poll with timeout)
+    max_wait_seconds = 30
+    poll_interval = 0.5
+    elapsed = 0
+    memories_found = False
+
+    while elapsed < max_wait_seconds:
+        # Check if memories are stored
+        response = await api_client.get(
+            f"/v1/default/banks/{test_bank_id}/memories/list",
+            params={"limit": 10}
+        )
+        assert response.status_code == 200
+        items = response.json()["items"]
+
+        if len(items) > 0:
+            memories_found = True
+            break
+
+        await asyncio.sleep(poll_interval)
+        elapsed += poll_interval
+
+    assert memories_found, f"Async retain did not complete within {max_wait_seconds} seconds"
+
+    # Verify we can recall the stored memory
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories/recall",
+        json={
+            "query": "Who works at TechCorp?",
+            "thinking_budget": 30
+        }
+    )
+    assert response.status_code == 200
+    search_results = response.json()
+    assert len(search_results["results"]) > 0, "Should find the asynchronously stored memory"
+
+    # Verify Alice is mentioned
+    found_alice = any("Alice" in r["text"] for r in search_results["results"])
+    assert found_alice, "Should find Alice in search results"
+
+
+@pytest.mark.asyncio
+async def test_async_retain_parallel(api_client):
+    """Test multiple async retain operations running in parallel.
+
+    Verifies that:
+    1. Multiple async operations can be submitted concurrently
+    2. All operations complete successfully
+    3. The exact number of documents are processed
+    """
+    import asyncio
+
+    test_bank_id = f"async_parallel_test_{datetime.now().timestamp()}"
+    num_documents = 5
+
+    # Prepare multiple documents to retain
+    documents = [
+        {
+            "content": f"Document {i}: This is test content about Person{i} who works at Company{i}.",
+            "context": f"test document {i}",
+            "document_id": f"doc_{i}"
+        }
+        for i in range(num_documents)
+    ]
+
+    # Submit all async retain operations in parallel
+    async def submit_async_retain(doc):
+        return await api_client.post(
+            f"/v1/default/banks/{test_bank_id}/memories",
+            json={
+                "async": True,
+                "items": [doc]
+            }
+        )
+
+    # Run all submissions concurrently
+    responses = await asyncio.gather(*[submit_async_retain(doc) for doc in documents])
+
+    # Verify all submissions succeeded
+    for i, response in enumerate(responses):
+        assert response.status_code == 200, f"Document {i} submission failed"
+        result = response.json()
+        assert result["success"] is True
+        assert result["async"] is True
+
+    # Check operations endpoint - should show pending operations
+    response = await api_client.get(f"/v1/default/banks/{test_bank_id}/operations")
+    assert response.status_code == 200
+
+    # Wait for all async operations to complete (poll with timeout)
+    max_wait_seconds = 60
+    poll_interval = 1.0
+    elapsed = 0
+    all_docs_processed = False
+
+    while elapsed < max_wait_seconds:
+        # Check document count
+        response = await api_client.get(f"/v1/default/banks/{test_bank_id}/documents")
+        assert response.status_code == 200
+        docs = response.json()["items"]
+
+        if len(docs) >= num_documents:
+            all_docs_processed = True
+            break
+
+        await asyncio.sleep(poll_interval)
+        elapsed += poll_interval
+
+    assert all_docs_processed, f"Expected {num_documents} documents, but only {len(docs)} were processed within {max_wait_seconds} seconds"
+
+    # Verify exact document count
+    response = await api_client.get(f"/v1/default/banks/{test_bank_id}/documents")
+    assert response.status_code == 200
+    final_docs = response.json()["items"]
+    assert len(final_docs) == num_documents, f"Expected exactly {num_documents} documents, got {len(final_docs)}"
+
+    # Verify each document exists
+    doc_ids = {doc["id"] for doc in final_docs}
+    for i in range(num_documents):
+        assert f"doc_{i}" in doc_ids, f"Document doc_{i} not found"
+
+    # Verify memories were created for all documents
+    response = await api_client.get(
+        f"/v1/default/banks/{test_bank_id}/memories/list",
+        params={"limit": 100}
+    )
+    assert response.status_code == 200
+    memories = response.json()["items"]
+    assert len(memories) >= num_documents, f"Expected at least {num_documents} memories, got {len(memories)}"
+
+    # Verify we can recall content from different documents
+    for i in [0, num_documents - 1]:  # Check first and last
+        response = await api_client.post(
+            f"/v1/default/banks/{test_bank_id}/memories/recall",
+            json={
+                "query": f"Who works at Company{i}?",
+                "thinking_budget": 30
+            }
+        )
+        assert response.status_code == 200
+        results = response.json()["results"]
+        assert len(results) > 0, f"Should find memories for document {i}"
+
+
+@pytest.mark.asyncio
+async def test_reflect_structured_output(api_client):
+    """Test reflect endpoint with structured output via response_schema.
+
+    When response_schema is provided, the reflect endpoint should return
+    both the natural language text response and a structured_output field
+    containing the response parsed according to the provided JSON schema.
+    """
+    test_bank_id = f"reflect_structured_test_{datetime.now().timestamp()}"
+
+    # Store some memories to reflect on
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "Alice is a senior machine learning engineer with 8 years of experience.",
+                    "context": "team member info"
+                },
+                {
+                    "content": "Bob is a junior data scientist who joined last month.",
+                    "context": "team member info"
+                },
+                {
+                    "content": "The team uses Python and TensorFlow for most projects.",
+                    "context": "tech stack"
+                }
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Define a JSON schema for structured output
+    response_schema = {
+        "type": "object",
+        "properties": {
+            "team_members": {
+                "type": "array",
+                "items": {
+                    "type": "object",
+                    "properties": {
+                        "name": {"type": "string"},
+                        "role": {"type": "string"},
+                        "experience_level": {"type": "string"}
+                    }
+                }
+            },
+            "technologies": {
+                "type": "array",
+                "items": {"type": "string"}
+            },
+            "summary": {"type": "string"}
+        },
+        "required": ["team_members", "summary"]
+    }
+
+    # Call reflect with response_schema
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/reflect",
+        json={
+            "query": "Give me an overview of the team and their tech stack",
+            "response_schema": response_schema
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+
+    # Verify text field exists (empty when using structured output)
+    assert "text" in result
+    assert result["text"] == ""
+
+    # Verify structured output exists and has expected structure
+    assert "structured_output" in result
+    assert result["structured_output"] is not None
+
+    structured = result["structured_output"]
+    assert "team_members" in structured
+    assert "summary" in structured
+    assert isinstance(structured["team_members"], list)
+    assert isinstance(structured["summary"], str)
+
+    # Verify team members have the expected fields
+    if len(structured["team_members"]) > 0:
+        member = structured["team_members"][0]
+        assert "name" in member or "role" in member  # At least some fields should be present
+
+
+@pytest.mark.asyncio
+async def test_reflect_without_structured_output(api_client):
+    """Test that reflect works normally without response_schema.
+
+    When response_schema is not provided, the structured_output field
+    should be null/None in the response.
+    """
+    test_bank_id = f"reflect_no_structured_test_{datetime.now().timestamp()}"
+
+    # Store a memory
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "The project deadline is next Friday.",
+                    "context": "project timeline"
+                }
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Call reflect without response_schema
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/reflect",
+        json={
+            "query": "When is the project deadline?"
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+
+    # Verify response has text but structured_output is null
+    assert "text" in result
+    assert len(result["text"]) > 0
+    assert result.get("structured_output") is None
+
+
+@pytest.mark.asyncio
+async def test_reflect_with_max_tokens(api_client):
+    """Test reflect endpoint with custom max_tokens parameter.
+
+    The max_tokens parameter controls the maximum tokens for the LLM response.
+    """
+    test_bank_id = f"reflect_max_tokens_test_{datetime.now().timestamp()}"
+
+    # Store a memory
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "Python is a popular programming language for data science and machine learning.",
+                    "context": "tech"
+                }
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Call reflect with custom max_tokens
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/reflect",
+        json={
+            "query": "What is Python used for?",
+            "max_tokens": 500
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+
+    # Verify response has text
+    assert "text" in result
+    assert len(result["text"]) > 0

hindsight-api/tests/test_llm_provider.py

@@ -1,9 +1,12 @@
 """
-Test LLM provider with different models and providers.
+Test LLM provider with different models using actual memory operations.
 """
 import os
+from datetime import datetime
 import pytest
 from hindsight_api.engine.llm_wrapper import LLMProvider
+from hindsight_api.engine.utils import extract_facts
+from hindsight_api.engine.search.think_utils import reflect
 
 
 # Model matrix: (provider, model)
@@ -15,13 +18,17 @@ MODEL_MATRIX = [
     ("openai", "gpt-5-mini"),
     ("openai", "gpt-5-nano"),
     ("openai", "gpt-5"),
+    ("openai", "gpt-5.2"),
     # Groq models
-    ("groq", "llama-3.3-70b-versatile"),
     ("groq", "openai/gpt-oss-120b"),
     ("groq", "openai/gpt-oss-20b"),
     # Gemini models
     ("gemini", "gemini-2.5-flash"),
     ("gemini", "gemini-2.5-flash-lite"),
+    ("gemini", "gemini-3-pro-preview"),
+    # Ollama models (local)
+    ("ollama", "gemma3:12b"),
+    ("ollama", "gemma3:1b"),
 ]
 
 
@@ -38,90 +45,75 @@ def get_api_key_for_provider(provider: str) -> str | None:
 
 @pytest.mark.parametrize("provider,model", MODEL_MATRIX)
 @pytest.mark.asyncio
-async def test_llm_provider_call(provider: str, model: str):
+async def test_llm_provider_memory_operations(provider: str, model: str):
     """
-    Test LLM provider can make a basic call with different models.
-    Skips if the required API key is not available.
+    Test LLM provider with actual memory operations: fact extraction and reflect.
+    All models must pass this test.
     """
     api_key = get_api_key_for_provider(provider)
-    if not api_key:
+
+    # Skip Ollama tests in CI (no models available)
+    if provider == "ollama" and os.getenv("CI"):
+        pytest.skip(f"Skipping {provider}/{model}: Ollama not available in CI")
+
+    # Other providers need an API key
+    if provider != "ollama" and not api_key:
         pytest.skip(f"Skipping {provider}/{model}: no API key available")
 
     llm = LLMProvider(
         provider=provider,
-        api_key=api_key,
+        api_key=api_key or "",
         base_url="",
         model=model,
     )
 
-    # Test basic call
-    response = await llm.call(
-        messages=[{"role": "user", "content": "Say 'hello' and nothing else."}],
-        max_completion_tokens=50,
-        temperature=0.1,
+    # Test 1: Fact extraction (structured output)
+    test_text = """
+    User: I just got back from my trip to Paris last week. The Eiffel Tower was amazing!
+    Assistant: That sounds wonderful! How long were you there?
+    User: About 5 days. I also visited the Louvre and saw the Mona Lisa.
+    """
+    event_date = datetime(2024, 12, 10)
+
+    facts, chunks = await extract_facts(
+        text=test_text,
+        event_date=event_date,
+        context="Travel conversation",
+        llm_config=llm,
     )
 
-    print(f"\n{provider}/{model} response: {response}")
-    assert response is not None, f"{provider}/{model} returned None"
+    print(f"\n{provider}/{model} - Fact extraction:")
+    print(f"  Extracted {len(facts)} facts from {len(chunks)} chunks")
+    for fact in facts:
+        print(f"  - {fact.fact}")
 
+    assert facts is not None, f"{provider}/{model} fact extraction returned None"
+    assert len(facts) > 0, f"{provider}/{model} should extract at least one fact"
 
-@pytest.mark.parametrize("provider,model", MODEL_MATRIX)
-@pytest.mark.asyncio
-async def test_llm_provider_verify_connection(provider: str, model: str):
-    """
-    Test LLM provider verify_connection method with different models.
-    Skips if the required API key is not available.
-    """
-    api_key = get_api_key_for_provider(provider)
-    if not api_key:
-        pytest.skip(f"Skipping {provider}/{model}: no API key available")
+    # Verify facts have required fields
+    for fact in facts:
+        assert fact.fact, f"{provider}/{model} fact missing text"
+        assert fact.fact_type in ["world", "experience", "opinion"], f"{provider}/{model} invalid fact_type: {fact.fact_type}"
 
-    llm = LLMProvider(
-        provider=provider,
-        api_key=api_key,
-        base_url="",
-        model=model,
+    # Test 2: Reflect (actual reflect function)
+    response = await reflect(
+        llm_config=llm,
+        query="What was the highlight of my Paris trip?",
+        experience_facts=[
+            "I visited Paris in December 2024",
+            "I saw the Eiffel Tower and it was amazing",
+            "I visited the Louvre and saw the Mona Lisa",
+            "The trip lasted 5 days",
+        ],
+        world_facts=[
+            "The Eiffel Tower is a famous landmark in Paris",
+            "The Mona Lisa is displayed at the Louvre museum",
+        ],
+        name="Traveler",
     )
 
-    # Test verify_connection
-    await llm.verify_connection()
-    print(f"\n{provider}/{model} connection verified")
+    print(f"\n{provider}/{model} - Reflect response:")
+    print(f"  {response[:200]}...")
 
-
-# Models that support large output (65000+ tokens)
-LARGE_OUTPUT_MODELS = [
-    ("openai", "gpt-5-mini"),
-    ("openai", "gpt-5-nano"),
-    ("openai", "gpt-5"),
-    ("gemini", "gemini-2.5-flash"),
-    ("gemini", "gemini-2.5-flash-lite"),
-]
-
-
-@pytest.mark.parametrize("provider,model", LARGE_OUTPUT_MODELS)
-@pytest.mark.asyncio
-async def test_llm_provider_large_output(provider: str, model: str):
-    """
-    Test LLM provider with large max_completion_tokens (65000).
-    Only tests models that support large outputs.
-    Skips if the required API key is not available.
-    """
-    api_key = get_api_key_for_provider(provider)
-    if not api_key:
-        pytest.skip(f"Skipping {provider}/{model}: no API key available")
-
-    llm = LLMProvider(
-        provider=provider,
-        api_key=api_key,
-        base_url="",
-        model=model,
-    )
-
-    # Test call with large max_completion_tokens
-    response = await llm.call(
-        messages=[{"role": "user", "content": "Say 'ok'"}],
-        max_completion_tokens=65000,
-    )
-
-    print(f"\n{provider}/{model} large output response: {response}")
-    assert response is not None, f"{provider}/{model} returned None"
+    assert response is not None, f"{provider}/{model} reflect returned None"
+    assert len(response) > 10, f"{provider}/{model} reflect response too short"

hindsight-api/tests/test_mcp_local.py

@@ -0,0 +1,162 @@
+"""Test local MCP server."""
+
+import asyncio
+import pytest
+from unittest.mock import AsyncMock, MagicMock
+
+
+@pytest.fixture
+def mock_memory():
+    """Create a mock MemoryEngine."""
+    memory = MagicMock()
+    memory._initialized = True
+    memory.retain_batch_async = AsyncMock()
+    memory.recall_async = AsyncMock(return_value=MagicMock(results=[]))
+    return memory
+
+
+@pytest.mark.asyncio
+async def test_local_mcp_server_retain(mock_memory):
+    """Test that retain tool fires async and returns immediately."""
+    from hindsight_api.mcp_local import create_local_mcp_server
+
+    bank_id = "test-bank"
+    mcp_server = create_local_mcp_server(bank_id, memory=mock_memory)
+
+    # Get the tools
+    tools = mcp_server._tool_manager._tools
+    assert "retain" in tools
+
+    # Call retain
+    retain_tool = tools["retain"]
+    result = await retain_tool.fn(content="test content", context="test_context")
+
+    # Returns immediately with accepted status
+    assert result["status"] == "accepted"
+
+    # Wait for background task to complete
+    await asyncio.sleep(0.1)
+
+    # Verify the memory was called correctly
+    mock_memory.retain_batch_async.assert_called_once()
+    call_kwargs = mock_memory.retain_batch_async.call_args.kwargs
+    assert call_kwargs["bank_id"] == "test-bank"
+    assert call_kwargs["contents"] == [{"content": "test content", "context": "test_context"}]
+
+
+@pytest.mark.asyncio
+async def test_local_mcp_server_recall(mock_memory):
+    """Test that recall tool calls memory.recall_async with correct params."""
+    from hindsight_api.mcp_local import create_local_mcp_server
+    from hindsight_api.engine.memory_engine import Budget
+
+    # Mock recall_async to return a proper pydantic model
+    mock_result = MagicMock()
+    mock_result.model_dump.return_value = {"results": []}
+    mock_memory.recall_async = AsyncMock(return_value=mock_result)
+
+    bank_id = "test-bank"
+    mcp_server = create_local_mcp_server(bank_id, memory=mock_memory)
+
+    # Get the tools
+    tools = mcp_server._tool_manager._tools
+    assert "recall" in tools
+
+    # Call recall with new params
+    recall_tool = tools["recall"]
+    result = await recall_tool.fn(query="test query", max_tokens=2048, budget="mid")
+
+    # Result is a dict
+    assert isinstance(result, dict)
+
+    # Verify the memory was called correctly
+    mock_memory.recall_async.assert_called_once()
+    call_kwargs = mock_memory.recall_async.call_args.kwargs
+    assert call_kwargs["bank_id"] == "test-bank"
+    assert call_kwargs["query"] == "test query"
+    assert call_kwargs["max_tokens"] == 2048
+    assert call_kwargs["budget"] == Budget.MID
+
+
+@pytest.mark.asyncio
+async def test_local_mcp_server_retain_with_default_context(mock_memory):
+    """Test that retain uses default context when not provided."""
+    from hindsight_api.mcp_local import create_local_mcp_server
+
+    bank_id = "test-bank"
+    mcp_server = create_local_mcp_server(bank_id, memory=mock_memory)
+
+    tools = mcp_server._tool_manager._tools
+    retain_tool = tools["retain"]
+
+    # Call retain without context
+    await retain_tool.fn(content="test content")
+
+    # Wait for background task
+    await asyncio.sleep(0.1)
+
+    call_kwargs = mock_memory.retain_batch_async.call_args.kwargs
+    assert call_kwargs["contents"] == [{"content": "test content", "context": "general"}]
+
+
+@pytest.mark.asyncio
+async def test_local_mcp_server_retain_error_handling(mock_memory):
+    """Test that retain errors are logged but don't affect response."""
+    from hindsight_api.mcp_local import create_local_mcp_server
+
+    mock_memory.retain_batch_async = AsyncMock(side_effect=Exception("Test error"))
+
+    mcp_server = create_local_mcp_server("test-bank", memory=mock_memory)
+
+    tools = mcp_server._tool_manager._tools
+    retain_tool = tools["retain"]
+
+    # Retain returns immediately with accepted status (fire and forget)
+    result = await retain_tool.fn(content="test content")
+    assert result["status"] == "accepted"
+
+    # Wait for background task to complete (and log error)
+    await asyncio.sleep(0.1)
+
+
+@pytest.mark.asyncio
+async def test_local_mcp_server_recall_error_handling(mock_memory):
+    """Test that recall handles errors gracefully."""
+    from hindsight_api.mcp_local import create_local_mcp_server
+
+    mock_memory.recall_async = AsyncMock(side_effect=Exception("Test error"))
+
+    mcp_server = create_local_mcp_server("test-bank", memory=mock_memory)
+
+    tools = mcp_server._tool_manager._tools
+    recall_tool = tools["recall"]
+
+    result = await recall_tool.fn(query="test query")
+
+    # Result is a dict with error
+    assert isinstance(result, dict)
+    assert "error" in result
+    assert result["results"] == []
+
+
+@pytest.mark.asyncio
+async def test_local_mcp_server_recall_with_defaults(mock_memory):
+    """Test that recall uses default max_tokens and budget."""
+    from hindsight_api.mcp_local import create_local_mcp_server
+    from hindsight_api.engine.memory_engine import Budget
+
+    mock_result = MagicMock()
+    mock_result.model_dump.return_value = {"results": []}
+    mock_memory.recall_async = AsyncMock(return_value=mock_result)
+
+    mcp_server = create_local_mcp_server("test-bank", memory=mock_memory)
+
+    tools = mcp_server._tool_manager._tools
+    recall_tool = tools["recall"]
+
+    # Call with defaults
+    await recall_tool.fn(query="test query")
+
+    call_kwargs = mock_memory.recall_async.call_args.kwargs
+    assert call_kwargs["max_tokens"] == 4096
+    assert call_kwargs["budget"] == Budget.LOW

hindsight-api/tests/test_mcp_routing.py

@@ -8,7 +8,7 @@ from unittest.mock import AsyncMock, MagicMock
 def mock_memory():
     """Create a mock MemoryEngine."""
     memory = MagicMock()
-    memory.put_batch_async = AsyncMock()
+    memory.retain_batch_async = AsyncMock()
     memory.recall_async = AsyncMock(return_value=MagicMock(results=[]))
     return memory
 
@@ -52,8 +52,8 @@ async def test_mcp_tools_use_context_bank_id(mock_memory):
         assert "successfully" in result.lower()
 
         # Verify the memory was called with the context bank_id
-        mock_memory.put_batch_async.assert_called_once()
-        call_kwargs = mock_memory.put_batch_async.call_args.kwargs
+        mock_memory.retain_batch_async.assert_called_once()
+        call_kwargs = mock_memory.retain_batch_async.call_args.kwargs
         assert call_kwargs["bank_id"] == "context-bank-id"
     finally:
         _current_bank_id.reset(token)

hindsight-api/tests/test_observations.py

@@ -3,11 +3,12 @@ Test observation generation and entity state functionality.
 """
 import pytest
 from hindsight_api.engine.memory_engine import Budget
+from hindsight_api import RequestContext
 from datetime import datetime, timezone
 
 
 @pytest.mark.asyncio
-async def test_observation_generation_on_put(memory):
+async def test_observation_generation_on_put(memory, request_context):
     """
     Test that observations are generated SYNCHRONOUSLY when new facts are added.
 
@@ -36,7 +37,8 @@ async def test_observation_generation_on_put(memory):
                 bank_id=bank_id,
                 content=content,
                 context="work info",
-                event_date=datetime(2024, 1, 15 + i, tzinfo=timezone.utc)
+                event_date=datetime(2024, 1, 15 + i, tzinfo=timezone.utc),
+                request_context=request_context,
             )
 
         # Observations are generated SYNCHRONOUSLY during retain,
@@ -75,7 +77,7 @@ async def test_observation_generation_on_put(memory):
         print(f"Entity: {entity_name} (id: {entity_id})")
 
         # Get observations for the entity - should be available immediately
-        observations = await memory.get_entity_observations(bank_id, entity_id, limit=10)
+        observations = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
 
         print(f"\n=== Observations for {entity_name} ===")
         print(f"Total observations: {len(observations)}")
@@ -102,7 +104,7 @@ async def test_observation_generation_on_put(memory):
 
 
 @pytest.mark.asyncio
-async def test_regenerate_entity_observations(memory):
+async def test_regenerate_entity_observations(memory, request_context):
     """
     Test explicit regeneration of observations for an entity.
     """
@@ -114,7 +116,8 @@ async def test_regenerate_entity_observations(memory):
             bank_id=bank_id,
             content="Sarah is a product manager who loves user research and data analysis.",
             context="work info",
-            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc)
+            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
+            request_context=request_context,
         )
 
         await memory.wait_for_background_tasks()
@@ -140,14 +143,15 @@ async def test_regenerate_entity_observations(memory):
             created_ids = await memory.regenerate_entity_observations(
                 bank_id=bank_id,
                 entity_id=entity_id,
-                entity_name=entity_name
+                entity_name=entity_name,
+                request_context=request_context,
             )
 
             print(f"\n=== Regenerated Observations ===")
             print(f"Created {len(created_ids)} observations for {entity_name}")
 
             # Get the observations
-            observations = await memory.get_entity_observations(bank_id, entity_id, limit=10)
+            observations = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
             for obs in observations:
                 print(f"  - {obs.text}")
 
@@ -170,7 +174,108 @@ async def test_regenerate_entity_observations(memory):
 
 
 @pytest.mark.asyncio
-async def test_search_with_include_entities(memory):
+async def test_manual_regenerate_with_few_facts(memory, request_context):
+    """
+    Test that manual regeneration works even with fewer than 5 facts.
+
+    This is important because:
+    - Automatic generation during retain requires MIN_FACTS_THRESHOLD (5)
+    - But manual regeneration via API should work with any number of facts
+    - The UI triggers manual regeneration, so it should work regardless of fact count
+    """
+    bank_id = f"test_manual_regen_{datetime.now(timezone.utc).timestamp()}"
+
+    try:
+        # Store only 2 facts - below the automatic threshold
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="Alice works at Google as a senior software engineer.",
+            context="work info",
+            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
+            request_context=request_context,
+        )
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="Alice loves hiking and outdoor photography.",
+            context="hobbies",
+            event_date=datetime(2024, 1, 16, tzinfo=timezone.utc),
+            request_context=request_context,
+        )
+
+        # Find the Alice entity
+        pool = await memory._get_pool()
+        async with pool.acquire() as conn:
+            entity_row = await conn.fetchrow(
+                """
+                SELECT id, canonical_name
+                FROM entities
+                WHERE bank_id = $1 AND LOWER(canonical_name) LIKE '%alice%'
+                LIMIT 1
+                """,
+                bank_id
+            )
+
+        assert entity_row is not None, "Alice entity should have been extracted"
+
+        entity_id = str(entity_row['id'])
+        entity_name = entity_row['canonical_name']
+
+        # Check fact count - should be < 5
+        async with pool.acquire() as conn:
+            fact_count = await conn.fetchval(
+                "SELECT COUNT(*) FROM unit_entities WHERE entity_id = $1",
+                entity_row['id']
+            )
+
+        print(f"\n=== Manual Regeneration Test ===")
+        print(f"Entity: {entity_name} (id: {entity_id})")
+        print(f"Linked facts: {fact_count}")
+
+        # Verify we're testing with fewer than the automatic threshold
+        assert fact_count < 5, f"Test requires < 5 facts, but entity has {fact_count}"
+
+        # Before regeneration - should have no observations (auto threshold not met)
+        obs_before = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
+        print(f"Observations before manual regenerate: {len(obs_before)}")
+
+        # Manually regenerate observations - this should work regardless of fact count
+        created_ids = await memory.regenerate_entity_observations(
+            bank_id=bank_id,
+            entity_id=entity_id,
+            entity_name=entity_name,
+            request_context=request_context,
+        )
+
+        print(f"Observations created by manual regenerate: {len(created_ids)}")
+
+        # Get observations after regeneration
+        observations = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
+        print(f"Observations after manual regenerate: {len(observations)}")
+        for obs in observations:
+            print(f"  - {obs.text}")
+
+        # Manual regeneration should create observations even with < 5 facts
+        assert len(observations) > 0, \
+            f"Manual regeneration should create observations even with only {fact_count} facts. " \
+            f"The LLM should synthesize at least 1 observation from the available facts."
+
+        # Verify observations contain relevant content
+        obs_texts = " ".join([o.text.lower() for o in observations])
+        assert any(keyword in obs_texts for keyword in ["google", "engineer", "hiking", "photography", "alice"]), \
+            "Observations should contain relevant information about Alice"
+
+        print(f"✓ Manual regeneration works with {fact_count} facts (below automatic threshold of 5)")
+
+    finally:
+        # Cleanup
+        pool = await memory._get_pool()
+        async with pool.acquire() as conn:
+            await conn.execute("DELETE FROM memory_units WHERE bank_id = $1", bank_id)
+            await conn.execute("DELETE FROM entities WHERE bank_id = $1", bank_id)
+
+
+@pytest.mark.asyncio
+async def test_search_with_include_entities(memory, request_context):
     """
     Test that search with include_entities=True returns entity observations.
 
@@ -196,7 +301,8 @@ async def test_search_with_include_entities(memory):
                 bank_id=bank_id,
                 content=content,
                 context="work info",
-                event_date=datetime(2024, 1, 15 + i, tzinfo=timezone.utc)
+                event_date=datetime(2024, 1, 15 + i, tzinfo=timezone.utc),
+                request_context=request_context,
             )
 
         # Observations are generated synchronously during retain, no need to wait
@@ -209,7 +315,8 @@ async def test_search_with_include_entities(memory):
             budget=Budget.LOW,
             max_tokens=2000,
             include_entities=True,
-            max_entity_tokens=500
+            max_entity_tokens=500,
+            request_context=request_context,
         )
 
         print(f"\n=== Search Results ===")
@@ -263,7 +370,7 @@ async def test_search_with_include_entities(memory):
 
 
 @pytest.mark.asyncio
-async def test_get_entity_state(memory):
+async def test_get_entity_state(memory, request_context):
     """
     Test getting the full state of an entity.
     """
@@ -275,7 +382,8 @@ async def test_get_entity_state(memory):
             bank_id=bank_id,
             content="Bob is a frontend developer who specializes in React and TypeScript.",
             context="work info",
-            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc)
+            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
+            request_context=request_context,
         )
 
         await memory.wait_for_background_tasks()
@@ -302,7 +410,8 @@ async def test_get_entity_state(memory):
                 bank_id=bank_id,
                 entity_id=entity_id,
                 entity_name=entity_name,
-                limit=10
+                limit=10,
+                request_context=request_context,
             )
 
             print(f"\n=== Entity State for {entity_name} ===")
@@ -324,7 +433,7 @@ async def test_get_entity_state(memory):
 
 
 @pytest.mark.asyncio
-async def test_observation_fact_type_in_database(memory):
+async def test_observation_fact_type_in_database(memory, request_context):
     """
     Test that observations are stored with correct fact_type in database.
     """
@@ -336,7 +445,8 @@ async def test_observation_fact_type_in_database(memory):
             bank_id=bank_id,
             content="Charlie is a DevOps engineer who manages the Kubernetes infrastructure.",
             context="work info",
-            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc)
+            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
+            request_context=request_context,
         )
 
         await memory.wait_for_background_tasks()
@@ -374,7 +484,7 @@ async def test_observation_fact_type_in_database(memory):
 
 
 @pytest.mark.asyncio
-async def test_user_entity_prioritized_for_observations(memory):
+async def test_user_entity_prioritized_for_observations(memory, request_context):
     """
     Test that the 'user' entity gets observations even when many other entities exist.
 
@@ -410,7 +520,8 @@ async def test_user_entity_prioritized_for_observations(memory):
                 bank_id=bank_id,
                 content=content,
                 context="personal info",
-                event_date=datetime(2024, 1, 15 + i, tzinfo=timezone.utc)
+                event_date=datetime(2024, 1, 15 + i, tzinfo=timezone.utc),
+                request_context=request_context,
             )
 
         # Observations are generated synchronously during retain
@@ -466,7 +577,7 @@ async def test_user_entity_prioritized_for_observations(memory):
             f"User entity should have at least 5 facts, but has {user_fact_count}"
 
         # Get observations for user entity
-        observations = await memory.get_entity_observations(bank_id, user_entity_id, limit=10)
+        observations = await memory.get_entity_observations(bank_id, user_entity_id, limit=10, request_context=request_context)
 
         print(f"\n=== User Entity Observations ===")
         print(f"Total observations: {len(observations)}")

hindsight-api/tests/test_retain.py

@@ -5,12 +5,13 @@ import pytest
 import logging
 from datetime import datetime, timezone, timedelta
 from hindsight_api.engine.memory_engine import Budget
+from hindsight_api import RequestContext
 
 logger = logging.getLogger(__name__)
 
 
 @pytest.mark.asyncio
-async def test_retain_with_chunks(memory):
+async def test_retain_with_chunks(memory, request_context):
     """
     Test that retain function:
     1. Stores facts with associated chunks
@@ -41,7 +42,8 @@ async def test_retain_with_chunks(memory):
             content=long_content,
             context="team overview",
             event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         print(f"\n=== Retained {len(unit_ids)} facts ===")
@@ -56,7 +58,8 @@ async def test_retain_with_chunks(memory):
             fact_type=["world"],  # Search for world facts
             include_entities=False,  # Disable entities for simpler test
             include_chunks=True,  # Enable chunks
-            max_chunk_tokens=8192
+            max_chunk_tokens=8192,
+            request_context=request_context,
         )
 
         print(f"\n=== Recall Results (with chunks) ===")
@@ -88,12 +91,12 @@ async def test_retain_with_chunks(memory):
 
     finally:
         # Cleanup - delete the test bank
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
         print(f"\n=== Cleaned up bank: {bank_id} ===")
 
 
 @pytest.mark.asyncio
-async def test_chunks_and_entities_follow_fact_order(memory):
+async def test_chunks_and_entities_follow_fact_order(memory, request_context):
     """
     Test that chunks and entities in recall results follow the same order as facts.
     This is critical because token limits may truncate later items.
@@ -130,7 +133,8 @@ async def test_chunks_and_entities_follow_fact_order(memory):
                 content=item["content"],
                 context=item["context"],
                 event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
-                document_id=item["document_id"]
+                document_id=item["document_id"],
+                request_context=request_context,
             )
 
         print("\n=== Stored 3 separate documents ===")
@@ -144,7 +148,8 @@ async def test_chunks_and_entities_follow_fact_order(memory):
             fact_type=["world"],
             include_entities=True,
             include_chunks=True,
-            max_chunk_tokens=8192
+            max_chunk_tokens=8192,
+            request_context=request_context,
         )
 
         print(f"\n=== Recall Results ===")
@@ -214,12 +219,12 @@ async def test_chunks_and_entities_follow_fact_order(memory):
 
     finally:
         # Cleanup
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
         print(f"\n=== Cleaned up bank: {bank_id} ===")
 
 
 @pytest.mark.asyncio
-async def test_event_date_storage(memory):
+async def test_event_date_storage(memory, request_context):
     """
     Test that event_date is correctly stored as occurred_start.
     Verifies that we can track when events actually happened vs when they were stored.
@@ -235,7 +240,8 @@ async def test_event_date_storage(memory):
             bank_id=bank_id,
             content="Alice completed the Q2 product launch on June 15th, 2023.",
             context="project history",
-            event_date=past_event_date
+            event_date=past_event_date,
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should have created at least one memory unit"
@@ -246,7 +252,8 @@ async def test_event_date_storage(memory):
             query="When did Alice complete the product launch?",
             budget=Budget.LOW,
             max_tokens=500,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall the stored fact"
@@ -268,11 +275,11 @@ async def test_event_date_storage(memory):
         print(f"\n✓ Event date correctly stored: {occurred_dt}")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_temporal_ordering(memory):
+async def test_temporal_ordering(memory, request_context):
     """
     Test that facts can be stored and retrieved with correct temporal ordering.
     Stores facts with different event_dates and verifies temporal relationships.
@@ -305,7 +312,8 @@ async def test_temporal_ordering(memory):
                 bank_id=bank_id,
                 content=event["content"],
                 context=event["context"],
-                event_date=event["event_date"]
+                event_date=event["event_date"],
+                request_context=request_context,
             )
 
         print("\n=== Stored 3 events with different temporal dates ===")
@@ -316,7 +324,8 @@ async def test_temporal_ordering(memory):
             query="Tell me about Alice's career progression",
             budget=Budget.MID,
             max_tokens=1000,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) >= 3, f"Should recall all 3 events, got {len(result.results)}"
@@ -345,11 +354,11 @@ async def test_temporal_ordering(memory):
         print(f"\n✓ Temporal ordering preserved: {min_date.date()} to {max_date.date()}")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_mentioned_at_vs_occurred(memory):
+async def test_mentioned_at_vs_occurred(memory, request_context):
     """
     Test distinction between when fact occurred vs when it was mentioned.
 
@@ -369,7 +378,8 @@ async def test_mentioned_at_vs_occurred(memory):
             bank_id=bank_id,
             content="Alice graduated from MIT in March 2020.",
             context="education history",
-            event_date=conversation_date  # When this conversation happened
+            event_date=conversation_date,  # When this conversation happened
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create memory unit"
@@ -380,7 +390,8 @@ async def test_mentioned_at_vs_occurred(memory):
             query="Where did Alice go to school?",
             budget=Budget.LOW,
             max_tokens=500,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall the fact"
@@ -415,11 +426,11 @@ async def test_mentioned_at_vs_occurred(memory):
         print(f"✓ Test passed: Historical conversation correctly ingested with event_date=2020")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_occurred_dates_not_defaulted(memory):
+async def test_occurred_dates_not_defaulted(memory, request_context):
     """
     Test that occurred_start and occurred_end are NOT defaulted to mentioned_at.
 
@@ -441,7 +452,8 @@ async def test_occurred_dates_not_defaulted(memory):
             bank_id=bank_id,
             content="Alice likes coffee. The weather is sunny today.",
             context="current observations",
-            event_date=event_date
+            event_date=event_date,
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create memory unit"
@@ -452,7 +464,8 @@ async def test_occurred_dates_not_defaulted(memory):
             query="What does Alice like?",
             budget=Budget.LOW,
             max_tokens=500,
-            fact_type=["world", "opinion"]
+            fact_type=["world", "opinion"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall the fact"
@@ -504,11 +517,11 @@ async def test_occurred_dates_not_defaulted(memory):
         print(f"✓ Test passed: occurred dates are not incorrectly defaulted to mentioned_at")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_mentioned_at_from_context_string(memory):
+async def test_mentioned_at_from_context_string(memory, request_context):
     """
     Test that mentioned_at is extracted from context string by LLM.
 
@@ -527,7 +540,8 @@ async def test_mentioned_at_from_context_string(memory):
             bank_id=bank_id,
             content="Alice mentioned she loves hiking in the mountains.",
             context=f"Session ABC123 - you are the assistant in this conversation - happened on {session_date.strftime('%Y-%m-%d %H:%M:%S')} UTC.",
-            event_date=None  # Not providing event_date - should default to now() if LLM doesn't extract
+            event_date=None,  # Not providing event_date - should default to now() if LLM doesn't extract
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create memory unit"
@@ -538,7 +552,8 @@ async def test_mentioned_at_from_context_string(memory):
             query="What does Alice like?",
             budget=Budget.LOW,
             max_tokens=500,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall the fact"
@@ -574,7 +589,7 @@ async def test_mentioned_at_from_context_string(memory):
         print(f"✓ mentioned_at is always set (never None)")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 # ============================================================
@@ -582,7 +597,7 @@ async def test_mentioned_at_from_context_string(memory):
 # ============================================================
 
 @pytest.mark.asyncio
-async def test_context_preservation(memory):
+async def test_context_preservation(memory, request_context):
     """
     Test that context is preserved and retrievable.
     Context helps understand why/how memory was formed.
@@ -597,7 +612,8 @@ async def test_context_preservation(memory):
             bank_id=bank_id,
             content="The team decided to prioritize mobile development for next quarter.",
             context=specific_context,
-            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc)
+            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create at least one memory unit"
@@ -608,7 +624,8 @@ async def test_context_preservation(memory):
             query="What did the team decide?",
             budget=Budget.LOW,
             max_tokens=500,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall the stored fact"
@@ -620,11 +637,11 @@ async def test_context_preservation(memory):
         print(f"  Retrieved {len(result.results)} facts")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_context_with_batch(memory):
+async def test_context_with_batch(memory, request_context):
     """
     Test that each item in a batch can have different contexts.
     """
@@ -650,7 +667,8 @@ async def test_context_with_batch(memory):
                     "context": "incident response",
                     "event_date": datetime(2024, 1, 12, tzinfo=timezone.utc)
                 }
-            ]
+            ],
+            request_context=request_context,
         )
 
         # Should have created facts from all items
@@ -661,7 +679,7 @@ async def test_context_with_batch(memory):
         print(f"  Created {total_units} total memory units")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 # ============================================================
@@ -669,7 +687,7 @@ async def test_context_with_batch(memory):
 # ============================================================
 
 @pytest.mark.asyncio
-async def test_metadata_storage_and_retrieval(memory):
+async def test_metadata_storage_and_retrieval(memory, request_context):
     """
     Test that user-defined metadata is preserved.
     Metadata allows arbitrary key-value data to be stored with facts.
@@ -692,7 +710,8 @@ async def test_metadata_storage_and_retrieval(memory):
             bank_id=bank_id,
             content="The product launch is scheduled for March 1st.",
             context="planning meeting",
-            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc)
+            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create memory units"
@@ -703,7 +722,8 @@ async def test_metadata_storage_and_retrieval(memory):
             query="When is the product launch?",
             budget=Budget.LOW,
             max_tokens=500,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall stored facts"
@@ -712,7 +732,7 @@ async def test_metadata_storage_and_retrieval(memory):
         print(f"  (Note: Metadata support depends on API implementation)")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 # ============================================================
@@ -720,7 +740,7 @@ async def test_metadata_storage_and_retrieval(memory):
 # ============================================================
 
 @pytest.mark.asyncio
-async def test_empty_batch(memory):
+async def test_empty_batch(memory, request_context):
     """
     Test that empty batch is handled gracefully without errors.
     """
@@ -730,7 +750,8 @@ async def test_empty_batch(memory):
         # Attempt to store empty batch
         unit_ids = await memory.retain_batch_async(
             bank_id=bank_id,
-            contents=[]
+            contents=[],
+            request_context=request_context,
         )
 
         # Should return empty list or handle gracefully
@@ -741,11 +762,11 @@ async def test_empty_batch(memory):
 
     finally:
         # Clean up (though nothing should be stored)
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_single_item_batch(memory):
+async def test_single_item_batch(memory, request_context):
     """
     Test that batch with one item works correctly.
     """
@@ -761,7 +782,8 @@ async def test_single_item_batch(memory):
                     "context": "deployment log",
                     "event_date": datetime(2024, 1, 15, tzinfo=timezone.utc)
                 }
-            ]
+            ],
+            request_context=request_context,
         )
 
         assert len(unit_ids) == 1, "Should return one list of unit IDs"
@@ -770,11 +792,11 @@ async def test_single_item_batch(memory):
         print(f"✓ Single-item batch created {len(unit_ids[0])} units")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_mixed_content_batch(memory):
+async def test_mixed_content_batch(memory, request_context):
     """
     Test batch with varying content sizes (short and long).
     """
@@ -798,7 +820,8 @@ async def test_mixed_content_batch(memory):
                 {"content": short_content, "context": "onboarding"},
                 {"content": long_content, "context": "performance review"},
                 {"content": "Charlie is on vacation this week.", "context": "team status"}
-            ]
+            ],
+            request_context=request_context,
         )
 
         # All items should be processed
@@ -813,11 +836,11 @@ async def test_mixed_content_batch(memory):
         print(f"  Long content: {long_units} units")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_batch_with_missing_optional_fields(memory):
+async def test_batch_with_missing_optional_fields(memory, request_context):
     """
     Test that batch handles items with missing optional fields.
     """
@@ -842,7 +865,8 @@ async def test_batch_with_missing_optional_fields(memory):
                     "context": "code review",
                     # No event_date
                 }
-            ]
+            ],
+            request_context=request_context,
         )
 
         # All items should be processed successfully
@@ -852,7 +876,7 @@ async def test_batch_with_missing_optional_fields(memory):
         print(f"✓ Batch with mixed optional fields created {total_units} total units")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 # ============================================================
@@ -860,7 +884,7 @@ async def test_batch_with_missing_optional_fields(memory):
 # ============================================================
 
 @pytest.mark.asyncio
-async def test_single_batch_multiple_documents(memory):
+async def test_single_batch_multiple_documents(memory, request_context):
     """
     Test storing multiple distinct documents in a single batch call.
     Each should be tracked separately.
@@ -876,21 +900,24 @@ async def test_single_batch_multiple_documents(memory):
             bank_id=bank_id,
             content="Alice's resume: 10 years Python experience, worked at Google.",
             context="resume review",
-            document_id="resume_alice"
+            document_id="resume_alice",
+            request_context=request_context,
         )
 
         doc2_units = await memory.retain_async(
             bank_id=bank_id,
             content="Bob's resume: 5 years JavaScript experience, worked at Meta.",
             context="resume review",
-            document_id="resume_bob"
+            document_id="resume_bob",
+            request_context=request_context,
         )
 
         doc3_units = await memory.retain_async(
             bank_id=bank_id,
             content="Charlie's resume: 8 years Go experience, worked at Amazon.",
             context="resume review",
-            document_id="resume_charlie"
+            document_id="resume_charlie",
+            request_context=request_context,
         )
 
         # All documents should be stored
@@ -907,17 +934,18 @@ async def test_single_batch_multiple_documents(memory):
             query="Who worked at Google?",
             budget=Budget.MID,
             max_tokens=1000,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should find facts about Alice"
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_document_upsert_behavior(memory):
+async def test_document_upsert_behavior(memory, request_context):
     """
     Test that upserting a document replaces the old content.
     """
@@ -930,7 +958,8 @@ async def test_document_upsert_behavior(memory):
             bank_id=bank_id,
             content="Project is in planning phase. Alice is the lead.",
             context="status update v1",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         assert len(v1_units) > 0, "Should create units for v1"
@@ -940,7 +969,8 @@ async def test_document_upsert_behavior(memory):
             bank_id=bank_id,
             content="Project is in development phase. Bob has joined as co-lead.",
             context="status update v2",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         assert len(v2_units) > 0, "Should create units for v2"
@@ -951,7 +981,8 @@ async def test_document_upsert_behavior(memory):
             query="What is the project status?",
             budget=Budget.MID,
             max_tokens=1000,
-            fact_type=["world"]
+            fact_type=["world"],
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall facts"
@@ -959,7 +990,7 @@ async def test_document_upsert_behavior(memory):
         print(f"✓ Document upsert created v1: {len(v1_units)} units, v2: {len(v2_units)} units")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 # ============================================================
@@ -967,7 +998,7 @@ async def test_document_upsert_behavior(memory):
 # ============================================================
 
 @pytest.mark.asyncio
-async def test_chunk_fact_mapping(memory):
+async def test_chunk_fact_mapping(memory, request_context):
     """
     Test that facts correctly reference their source chunks via chunk_id.
     """
@@ -990,7 +1021,8 @@ async def test_chunk_fact_mapping(memory):
             bank_id=bank_id,
             content=content,
             context="technical documentation",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create memory units"
@@ -1003,7 +1035,8 @@ async def test_chunk_fact_mapping(memory):
             max_tokens=1000,
             fact_type=["world"],
             include_chunks=True,
-            max_chunk_tokens=8192
+            max_chunk_tokens=8192,
+            request_context=request_context,
         )
 
         assert len(result.results) > 0, "Should recall facts"
@@ -1026,11 +1059,11 @@ async def test_chunk_fact_mapping(memory):
             print(f"  Returned {len(result.chunks)} chunks matching fact references")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_chunk_ordering_preservation(memory):
+async def test_chunk_ordering_preservation(memory, request_context):
     """
     Test that chunk_index reflects the correct order within a document.
     """
@@ -1070,7 +1103,8 @@ async def test_chunk_ordering_preservation(memory):
             bank_id=bank_id,
             content=content,
             context="multi-section document",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create units"
@@ -1083,7 +1117,8 @@ async def test_chunk_ordering_preservation(memory):
             max_tokens=2000,
             fact_type=["world"],
             include_chunks=True,
-            max_chunk_tokens=8192
+            max_chunk_tokens=8192,
+            request_context=request_context,
         )
 
         if result.chunks:
@@ -1103,11 +1138,11 @@ async def test_chunk_ordering_preservation(memory):
             print("✓ Content stored (may have created single chunk or no chunks returned)")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_chunks_truncation_behavior(memory):
+async def test_chunks_truncation_behavior(memory, request_context):
     """
     Test that when chunks exceed max_chunk_tokens, truncation is indicated.
     """
@@ -1165,7 +1200,8 @@ async def test_chunks_truncation_behavior(memory):
             bank_id=bank_id,
             content=large_content,
             context="large document test",
-            document_id=document_id
+            document_id=document_id,
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should create units"
@@ -1178,7 +1214,8 @@ async def test_chunks_truncation_behavior(memory):
             max_tokens=1000,
             fact_type=["world"],
             include_chunks=True,
-            max_chunk_tokens=500  # Small limit to test truncation
+            max_chunk_tokens=500,  # Small limit to test truncation
+            request_context=request_context,
         )
 
         if result.chunks:
@@ -1198,7 +1235,7 @@ async def test_chunks_truncation_behavior(memory):
             print("✓ No chunks returned (may be under token limit)")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 # ============================================================
@@ -1206,7 +1243,7 @@ async def test_chunks_truncation_behavior(memory):
 # ============================================================
 
 @pytest.mark.asyncio
-async def test_temporal_links_creation(memory):
+async def test_temporal_links_creation(memory, request_context):
     """
     Test that temporal links are created between facts with nearby event dates.
 
@@ -1223,7 +1260,8 @@ async def test_temporal_links_creation(memory):
             bank_id=bank_id,
             content="Alice started working on the authentication module.",
             context="daily standup",
-            event_date=base_date
+            event_date=base_date,
+            request_context=request_context,
         )
 
         # Fact 2 at 2:00 PM same day (4 hours later)
@@ -1231,7 +1269,8 @@ async def test_temporal_links_creation(memory):
             bank_id=bank_id,
             content="Bob reviewed the API design document.",
             context="daily standup",
-            event_date=base_date.replace(hour=14)
+            event_date=base_date.replace(hour=14),
+            request_context=request_context,
         )
 
         # Fact 3 at 9:00 AM next day (23 hours later)
@@ -1239,7 +1278,8 @@ async def test_temporal_links_creation(memory):
             bank_id=bank_id,
             content="Charlie deployed the new database schema.",
             context="daily standup",
-            event_date=base_date.replace(day=16, hour=9)
+            event_date=base_date.replace(day=16, hour=9),
+            request_context=request_context,
         )
 
         assert len(unit_ids_1) > 0 and len(unit_ids_2) > 0 and len(unit_ids_3) > 0
@@ -1278,11 +1318,11 @@ async def test_temporal_links_creation(memory):
             logger.info("Temporal links created successfully with proper weights")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_semantic_links_creation(memory):
+async def test_semantic_links_creation(memory, request_context):
     """
     Test that semantic links are created between facts with similar content.
 
@@ -1295,21 +1335,24 @@ async def test_semantic_links_creation(memory):
         unit_ids_1 = await memory.retain_async(
             bank_id=bank_id,
             content="Alice is an expert in Python programming and has built many web applications.",
-            context="team skills"
+            context="team skills",
+            request_context=request_context,
         )
 
         # Similar content - should create semantic link
         unit_ids_2 = await memory.retain_async(
             bank_id=bank_id,
             content="Bob is proficient in Python development and specializes in building APIs.",
-            context="team skills"
+            context="team skills",
+            request_context=request_context,
         )
 
         # Different content - less likely to create strong semantic link
         unit_ids_3 = await memory.retain_async(
             bank_id=bank_id,
             content="The quarterly sales meeting is scheduled for next Tuesday at 3 PM.",
-            context="calendar events"
+            context="calendar events",
+            request_context=request_context,
         )
 
         assert len(unit_ids_1) > 0 and len(unit_ids_2) > 0 and len(unit_ids_3) > 0
@@ -1349,11 +1392,11 @@ async def test_semantic_links_creation(memory):
             logger.info("Semantic links created successfully between similar content")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_entity_links_creation(memory):
+async def test_entity_links_creation(memory, request_context):
     """
     Test that entity links are created between facts that mention the same entities.
 
@@ -1367,28 +1410,32 @@ async def test_entity_links_creation(memory):
         unit_ids_1 = await memory.retain_async(
             bank_id=bank_id,
             content="Alice joined Google as a software engineer in 2020.",
-            context="career history"
+            context="career history",
+            request_context=request_context,
         )
 
         # Mentions same entity (Alice) - should create entity link
         unit_ids_2 = await memory.retain_async(
             bank_id=bank_id,
             content="Alice led the development of the new authentication system.",
-            context="project updates"
+            context="project updates",
+            request_context=request_context,
         )
 
         # Mentions same entity (Google) - should create entity link
         unit_ids_3 = await memory.retain_async(
             bank_id=bank_id,
             content="Google announced new cloud services at their annual conference.",
-            context="tech news"
+            context="tech news",
+            request_context=request_context,
         )
 
         # Different entities - no entity link expected
         unit_ids_4 = await memory.retain_async(
             bank_id=bank_id,
             content="Bob works at Meta on machine learning infrastructure.",
-            context="career history"
+            context="career history",
+            request_context=request_context,
         )
 
         assert len(unit_ids_1) > 0 and len(unit_ids_2) > 0 and len(unit_ids_3) > 0 and len(unit_ids_4) > 0
@@ -1445,11 +1492,11 @@ async def test_entity_links_creation(memory):
             logger.info("Entity links are properly bidirectional")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_causal_links_creation(memory):
+async def test_causal_links_creation(memory, request_context):
     """
     Test that causal links are created between facts with causal relationships.
 
@@ -1471,7 +1518,8 @@ async def test_causal_links_creation(memory):
         unit_ids = await memory.retain_async(
             bank_id=bank_id,
             content=content,
-            context="project timeline"
+            context="project timeline",
+            request_context=request_context,
         )
 
         assert len(unit_ids) > 0, "Should have created facts"
@@ -1517,11 +1565,11 @@ async def test_causal_links_creation(memory):
         logger.info("Test completed (causal link extraction is LLM-dependent)")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_all_link_types_together(memory):
+async def test_all_link_types_together(memory, request_context):
     """
     Integration test: Verify all link types can be created in a single retain operation.
 
@@ -1539,7 +1587,8 @@ async def test_all_link_types_together(memory):
             bank_id=bank_id,
             content="Alice completed the Python backend service for the authentication system.",
             context="sprint review",
-            event_date=base_date
+            event_date=base_date,
+            request_context=request_context,
         )
 
         # Fact 2: Related to Alice, similar topic (Python), close in time
@@ -1547,7 +1596,8 @@ async def test_all_link_types_together(memory):
             bank_id=bank_id,
             content="Alice optimized the Python code and improved the authentication performance by 40%.",
             context="sprint review",
-            event_date=base_date.replace(hour=14)  # Same day, 4 hours later
+            event_date=base_date.replace(hour=14),  # Same day, 4 hours later
+            request_context=request_context,
         )
 
         # Fact 3: Related to Alice, different topic but same entity
@@ -1555,7 +1605,8 @@ async def test_all_link_types_together(memory):
             bank_id=bank_id,
             content="Alice presented the security architecture at the team meeting.",
             context="team meeting",
-            event_date=base_date.replace(day=16)  # Next day
+            event_date=base_date.replace(day=16),  # Next day
+            request_context=request_context,
         )
 
         assert len(unit_ids_1) > 0 and len(unit_ids_2) > 0 and len(unit_ids_3) > 0
@@ -1594,11 +1645,11 @@ async def test_all_link_types_together(memory):
             logger.info("All major link types (temporal, semantic, entity) are working correctly")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_semantic_links_within_same_batch(memory):
+async def test_semantic_links_within_same_batch(memory, request_context):
     """
     Test that semantic links are created between facts retained in the SAME batch.
 
@@ -1617,7 +1668,8 @@ async def test_semantic_links_within_same_batch(memory):
 
         result = await memory.retain_batch_async(
             bank_id=bank_id,
-            contents=contents
+            contents=contents,
+            request_context=request_context,
         )
 
         # Flatten the list of lists
@@ -1652,11 +1704,11 @@ async def test_semantic_links_within_same_batch(memory):
                 logger.info(f"  Semantic link: {str(link['from_unit_id'])[:8]}... -> {str(link['to_unit_id'])[:8]}... (weight: {link['weight']:.3f})")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_temporal_links_within_same_batch(memory):
+async def test_temporal_links_within_same_batch(memory, request_context):
     """
     Test that temporal links are created between facts retained in the SAME batch.
 
@@ -1689,7 +1741,8 @@ async def test_temporal_links_within_same_batch(memory):
 
         result = await memory.retain_batch_async(
             bank_id=bank_id,
-            contents=contents
+            contents=contents,
+            request_context=request_context,
         )
 
         # Flatten the list of lists
@@ -1724,4 +1777,4 @@ async def test_temporal_links_within_same_batch(memory):
                 logger.info(f"  Temporal link: {str(link['from_unit_id'])[:8]}... -> {str(link['to_unit_id'])[:8]}... (weight: {link['weight']:.3f})")
 
     finally:
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)

hindsight-api/tests/test_schema_isolation.py

@@ -0,0 +1,406 @@
+"""
+Tests for multi-tenant schema isolation.
+
+Verifies that concurrent retain operations from different tenants
+are properly isolated in their respective PostgreSQL schemas.
+"""
+
+import asyncio
+import uuid
+
+import pytest
+import pytest_asyncio
+
+from hindsight_api.extensions import RequestContext, TenantContext, TenantExtension
+from hindsight_api.engine.memory_engine import _current_schema, fq_table
+from hindsight_api.migrations import run_migrations
+
+
+class MultiSchemaTestTenantExtension(TenantExtension):
+    """
+    Test tenant extension that maps API keys to schema names.
+
+    API keys are in format: "key-{schema_name}"
+    Provisions schemas on first access using run_migrations(schema=name).
+    """
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+        self.db_url = config.get("db_url")
+        # Pre-configured valid schemas for test
+        self.valid_schemas = config.get("valid_schemas", set())
+        # Track provisioned schemas
+        self._provisioned: set[str] = set()
+
+    async def authenticate(self, context: RequestContext) -> TenantContext:
+        if not context.api_key:
+            from hindsight_api.extensions import AuthenticationError
+
+            raise AuthenticationError("API key required")
+
+        # Parse schema from API key (format: "key-{schema}")
+        if context.api_key.startswith("key-"):
+            schema = context.api_key[4:]  # Remove "key-" prefix
+            if schema in self.valid_schemas:
+                # Provision schema on first access
+                if schema not in self._provisioned and self.db_url:
+                    run_migrations(self.db_url, schema=schema)
+                    self._provisioned.add(schema)
+                return TenantContext(schema_name=schema)
+
+        from hindsight_api.extensions import AuthenticationError
+
+        raise AuthenticationError(f"Unknown API key: {context.api_key}")
+
+
+async def drop_schema(conn, schema_name: str) -> None:
+    """Drop a schema and all its contents."""
+    await conn.execute(f'DROP SCHEMA IF EXISTS "{schema_name}" CASCADE')
+
+
+async def count_memories_in_schema(conn, schema_name: str, bank_id: str) -> int:
+    """Count memory units in a specific schema for a bank."""
+    result = await conn.fetchval(
+        f'SELECT COUNT(*) FROM "{schema_name}".memory_units WHERE bank_id = $1',
+        bank_id,
+    )
+    return result or 0
+
+
+async def get_memory_texts_in_schema(conn, schema_name: str, bank_id: str) -> list[str]:
+    """Get all memory texts in a specific schema for a bank."""
+    rows = await conn.fetch(
+        f'SELECT text FROM "{schema_name}".memory_units WHERE bank_id = $1 ORDER BY text',
+        bank_id,
+    )
+    return [row["text"] for row in rows]
+
+
+class TestSchemaIsolation:
+    """Tests for multi-tenant schema isolation."""
+
+    @pytest.mark.asyncio
+    async def test_concurrent_inserts_isolated_by_schema(self, memory, pg0_db_url):
+        """
+        Multiple concurrent database operations from different tenants
+        should store data in their respective schemas without cross-contamination.
+
+        Uses run_migrations(schema=x) to provision schemas like a real extension.
+        """
+        import asyncpg
+
+        # Test schemas
+        schemas = ["tenant_alpha", "tenant_beta", "tenant_gamma"]
+        bank_id = f"test-isolation-{uuid.uuid4().hex[:8]}"
+
+        # Clean up any existing schemas
+        conn = await asyncpg.connect(pg0_db_url)
+        try:
+            for schema in schemas:
+                await drop_schema(conn, schema)
+        finally:
+            await conn.close()
+
+        # Configure tenant extension that provisions schemas via run_migrations
+        tenant_ext = MultiSchemaTestTenantExtension({
+            "db_url": pg0_db_url,
+            "valid_schemas": set(schemas),
+        })
+        memory._tenant_extension = tenant_ext
+
+        # Define concurrent insert tasks for each tenant
+        async def insert_for_tenant(schema_name: str, content_prefix: str):
+            """Insert memories for a specific tenant using schema context."""
+            # Authenticate to set the schema context
+            tenant_request = RequestContext(api_key=f"key-{schema_name}")
+            await memory._authenticate_tenant(tenant_request)
+
+            # Now fq_table will use the correct schema
+            pool = await memory._get_pool()
+            from hindsight_api.engine.db_utils import acquire_with_retry
+
+            async with acquire_with_retry(pool) as conn:
+                # Insert 3 memories for this tenant
+                for i in range(3):
+                    await conn.execute(
+                        f"""
+                        INSERT INTO {fq_table('memory_units')} (bank_id, text, event_date, fact_type)
+                        VALUES ($1, $2, now(), 'world')
+                        """,
+                        bank_id,
+                        f"MARKER_{content_prefix}_DOC{i}: Memory for {schema_name}",
+                    )
+
+        # Run concurrent inserts for all tenants
+        await asyncio.gather(
+            insert_for_tenant("tenant_alpha", "ALPHA"),
+            insert_for_tenant("tenant_beta", "BETA"),
+            insert_for_tenant("tenant_gamma", "GAMMA"),
+        )
+
+        # Verify isolation - each schema should only have its own data
+        conn = await asyncpg.connect(pg0_db_url)
+        try:
+            for schema in schemas:
+                texts = await get_memory_texts_in_schema(conn, schema, bank_id)
+                prefix = schema.replace("tenant_", "").upper()
+
+                # Should have exactly 3 memories
+                assert len(texts) == 3, f"Schema {schema} should have 3 memories, got {len(texts)}"
+
+                # All texts should contain the schema's marker
+                for text in texts:
+                    assert f"MARKER_{prefix}" in text, (
+                        f"Memory in {schema} missing its marker: {text}"
+                    )
+
+                # Should NOT contain other tenants' markers
+                other_prefixes = ["ALPHA", "BETA", "GAMMA"]
+                other_prefixes.remove(prefix)
+                for other in other_prefixes:
+                    for text in texts:
+                        assert f"MARKER_{other}" not in text, (
+                            f"Cross-contamination! Schema {schema} has {other}'s marker: {text}"
+                        )
+
+        finally:
+            # Cleanup
+            for schema in schemas:
+                await drop_schema(conn, schema)
+            await conn.close()
+
+        # Reset tenant extension
+        memory._tenant_extension = None
+        _current_schema.set("public")
+
+    @pytest.mark.asyncio
+    async def test_schema_context_isolation_in_concurrent_tasks(self, pg0_db_url):
+        """
+        Verify that _current_schema contextvar is properly isolated
+        between concurrent async tasks.
+        """
+        results = {}
+        errors = []
+
+        async def check_schema_context(schema_name: str, delay: float):
+            """Set schema context, wait, then verify it's still correct."""
+            try:
+                # Set the schema
+                _current_schema.set(schema_name)
+
+                # Small delay to allow interleaving
+                await asyncio.sleep(delay)
+
+                # Verify schema is still correct
+                current = _current_schema.get()
+                if current != schema_name:
+                    errors.append(f"Expected {schema_name}, got {current}")
+
+                # Verify fq_table uses correct schema
+                table = fq_table("memory_units")
+                expected = f"{schema_name}.memory_units"
+                if table != expected:
+                    errors.append(f"Expected {expected}, got {table}")
+
+                results[schema_name] = current
+
+            except Exception as e:
+                errors.append(f"Error in {schema_name}: {e}")
+
+        # Run many concurrent tasks with different schemas
+        tasks = []
+        for i in range(10):
+            for schema in ["schema_a", "schema_b", "schema_c"]:
+                # Vary delays to create interleaving
+                delay = 0.01 * (i % 3)
+                tasks.append(check_schema_context(f"{schema}_{i}", delay))
+
+        await asyncio.gather(*tasks)
+
+        # No errors should have occurred
+        assert not errors, f"Schema context isolation errors: {errors}"
+
+    @pytest.mark.asyncio
+    async def test_list_memories_respects_schema(self, memory, pg0_db_url):
+        """
+        list_memory_units should only return memories from the current schema.
+
+        Uses run_migrations(schema=x) to provision schemas.
+        """
+        import asyncpg
+
+        schemas = ["tenant_list_a", "tenant_list_b"]
+        bank_id = f"test-list-{uuid.uuid4().hex[:8]}"
+
+        # Clean up any existing schemas and provision via migrations
+        conn = await asyncpg.connect(pg0_db_url)
+        try:
+            for schema in schemas:
+                await drop_schema(conn, schema)
+        finally:
+            await conn.close()
+
+        # Provision schemas using run_migrations
+        for schema in schemas:
+            run_migrations(pg0_db_url, schema=schema)
+
+        # Insert test data directly into each schema
+        conn = await asyncpg.connect(pg0_db_url)
+        try:
+            for schema in schemas:
+                await conn.execute(
+                    f"""
+                    INSERT INTO "{schema}".memory_units (bank_id, text, event_date, fact_type)
+                    VALUES ($1, $2, now(), 'world')
+                    """,
+                    bank_id,
+                    f"Direct insert for {schema}",
+                )
+        finally:
+            await conn.close()
+
+        # Configure tenant extension
+        tenant_ext = MultiSchemaTestTenantExtension({
+            "db_url": pg0_db_url,
+            "valid_schemas": set(schemas),
+        })
+        memory._tenant_extension = tenant_ext
+
+        try:
+            # Query as tenant_list_a - should only see tenant_list_a's data
+            tenant_a_request = RequestContext(api_key="key-tenant_list_a")
+            await memory._authenticate_tenant(tenant_a_request)
+
+            result_a = await memory.list_memory_units(bank_id=bank_id, request_context=tenant_a_request)
+            texts_a = [item["text"] for item in result_a.get("items", [])]
+
+            assert len(texts_a) == 1, f"Expected 1 memory for tenant_list_a, got {len(texts_a)}"
+            assert "tenant_list_a" in texts_a[0], f"Wrong content: {texts_a[0]}"
+
+            # Query as tenant_list_b - should only see tenant_list_b's data
+            tenant_b_request = RequestContext(api_key="key-tenant_list_b")
+            await memory._authenticate_tenant(tenant_b_request)
+
+            result_b = await memory.list_memory_units(bank_id=bank_id, request_context=tenant_b_request)
+            texts_b = [item["text"] for item in result_b.get("items", [])]
+
+            assert len(texts_b) == 1, f"Expected 1 memory for tenant_list_b, got {len(texts_b)}"
+            assert "tenant_list_b" in texts_b[0], f"Wrong content: {texts_b[0]}"
+
+        finally:
+            # Cleanup
+            conn = await asyncpg.connect(pg0_db_url)
+            try:
+                for schema in schemas:
+                    await drop_schema(conn, schema)
+            finally:
+                await conn.close()
+
+            memory._tenant_extension = None
+            _current_schema.set("public")
+
+    @pytest.mark.asyncio
+    async def test_high_concurrency_schema_isolation(self, memory, pg0_db_url):
+        """
+        Stress test: Many concurrent operations across multiple schemas
+        should maintain perfect isolation.
+
+        Uses run_migrations(schema=x) to provision schemas like a real extension.
+        """
+        import asyncpg
+
+        # Create more schemas for stress test
+        num_schemas = 5
+        ops_per_schema = 10
+        schemas = [f"stress_tenant_{i}" for i in range(num_schemas)]
+        bank_id = f"test-stress-{uuid.uuid4().hex[:8]}"
+
+        # Clean up any existing schemas first
+        conn = await asyncpg.connect(pg0_db_url)
+        try:
+            for schema in schemas:
+                await drop_schema(conn, schema)
+        finally:
+            await conn.close()
+
+        # Provision schemas using run_migrations
+        for schema in schemas:
+            run_migrations(pg0_db_url, schema=schema)
+
+        # Configure tenant extension (schemas already provisioned)
+        tenant_ext = MultiSchemaTestTenantExtension({
+            "db_url": pg0_db_url,
+            "valid_schemas": set(schemas),
+        })
+        # Mark schemas as already provisioned so extension doesn't re-run migrations
+        tenant_ext._provisioned = set(schemas)
+        memory._tenant_extension = tenant_ext
+
+        errors = []
+
+        async def insert_one(schema: str, item_id: int):
+            """Single insert operation for tracking."""
+            try:
+                # Authenticate to set the schema context
+                tenant_request = RequestContext(api_key=f"key-{schema}")
+                await memory._authenticate_tenant(tenant_request)
+
+                # Insert using fq_table
+                pool = await memory._get_pool()
+                from hindsight_api.engine.db_utils import acquire_with_retry
+
+                async with acquire_with_retry(pool) as conn:
+                    await conn.execute(
+                        f"""
+                        INSERT INTO {fq_table('memory_units')} (bank_id, text, event_date, fact_type)
+                        VALUES ($1, $2, now(), 'world')
+                        """,
+                        bank_id,
+                        f"STRESS_MARKER_{schema}_ITEM{item_id}: Memory for {schema}",
+                    )
+            except Exception as e:
+                errors.append(f"Insert error for {schema}: {e}")
+
+        # Run many concurrent operations
+        tasks = []
+        for i in range(ops_per_schema):
+            for schema in schemas:
+                tasks.append(insert_one(schema, i))
+
+        await asyncio.gather(*tasks)
+
+        # Check for errors during insert
+        assert not errors, f"Errors during insert: {errors}"
+
+        # Verify no cross-contamination
+        conn = await asyncpg.connect(pg0_db_url)
+        try:
+            for schema in schemas:
+                texts = await get_memory_texts_in_schema(conn, schema, bank_id)
+
+                # Should have exactly ops_per_schema memories
+                assert len(texts) == ops_per_schema, (
+                    f"Schema {schema} should have {ops_per_schema} memories, got {len(texts)}"
+                )
+
+                # All memories should reference this schema only
+                for text in texts:
+                    # Check it contains our schema marker
+                    assert f"STRESS_MARKER_{schema}" in text, (
+                        f"Memory in {schema} doesn't contain schema marker: {text}"
+                    )
+
+                    # Check it doesn't contain other schema markers
+                    for other_schema in schemas:
+                        if other_schema != schema:
+                            assert f"STRESS_MARKER_{other_schema}" not in text, (
+                                f"Cross-contamination! {schema} has {other_schema}'s data: {text}"
+                            )
+        finally:
+            # Cleanup
+            for schema in schemas:
+                await drop_schema(conn, schema)
+            await conn.close()
+
+        memory._tenant_extension = None
+        _current_schema.set("public")

hindsight-api/tests/test_search_trace.py

@@ -3,12 +3,12 @@ Test search tracing functionality.
 """
 import pytest
 from hindsight_api.engine.memory_engine import Budget
-from hindsight_api import SearchTrace
+from hindsight_api import SearchTrace, RequestContext
 from datetime import datetime, timezone
 
 
 @pytest.mark.asyncio
-async def test_search_with_trace(memory):
+async def test_search_with_trace(memory, request_context):
     """Test that search with enable_trace=True returns a valid SearchTrace."""
     # Generate a unique agent ID for this test
     bank_id = f"test_trace_{datetime.now(timezone.utc).timestamp()}"
@@ -20,16 +20,19 @@ async def test_search_with_trace(memory):
             bank_id=bank_id,
             content="Alice works at Google in Mountain View",
             context="test context",
+            request_context=request_context,
         )
         await memory.retain_async(
             bank_id=bank_id,
             content="Bob also works at Google but in New York",
             context="test context",
+            request_context=request_context,
         )
         await memory.retain_async(
             bank_id=bank_id,
             content="Charlie founded a startup called TechCorp",
             context="test context",
+            request_context=request_context,
         )
 
         # Search with tracing enabled
@@ -40,6 +43,7 @@ async def test_search_with_trace(memory):
             budget=Budget.LOW, # 20,
             max_tokens=512,
             enable_trace=True,
+            request_context=request_context,
         )
 
         # Verify results
@@ -102,11 +106,11 @@ async def test_search_with_trace(memory):
 
     finally:
         # Cleanup
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)
 
 
 @pytest.mark.asyncio
-async def test_search_without_trace(memory):
+async def test_search_without_trace(memory, request_context):
     """Test that search with enable_trace=False returns None for trace."""
     bank_id = f"test_no_trace_{datetime.now(timezone.utc).timestamp()}"
 
@@ -117,6 +121,7 @@ async def test_search_without_trace(memory):
             bank_id=bank_id,
             content="Test memory without trace",
             context="test",
+            request_context=request_context,
         )
 
         # Search without tracing
@@ -127,6 +132,7 @@ async def test_search_without_trace(memory):
             budget=Budget.LOW, # 10,
             max_tokens=512,
             enable_trace=False,
+            request_context=request_context,
         )
 
         # Verify trace is None
@@ -137,4 +143,4 @@ async def test_search_without_trace(memory):
 
     finally:
         # Cleanup
-        await memory.delete_bank(bank_id)
+        await memory.delete_bank(bank_id, request_context=request_context)

hindsight-api/tests/test_sql_schema_safety.py

@@ -0,0 +1,136 @@
+"""
+Safety tests to ensure all SQL queries use fully-qualified table names.
+
+This prevents cross-tenant data access by ensuring every table reference
+includes the schema prefix (e.g., public.memory_units instead of just memory_units).
+"""
+
+import re
+from pathlib import Path
+
+import pytest
+
+# All tables that MUST be schema-qualified in SQL queries
+TABLES = [
+    "memory_units",
+    "memory_links",
+    "unit_entities",
+    "entities",
+    "entity_cooccurrences",
+    "banks",
+    "documents",
+    "chunks",
+    "async_operations",
+]
+
+# Files to scan for SQL queries
+SCAN_PATHS = [
+    "hindsight_api/engine",
+    "hindsight_api/api",
+]
+
+# Files to exclude (e.g., migrations, tests)
+EXCLUDE_PATTERNS = [
+    "alembic",
+    "__pycache__",
+    "test_",
+]
+
+
+def get_python_files() -> list[Path]:
+    """Get all Python files to scan."""
+    root = Path(__file__).parent.parent
+    files = []
+    for scan_path in SCAN_PATHS:
+        path = root / scan_path
+        if path.exists():
+            for py_file in path.rglob("*.py"):
+                # Check exclusions
+                if any(excl in str(py_file) for excl in EXCLUDE_PATTERNS):
+                    continue
+                files.append(py_file)
+    return files
+
+
+def find_unqualified_table_refs(content: str, filename: str) -> list[tuple[int, str, str]]:
+    """
+    Find SQL statements with unqualified table references.
+
+    Returns list of (line_number, table_name, line_content).
+    """
+    violations = []
+
+    # Patterns that indicate SQL context
+    sql_keywords = r"(?:FROM|JOIN|INTO|UPDATE|DELETE\s+FROM)\s+"
+
+    # Additional SQL indicators to confirm this is actually SQL, not prose
+    sql_indicators = re.compile(
+        r"(SELECT|INSERT|DELETE|UPDATE|CREATE|ALTER|DROP|WHERE|SET|VALUES|"
+        r'f"""|f\'\'\'|""".*SELECT|\'\'\'.*SELECT)',
+        re.IGNORECASE,
+    )
+
+    lines = content.split("\n")
+    for line_num, line in enumerate(lines, 1):
+        # Skip comments and strings that are clearly not SQL
+        stripped = line.strip()
+        if stripped.startswith("#"):
+            continue
+
+        for table in TABLES:
+            # Pattern: SQL keyword followed by unqualified table name
+            # Should match: FROM memory_units, JOIN memory_units, INTO memory_units
+            # Should NOT match: FROM public.memory_units, FROM {schema}.memory_units
+            # Should NOT match: fq_table("memory_units")
+
+            # Check for unqualified table after SQL keyword
+            pattern = rf"{sql_keywords}{table}(?:\s|$|,|\))"
+
+            if re.search(pattern, line, re.IGNORECASE):
+                # Check if it's actually qualified (has schema prefix)
+                qualified_pattern = rf"\.\s*{table}(?:\s|$|,|\))"
+                fq_table_pattern = rf'fq_table\s*\(\s*["\']?{table}'
+
+                if not re.search(qualified_pattern, line) and not re.search(
+                    fq_table_pattern, line
+                ):
+                    # Additional check: line must have SQL indicators
+                    # This avoids false positives in docstrings like "split into chunks"
+                    if sql_indicators.search(line):
+                        violations.append((line_num, table, stripped))
+
+    return violations
+
+
+class TestSQLSchemaSafety:
+    """Ensure all SQL uses schema-qualified table names."""
+
+    def test_no_unqualified_table_references(self):
+        """All SQL queries must use fq_table() or schema.table format."""
+        all_violations = []
+
+        for py_file in get_python_files():
+            content = py_file.read_text()
+            violations = find_unqualified_table_refs(content, py_file.name)
+
+            for line_num, table, line in violations:
+                all_violations.append(
+                    f"{py_file.relative_to(py_file.parent.parent)}:{line_num} - "
+                    f"unqualified '{table}': {line[:80]}..."
+                )
+
+        if all_violations:
+            msg = (
+                f"Found {len(all_violations)} unqualified table references!\n"
+                "These could cause cross-tenant data access.\n"
+                "Use fq_table('table_name') for all table references.\n\n"
+                + "\n".join(all_violations[:20])  # Show first 20
+            )
+            if len(all_violations) > 20:
+                msg += f"\n... and {len(all_violations) - 20} more"
+            pytest.fail(msg)
+
+    def test_tables_list_is_complete(self):
+        """Verify we're checking for all tables (sanity check)."""
+        # This is a sanity check - if you add a new table, add it to TABLES
+        assert len(TABLES) >= 9, "Update TABLES list if you added new tables"