chore: update Rust client Cargo.lock for CI verification

Run generate-clients.sh to sync Cargo.lock with current dependencies.

.env.example

@@ -2,10 +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=gpt-4o-mini
+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
@@ -13,3 +26,19 @@ HINDSIGHT_API_LOG_LEVEL=info
 
 # Database (Optional - uses embedded pg0 by default)
 # HINDSIGHT_API_DATABASE_URL=postgresql://user:pass@host:5432/db
+
+# Embeddings Configuration (Optional - uses local by default)
+# Provider: "local" (default) or "tei" (HuggingFace Text Embeddings Inference)
+# HINDSIGHT_API_EMBEDDINGS_PROVIDER=local
+# For local provider:
+# HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL=BAAI/bge-small-en-v1.5
+# For TEI provider:
+# HINDSIGHT_API_EMBEDDINGS_TEI_URL=http://localhost:8080
+
+# Reranker Configuration (Optional - uses local by default)
+# Provider: "local" (default) or "tei" (HuggingFace Text Embeddings Inference)
+# HINDSIGHT_API_RERANKER_PROVIDER=local
+# For local provider:
+# HINDSIGHT_API_RERANKER_LOCAL_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2
+# For TEI provider:
+# HINDSIGHT_API_RERANKER_TEI_URL=http://localhost:8081

.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/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,71 @@
+name: Bug Report
+description: Report a bug or unexpected behavior
+labels: ["bug", "triage"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to report a bug! Please fill out the sections below.
+
+  - type: textarea
+    id: description
+    attributes:
+      label: Bug Description
+      description: A clear and concise description of the bug
+      placeholder: What happened?
+    validations:
+      required: true
+
+  - type: textarea
+    id: reproduction
+    attributes:
+      label: Steps to Reproduce
+      description: Steps to reproduce the behavior
+      placeholder: |
+        1. Configure '...'
+        2. Call '...'
+        3. See error
+    validations:
+      required: true
+
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected Behavior
+      description: What did you expect to happen?
+    validations:
+      required: true
+
+  - type: textarea
+    id: actual
+    attributes:
+      label: Actual Behavior
+      description: What actually happened?
+    validations:
+      required: true
+
+  - type: input
+    id: version
+    attributes:
+      label: Version
+      description: What version are you using?
+      placeholder: e.g., 0.1.0 or commit hash
+    validations:
+      required: false
+
+  - type: dropdown
+    id: llm-provider
+    attributes:
+      label: LLM Provider
+      description: Which LLM provider are you using?
+      options:
+        - OpenAI
+        - Anthropic
+        - Gemini
+        - Groq
+        - Ollama
+        - LM Studio
+        - Other
+    validations:
+      required: false
+

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,8 @@
+blank_issues_enabled: false
+contact_links:
+  - name: Questions & Help
+    url: https://github.com/vectorize-io/hindsight/discussions/categories/q-a
+    about: Please ask questions and get help in Discussions instead of opening an issue.
+  - name: Ideas & Feedback
+    url: https://github.com/vectorize-io/hindsight/discussions/categories/ideas
+    about: Share ideas or give feedback in Discussions.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,82 @@
+name: Feature Request
+description: Suggest a new feature or enhancement
+labels: ["enhancement", "triage"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for suggesting a feature! Please describe what you'd like to see added.
+
+  - type: textarea
+    id: use-case
+    attributes:
+      label: Use Case
+      description: Describe your specific use case. What are you building? What's your goal?
+      placeholder: |
+        I'm building an AI agent that needs to...
+        My application handles...
+    validations:
+      required: true
+
+  - type: textarea
+    id: problem
+    attributes:
+      label: Problem Statement
+      description: What problem are you facing? What's missing or difficult today?
+      placeholder: Currently I have to... which causes...
+    validations:
+      required: true
+
+  - type: textarea
+    id: benefit
+    attributes:
+      label: How This Feature Would Help
+      description: Explain how this feature would improve your workflow or solve your problem
+      placeholder: With this feature, I would be able to...
+    validations:
+      required: true
+
+  - type: textarea
+    id: solution
+    attributes:
+      label: Proposed Solution
+      description: Describe your ideal solution (optional - we may have ideas too!)
+      placeholder: It would be great if Hindsight could...
+    validations:
+      required: false
+
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives Considered
+      description: Have you considered any alternative solutions or workarounds?
+    validations:
+      required: false
+
+  - type: dropdown
+    id: priority
+    attributes:
+      label: Priority
+      description: How important is this feature to you?
+      options:
+        - Nice to have
+        - Important - affects my workflow
+        - Critical - blocking my use case
+    validations:
+      required: true
+
+  - type: textarea
+    id: additional
+    attributes:
+      label: Additional Context
+      description: Any other context, mockups, or examples?
+    validations:
+      required: false
+
+  - type: checkboxes
+    id: checklist
+    attributes:
+      label: Checklist
+      options:
+        - label: I would be willing to contribute this feature
+          required: false

.github/workflows/deploy-docs.yml

@@ -20,18 +20,17 @@ concurrency:
 jobs:
   build:
     runs-on: ubuntu-latest
-    defaults:
-      run:
-        working-directory: hindsight-docs
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-node@v4
         with:
           node-version: 20
           cache: npm
-          cache-dependency-path: hindsight-docs/package-lock.json
-      - run: npm ci
-      - run: npm run build
+          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:
           path: hindsight-docs/build

.github/workflows/release.yml

@@ -38,6 +38,14 @@ jobs:
       working-directory: ./hindsight
       run: uv build --out-dir dist
 
+    - name: Build hindsight-litellm
+      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
@@ -57,6 +65,18 @@ jobs:
         packages-dir: ./hindsight/dist
         skip-existing: true
 
+    - name: Publish hindsight-litellm to PyPI
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        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
@@ -66,6 +86,8 @@ jobs:
           hindsight-clients/python/dist/*
           hindsight-api/dist/*
           hindsight/dist/*
+          hindsight-integrations/litellm/dist/*
+          hindsight-embed/dist/*
         retention-days: 1
 
   release-typescript-client:
@@ -80,18 +102,29 @@ jobs:
       with:
         node-version: '20'
         registry-url: 'https://registry.npmjs.org'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
 
     - name: Install dependencies
-      working-directory: ./hindsight-clients/typescript
-      run: npm ci
+      run: npm ci --workspace=hindsight-clients/typescript
 
     - name: Build
-      working-directory: ./hindsight-clients/typescript
-      run: npm run build
+      run: npm run build --workspace=hindsight-clients/typescript
 
     - 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 }}
 
@@ -106,6 +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: ${{ matrix.os }}
     strategy:
@@ -170,7 +262,7 @@ jobs:
     - name: Free Disk Space
       uses: jlumbroso/free-disk-space@main
       with:
-        tool-cache: false
+        tool-cache: true
         android: true
         dotnet: true
         haskell: true
@@ -195,7 +287,7 @@ jobs:
       id: get_version
       run: echo "VERSION=${GITHUB_REF#refs/tags/v}" >> $GITHUB_OUTPUT
 
-    - name: Extract metadata
+    - name: Extract metadata for release tags
       id: meta
       uses: docker/metadata-action@v5
       with:
@@ -206,7 +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: .
@@ -219,6 +333,9 @@ jobs:
 
   release-helm-chart:
     runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      packages: write
 
     steps:
     - uses: actions/checkout@v4
@@ -228,12 +345,18 @@ jobs:
       with:
         version: 'latest'
 
+    - name: Log in to GHCR
+      run: echo "${{ secrets.GITHUB_TOKEN }}" | helm registry login ghcr.io -u ${{ github.actor }} --password-stdin
+
     - name: Lint Helm chart
       run: helm lint helm/hindsight
 
     - name: Package Helm chart
       run: helm package helm/hindsight --destination ./helm-packages
 
+    - name: Push to GHCR OCI
+      run: helm push helm-packages/*.tgz oci://ghcr.io/${{ github.repository_owner }}/charts
+
     - name: Upload artifacts
       uses: actions/upload-artifact@v4
       with:
@@ -243,7 +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
 
@@ -266,6 +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:
@@ -297,8 +426,12 @@ jobs:
         cp artifacts/python-packages/hindsight-clients/python/dist/* release-assets/ || true
         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
@@ -307,54 +440,11 @@ jobs:
         cp artifacts/helm-chart/*.tgz release-assets/ || true
         ls -la release-assets/
 
-    - name: Generate release notes
-      run: |
-        cat << 'EOF' > release-notes.md
-        ## Quick Start
-
-        ```bash
-        # Install the CLI
-        curl -fsSL https://raw.githubusercontent.com/vectorize-io/hindsight/refs/heads/main/hindsight-cli/install.sh | bash
-
-        # Start the server
-        docker run -p 8888:8888 -p 9999:9999 \
-          -e HINDSIGHT_API_LLM_PROVIDER=openai \
-          -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
-          -e HINDSIGHT_API_LLM_MODEL=gpt-4o-mini \
-          ghcr.io/${{ github.repository_owner }}/hindsight:${{ steps.get_version.outputs.VERSION }}
-        ```
-
-        ## Docker Images
-        - `ghcr.io/${{ github.repository_owner }}/hindsight:${{ steps.get_version.outputs.VERSION }}` - Standalone (recommended)
-        - `ghcr.io/${{ github.repository_owner }}/hindsight-api:${{ steps.get_version.outputs.VERSION }}` - API only
-        - `ghcr.io/${{ github.repository_owner }}/hindsight-control-plane:${{ steps.get_version.outputs.VERSION }}` - Web UI only
-
-        ## CLI
-        ```bash
-        curl -fsSL https://raw.githubusercontent.com/vectorize-io/hindsight/refs/heads/main/hindsight-cli/install.sh | bash
-        ```
-
-        ## Python
-        ```bash
-        pip install hindsight-all  # or hindsight-api, hindsight-client
-        ```
-
-        ## TypeScript/JavaScript
-        ```bash
-        npm install @vectorize-io/hindsight-client
-        ```
-
-        ## Helm
-        ```bash
-        helm install hindsight oci://ghcr.io/${{ github.repository_owner }}/charts/hindsight --version ${{ steps.get_version.outputs.VERSION }}
-        ```
-        EOF
-
     - name: Create GitHub Release
       uses: softprops/action-gh-release@v2
       with:
         files: release-assets/*
-        body_path: release-notes.md
+        generate_release_notes: true
         draft: false
         prerelease: false
       env:

.github/workflows/test.yml

@@ -4,6 +4,10 @@ on:
   pull_request:
     branches: [ main ]
 
+concurrency:
+  group: ci-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   build-python-packages:
     runs-on: ubuntu-latest
@@ -16,6 +20,8 @@ jobs:
             path: hindsight-api
           - name: hindsight-client
             path: hindsight-clients/python
+          - name: hindsight-embed
+            path: hindsight-embed
 
     steps:
     - uses: actions/checkout@v4
@@ -34,6 +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
 
@@ -44,14 +73,66 @@ jobs:
       uses: actions/setup-node@v4
       with:
         node-version: '20'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
 
     - name: Install dependencies
-      working-directory: ./hindsight-clients/typescript
-      run: npm ci
+      run: npm ci --workspace=hindsight-clients/typescript
 
     - name: Build TypeScript client
-      working-directory: ./hindsight-clients/typescript
-      run: npm run build
+      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
@@ -63,14 +144,14 @@ jobs:
       uses: actions/setup-node@v4
       with:
         node-version: '20'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
 
     - name: Install dependencies
-      working-directory: ./hindsight-docs
-      run: npm ci
+      run: npm ci --workspace=hindsight-docs
 
     - name: Build docs
-      working-directory: ./hindsight-docs
-      run: npm run build
+      run: npm run build --workspace=hindsight-docs
 
   build-rust-cli:
     runs-on: ubuntu-latest
@@ -94,6 +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
 
@@ -126,7 +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
@@ -144,14 +309,587 @@ 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
-    needs: [build-python-packages]
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
+      OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+      HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
+      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: Build API
+      working-directory: ./hindsight-api
+      run: uv build
+
+    - name: Install dependencies
+      working-directory: ./hindsight-api
+      run: uv sync --extra test --no-install-project --index-strategy unsafe-best-match
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-
+
+    - name: Pre-download models
+      working-directory: ./hindsight-api
+      run: |
+        uv run 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 downloaded successfully')
+        "
+
+    - name: Run tests
+      working-directory: ./hindsight-api
+      run: uv run pytest tests -v
+
+  test-python-client:
+    runs-on: ubuntu-latest
     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: ${{ secrets.GITHUB_TOKEN }}
+      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Build API
+      working-directory: ./hindsight-api
+      run: uv build
+
+    - name: Build Python client
+      working-directory: ./hindsight-clients/python
+      run: uv build
+
+    - name: Install client test dependencies
+      working-directory: ./hindsight-clients/python
+      run: uv sync --extra test --index-strategy unsafe-best-match
+
+    - name: Install API dependencies
+      working-directory: ./hindsight-api
+      run: uv sync --no-install-project --index-strategy unsafe-best-match
+
+    - name: Create .env file
+      run: |
+        cat > .env << EOF
+        HINDSIGHT_API_LLM_PROVIDER=${{ 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 Python client tests
+      working-directory: ./hindsight-clients/python
+      run: uv run pytest tests -v
+
+    - name: Show API server logs
+      if: always()
+      run: |
+        echo "=== API Server Logs ==="
+        cat /tmp/api-server.log || echo "No API server log found"
+
+  test-typescript-client:
+    runs-on: ubuntu-latest
+    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: ${{ secrets.GITHUB_TOKEN }}
+      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+
+    - name: Build API
+      working-directory: ./hindsight-api
+      run: uv build
+
+    - name: Install API dependencies
+      working-directory: ./hindsight-api
+      run: uv sync --no-install-project --index-strategy unsafe-best-match
+
+    - name: Install TypeScript client dependencies
+      working-directory: ./hindsight-clients/typescript
+      run: npm ci
+
+    - name: Build TypeScript client
+      working-directory: ./hindsight-clients/typescript
+      run: npm run build
+
+    - name: Create .env file
+      run: |
+        cat > .env << EOF
+        HINDSIGHT_API_LLM_PROVIDER=${{ 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 TypeScript client tests
+      working-directory: ./hindsight-clients/typescript
+      run: npm test
+
+    - name: Show API server logs
+      if: always()
+      run: |
+        echo "=== API Server Logs ==="
+        cat /tmp/api-server.log || echo "No API server log found"
+
+  test-rust-client:
+    runs-on: ubuntu-latest
+    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 }}
+      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
+      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 Rust
+      uses: dtolnay/rust-toolchain@stable
+
+    - name: Cache cargo
+      uses: actions/cache@v4
+      with:
+        path: |
+          ~/.cargo/registry
+          ~/.cargo/git
+          hindsight-clients/rust/target
+        key: ${{ runner.os }}-cargo-client-${{ hashFiles('hindsight-clients/rust/Cargo.lock') }}
+
+    - 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 Rust client tests
+      working-directory: ./hindsight-clients/rust
+      run: cargo test --lib
+
+    - name: Show API server logs
+      if: always()
+      run: |
+        echo "=== API Server Logs ==="
+        cat /tmp/api-server.log || echo "No API server log found"
+
+  test-integration:
+    runs-on: ubuntu-latest
+    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: 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: Install integration test dependencies
+      working-directory: ./hindsight-integration-tests
+      run: uv sync
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-
+
+    - name: Pre-download models
+      working-directory: ./hindsight-api
+      run: |
+        uv run python -c "
+        from sentence_transformers import SentenceTransformer, CrossEncoder
+        print('Downloading embedding model...')
+        SentenceTransformer('BAAI/bge-small-en-v1.5')
+        print('Downloading cross-encoder model...')
+        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
+        print('Models downloaded successfully')
+        "
+
+    - name: Create .env file
+      run: |
+        cat > .env << EOF
+        HINDSIGHT_API_LLM_PROVIDER=${{ 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 integration tests
+      working-directory: ./hindsight-integration-tests
+      run: uv run pytest tests/ -v
+
+    - name: Show API server logs
+      if: always()
+      run: |
+        echo "=== API Server Logs ==="
+        cat /tmp/api-server.log || echo "No API server log found"
+
+  test-litellm-integration:
+    runs-on: ubuntu-latest
+
+    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: Build litellm integration
+      working-directory: ./hindsight-integrations/litellm
+      run: uv build
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/litellm
+      run: uv sync --extra dev
+
+    - name: Run tests
+      working-directory: ./hindsight-integrations/litellm
+      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"
+
+  verify-generated-files:
+    runs-on: ubuntu-latest
+    env:
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
 
     steps:
     - uses: actions/checkout@v4
@@ -166,10 +904,56 @@ jobs:
       with:
         python-version-file: ".python-version"
 
-    - name: Install dependencies
-      working-directory: ./hindsight-api
-      run: uv sync --extra test
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
 
-    - name: Run tests
-      working-directory: ./hindsight-api
-      run: uv run pytest tests -v
+    - name: Install Rust
+      uses: dtolnay/rust-toolchain@stable
+
+    - name: Cache cargo
+      uses: actions/cache@v4
+      with:
+        path: |
+          ~/.cargo/registry
+          ~/.cargo/git
+        key: ${{ runner.os }}-cargo-gen-${{ hashFiles('**/Cargo.lock') }}
+
+    - name: Install Node dependencies
+      run: npm ci
+
+    - name: Install Python dependencies
+      run: |
+        cd hindsight-dev && uv sync --index-strategy unsafe-best-match
+        cd ../hindsight-api && uv sync --index-strategy unsafe-best-match
+        cd ../hindsight-embed && uv sync --index-strategy unsafe-best-match
+
+    - name: Run generate-openapi
+      run: ./scripts/generate-openapi.sh
+
+    - name: Run generate-clients
+      run: ./scripts/generate-clients.sh
+
+    - name: Run lint
+      run: ./scripts/hooks/lint.sh
+
+    - name: Verify no uncommitted changes
+      run: |
+        if [ -n "$(git status --porcelain)" ]; then
+          echo "❌ Error: Generated files are out of sync with committed files."
+          echo ""
+          echo "The following files have changed after running generation scripts:"
+          git status --porcelain
+          echo ""
+          echo "Please run the following commands locally and commit the changes:"
+          echo "  ./scripts/generate-openapi.sh"
+          echo "  ./scripts/generate-clients.sh"
+          echo "  ./scripts/hooks/lint.sh"
+          echo ""
+          git diff --stat
+          exit 1
+        fi
+        echo "✓ All generated files are up to date"

.gitignore

@@ -9,8 +9,13 @@ wheels/
 # Virtual environments
 .venv
 
-# Environment variables
+# Node
+node_modules/
+
+# Environment variables and local config
 .env
+docker-compose.yml
+docker-compose.override.yml
 
 # IDE
 .idea/
@@ -29,6 +34,9 @@ logs/
 
 .DS_Store
 
+# Generated docs files
+hindsight-docs/static/llms-full.txt
+
 
 hindsight-dev/benchmarks/locomo/results/
 hindsight-dev/benchmarks/longmemeval/results/

AGENTS.md

@@ -0,0 +1,153 @@
+# AGENTS.md
+
+This document captures architectural decisions and coding conventions for the Hindsight project.
+
+## Documentation
+
+- **Main documentation**: [hindsight-docs/docs/developer/](./hindsight-docs/docs/developer/)
+- **Use case patterns**: [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/)
+- **API reference**: Auto-generated from OpenAPI spec
+
+## Project Structure
+
+```
+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
+hindsight-integrations/ # Framework integrations (LangChain, etc.)
+hindsight-clients/      # Generated API clients (Python, TypeScript, Rust)
+```
+
+## Core Concepts
+
+### Memory Banks
+- Each bank is an isolated memory store (like a "brain" for one user/agent)
+- Banks contain: memory units (facts), entities, documents, entity links
+- Banks have a **disposition** (personality traits) and **background** (context)
+- Bank isolation is strict - no cross-bank data leakage
+
+### Memory Types
+- **World facts**: General knowledge ("The sky is blue")
+- **Experience facts**: Personal experiences ("I visited Paris in 2023")
+- **Opinion facts**: Beliefs with confidence scores ("Paris is beautiful" - 0.9 confidence)
+
+### Operations
+- **Retain**: Store new memories (extracts facts, entities, relationships)
+- **Recall**: Retrieve memories (semantic, BM25, graph, temporal search)
+- **Reflect**: Deep analysis to form new insights/opinions
+
+## API Design Decisions
+
+### Single Bank Per Request
+- All API endpoints (`recall`, `reflect`, `retain`) operate on a single bank
+- Multi-bank queries are the **client/agent's responsibility** to orchestrate
+- This keeps the API simple and the isolation model clear
+
+### Disposition Traits (3-trait system)
+- **Skepticism** (1-5): How skeptical vs trusting when forming opinions
+- **Literalism** (1-5): How literally to interpret information
+- **Empathy** (1-5): How much to consider emotional context
+- These influence the `reflect` operation, not `recall`
+- Background info also only affects `reflect` (opinion formation)
+
+## Multi-Bank Architecture Patterns
+
+See [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/) for detailed guides:
+
+- **Per-User Memory**: One bank per user, simplest pattern
+- **Support Agent + Shared Knowledge**: User bank + shared docs bank, client orchestrates
+
+## Developer Guide
+
+### Running the API Server
+
+```bash
+# From project root
+./scripts/dev/start-api.sh
+
+# With options
+./scripts/dev/start-api.sh --reload --port 8888 --log-level debug
+```
+
+### Running Tests
+
+```bash
+# API tests
+cd hindsight-api
+uv run pytest tests/
+
+# Specific test
+uv run pytest tests/test_http_api_integration.py -v
+```
+
+### Generating OpenAPI Spec
+
+After changing API endpoints, regenerate the OpenAPI spec and docs:
+
+```bash
+./scripts/generate-openapi.sh
+```
+
+This will:
+1. Generate `openapi.json` at project root
+2. Copy to `hindsight-docs/openapi.json`
+3. Regenerate API reference documentation
+
+### Generating API Clients
+
+After updating the OpenAPI spec, regenerate all clients:
+
+```bash
+./scripts/generate-clients.sh
+```
+
+This generates:
+- **Rust client**: `hindsight-clients/rust/` (via progenitor in build.rs)
+- **Python client**: `hindsight-clients/python/` (via openapi-generator Docker)
+- **TypeScript client**: `hindsight-clients/typescript/` (via @hey-api/openapi-ts)
+
+Note: The maintained wrapper `hindsight_client.py` and `README.md` are preserved during regeneration.
+
+### Running the Documentation Site
+
+```bash
+./scripts/dev/start-docs.sh
+```
+
+### Running the Control Plane
+
+```bash
+./scripts/dev/start-control-plane.sh
+```
+
+## Code Style
+
+### Python (hindsight-api)
+- Use `uv` for package management
+- Async throughout (asyncpg, async FastAPI endpoints)
+- Pydantic models for request/response validation
+- No py files at project root - maintain clean directory structure
+
+### TypeScript (control-plane, clients)
+- Next.js with App Router for control plane
+- Tailwind CSS with shadcn/ui components
+
+### Rust (CLI)
+- Async with tokio
+- reqwest for HTTP client
+- progenitor for API client generation
+
+## Database
+
+- PostgreSQL with pgvector extension
+- Schema managed via Alembic migrations in `hindsight-api/alembic/`, db migrations happen during api startup, no manual commands
+- Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
+
+# Branding
+## Colors
+- Primary: gradient from #0074d9 to #009296
+

CLAUDE.md

@@ -0,0 +1,146 @@
+# 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`
+
+### Database Backups (IMPORTANT)
+**Before any operation that may affect the database, run a backup:**
+```bash
+docker exec hindsight /backups/backup.sh
+```
+
+Operations requiring backup:
+- Running database migrations
+- Modifying Alembic migration files
+- Rebuilding Docker images
+- Resetting or recreating containers
+- Any schema changes
+- Bulk data operations
+
+Backups are stored in `~/hindsight-backups/` on the host.
+
+To restore:
+```bash
+docker exec -it hindsight /backups/restore.sh <backup-file.sql.gz>
+```
+
+## 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)

CONTRIBUTING.md

@@ -5,13 +5,23 @@ Thanks for your interest in contributing to Hindsight!
 ## Getting Started
 
 1. Fork and clone the repository
-2. Install dependencies:
    ```bash
-   cd hindsight-api && uv sync
+   git clone [email protected]:vectorize-io/hindsight.git
+   cd hindsight
    ```
-3. Set up your environment:
+2. Set up your environment:
    ```bash
-   export OPENAI_API_KEY=your-key
+   cp .env.example .env
+   ```
+   Edit the .env to add LLM API key and config as required
+
+3. Install dependencies:
+   ```bash
+   # Python dependencies
+   uv sync --directory hindsight-api/
+
+   # Node dependencies (uses npm workspaces)
+   npm install
    ```
 
 ## Development
@@ -41,7 +51,36 @@ cd hindsight-api
 uv run pytest tests/
 ```
 
-### Code style
+### Code Style
+
+We use [Ruff](https://docs.astral.sh/ruff/) for Python linting and formatting, and ESLint/Prettier for TypeScript.
+
+#### Setting up git hooks (recommended)
+
+Set up git hooks to automatically lint and format code before each commit:
+
+```bash
+./scripts/setup-hooks.sh
+```
+
+This configures git to use the hooks in `.githooks/`, which run all scripts in `scripts/hooks/` on commit. The lint hook runs in parallel:
+- **Python**: `ruff check --fix`, `ruff format`, `ty check`
+- **TypeScript**: `eslint --fix`, `prettier`
+
+#### Manual linting and formatting
+
+```bash
+# Run all lints (same as pre-commit)
+./scripts/hooks/lint.sh
+
+# Or run individually for Python:
+cd hindsight-api
+uv run ruff check --fix .   # Lint and auto-fix
+uv run ruff format .        # Format code
+uv run ty check hindsight_api  # Type check
+```
+
+#### Style guidelines
 
 - Use Python type hints
 - Follow existing code patterns

NEW.md

@@ -1,268 +0,0 @@
-<div align="center">
-
-# Hindsight
-
-**Agent Memory that Works Like Human Memory**
-
-[![CI](https://github.com/vectorize-io/hindsight/actions/workflows/test.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/test.yml)
-[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![PyPI - hindsight-client](https://img.shields.io/pypi/v/hindsight-client?label=hindsight-client)](https://pypi.org/project/hindsight-client/)
-[![npm](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)
-
-[Documentation](https://vectorize-io.github.io/hindsight) • [Paper](./Hindsight.pdf) • [Examples](./examples)
-
-</div>
-
----
-
-## 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 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.
-
-- **Inconsistency:** Agents complete tasks successfully one time, then fail when asked to complete the same task again. Memory gives the agent a mechanism to remember what worked and what didn't and to use that information to reduce errors and improve consistency.
-- **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
-
-![Overview](./hindsight-docs/static/img/hindsight-overview.png)
-
-Hindsight organizes memory into four networks to mimic the way 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.")
-
-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.
-
----
-
-## Quick Start
-
-### Docker (recommended)
-
-```bash
-export OPENAI_API_KEY=your-key
-
-docker run -p 8888:8888 -p 9999:9999 \
-  -e HINDSIGHT_API_LLM_PROVIDER=openai \
-  -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
-  -e HINDSIGHT_API_LLM_MODEL=gpt-4o-mini \
-  -v $HOME/.hindsight-docker:/home/hindsight/.pg0 \
-  ghcr.io/vectorize-io/hindsight
-```
-
-API: http://localhost:8888  
-UI: http://localhost:9999
-
-Install client:
-
-```bash
-pip install hindsight-client
-# or
-npm install @vectorize-io/hindsight-client
-```
-
-Python example:
-
-```python
-from hindsight import HindsightClient
-
-client = HindsightClient(base_url="http://localhost:8888")
-
-# Store
-client.retain(bank_id="my-agent", content="Alice works at Google as a software engineer")
-
-# Query
-results = client.recall(bank_id="my-agent", query="What does Alice do?")
-
-# Reflect
-response = client.reflect(bank_id="my-agent", query="Tell me about Alice")
-print(response.text)
-```
-
-### Python (embedded, no Docker)
-
-```bash
-pip install hindsight-all
-```
-
-```python
-import os
-from hindsight import HindsightServer, HindsightClient
-
-with HindsightServer(
-    llm_provider="openai",
-    llm_model="gpt-4o-mini", 
-    llm_api_key=os.environ["OPENAI_API_KEY"]
-) as server:
-    client = HindsightClient(base_url=server.url)
-    client.retain(bank_id="my-agent", content="Alice works at Google")
-    results = client.recall(bank_id="my-agent", query="Where does Alice work?")
-```
-
-### TypeScript
-
-```bash
-npm install @vectorize-io/hindsight-client
-```
-
-```typescript
-import { HindsightClient } from '@vectorize-io/hindsight-client';
-
-const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });
-
-await client.retain('my-agent', 'Alice loves hiking in Yosemite');
-const response = await client.recall('my-agent', 'What does Alice like?');
-```
-
----
-
-## Architecture & Operations
-
-### Retain
-
-The `retain` operation is used to push new memories into Hindsight. It tells Hindsight to _retain_ the information you pass in as an input.
-
-```python
-from hindsight_client import Hindsight
-
-client = Hindsight(base_url="http://localhost:8888")
-
-# Simple
-client.retain(
-    bank_id="my-bank",
-    content="Alice works at Google as a software engineer"
-)
-
-# With context and timestamp
-client.retain(
-    bank_id="my-bank",
-    content="Alice got promoted to senior engineer",
-    context="career update",
-    timestamp="2025-06-15T10:00:00Z"
-)
-```
-
-Behind the scenes, the retain operation uses an LLM to extract key facts, temporal data, entities, and relationships. It passes these through a normalization process to transform extracted data into canonical entities, time series, and search indexes along with metadata. These representations create the pathways for accurate memory retrieval in the recall and reflect operations. 
-
-![Retain Operation](hindsight-docs/static/img/retain-operation.png)
-
-### Recall
-
-The recall operation is used to retrieve memories. These memories can come from any of the memory types (world, experiences, etc.)
-
-```python
-from hindsight_client import Hindsight
-
-client = Hindsight(base_url="http://localhost:8888")
-
-# Simple
-client.recall(bank_id="my-bank", query="What does Alice do?")
-
-# Temporal
-results = client.recall(bank_id="my-bank", query="What happened in June?")
-
-
-```
-
-Recall performs 4 retrieval strategies in parallel:
-- Semantic: Vector similarity
-- Keyword: BM25 exact matching
-- Graph: Entity/temporal/causal links
-- Temporal: Time range filtering
-
-![Retain Operation](hindsight-docs/static/img/recall-operation.png)
-
-The individual results from the retrievals are merged, then ordered by relevance using reciprocal rank fusion and a cross-encoder reranking model.
-
-The final output is trimmed as needed to fit within the token limit.
-
-### Reflect
-
-The reflect operation is used to perform a more thorough analysis of existing memories. This allows the agent to form new connections between memories which are then persisted as opinions and/or observations. When building agents, the reflect operation is a key capability to enable the agent to learn from its experiences. 
-
-For example, the `reflect` operation can be used to support use cases such as:
-
-- An **AI Project Manager** reflecting on what risks need to be mitigated on a project.
-- A **Sales Agent** reflecting on why certain outreach messages have gotten responses while others haven't.
-- A **Support Agent** reflecting on opportunities where customers have questions not answered by current product documentation.
-
-The `reflect` operation can also be used to handle on-demand question answering or analysis which require more deep thinking.
-
-```python
-from hindsight_client import Hindsight
-
-client = Hindsight(base_url="http://localhost:8888")
-
-client.reflect(bank_id="my-bank", query="What should I know about Alice?")
-```
-
-![Retain Operation](hindsight-docs/static/img/reflect-operation.png)
-
-## Integrations
-
-### OpenAI Drop-in Replacement
-
-```python
-from hindsight_openai import configure, OpenAI
-
-configure(hindsight_api_url="http://localhost:8888", agent_id="my-assistant")
-
-client = OpenAI(api_key="sk-...")
-response = client.chat.completions.create(
-    model="gpt-4",
-    messages=[{"role": "user", "content": "What did we discuss?"}]
-)
-# Memory automatically recalled and stored
-```
-
-### Examples
-
-[Examples directory](./examples) includes:
-
-- Basic usage
-- Multi-session conversations
-- Temporal queries
-- Entity reasoning
-- Opinion tracking
-- Production setup (Docker Compose + monitoring)
-
----
-
-## Resources
-
-**Documentation:** [vectorize-io.github.io/hindsight](https://vectorize-io.github.io/hindsight)
-
-**Clients:**
-- [Python](http://hindsight.vectorize.io/sdks/python)
-- [Node.js](http://hindsight.vectorize.io/sdks/nodejs)
-- [REST API](http://hindsight.vectorize.io/api-reference)
-
-**Community:**
-- [Slack](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
-- [GitHub Issues](https://github.com/vectorize-io/hindsight/issues)
-
----
-
-## Contributing
-
-See [CONTRIBUTING.md](./CONTRIBUTING.md).
-
-## License
-
-MIT — see [LICENSE](./LICENSE)
-
----
-
-Built by [Vectorize.io](https://vectorize.io)

README.md

@@ -1,101 +1,264 @@
-# Hindsight
+<div align="center">
 
-[![CI](https://github.com/vectorize-io/hindsight/actions/workflows/test.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/test.yml)
+![Hindsight Banner](./hindsight-docs/static/img/banner.svg)
+
+[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)
+[![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 - hindsight-client](https://img.shields.io/pypi/v/hindsight-client?label=hindsight-client)](https://pypi.org/project/hindsight-client/)
-[![PyPI - hindsight-api](https://img.shields.io/pypi/v/hindsight-api?label=hindsight-api)](https://pypi.org/project/hindsight-api/)
-[![PyPI - hindsight-all](https://img.shields.io/pypi/v/hindsight-all?label=hindsight-all)](https://pypi.org/project/hindsight-all/)
-[![npm](https://img.shields.io/npm/v/@vectorize-io/hindsight-client)](https://www.npmjs.com/package/@vectorize-io/hindsight-client)
+![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)
 
-**Long-term memory for AI agents.**
 
-## Why Hindsight?
+</div>
 
-AI assistants forget everything between sessions. Every conversation starts from zero—no context about who you are, what you've discussed, or what the memory bank has learned. This isn't just inconvenient; it fundamentally limits what AI memory banks can do.
+---
 
-**The problem is harder than it looks:**
+## What is Hindsight?
 
-- **Simple vector search isn't enough** — "What did Alice do last spring?" requires temporal reasoning, not just semantic similarity
-- **Facts get disconnected** — Knowing "Alice works at Google" and "Google is in Mountain View" should let you answer "Where does Alice work?" even if you never stored that directly
-- **Memory banks need opinions** — A coding assistant that remembers "the user prefers functional programming" should weigh that when making recommendations
-- **Context matters** — The same information means different things to different memory banks with different personalities
+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 solves these problems with a memory system designed specifically for AI memory banks.
+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.
 
+- **Inconsistency:** Agents complete tasks successfully one time, then fail when asked to complete the same task again. Memory gives the agent a mechanism to remember what worked and what didn't and to use that information to reduce errors and improve consistency.
+- **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 is Hindsight Different From Other Memory Systems?
+
+![Overview](./hindsight-docs/static/img/hindsight-overview.webp)
+
+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.
+
+### 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
 
-### Option 1: Docker (recommended)
-
-Get the full experience with the API and Control Plane UI:
+### Docker (recommended)
 
 ```bash
 export OPENAI_API_KEY=your-key
-docker run -p 8888:8888 -p 9999:9999 \
-  -e HINDSIGHT_API_LLM_PROVIDER=openai \
+
+docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
   -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
-  -e HINDSIGHT_API_LLM_MODEL=gpt-4o-mini \
+  -e HINDSIGHT_API_LLM_MODEL=o3-mini \
   -v $HOME/.hindsight-docker:/home/hindsight/.pg0 \
-  ghcr.io/vectorize-io/hindsight
+  ghcr.io/vectorize-io/hindsight:latest
 ```
 
-- **API**: http://localhost:8888
-- **Control Plane UI**: http://localhost:9999
+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).
 
-Then use the Python client:
+API: http://localhost:8888  
+UI: http://localhost:9999
+
+Install client:
 
 ```bash
-pip install hindsight-client
+pip install hindsight-client -U
+# or
+npm install @vectorize-io/hindsight-client
 ```
 
+Python example:
+
 ```python
-from hindsight import HindsightClient
+from hindsight_client import Hindsight
 
-client = HindsightClient(base_url="http://localhost:8888")
+client = Hindsight(base_url="http://localhost:8888")
 
-# Store memories
-client.retain(bank_id="my-agent", content="Alice works at Google as a software engineer")
-client.retain(bank_id="my-agent", content="Alice mentioned she loves hiking in the mountains")
+# Retain: Store information
+client.retain(bank_id="my-bank", content="Alice works at Google as a software engineer")
 
-# Query with temporal reasoning
-results = client.recall(bank_id="my-agent", query="What does Alice do for work?")
+# Recall: Search memories
+client.recall(bank_id="my-bank", query="What does Alice do?")
 
-# Get a synthesized perspective
-response = client.reflect(bank_id="my-agent", query="Tell me about Alice")
-print(response.text)
+# Reflect: Generate disposition-aware response
+client.reflect(bank_id="my-bank", query="Tell me about Alice")
 ```
 
-### Option 2: Embedded (no docker/server required)
-
-For quick prototyping, run everything in-process:
+### Python (embedded, no Docker)
 
 ```bash
-pip install hindsight-all
-export OPENAI_API_KEY=your-key
+pip install hindsight-all -U
 ```
 
 ```python
 import os
 from hindsight import HindsightServer, HindsightClient
 
-with HindsightServer(llm_provider="openai", llm_model="gpt-4o-mini", llm_api_key=os.environ["OPENAI_API_KEY"]) as server:
+with HindsightServer(
+    llm_provider="openai",
+    llm_model="gpt-5-mini", 
+    llm_api_key=os.environ["OPENAI_API_KEY"]
+) as server:
     client = HindsightClient(base_url=server.url)
-
-    client.retain(bank_id="my-user", content="User prefers functional programming")
-    response = client.reflect(bank_id="my-user", query="What coding style should I use?")
-    print(response.text)
+    client.retain(bank_id="my-bank", content="Alice works at Google")
+    results = client.recall(bank_id="my-bank", query="Where does Alice work?")
 ```
 
+### Node.js / TypeScript
 
+```bash
+npm install @vectorize-io/hindsight-client
+```
 
-## Documentation
+```javascript
+const { HindsightClient } = require('@vectorize-io/hindsight-client');
 
-Full documentation: [hindsight.vectorize.io](https://hindsight.vectorize.io)
+const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });
+
+await client.retain('my-bank', 'Alice loves hiking in Yosemite');
+await client.recall('my-bank', 'What does Alice like?');
+```
+
+---
+
+## Architecture & Operations
+
+### Retain
+
+The `retain` operation is used to push new memories into Hindsight. It tells Hindsight to _retain_ the information you pass in as an input.
+
+```python
+from hindsight_client import Hindsight
+
+client = Hindsight(base_url="http://localhost:8888")
+
+# Simple
+client.retain(
+    bank_id="my-bank",
+    content="Alice works at Google as a software engineer"
+)
+
+# With context and timestamp
+client.retain(
+    bank_id="my-bank",
+    content="Alice got promoted to senior engineer",
+    context="career update",
+    timestamp="2025-06-15T10:00:00Z"
+)
+```
+
+Behind the scenes, the retain operation uses an LLM to extract key facts, temporal data, entities, and relationships. It passes these through a normalization process to transform extracted data into canonical entities, time series, and search indexes along with metadata. These representations create the pathways for accurate memory retrieval in the recall and reflect operations. 
+
+![Retain Operation](hindsight-docs/static/img/retain-operation.webp)
+
+### Recall
+
+The recall operation is used to retrieve memories. These memories can come from any of the memory types (world, experiences, etc.)
+
+```python
+from hindsight_client import Hindsight
+
+client = Hindsight(base_url="http://localhost:8888")
+
+# Simple
+client.recall(bank_id="my-bank", query="What does Alice do?")
+
+# Temporal
+client.recall(bank_id="my-bank", query="What happened in June?")
+```
+
+Recall performs 4 retrieval strategies in parallel:
+- Semantic: Vector similarity
+- Keyword: BM25 exact matching
+- Graph: Entity/temporal/causal links
+- Temporal: Time range filtering
+
+![Retain Operation](hindsight-docs/static/img/recall-operation.webp)
+
+The individual results from the retrievals are merged, then ordered by relevance using reciprocal rank fusion and a cross-encoder reranking model.
+
+The final output is trimmed as needed to fit within the token limit.
+
+### Reflect
+
+The reflect operation is used to perform a more thorough analysis of existing memories. This allows the agent to form new connections between memories which are then persisted as opinions and/or observations. When building agents, the reflect operation is a key capability to enable the agent to learn from its experiences. 
+
+For example, the `reflect` operation can be used to support use cases such as:
+
+- An **AI Project Manager** reflecting on what risks need to be mitigated on a project.
+- A **Sales Agent** reflecting on why certain outreach messages have gotten responses while others haven't.
+- A **Support Agent** reflecting on opportunities where customers have questions not answered by current product documentation.
+
+The `reflect` operation can also be used to handle on-demand question answering or analysis which require more deep thinking.
+
+```python
+from hindsight_client import Hindsight
+
+client = Hindsight(base_url="http://localhost:8888")
+
+client.reflect(bank_id="my-bank", query="What should I know about Alice?")
+```
+
+![Retain Operation](hindsight-docs/static/img/reflect-operation.webp)
+
+---
+
+## Resources
+
+**Documentation:** 
+- [https://hindsight.vectorize.io](https://hindsight.vectorize.io)
+
+**Clients:**
+- [Python](http://hindsight.vectorize.io/sdks/python)
+- [Node.js](http://hindsight.vectorize.io/sdks/nodejs)
+- [REST API](https://hindsight.vectorize.io/api-reference)
+- [CLI](https://hindsight.vectorize.io/sdks/cli)
+
+**Community:**
+- [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
 
-We welcome contributions! See [CONTRIBUTING.md](./CONTRIBUTING.md) for guidelines.
+See [CONTRIBUTING.md](./CONTRIBUTING.md).
 
 ## License
 
-MIT
+MIT — see [LICENSE](./LICENSE)
+
+---
+
+Built by [Vectorize.io](https://vectorize.io)

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
@@ -40,9 +43,11 @@ WORKDIR /app/api
 # Sync dependencies (will create lock file if needed)
 RUN uv sync
 
-# Copy source code and alembic migrations
+# Copy source code (alembic migrations are inside hindsight_api/)
 COPY hindsight-api/hindsight_api ./hindsight_api
-COPY hindsight-api/alembic ./alembic
+
+# Install the local package (uv sync only installed dependencies, not the package itself)
+RUN uv pip install -e .
 
 # =============================================================================
 # Stage: SDK Builder (needed for Control Plane)
@@ -52,13 +57,15 @@ FROM node:20-slim AS sdk-builder
 ARG INCLUDE_CP
 RUN if [ "$INCLUDE_CP" != "true" ]; then echo "Skipping SDK build" && exit 0; fi
 
-WORKDIR /app/sdk
+WORKDIR /app
 
-COPY hindsight-clients/typescript/package*.json ./
-RUN npm ci
+# Copy root package files for npm workspaces
+COPY package.json package-lock.json ./
+COPY hindsight-clients/typescript/ ./hindsight-clients/typescript/
 
-COPY hindsight-clients/typescript/ ./
-RUN npm run build
+# 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
 
 # =============================================================================
 # Stage: Control Plane Builder
@@ -68,30 +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/sdk /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
@@ -100,18 +125,18 @@ FROM python:3.11-slim AS api-only
 
 WORKDIR /app
 
-# Install pg0 dependencies
+# 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
 
-# Create non-root user (PostgreSQL cannot run as root)
 RUN useradd -m -s /bin/bash hindsight
 
 # Copy API with virtual environment from builder
@@ -121,52 +146,23 @@ COPY --from=api-builder /app/api /app/api
 COPY docker/standalone/start-all.sh /app/start-all.sh
 RUN chmod +x /app/start-all.sh
 
-# Create data directory for pg0 and set ownership
-RUN mkdir -p /app/data && chown -R hindsight:hindsight /app
+RUN chown -R hindsight:hindsight /app
 
-# Switch to non-root user
 USER hindsight
 
-# Set PATH for hindsight user
-ENV PATH="/home/hindsight/.hindsight/bin:/app/api/.venv/bin:${PATH}"
+ENV PATH="/app/api/.venv/bin:${PATH}"
 
-# Install pg0 binary
-RUN mkdir -p /home/hindsight/.hindsight/bin && \
-    ARCH=$(uname -m) && \
-    if [ "$ARCH" = "aarch64" ] || [ "$ARCH" = "arm64" ]; then \
-        PG0_BINARY="pg0-linux-aarch64-gnu"; \
-    elif [ "$ARCH" = "x86_64" ]; then \
-        PG0_BINARY="pg0-linux-x86_64-gnu"; \
-    else \
-        echo "Unsupported architecture: $ARCH" && exit 1; \
-    fi && \
-    echo "Installing pg0 binary: $PG0_BINARY" && \
-    for i in 1 2 3 4 5; do \
-        curl -fsSL -o /home/hindsight/.hindsight/bin/pg0 \
-            "https://github.com/vectorize-io/pg0/releases/latest/download/$PG0_BINARY" && \
-        chmod +x /home/hindsight/.hindsight/bin/pg0 && \
-        break || (echo "Retry $i failed, waiting..." && sleep 10); \
-    done && \
-    /home/hindsight/.hindsight/bin/pg0 --version
-
-# Pre-download PostgreSQL binaries
-ENV PG0_HOME=/home/hindsight/.pg0-cache
-RUN pg0 start --help && \
-    (pg0 start --name hindsight --port 5555 --username hindsight --password hindsight --database hindsight && \
-    sleep 2 && \
-    pg0 stop --name hindsight && \
-    echo "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
 
@@ -175,6 +171,7 @@ ENV HINDSIGHT_API_PORT=8888
 ENV HINDSIGHT_API_LOG_LEVEL=info
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=false
+ENV PYTHONUNBUFFERED=1
 
 CMD ["/app/start-all.sh"]
 
@@ -186,13 +183,13 @@ FROM node:20-alpine AS cp-only
 WORKDIR /app
 
 # Copy built SDK
-COPY --from=sdk-builder /app/sdk /app/sdk
+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
 
@@ -219,33 +216,34 @@ FROM python:3.11-slim AS standalone
 
 WORKDIR /app
 
-# Install Node.js, curl, uv, and pg0 dependencies
+# Install Node.js, curl, uv, and system dependencies
+# 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/* \
     && pip install --no-cache-dir uv
 
-# Create non-root user (PostgreSQL cannot run as root)
 RUN useradd -m -s /bin/bash hindsight
 
 # Copy API with virtual environment from builder
 COPY --from=api-builder /app/api /app/api
 
 # Copy built SDK
-COPY --from=sdk-builder /app/sdk /app/sdk
+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
 
@@ -253,52 +251,23 @@ WORKDIR /app
 COPY docker/standalone/start-all.sh /app/start-all.sh
 RUN chmod +x /app/start-all.sh
 
-# Create data directory for pg0 and set ownership
-RUN mkdir -p /app/data && chown -R hindsight:hindsight /app
+RUN chown -R hindsight:hindsight /app
 
-# Switch to non-root user
 USER hindsight
 
-# Set PATH for hindsight user
-ENV PATH="/home/hindsight/.hindsight/bin:/app/api/.venv/bin:${PATH}"
+ENV PATH="/app/api/.venv/bin:${PATH}"
 
-# Install pg0 binary
-RUN mkdir -p /home/hindsight/.hindsight/bin && \
-    ARCH=$(uname -m) && \
-    if [ "$ARCH" = "aarch64" ] || [ "$ARCH" = "arm64" ]; then \
-        PG0_BINARY="pg0-linux-aarch64-gnu"; \
-    elif [ "$ARCH" = "x86_64" ]; then \
-        PG0_BINARY="pg0-linux-x86_64-gnu"; \
-    else \
-        echo "Unsupported architecture: $ARCH" && exit 1; \
-    fi && \
-    echo "Installing pg0 binary: $PG0_BINARY" && \
-    for i in 1 2 3 4 5; do \
-        curl -fsSL -o /home/hindsight/.hindsight/bin/pg0 \
-            "https://github.com/vectorize-io/pg0/releases/latest/download/$PG0_BINARY" && \
-        chmod +x /home/hindsight/.hindsight/bin/pg0 && \
-        break || (echo "Retry $i failed, waiting..." && sleep 10); \
-    done && \
-    /home/hindsight/.hindsight/bin/pg0 --version
-
-# Pre-download PostgreSQL binaries
-ENV PG0_HOME=/home/hindsight/.pg0-cache
-RUN pg0 start --help && \
-    (pg0 start --name hindsight --port 5555 --username hindsight --password hindsight --database hindsight && \
-    sleep 2 && \
-    pg0 stop --name hindsight && \
-    echo "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
 
@@ -309,6 +278,7 @@ ENV NODE_ENV=production
 ENV HINDSIGHT_CP_DATAPLANE_API_URL=http://localhost:8888
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=true
+ENV PYTHONUNBUFFERED=1
 
 CMD ["/app/start-all.sh"]
 

docker/standalone/start-all.sh

@@ -1,23 +1,74 @@
 #!/bin/bash
 set -e
 
-echo "🚀 Starting Hindsight..."
-echo ""
-
 # Service flags (default to true if not set)
 ENABLE_API="${HINDSIGHT_ENABLE_API:-true}"
 ENABLE_CP="${HINDSIGHT_ENABLE_CP:-true}"
 
-# Copy pre-cached PostgreSQL data if runtime directory is empty (first run with volume)
-if [ "$ENABLE_API" = "true" ]; then
-    PG0_CACHE="/home/hindsight/.pg0-cache"
-    PG0_HOME="/home/hindsight/.pg0"
-    if [ -d "$PG0_CACHE" ] && [ "$(ls -A $PG0_CACHE 2>/dev/null)" ]; then
-        if [ ! "$(ls -A $PG0_HOME 2>/dev/null)" ]; then
-            echo "📦 Copying pre-cached PostgreSQL data..."
-            cp -r "$PG0_CACHE"/* "$PG0_HOME"/ 2>/dev/null || true
-        fi
+# =============================================================================
+# Dependency waiting (opt-in via HINDSIGHT_WAIT_FOR_DEPS=true)
+#
+# Problem: When running with LM Studio, the LLM may take time to load models.
+# If Hindsight starts before LM Studio is ready, it fails on LLM verification.
+# This wait loop ensures dependencies are ready before starting.
+# =============================================================================
+if [ "${HINDSIGHT_WAIT_FOR_DEPS:-false}" = "true" ]; then
+    LLM_BASE_URL="${HINDSIGHT_API_LLM_BASE_URL:-http://host.docker.internal:1234/v1}"
+    MAX_RETRIES="${HINDSIGHT_RETRY_MAX:-0}"  # 0 = infinite
+    RETRY_INTERVAL="${HINDSIGHT_RETRY_INTERVAL:-10}"
+
+    # Check if external database is configured (skip check for embedded pg0)
+    SKIP_DB_CHECK=false
+    if [ -z "${HINDSIGHT_API_DATABASE_URL}" ]; then
+        SKIP_DB_CHECK=true
+    else
+        DB_CHECK_HOST=$(echo "$HINDSIGHT_API_DATABASE_URL" | sed -E 's|.*@([^:/]+):([0-9]+)/.*|\1 \2|')
     fi
+
+    check_db() {
+        if $SKIP_DB_CHECK; then
+            return 0
+        fi
+        if command -v pg_isready &> /dev/null; then
+            pg_isready -h $(echo $DB_CHECK_HOST | cut -d' ' -f1) -p $(echo $DB_CHECK_HOST | cut -d' ' -f2) &>/dev/null
+        else
+            python3 -c "import socket; s=socket.socket(); s.settimeout(5); exit(0 if s.connect_ex(('$(echo $DB_CHECK_HOST | cut -d' ' -f1)', $(echo $DB_CHECK_HOST | cut -d' ' -f2))) == 0 else 1)" 2>/dev/null
+        fi
+    }
+
+    check_llm() {
+        curl -sf "${LLM_BASE_URL}/models" --connect-timeout 5 &>/dev/null
+    }
+
+    echo "⏳ Waiting for dependencies to be ready..."
+    attempt=1
+
+    while true; do
+        db_ok=false
+        llm_ok=false
+
+        if check_db; then
+            db_ok=true
+        fi
+
+        if check_llm; then
+            llm_ok=true
+        fi
+
+        if $db_ok && $llm_ok; then
+            echo "✅ Dependencies ready!"
+            break
+        fi
+
+        if [ "$MAX_RETRIES" -ne 0 ] && [ "$attempt" -ge "$MAX_RETRIES" ]; then
+            echo "❌ Max retries ($MAX_RETRIES) reached. Dependencies not available."
+            exit 1
+        fi
+
+        echo "   Attempt $attempt: DB=$( $db_ok && echo 'ok' || echo 'waiting' ), LLM=$( $llm_ok && echo 'ok' || echo 'waiting' )"
+        sleep "$RETRY_INTERVAL"
+        ((attempt++))
+    done
 fi
 
 # Track PIDs for wait
@@ -26,32 +77,31 @@ PIDS=()
 # Start API if enabled
 if [ "$ENABLE_API" = "true" ]; then
     cd /app/api
-    python -m hindsight_api.web.server 2>&1 | sed -u 's/^/[api] /' &
+    # Run API directly - Python's PYTHONUNBUFFERED=1 handles output buffering
+    hindsight-api &
     API_PID=$!
     PIDS+=($API_PID)
 
     # Wait for API to be ready
-    echo "⏳ Waiting for API..."
     for i in {1..60}; do
         if curl -sf http://localhost:8888/health &>/dev/null; then
-            echo "✅ API is ready"
             break
         fi
         sleep 1
     done
 else
-    echo "⏭️  API disabled (HINDSIGHT_ENABLE_API=false)"
+    echo "API disabled (HINDSIGHT_ENABLE_API=false)"
 fi
 
 # Start Control Plane if enabled
 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
-    echo "⏭️  Control Plane disabled (HINDSIGHT_ENABLE_CP=false)"
+    echo "Control Plane disabled (HINDSIGHT_ENABLE_CP=false)"
 fi
 
 # Print status

helm/INSTALL.txt

@@ -1,135 +0,0 @@
-HINDSIGHT HELM CHART INSTALLATION GUIDE
-=====================================
-
-PREREQUISITES
--------------
-- Kubernetes cluster (1.19+)
-- kubectl configured
-- Helm 3.x installed
-- PostgreSQL database with pgvector extension (if not using bundled PostgreSQL)
-
-BASIC INSTALLATION
-------------------
-
-1. Install with default values (requires external PostgreSQL):
-
-   helm install hindsight ./hindsight \
-     --set postgresql.external.host=your-postgres-host \
-     --set postgresql.external.password=your-password \
-     --set api.secrets.MEMORY_LLM_API_KEY=your-api-key
-
-2. Install with custom values file:
-
-   helm install hindsight ./hindsight -f hindsight/values-production.yaml
-
-3. Install in a specific namespace:
-
-   kubectl create namespace hindsight
-   helm install hindsight ./hindsight -n hindsight
-
-CONFIGURATION OPTIONS
----------------------
-
-Development setup (using values-development.yaml):
-   helm install hindsight ./hindsight -f hindsight/values-development.yaml
-
-Production setup (using values-production.yaml):
-   helm install hindsight ./hindsight -f hindsight/values-production.yaml
-
-Custom LLM provider:
-   helm install hindsight ./hindsight \
-     --set api.env.MEMORY_LLM_PROVIDER=openai \
-     --set api.env.MEMORY_LLM_MODEL=gpt-4 \
-     --set api.secrets.MEMORY_LLM_API_KEY=sk-your-key
-
-Enable ingress:
-   helm install hindsight ./hindsight \
-     --set ingress.enabled=true \
-     --set ingress.hosts[0].host=hindsight.example.com
-
-Enable autoscaling:
-   helm install hindsight ./hindsight \
-     --set autoscaling.enabled=true \
-     --set autoscaling.minReplicas=2 \
-     --set autoscaling.maxReplicas=10
-
-UPGRADE
--------
-
-Upgrade existing installation:
-   helm upgrade hindsight ./hindsight
-
-Upgrade with new values:
-   helm upgrade hindsight ./hindsight -f hindsight/values-production.yaml
-
-UNINSTALL
----------
-
-Remove the Helm release:
-   helm uninstall hindsight
-
-Remove with namespace:
-   helm uninstall hindsight -n hindsight
-
-TESTING
--------
-
-Test the installation with dry-run:
-   helm install hindsight ./hindsight --dry-run --debug
-
-Validate templates:
-   helm template hindsight ./hindsight
-
-Lint the chart:
-   helm lint ./hindsight
-
-ACCESSING THE SERVICES
-----------------------
-
-Port-forward control plane:
-   kubectl port-forward svc/hindsight-control-plane 3000:3000
-
-Port-forward API:
-   kubectl port-forward svc/hindsight-api 8888:8888
-
-Get service URLs:
-   helm status hindsight
-
-DATABASE INITIALIZATION
------------------------
-
-NOTE: Database migrations now run automatically when the API service starts.
-      You typically don't need to run migrations manually.
-
-If you want to pre-initialize the database before deploying (optional):
-   kubectl run hindsight-init --rm -it --restart=Never \
-     --image=hindsight/api:latest \
-     --env="DATABASE_URL=postgresql://user:pass@host:5432/hindsight" \
-     -- python -c "from hindsight.migrations import run_migrations; run_migrations()"
-
-TROUBLESHOOTING
----------------
-
-Check pod status:
-   kubectl get pods -l app.kubernetes.io/name=hindsight
-
-View logs for API:
-   kubectl logs -l app.kubernetes.io/component=api
-
-View logs for control plane:
-   kubectl logs -l app.kubernetes.io/component=control-plane
-
-Describe a pod:
-   kubectl describe pod <pod-name>
-
-Check configuration:
-   kubectl get configmap hindsight-config -o yaml
-   kubectl get secret hindsight-secret -o yaml
-
-NOTES
------
-- Make sure PostgreSQL has pgvector extension enabled
-- Run database migrations before first use
-- Configure proper resource limits for production
-- Use external secrets management for production
-- Enable TLS/SSL for production deployments

helm/hindsight/Chart.lock

@@ -0,0 +1,6 @@
+dependencies:
+- name: postgresql
+  repository: https://charts.bitnami.com/bitnami
+  version: 15.5.38
+digest: sha256:f67c7612736803ece8a669f8ca6b0555f3b78557bc0ecb732aa2e43f0df7750d
+generated: "2025-12-10T17:20:57.058794+01:00"

helm/hindsight/Chart.yaml

@@ -1,9 +1,9 @@
 apiVersion: v2
 name: hindsight
-description: A Helm chart for Hindsight - temporal-semantic-entity memory system for AI agents
+description: Hindsight helm chart
 type: application
-version: 0.0.21
-appVersion: "0.0.21"
+version: 0.2.1
+appVersion: "0.2.1"
 keywords:
   - ai
   - memory

helm/hindsight/README.md

@@ -0,0 +1,182 @@
+# Hindsight Helm Chart
+
+Helm chart for deploying Hindsight - a temporal-semantic-entity memory system for AI agents.
+
+## Prerequisites
+
+- Kubernetes 1.19+
+- Helm 3.0+
+- PostgreSQL database (external or bundled)
+
+## Quick Start
+
+```bash
+# Update dependencies first
+helm dependency update ./helm/hindsight
+
+# Install (PostgreSQL included by default)
+export OPENAI_API_KEY="sk-your-openai-key"
+helm upgrade hindsight --install ./helm/hindsight -n hindsight --create-namespace \
+  --set api.secrets.HINDSIGHT_API_LLM_API_KEY="$OPENAI_API_KEY"
+```
+
+To use an external database instead:
+
+```bash
+helm install hindsight ./helm/hindsight -n hindsight --create-namespace \
+  --set api.secrets.HINDSIGHT_API_LLM_API_KEY="sk-your-openai-key" \
+  --set postgresql.enabled=false \
+  --set postgresql.external.host=my-postgres.example.com \
+  --set postgresql.external.password=mypassword
+```
+
+## Installation
+
+### Add the repository (if published)
+
+```bash
+helm repo add hindsight https://your-helm-repo.com
+helm repo update
+```
+
+### Install with custom values file
+
+Create a `values-override.yaml`:
+
+```yaml
+api:
+  secrets:
+    HINDSIGHT_API_LLM_API_KEY: "sk-your-openai-key"
+
+postgresql:
+  external:
+    host: "my-postgres.example.com"
+    password: "mypassword"
+```
+
+Then install:
+
+```bash
+helm install hindsight ./helm/hindsight -n hindsight --create-namespace -f values-override.yaml
+```
+
+## Configuration
+
+### Key Values
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `version` | Default image tag for all components | `0.1.0` |
+| `api.enabled` | Enable the API component | `true` |
+| `api.image.repository` | API image repository | `hindsight/api` |
+| `api.image.tag` | API image tag (defaults to `version`) | - |
+| `api.service.port` | API service port | `8888` |
+| `controlPlane.enabled` | Enable the control plane | `true` |
+| `controlPlane.image.repository` | Control plane image repository | `hindsight/control-plane` |
+| `controlPlane.image.tag` | Control plane image tag (defaults to `version`) | - |
+| `controlPlane.service.port` | Control plane service port | `3000` |
+| `postgresql.enabled` | Deploy PostgreSQL as subchart | `true` |
+| `postgresql.external.host` | External PostgreSQL host | `postgresql` |
+| `postgresql.external.port` | External PostgreSQL port | `5432` |
+| `postgresql.external.database` | Database name | `hindsight` |
+| `postgresql.external.username` | Database username | `hindsight` |
+| `ingress.enabled` | Enable ingress | `false` |
+| `autoscaling.enabled` | Enable HPA | `false` |
+
+### Environment Variables
+
+All environment variables in `api.env` and `controlPlane.env` are automatically added to the respective pods. Sensitive values should go in `api.secrets` or `controlPlane.secrets`.
+
+```yaml
+api:
+  env:
+    HINDSIGHT_API_LLM_PROVIDER: "openai"
+    HINDSIGHT_API_LLM_MODEL: "gpt-4"
+  secrets:
+    HINDSIGHT_API_LLM_API_KEY: "your-api-key"
+    HINDSIGHT_API_LLM_BASE_URL: "https://api.openai.com/v1"
+
+controlPlane:
+  env:
+    NODE_ENV: "production"
+  secrets: {}
+```
+
+### External Database
+
+To connect to an external PostgreSQL database:
+
+```yaml
+postgresql:
+  enabled: false
+  external:
+    host: "my-postgres.example.com"
+    port: 5432
+    database: "hindsight"
+    username: "hindsight"
+    password: "your-password"
+```
+
+### Ingress
+
+To expose the services via ingress:
+
+```yaml
+ingress:
+  enabled: true
+  className: "nginx"
+  annotations:
+    cert-manager.io/cluster-issuer: "letsencrypt-prod"
+  hosts:
+    - host: hindsight.example.com
+      paths:
+        - path: /
+          pathType: Prefix
+          service: controlPlane
+        - path: /api
+          pathType: Prefix
+          service: api
+  tls:
+    - secretName: hindsight-tls
+      hosts:
+        - hindsight.example.com
+```
+
+## Upgrading
+
+```bash
+helm upgrade hindsight ./helm/hindsight -n hindsight
+```
+
+## Uninstalling
+
+```bash
+helm uninstall hindsight -n hindsight
+```
+
+## Components
+
+The chart deploys:
+
+- **API**: The main Hindsight API server for memory operations
+- **Control Plane**: Web UI for managing agents and viewing memories
+
+## Development
+
+### Lint the chart
+
+```bash
+helm lint ./helm/hindsight
+```
+
+### Template locally
+
+```bash
+helm template hindsight ./helm/hindsight --debug
+```
+
+### Dry run installation
+
+```bash
+helm install hindsight ./helm/hindsight --dry-run --debug
+```

helm/hindsight/templates/NOTES.txt

@@ -1,71 +1,2 @@
-Thank you for installing {{ .Chart.Name }}!
-
-Your release is named {{ .Release.Name }}.
-
-To learn more about the release, try:
-
-  $ helm status {{ .Release.Name }}
-  $ helm get all {{ .Release.Name }}
-
-{{- if .Values.ingress.enabled }}
-
-The application is accessible via the following URL(s):
-{{- range .Values.ingress.hosts }}
-  - http{{ if $.Values.ingress.tls }}s{{ end }}://{{ .host }}
-{{- end }}
-
-{{- else }}
-
-1. Get the Control Plane URL by running these commands:
-{{- if contains "NodePort" .Values.controlPlane.service.type }}
-  export NODE_PORT=$(kubectl get --namespace {{ .Release.Namespace }} -o jsonpath="{.spec.ports[0].nodePort}" services {{ include "hindsight.fullname" . }}-control-plane)
-  export NODE_IP=$(kubectl get nodes --namespace {{ .Release.Namespace }} -o jsonpath="{.items[0].status.addresses[0].address}")
-  echo "Control Plane URL: http://$NODE_IP:$NODE_PORT"
-{{- else if contains "LoadBalancer" .Values.controlPlane.service.type }}
-  NOTE: It may take a few minutes for the LoadBalancer IP to be available.
-  You can watch the status by running 'kubectl get --namespace {{ .Release.Namespace }} svc -w {{ include "hindsight.fullname" . }}-control-plane'
-  export SERVICE_IP=$(kubectl get svc --namespace {{ .Release.Namespace }} {{ include "hindsight.fullname" . }}-control-plane --template "{{"{{ range (index .status.loadBalancer.ingress 0) }}{{.}}{{ end }}"}}")
-  echo "Control Plane URL: http://$SERVICE_IP:{{ .Values.controlPlane.service.port }}"
-{{- else if contains "ClusterIP" .Values.controlPlane.service.type }}
-  export POD_NAME=$(kubectl get pods --namespace {{ .Release.Namespace }} -l "app.kubernetes.io/component=control-plane,app.kubernetes.io/instance={{ .Release.Name }}" -o jsonpath="{.items[0].metadata.name}")
-  export CONTAINER_PORT=$(kubectl get pod --namespace {{ .Release.Namespace }} $POD_NAME -o jsonpath="{.spec.containers[0].ports[0].containerPort}")
-  echo "Control Plane URL: http://127.0.0.1:3000"
-  kubectl --namespace {{ .Release.Namespace }} port-forward $POD_NAME 3000:$CONTAINER_PORT
-{{- end }}
-
-2. Get the API URL by running these commands:
-{{- if contains "NodePort" .Values.api.service.type }}
-  export NODE_PORT=$(kubectl get --namespace {{ .Release.Namespace }} -o jsonpath="{.spec.ports[0].nodePort}" services {{ include "hindsight.fullname" . }}-api)
-  export NODE_IP=$(kubectl get nodes --namespace {{ .Release.Namespace }} -o jsonpath="{.items[0].status.addresses[0].address}")
-  echo "API URL: http://$NODE_IP:$NODE_PORT"
-{{- else if contains "LoadBalancer" .Values.api.service.type }}
-  NOTE: It may take a few minutes for the LoadBalancer IP to be available.
-  You can watch the status by running 'kubectl get --namespace {{ .Release.Namespace }} svc -w {{ include "hindsight.fullname" . }}-api'
-  export SERVICE_IP=$(kubectl get svc --namespace {{ .Release.Namespace }} {{ include "hindsight.fullname" . }}-api --template "{{"{{ range (index .status.loadBalancer.ingress 0) }}{{.}}{{ end }}"}}")
-  echo "API URL: http://$SERVICE_IP:{{ .Values.api.service.port }}"
-{{- else if contains "ClusterIP" .Values.api.service.type }}
-  export POD_NAME=$(kubectl get pods --namespace {{ .Release.Namespace }} -l "app.kubernetes.io/component=api,app.kubernetes.io/instance={{ .Release.Name }}" -o jsonpath="{.items[0].metadata.name}")
-  export CONTAINER_PORT=$(kubectl get pod --namespace {{ .Release.Namespace }} $POD_NAME -o jsonpath="{.spec.containers[0].ports[0].containerPort}")
-  echo "API URL: http://127.0.0.1:8888"
-  kubectl --namespace {{ .Release.Namespace }} port-forward $POD_NAME 8888:$CONTAINER_PORT
-{{- end }}
-
-{{- end }}
-
-{{- if not .Values.postgresql.enabled }}
-
-NOTE: You are using an external PostgreSQL database.
-Please ensure that:
-  1. The database is accessible from the cluster
-  2. The pgvector extension is enabled
-
-Database migrations run automatically when the API service starts.
-
-If you want to pre-initialize the database before deploying (optional):
-  kubectl run --namespace {{ .Release.Namespace }} hindsight-init --rm -it --restart=Never \
-    --image={{ .Values.api.image.repository }}:{{ .Values.api.image.tag }} \
-    --env="DATABASE_URL={{ include "hindsight.databaseUrl" . }}" \
-    -- python -c "from hindsight.migrations import run_migrations; run_migrations()"
-{{- end }}
-
-For more information, visit: https://github.com/yourusername/hindsight
+Hindsight installed. Access the control plane:
+  kubectl port-forward -n {{ .Release.Namespace }} svc/{{ include "hindsight.fullname" . }}-control-plane 3000:3000

helm/hindsight/templates/_helpers.tpl

@@ -98,7 +98,7 @@ Generate database URL
 {{- if .Values.databaseUrl }}
 {{- .Values.databaseUrl }}
 {{- else if .Values.postgresql.enabled }}
-{{- printf "postgresql://%s:%s@%s-postgresql:%d/%s" .Values.postgresql.auth.username .Values.postgresql.auth.password (include "hindsight.fullname" .) (.Values.postgresql.primary.service.port | int) .Values.postgresql.auth.database }}
+{{- printf "postgresql://%s:%s@%s-postgresql:%d/%s" .Values.postgresql.auth.username .Values.postgresql.auth.password (include "hindsight.fullname" .) (.Values.postgresql.service.port | int) .Values.postgresql.auth.database }}
 {{- else }}
 {{- printf "postgresql://%s:$(POSTGRES_PASSWORD)@%s:%d/%s" .Values.postgresql.external.username .Values.postgresql.external.host (.Values.postgresql.external.port | int) .Values.postgresql.external.database }}
 {{- end }}

helm/hindsight/templates/api-deployment.yaml

@@ -15,7 +15,6 @@ spec:
   template:
     metadata:
       annotations:
-        checksum/config: {{ include (print $.Template.BasePath "/configmap.yaml") . | sha256sum }}
         checksum/secret: {{ include (print $.Template.BasePath "/secret.yaml") . | sha256sum }}
         {{- with .Values.podAnnotations }}
         {{- toYaml . | nindent 8 }}
@@ -32,7 +31,7 @@ spec:
       - name: api
         securityContext:
           {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag }}"
+        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag | default .Values.version }}"
         imagePullPolicy: {{ .Values.api.image.pullPolicy }}
         ports:
         - name: http
@@ -48,29 +47,16 @@ spec:
               name: {{ include "hindsight.fullname" . }}-secret
               key: postgres-password
         {{- end }}
-        - name: HINDSIGHT_API_LLM_PROVIDER
-          valueFrom:
-            configMapKeyRef:
-              name: {{ include "hindsight.fullname" . }}-config
-              key: llm-provider
-        - name: HINDSIGHT_API_LLM_MODEL
-          valueFrom:
-            configMapKeyRef:
-              name: {{ include "hindsight.fullname" . }}-config
-              key: llm-model
-        {{- if and .Values.api.secrets (hasKey .Values.api.secrets "HINDSIGHT_API_LLM_API_KEY") }}
-        - name: HINDSIGHT_API_LLM_API_KEY
-          valueFrom:
-            secretKeyRef:
-              name: {{ include "hindsight.fullname" . }}-secret
-              key: llm-api-key
+        {{- range $key, $value := .Values.api.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
         {{- end }}
-        {{- if and .Values.api.secrets (hasKey .Values.api.secrets "HINDSIGHT_API_LLM_BASE_URL") }}
-        - name: HINDSIGHT_API_LLM_BASE_URL
+        {{- range $key, $value := .Values.api.secrets }}
+        - name: {{ $key }}
           valueFrom:
             secretKeyRef:
-              name: {{ include "hindsight.fullname" . }}-secret
-              key: llm-base-url
+              name: {{ include "hindsight.fullname" $ }}-secret
+              key: {{ $key }}
         {{- end }}
         livenessProbe:
           {{- toYaml .Values.api.livenessProbe | nindent 10 }}

helm/hindsight/templates/configmap.yaml

@@ -1,15 +0,0 @@
-apiVersion: v1
-kind: ConfigMap
-metadata:
-  name: {{ include "hindsight.fullname" . }}-config
-  labels:
-    {{- include "hindsight.labels" . | nindent 4 }}
-data:
-  # API configuration
-  llm-provider: {{ .Values.api.env.HINDSIGHT_API_LLM_PROVIDER | quote }}
-  llm-model: {{ .Values.api.env.HINDSIGHT_API_LLM_MODEL | quote }}
-
-  # Control plane configuration
-  node-env: {{ .Values.controlPlane.env.NODE_ENV | quote }}
-  hostname: {{ .Values.controlPlane.env.HINDSIGHT_CP_HOSTNAME | quote }}
-  control-plane-port: {{ .Values.controlPlane.env.HINDSIGHT_CP_PORT | quote }}

helm/hindsight/templates/controlplane-deployment.yaml

@@ -15,7 +15,7 @@ spec:
   template:
     metadata:
       annotations:
-        checksum/config: {{ include (print $.Template.BasePath "/configmap.yaml") . | sha256sum }}
+        checksum/secret: {{ include (print $.Template.BasePath "/secret.yaml") . | sha256sum }}
         {{- with .Values.podAnnotations }}
         {{- toYaml . | nindent 8 }}
         {{- end }}
@@ -31,30 +31,26 @@ spec:
       - name: control-plane
         securityContext:
           {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag }}"
+        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag | default .Values.version }}"
         imagePullPolicy: {{ .Values.controlPlane.image.pullPolicy }}
         ports:
         - name: http
           containerPort: {{ .Values.controlPlane.service.targetPort }}
           protocol: TCP
         env:
-        - name: NODE_ENV
-          valueFrom:
-            configMapKeyRef:
-              name: {{ include "hindsight.fullname" . }}-config
-              key: node-env
-        - name: HINDSIGHT_CP_HOSTNAME
-          valueFrom:
-            configMapKeyRef:
-              name: {{ include "hindsight.fullname" . }}-config
-              key: hostname
-        - name: HINDSIGHT_CP_PORT
-          valueFrom:
-            configMapKeyRef:
-              name: {{ include "hindsight.fullname" . }}-config
-              key: control-plane-port
         - name: HINDSIGHT_CP_DATAPLANE_API_URL
           value: {{ include "hindsight.apiUrl" . | quote }}
+        {{- range $key, $value := .Values.controlPlane.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        {{- range $key, $value := .Values.controlPlane.secrets }}
+        - name: {{ $key }}
+          valueFrom:
+            secretKeyRef:
+              name: {{ include "hindsight.fullname" $ }}-secret
+              key: {{ $key }}
+        {{- end }}
         livenessProbe:
           {{- toYaml .Values.controlPlane.livenessProbe | nindent 10 }}
         readinessProbe:

helm/hindsight/templates/postgresql-service.yaml

@@ -0,0 +1,19 @@
+{{- if .Values.postgresql.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-postgresql
+  labels:
+    {{- include "hindsight.labels" . | nindent 4 }}
+    app.kubernetes.io/component: postgresql
+spec:
+  type: ClusterIP
+  ports:
+  - port: {{ .Values.postgresql.service.port }}
+    targetPort: postgresql
+    protocol: TCP
+    name: postgresql
+  selector:
+    {{- include "hindsight.selectorLabels" . | nindent 4 }}
+    app.kubernetes.io/component: postgresql
+{{- end }}

helm/hindsight/templates/postgresql-statefulset.yaml

@@ -0,0 +1,85 @@
+{{- if .Values.postgresql.enabled }}
+apiVersion: apps/v1
+kind: StatefulSet
+metadata:
+  name: {{ include "hindsight.fullname" . }}-postgresql
+  labels:
+    {{- include "hindsight.labels" . | nindent 4 }}
+    app.kubernetes.io/component: postgresql
+spec:
+  serviceName: {{ include "hindsight.fullname" . }}-postgresql
+  replicas: 1
+  selector:
+    matchLabels:
+      {{- include "hindsight.selectorLabels" . | nindent 6 }}
+      app.kubernetes.io/component: postgresql
+  template:
+    metadata:
+      labels:
+        {{- include "hindsight.selectorLabels" . | nindent 8 }}
+        app.kubernetes.io/component: postgresql
+    spec:
+      containers:
+      - name: postgresql
+        image: "{{ .Values.postgresql.image.repository }}:{{ .Values.postgresql.image.tag }}"
+        imagePullPolicy: {{ .Values.postgresql.image.pullPolicy }}
+        ports:
+        - name: postgresql
+          containerPort: 5432
+          protocol: TCP
+        env:
+        - name: POSTGRES_USER
+          value: {{ .Values.postgresql.auth.username | quote }}
+        - name: POSTGRES_PASSWORD
+          value: {{ .Values.postgresql.auth.password | quote }}
+        - name: POSTGRES_DB
+          value: {{ .Values.postgresql.auth.database | quote }}
+        - name: PGDATA
+          value: /var/lib/postgresql/data/pgdata
+        livenessProbe:
+          exec:
+            command:
+            - pg_isready
+            - -U
+            - {{ .Values.postgresql.auth.username }}
+            - -d
+            - {{ .Values.postgresql.auth.database }}
+          initialDelaySeconds: 30
+          periodSeconds: 10
+          timeoutSeconds: 5
+          failureThreshold: 3
+        readinessProbe:
+          exec:
+            command:
+            - pg_isready
+            - -U
+            - {{ .Values.postgresql.auth.username }}
+            - -d
+            - {{ .Values.postgresql.auth.database }}
+          initialDelaySeconds: 5
+          periodSeconds: 5
+          timeoutSeconds: 3
+          failureThreshold: 3
+        resources:
+          {{- toYaml .Values.postgresql.resources | nindent 10 }}
+        volumeMounts:
+        - name: data
+          mountPath: /var/lib/postgresql/data
+  {{- if .Values.postgresql.persistence.enabled }}
+  volumeClaimTemplates:
+  - metadata:
+      name: data
+    spec:
+      accessModes: ["ReadWriteOnce"]
+      {{- if .Values.postgresql.persistence.storageClass }}
+      storageClassName: {{ .Values.postgresql.persistence.storageClass | quote }}
+      {{- end }}
+      resources:
+        requests:
+          storage: {{ .Values.postgresql.persistence.size }}
+  {{- else }}
+      volumes:
+      - name: data
+        emptyDir: {}
+  {{- end }}
+{{- end }}

helm/hindsight/templates/secret.yaml

@@ -6,14 +6,12 @@ metadata:
     {{- include "hindsight.labels" . | nindent 4 }}
 type: Opaque
 data:
-  {{- if and .Values.api.secrets (hasKey .Values.api.secrets "MEMORY_LLM_API_KEY") }}
-  llm-api-key: {{ .Values.api.secrets.MEMORY_LLM_API_KEY | b64enc | quote }}
+  {{- range $key, $value := .Values.api.secrets }}
+  {{ $key }}: {{ $value | b64enc | quote }}
   {{- end }}
-  {{- if and .Values.api.secrets (hasKey .Values.api.secrets "MEMORY_LLM_BASE_URL") }}
-  llm-base-url: {{ .Values.api.secrets.MEMORY_LLM_BASE_URL | b64enc | quote }}
+  {{- range $key, $value := .Values.controlPlane.secrets }}
+  {{ $key }}: {{ $value | b64enc | quote }}
   {{- end }}
-  {{- if not .Values.postgresql.enabled }}
-  {{- if .Values.postgresql.external.password }}
+  {{- if and (not .Values.postgresql.enabled) .Values.postgresql.external.password }}
   postgres-password: {{ .Values.postgresql.external.password | b64enc | quote }}
   {{- end }}
-  {{- end }}

helm/hindsight/values.yaml

@@ -1,5 +1,8 @@
 # Default values for hindsight
 
+# Chart version - use this to set a consistent image tag across all components
+version: "0.1.1"
+
 # Global settings
 replicaCount: 1
 
@@ -8,9 +11,9 @@ api:
   enabled: true
   replicaCount: 1
   image:
-    repository: hindsight/api
+    repository: ghcr.io/vectorize-io/hindsight-api
     pullPolicy: IfNotPresent
-    tag: "latest"
+    # tag defaults to .Values.version if not specified
 
   service:
     type: ClusterIP
@@ -29,7 +32,7 @@ api:
   # Liveness and readiness probes
   livenessProbe:
     httpGet:
-      path: /
+      path: /health
       port: 8888
     initialDelaySeconds: 30
     periodSeconds: 10
@@ -38,7 +41,7 @@ api:
 
   readinessProbe:
     httpGet:
-      path: /
+      path: /health
       port: 8888
     initialDelaySeconds: 10
     periodSeconds: 5
@@ -47,7 +50,7 @@ api:
 
   # Environment variables
   env:
-    HINDSIGHT_API_LLM_PROVIDER: "groq"
+    #HINDSIGHT_API_LLM_PROVIDER: "groq"
     HINDSIGHT_API_LLM_MODEL: "openai/gpt-oss-120b"
 
   # Secret environment variables
@@ -60,9 +63,9 @@ controlPlane:
   enabled: true
   replicaCount: 1
   image:
-    repository: hindsight/hindsight-control-plane
+    repository: ghcr.io/vectorize-io/hindsight-control-plane
     pullPolicy: IfNotPresent
-    tag: "latest"
+    # tag defaults to .Values.version if not specified
 
   service:
     type: ClusterIP
@@ -78,10 +81,9 @@ controlPlane:
       cpu: 250m
       memory: 512Mi
 
-  # Liveness and readiness probes
+  # Liveness and readiness probes (TCP check)
   livenessProbe:
-    httpGet:
-      path: /
+    tcpSocket:
       port: 3000
     initialDelaySeconds: 30
     periodSeconds: 10
@@ -89,8 +91,7 @@ controlPlane:
     failureThreshold: 3
 
   readinessProbe:
-    httpGet:
-      path: /
+    tcpSocket:
       port: 3000
     initialDelaySeconds: 10
     periodSeconds: 5
@@ -106,21 +107,43 @@ controlPlane:
 # PostgreSQL configuration
 postgresql:
   # Set to true to deploy PostgreSQL as part of this chart
-  enabled: false
+  enabled: true
+
+  image:
+    repository: ankane/pgvector
+    tag: latest
+    pullPolicy: IfNotPresent
+
+  auth:
+    username: "hindsight"
+    password: "hindsight"
+    database: "hindsight"
+
+  service:
+    port: 5432
+
+  persistence:
+    enabled: true
+    size: 8Gi
+    # storageClass: ""
+
+  resources:
+    limits:
+      cpu: 1000m
+      memory: 1Gi
+    requests:
+      cpu: 250m
+      memory: 256Mi
 
   # External PostgreSQL connection details
-  # If postgresql.enabled is false, provide external database details
+  # Only used if postgresql.enabled is false
   external:
     host: "postgresql"
     port: 5432
     database: "hindsight"
     username: "hindsight"
-    # Password should be provided via secret
     # password: ""
 
-# Database URL (auto-generated from postgresql config if not provided)
-# databaseUrl: "postgresql://user:pass@host:5432/database"
-
 # Ingress configuration
 ingress:
   enabled: false

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/alembic/versions/5a366d414dce_initial_schema.py

@@ -1,274 +0,0 @@
-"""initial_schema
-
-Revision ID: 5a366d414dce
-Revises:
-Create Date: 2025-11-27 11:54:19.228030
-
-"""
-from typing import Sequence, Union
-
-from alembic import op
-import sqlalchemy as sa
-from sqlalchemy.dialects import postgresql
-from pgvector.sqlalchemy import Vector
-
-
-# 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
-
-
-def upgrade() -> None:
-    """Upgrade schema - create all tables from scratch."""
-
-    # Enable required extensions
-    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'))
-    )
-
-    # 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'))
-    )
-    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')
-    )
-    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'))
-    )
-    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))')
-
-    # 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'),
-        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'
-        )
-    )
-
-    # Add search_vector column for full-text search
-    op.execute("""
-        ALTER TABLE memory_units
-        ADD COLUMN search_vector tsvector
-        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'})
-
-    # Create BM25 full-text search index on search_vector
-    op.execute("""
-        CREATE INDEX idx_memory_units_text_search ON memory_units
-        USING gin(search_vector)
-    """)
-
-    op.execute("""
-        CREATE MATERIALIZED VIEW memory_units_bm25 AS
-        SELECT
-            id,
-            bank_id,
-            text,
-            to_tsvector('english', text) AS text_vector,
-            log(1.0 + length(text)::float / (SELECT avg(length(text)) FROM memory_units)) AS doc_length_factor
-        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')
-
-    # 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')
-    )
-    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')
-    )
-    # 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'])
-
-    # 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'))
-    )
-    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_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')
-
-    # 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_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.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_table('banks')
-
-    # Drop extensions (optional - comment out if you want to keep them)
-    # op.execute('DROP EXTENSION IF EXISTS vector')
-    # op.execute('DROP EXTENSION IF EXISTS "uuid-ossp"')

hindsight-api/alembic/versions/b7c4d8e9f1a2_add_chunks_table.py

@@ -1,70 +0,0 @@
-"""add_chunks_table
-
-Revision ID: b7c4d8e9f1a2
-Revises: 5a366d414dce
-Create Date: 2025-11-28 00:00:00.000000
-
-"""
-from typing import Sequence, Union
-
-from alembic import op
-import sqlalchemy as sa
-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
-
-
-def upgrade() -> None:
-    """Add chunks table and link memory_units to chunks."""
-
-    # 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'))
-    )
-
-    # 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'])
-
-    # 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))
-
-    # 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'
-    )
-
-    # Add index on chunk_id for efficient lookups
-    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')
-
-    # 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')

hindsight-api/alembic/versions/d9f6a3b4c5e2_rename_bank_to_interactions.py

@@ -1,48 +0,0 @@
-"""Rename fact_type 'bank' to 'experience'
-
-Revision ID: d9f6a3b4c5e2
-Revises: c8e5f2a3b4d1
-Create Date: 2024-12-04 15:00:00.000000
-
-"""
-from alembic import op
-import sqlalchemy as sa
-
-
-# revision identifiers, used by Alembic.
-revision = 'd9f6a3b4c5e2'
-down_revision = 'c8e5f2a3b4d1'
-branch_labels = None
-depends_on = None
-
-
-def upgrade():
-    # Drop old check constraint FIRST (before updating data)
-    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'")
-    # Also update any 'interactions' values (in case of partial migration)
-    op.execute("UPDATE 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')"
-    )
-
-
-def downgrade():
-    # Drop new check constraint FIRST
-    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'")
-
-    # Recreate old check constraint
-    op.create_check_constraint(
-        'memory_units_fact_type_check',
-        'memory_units',
-        "fact_type IN ('world', 'bank', 'opinion', 'observation')"
-    )

hindsight-api/alembic/versions/e0a1b2c3d4e5_disposition_to_3_traits.py

@@ -1,62 +0,0 @@
-"""disposition_to_3_traits
-
-Revision ID: e0a1b2c3d4e5
-Revises: rename_personality
-Create Date: 2024-12-08
-
-Migrate disposition traits from Big Five (openness, conscientiousness, extraversion,
-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
-import sqlalchemy as sa
-
-
-# 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
-
-
-def upgrade() -> None:
-    """Convert Big Five disposition to 3-trait disposition."""
-    conn = op.get_bind()
-
-    # Update all existing banks to use the new disposition format
-    # Convert from old format to new format with reasonable mappings:
-    # - skepticism: derived from inverse of agreeableness (skeptical people are less agreeable)
-    # - 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
-        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
-    """))
-
-
-def downgrade() -> None:
-    """Convert back to Big Five disposition."""
-    conn = op.get_bind()
-
-    # 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
-        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
-    """))

hindsight-api/alembic/versions/rename_personality_to_disposition.py

@@ -1,65 +0,0 @@
-"""rename_personality_to_disposition
-
-Revision ID: rename_personality
-Revises: d9f6a3b4c5e2
-Create Date: 2024-12-04
-
-"""
-from typing import Sequence, Union
-
-from alembic import op
-import sqlalchemy as sa
-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
-
-
-def upgrade() -> None:
-    """Rename personality column to disposition in banks table (if it exists)."""
-    conn = op.get_bind()
-
-    # Check if 'personality' column exists (old database)
-    result = conn.execute(sa.text("""
-        SELECT column_name
-        FROM information_schema.columns
-        WHERE table_name = 'banks' AND column_name = 'personality'
-    """))
-    has_personality = result.fetchone() is not None
-
-    # Check if 'disposition' column exists (new database)
-    result = conn.execute(sa.text("""
-        SELECT column_name
-        FROM information_schema.columns
-        WHERE table_name = 'banks' AND column_name = 'disposition'
-    """))
-    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')
-    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
-        ))
-    # 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("""
-        SELECT column_name
-        FROM information_schema.columns
-        WHERE table_name = 'banks' AND column_name = 'disposition'
-    """))
-    if result.fetchone():
-        op.alter_column('banks', 'disposition', new_column_name='personality')

hindsight-api/hindsight_api/__init__.py

@@ -3,24 +3,31 @@ 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, SentenceTransformersEmbeddings
-from .engine.llm_wrapper import LLMConfig
+from .models import RequestContext
 
 __all__ = [
     "MemoryEngine",
+    "RequestContext",
+    "HindsightConfig",
+    "get_config",
     "SearchTrace",
     "SearchTracer",
     "QueryInfo",
@@ -32,7 +39,11 @@ __all__ = [
     "SearchSummary",
     "SearchPhaseMetrics",
     "Embeddings",
-    "SentenceTransformersEmbeddings",
+    "LocalSTEmbeddings",
+    "RemoteTEIEmbeddings",
+    "CrossEncoderModel",
+    "LocalSTCrossEncoder",
+    "RemoteTEICrossEncoder",
     "LLMConfig",
 ]
 __version__ = "0.1.0"

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

@@ -0,0 +1,360 @@
+"""initial_schema
+
+Revision ID: 5a366d414dce
+Revises:
+Create Date: 2025-11-27 11:54:19.228030
+
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import op
+from pgvector.sqlalchemy import Vector
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+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")
+
+    # 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")),
+    )
+
+    # 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")),
+    )
+    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"
+        ),
+    )
+    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")),
+    )
+    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))")
+
+    # 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",
+        ),
+        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",
+        ),
+    )
+
+    # Add search_vector column for full-text search
+    op.execute("""
+        ALTER TABLE memory_units
+        ADD COLUMN search_vector tsvector
+        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"},
+    )
+
+    # Create BM25 full-text search index on search_vector
+    op.execute("""
+        CREATE INDEX idx_memory_units_text_search ON memory_units
+        USING gin(search_vector)
+    """)
+
+    op.execute("""
+        CREATE MATERIALIZED VIEW memory_units_bm25 AS
+        SELECT
+            id,
+            bank_id,
+            text,
+            to_tsvector('english', text) AS text_vector,
+            log(1.0 + length(text)::float / (SELECT avg(length(text)) FROM memory_units)) AS doc_length_factor
+        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")
+
+    # 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"),
+    )
+    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"),
+    )
+    # 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"])
+
+    # 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")),
+    )
+    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_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")
+
+    # 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_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.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_table("banks")
+
+    # Drop extensions (optional - comment out if you want to keep them)
+    # op.execute('DROP EXTENSION IF EXISTS vector')
+    # op.execute('DROP EXTENSION IF EXISTS "uuid-ossp"')

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

@@ -0,0 +1,70 @@
+"""add_chunks_table
+
+Revision ID: b7c4d8e9f1a2
+Revises: 5a366d414dce
+Create Date: 2025-11-28 00:00:00.000000
+
+"""
+
+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: str | Sequence[str] | None = "5a366d414dce"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    """Add chunks table and link memory_units to chunks."""
+
+    # 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")),
+    )
+
+    # 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"])
+
+    # 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))
+
+    # 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"
+    )
+
+    # Add index on chunk_id for efficient lookups
+    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")
+
+    # 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")

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

@@ -0,0 +1,53 @@
+"""Rename fact_type 'bank' to 'experience'
+
+Revision ID: d9f6a3b4c5e2
+Revises: c8e5f2a3b4d1
+Create Date: 2024-12-04 15:00:00.000000
+
+"""
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+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")
+
+    # Update existing 'bank' values to 'experience'
+    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(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')"
+    )
+
+
+def downgrade():
+    schema = _get_schema_prefix()
+
+    # Drop new check constraint FIRST
+    op.drop_constraint("memory_units_fact_type_check", "memory_units", type_="check")
+
+    # Update 'experience' back to 'bank'
+    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')"
+    )

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

@@ -0,0 +1,111 @@
+"""disposition_to_3_traits
+
+Revision ID: e0a1b2c3d4e5
+Revises: rename_personality
+Create Date: 2024-12-08
+
+Migrate disposition traits from Big Five (openness, conscientiousness, extraversion,
+agreeableness, neuroticism, bias_strength with 0-1 float values) to the new 3-trait
+system (skepticism, literalism, empathy with 1-5 integer values).
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+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:
+    # - skepticism: derived from inverse of agreeableness (skeptical people are less agreeable)
+    # - 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(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(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(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(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

@@ -0,0 +1,85 @@
+"""rename_personality_to_disposition
+
+Revision ID: rename_personality
+Revises: d9f6a3b4c5e2
+Create Date: 2024-12-04
+
+"""
+
+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: 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("""
+        SELECT column_name
+        FROM information_schema.columns
+        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("""
+        SELECT column_name
+        FROM information_schema.columns
+        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")
+    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,
+            ),
+        )
+    # else: disposition already exists, nothing to do
+
+
+def downgrade() -> None:
+    """Revert disposition column back to personality."""
+    conn = op.get_bind()
+    target_schema = _get_target_schema()
+    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 result.fetchone():
+        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
@@ -17,18 +20,17 @@ def create_app(
     http_api_enabled: bool = True,
     mcp_api_enabled: bool = False,
     mcp_mount_path: str = "/mcp",
-    run_migrations: bool = True,
-    initialize_memory: bool = True
+    initialize_memory: bool = True,
 ) -> FastAPI:
     """
     Create and configure the unified Hindsight API application.
 
     Args:
-        memory: MemoryEngine instance (already initialized with required parameters)
+        memory: MemoryEngine instance (already initialized with required parameters).
+                Migrations are controlled by the MemoryEngine's run_migrations parameter.
         http_api_enabled: Whether to enable HTTP REST API endpoints (default: True)
         mcp_api_enabled: Whether to enable MCP server (default: False)
         mcp_mount_path: Path to mount MCP server (default: /mcp)
-        run_migrations: Whether to run database migrations on startup (default: True)
         initialize_memory: Whether to initialize memory system on startup (default: True)
 
     Returns:
@@ -44,50 +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,
-            run_migrations=run_migrations,
-            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}/")
 
     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

@@ -4,28 +4,39 @@ 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.api.http import BankListItem, BankListResponse, BankProfileResponse, DispositionTraits
 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)
+# Default bank_id from environment variable
+DEFAULT_BANK_ID = os.environ.get("HINDSIGHT_MCP_BANK_ID", "default")
+
+# Context variable to hold the current bank_id
+_current_bank_id: ContextVar[str | None] = ContextVar("current_bank_id", default=None)
 
 
-def get_current_bank_id() -> Optional[str]:
-    """Get the current bank_id from context (set from URL path)."""
+def get_current_bank_id() -> str | None:
+    """Get the current bank_id from context."""
     return _current_bank_id.get()
 
 
@@ -37,12 +48,13 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
         memory: MemoryEngine instance (required)
 
     Returns:
-        Configured FastMCP server instance
+        Configured FastMCP server instance with stateless_http enabled
     """
-    mcp = FastMCP("hindsight-mcp-server")
+    # Use stateless_http=True for Claude Code compatibility
+    mcp = FastMCP("hindsight-mcp-server", stateless_http=True)
 
     @mcp.tool()
-    async def retain(content: str, context: str = "general") -> str:
+    async def retain(content: str, context: str = "general", bank_id: str | None = None) -> str:
         """
         Store important information to long-term memory.
 
@@ -58,20 +70,24 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
         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'
+            bank_id: Optional bank to store in (defaults to session bank). Use for cross-bank operations.
         """
         try:
-            bank_id = get_current_bank_id()
-            await memory.put_batch_async(
-                bank_id=bank_id,
-                contents=[{"content": content, "context": context}]
+            target_bank = bank_id or get_current_bank_id()
+            if target_bank is None:
+                return "Error: No bank_id configured"
+            await memory.retain_batch_async(
+                bank_id=target_bank,
+                contents=[{"content": content, "context": context}],
+                request_context=RequestContext(),
             )
-            return "Memory stored successfully"
+            return f"Memory stored successfully in bank '{target_bank}'"
         except Exception as e:
             logger.error(f"Error storing memory: {e}", exc_info=True)
             return f"Error: {str(e)}"
 
     @mcp.tool()
-    async def recall(query: str, max_results: int = 10) -> str:
+    async def recall(query: str, max_tokens: int = 4096, bank_id: str | None = None) -> str:
         """
         Search memories to provide personalized, context-aware responses.
 
@@ -83,49 +99,184 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
 
         Args:
             query: Natural language search query (e.g., "user's food preferences", "what projects is user working on")
-            max_results: Maximum number of results to return (default: 10)
+            max_tokens: Maximum tokens in the response (default: 4096)
+            bank_id: Optional bank to search in (defaults to session bank). Use for cross-bank operations.
         """
         try:
-            bank_id = get_current_bank_id()
+            target_bank = bank_id or get_current_bank_id()
+            if target_bank 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,
+
+            recall_result = await memory.recall_async(
+                bank_id=target_bank,
                 query=query,
                 fact_type=list(VALID_RECALL_FACT_TYPES),
-                budget=Budget.LOW
+                budget=Budget.HIGH,
+                max_tokens=max_tokens,
+                request_context=RequestContext(),
             )
 
-            results = [
-                {
-                    "id": fact.id,
-                    "text": fact.text,
-                    "type": fact.fact_type,
-                    "context": fact.context,
-                    "event_date": fact.event_date,
-                }
-                for fact in search_result.results[:max_results]
-            ]
-
-            return json.dumps({"results": results}, indent=2)
+            # Use model's JSON serialization
+            return recall_result.model_dump_json(indent=2)
         except Exception as e:
             logger.error(f"Error searching: {e}", exc_info=True)
-            return json.dumps({"error": str(e), "results": []})
+            return f'{{"error": "{e}", "results": []}}'
+
+    @mcp.tool()
+    async def reflect(query: str, context: str | None = None, budget: str = "low", bank_id: str | None = None) -> str:
+        """
+        Generate thoughtful analysis by synthesizing stored memories with the bank's personality.
+
+        WHEN TO USE THIS TOOL:
+        Use reflect when you need reasoned analysis, not just fact retrieval. This tool
+        thinks through the question using everything the bank knows and its personality traits.
+
+        EXAMPLES OF GOOD QUERIES:
+        - "What patterns have emerged in how I approach debugging?"
+        - "Based on my past decisions, what architectural style do I prefer?"
+        - "What might be the best approach for this problem given what you know about me?"
+        - "How should I prioritize these tasks based on my goals?"
+
+        HOW IT DIFFERS FROM RECALL:
+        - recall: Returns raw facts matching your search (fast lookup)
+        - reflect: Reasons across memories to form a synthesized answer (deeper analysis)
+
+        Use recall for "what did I say about X?" and reflect for "what should I do about X?"
+
+        Args:
+            query: The question or topic to reflect on
+            context: Optional context about why this reflection is needed
+            budget: Search budget - 'low', 'mid', or 'high' (default: 'low')
+            bank_id: Optional bank to reflect in (defaults to session bank). Use for cross-bank operations.
+        """
+        try:
+            target_bank = bank_id or get_current_bank_id()
+            if target_bank is None:
+                return "Error: No bank_id configured"
+            from hindsight_api.engine.memory_engine import Budget
+
+            # 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)
+
+            reflect_result = await memory.reflect_async(
+                bank_id=target_bank,
+                query=query,
+                budget=budget_enum,
+                context=context,
+                request_context=RequestContext(),
+            )
+
+            return reflect_result.model_dump_json(indent=2)
+        except Exception as e:
+            logger.error(f"Error reflecting: {e}", exc_info=True)
+            return f'{{"error": "{e}", "text": ""}}'
+
+    @mcp.tool()
+    async def list_banks() -> str:
+        """
+        List all available memory banks.
+
+        Use this to discover banks for orchestration or to find
+        the correct bank_id for cross-bank operations.
+
+        Returns:
+            JSON object with banks array containing bank_id, name, disposition, background, and timestamps
+        """
+        try:
+            banks = await memory.list_banks(request_context=RequestContext())
+            bank_items = [
+                BankListItem(
+                    bank_id=b.get("bank_id") or b.get("id"),
+                    name=b.get("name"),
+                    disposition=DispositionTraits(
+                        **b.get("disposition", {"skepticism": 3, "literalism": 3, "empathy": 3})
+                    ),
+                    background=b.get("background"),
+                    created_at=str(b.get("created_at")) if b.get("created_at") else None,
+                    updated_at=str(b.get("updated_at")) if b.get("updated_at") else None,
+                )
+                for b in banks
+            ]
+            return BankListResponse(banks=bank_items).model_dump_json(indent=2)
+        except Exception as e:
+            logger.error(f"Error listing banks: {e}", exc_info=True)
+            return f'{{"error": "{e}", "banks": []}}'
+
+    @mcp.tool()
+    async def create_bank(bank_id: str, name: str | None = None, background: str | None = None) -> str:
+        """
+        Create or update a memory bank.
+
+        Use this to create new banks for different agents, sessions, or purposes.
+        Banks are isolated memory stores - each bank has its own memories and personality.
+
+        Args:
+            bank_id: Unique identifier for the bank (e.g., 'orchestrator-memory', 'agent-1')
+            name: Human-readable name for the bank
+            background: Context about what this bank stores or its purpose
+        """
+        try:
+            # Get or create the bank profile (auto-creates with defaults)
+            await memory.get_bank_profile(bank_id, request_context=RequestContext())
+
+            # Update name and/or background if provided
+            if name is not None or background is not None:
+                await memory.update_bank(bank_id, name=name, background=background, request_context=RequestContext())
+
+            # Get final profile and return using BankProfileResponse model
+            profile = await memory.get_bank_profile(bank_id, request_context=RequestContext())
+            disposition = profile.get("disposition")
+            if hasattr(disposition, "model_dump"):
+                disposition_traits = DispositionTraits(**disposition.model_dump())
+            else:
+                disposition_traits = DispositionTraits(
+                    **dict(disposition or {"skepticism": 3, "literalism": 3, "empathy": 3})
+                )
+
+            response = BankProfileResponse(
+                bank_id=bank_id,
+                name=profile.get("name") or "",
+                disposition=disposition_traits,
+                background=profile.get("background") or "",
+            )
+            return response.model_dump_json(indent=2)
+        except Exception as e:
+            logger.error(f"Error creating bank: {e}", exc_info=True)
+            return json.dumps({"error": str(e)})
 
     return mcp
 
 
 class MCPMiddleware:
-    """ASGI middleware that extracts bank_id from path and sets context."""
+    """ASGI middleware that extracts bank_id from header or path and sets context.
+
+    Bank ID can be provided via:
+    1. X-Bank-Id header (recommended for Claude Code)
+    2. URL path: /mcp/{bank_id}/
+    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback default)
+
+    For Claude Code, configure with:
+        claude mcp add --transport http hindsight http://localhost:8888/mcp \\
+            --header "X-Bank-Id: my-bank"
+    """
 
     def __init__(self, app, memory: MemoryEngine):
         self.app = app
         self.memory = memory
         self.mcp_server = create_mcp_server(memory)
-        # Use sse_app - http_app requires lifespan management that's complex with middleware
-        import warnings
-        with warnings.catch_warnings():
-            warnings.simplefilter("ignore", DeprecationWarning)
-            self.mcp_app = self.mcp_server.sse_app()
+        self.mcp_app = self.mcp_server.http_app(path="/")
+        # Expose the lifespan for the parent app to chain
+        self.lifespan = self.mcp_app.lifespan_handler if hasattr(self.mcp_app, "lifespan_handler") else None
+
+    def _get_header(self, scope: dict, name: str) -> str | None:
+        """Extract a header value from ASGI scope."""
+        name_lower = name.lower().encode()
+        for header_name, header_value in scope.get("headers", []):
+            if header_name.lower() == name_lower:
+                return header_value.decode()
+        return None
 
     async def __call__(self, scope, receive, send):
         if scope["type"] != "http":
@@ -137,46 +288,50 @@ 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/"):
             path = path[4:]  # Remove /mcp prefix
+        elif path == "/mcp":
+            path = "/"
 
-        # Extract bank_id from path: /{bank_id}/ or /{bank_id}
-        # http_app expects requests at /
-        if not path.startswith("/") or len(path) <= 1:
-            # No bank_id in path - return error
-            await self._send_error(send, 400, "bank_id required in path: /mcp/{bank_id}/")
-            return
+        # Try to get bank_id from header first (for Claude Code compatibility)
+        bank_id = self._get_header(scope, "X-Bank-Id")
 
-        # Extract bank_id from first path segment
-        parts = path[1:].split("/", 1)
-        if not parts[0]:
-            await self._send_error(send, 400, "bank_id required in path: /mcp/{bank_id}/")
-            return
+        # MCP endpoint paths that should not be treated as bank_ids
+        MCP_ENDPOINTS = {"sse", "messages"}
 
-        bank_id = parts[0]
-        new_path = "/" + parts[1] if len(parts) > 1 else "/"
+        # If no header, try to extract from path: /{bank_id}/...
+        new_path = path
+        if not bank_id and path.startswith("/") and len(path) > 1:
+            parts = path[1:].split("/", 1)
+            # Don't treat MCP endpoints as bank_ids
+            if parts[0] and parts[0] not in MCP_ENDPOINTS:
+                # First segment looks like a bank_id
+                bank_id = parts[0]
+                new_path = "/" + parts[1] if len(parts) > 1 else "/"
+
+        # Fall back to default bank_id
+        if not bank_id:
+            bank_id = DEFAULT_BANK_ID
+            logger.debug(f"Using default bank_id: {bank_id}")
 
         # Set bank_id context
         token = _current_bank_id.set(bank_id)
         try:
             new_scope = scope.copy()
             new_scope["path"] = new_path
+            # Clear root_path since we're passing directly to the app
+            new_scope["root_path"] = ""
 
-            # Wrap send to rewrite the SSE endpoint URL to include bank_id
-            # The SSE app sends "event: endpoint\ndata: /messages\n" but we need
-            # the client to POST to /{bank_id}/messages instead
+            # Wrap send to rewrite the SSE endpoint URL to include bank_id if using path-based routing
             async def send_wrapper(message):
                 if message["type"] == "http.response.body":
                     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)
 
@@ -187,24 +342,29 @@ 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):
     """
     Create an ASGI app that handles MCP requests.
 
-    URL pattern: /mcp/{bank_id}/
-
-    The bank_id is extracted from the URL path and made available to tools.
+    Bank ID can be provided via:
+    1. X-Bank-Id header: claude mcp add --transport http hindsight http://localhost:8888/mcp --header "X-Bank-Id: my-bank"
+    2. URL path: /mcp/{bank_id}/
+    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback, default: "default")
 
     Args:
         memory: MemoryEngine instance

hindsight-api/hindsight_api/banner.py

@@ -0,0 +1,96 @@
+"""
+Banner display for Hindsight API startup.
+
+Shows the logo and tagline with gradient colors.
+"""
+
+# Gradient colors: #0074d9 -> #009296
+GRADIENT_START = (0, 116, 217)  # #0074d9
+GRADIENT_END = (0, 146, 150)  # #009296
+
+# Pre-generated logo (generated by test-logo.py)
+LOGO = """\
+  \033[38;2;9;127;184m\u2584\033[0m\033[48;2;8;130;178m\033[38;2;5;133;186m\u2584\033[0m       \033[48;2;10;143;160m\033[38;2;10;143;165m\u2584\033[0m\033[38;2;7;140;156m\u2584\033[0m
+ \033[38;2;8;125;192m\u2584\033[0m \033[38;2;3;132;191m\u2580\033[0m\033[38;2;2;133;192m\u2584\033[0m \033[38;2;3;132;180m\u2584\033[0m\033[38;2;1;137;184m\u2584\033[0m\033[38;2;3;133;174m\u2584\033[0m \033[38;2;3;142;176m\u2584\033[0m\033[38;2;4;142;169m\u2580\033[0m \033[38;2;10;144;164m\u2584\033[0m
+\033[38;2;6;121;195m\u2580\033[0m\033[38;2;5;128;203m\u2580\033[0m\033[48;2;5;124;195m\033[38;2;3;125;200m\u2584\033[0m\033[38;2;2;126;196m\u2584\033[0m\033[48;2;3;128;188m\033[38;2;1;131;196m\u2584\033[0m\033[48;2;0;152;219m\033[38;2;2;131;191m\u2584\033[0m\033[38;2;1;141;196m\u2580\033[0m\033[38;2;1;135;183m\u2580\033[0m\033[38;2;1;148;198m\u2580\033[0m\033[48;2;1;156;202m\033[38;2;2;135;180m\u2584\033[0m\033[48;2;4;134;169m\033[38;2;1;137;177m\u2584\033[0m\033[38;2;3;138;173m\u2584\033[0m\033[48;2;6;137;165m\033[38;2;2;140;170m\u2584\033[0m\033[38;2;7;144;169m\u2580\033[0m\033[38;2;7;139;158m\u2580\033[0m
+   \033[48;2;2;128;202m\033[38;2;2;124;201m\u2584\033[0m\033[48;2;1;130;201m\033[38;2;0;135;212m\u2584\033[0m\033[38;2;2;128;196m\u2584\033[0m \033[48;2;2;142;204m\033[38;2;7;138;199m\u2584\033[0m \033[38;2;1;135;186m\u2584\033[0m\033[48;2;1;142;186m\033[38;2;2;144;194m\u2584\033[0m\033[48;2;3;138;176m\033[38;2;2;134;176m\u2584\033[0m
+ \033[48;2;8;118;200m\033[38;2;8;121;209m\u2584\033[0m\033[38;2;3;121;203m\u2580\033[0m \033[38;2;3;122;192m\u2580\033[0m\033[38;2;1;138;216m\u2580\033[0m\033[48;2;0;138;210m\033[38;2;3;128;198m\u2584\033[0m\033[48;2;0;126;188m\033[38;2;2;131;198m\u2584\033[0m\033[48;2;0;142;205m\033[38;2;3;132;193m\u2584\033[0m\033[38;2;1;140;196m\u2580\033[0m  \033[38;2;4;134;175m\u2580\033[0m\033[48;2;13;135;167m\033[38;2;8;136;174m\u2584\033[0m """
+
+
+def _interpolate_color(start: tuple, end: tuple, t: float) -> tuple:
+    """Interpolate between two RGB colors."""
+    return (
+        int(start[0] + (end[0] - start[0]) * t),
+        int(start[1] + (end[1] - start[1]) * t),
+        int(start[2] + (end[2] - start[2]) * t),
+    )
+
+
+def gradient_text(text: str, start: tuple = GRADIENT_START, end: tuple = GRADIENT_END) -> str:
+    """Render text with a gradient color effect."""
+    result = []
+    length = len(text)
+    for i, char in enumerate(text):
+        if char == " ":
+            result.append(" ")
+        else:
+            t = i / max(length - 1, 1)
+            r, g, b = _interpolate_color(start, end, t)
+            result.append(f"\033[38;2;{r};{g};{b}m{char}")
+    result.append("\033[0m")
+    return "".join(result)
+
+
+def print_banner():
+    """Print the Hindsight startup banner."""
+    print(LOGO)
+    tagline = gradient_text("Hindsight: Agent Memory That Works Like Human Memory")
+    print(f"\n  {tagline}\n")
+
+
+def color(text: str, t: float = 0.0) -> str:
+    """Color text using gradient position (0.0 = start, 1.0 = end)."""
+    r, g, b = _interpolate_color(GRADIENT_START, GRADIENT_END, t)
+    return f"\033[38;2;{r};{g};{b}m{text}\033[0m"
+
+
+def color_start(text: str) -> str:
+    """Color text with gradient start color (#0074d9)."""
+    return color(text, 0.0)
+
+
+def color_end(text: str) -> str:
+    """Color text with gradient end color (#009296)."""
+    return color(text, 1.0)
+
+
+def color_mid(text: str) -> str:
+    """Color text with gradient middle color."""
+    return color(text, 0.5)
+
+
+def dim(text: str) -> str:
+    """Dim/gray text."""
+    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,
+):
+    """Print styled startup information."""
+    print(color_start("Starting Hindsight API..."))
+    print(f"  {dim('URL:')} {color(f'http://{host}:{port}', 0.2)}")
+    print(f"  {dim('Database:')} {color(database_url, 0.4)}")
+    print(f"  {dim('LLM:')} {color(f'{llm_provider} / {llm_model}', 0.6)}")
+    print(f"  {dim('Embeddings:')} {color(embeddings_provider, 0.8)}")
+    print(f"  {dim('Reranker:')} {color(reranker_provider, 1.0)}")
+    if mcp_enabled:
+        print(f"  {dim('MCP:')} {color_end('enabled at /mcp')}")
+    print()

hindsight-api/hindsight_api/cli.py

@@ -1,128 +0,0 @@
-"""
-Command-line interface for Hindsight API.
-
-Run the server with:
-    hindsight-api
-
-Stop with Ctrl+C.
-"""
-import argparse
-import asyncio
-import atexit
-import os
-import signal
-import sys
-from typing import Optional
-
-import uvicorn
-
-from . import MemoryEngine
-from .api import create_app
-
-
-# Disable tokenizers parallelism to avoid warnings
-os.environ["TOKENIZERS_PARALLELISM"] = "false"
-
-# Global reference for cleanup
-_memory: Optional[MemoryEngine] = None
-
-
-def _cleanup():
-    """Synchronous cleanup function to stop resources on exit."""
-    global _memory
-    if _memory is not None and _memory._pg0 is not None:
-        try:
-            loop = asyncio.new_event_loop()
-            loop.run_until_complete(_memory._pg0.stop())
-            loop.close()
-            print("\npg0 stopped.")
-        except Exception as e:
-            print(f"\nError stopping pg0: {e}")
-
-
-def _signal_handler(signum, frame):
-    """Handle SIGINT/SIGTERM to ensure cleanup."""
-    print(f"\nReceived signal {signum}, shutting down...")
-    _cleanup()
-    sys.exit(0)
-
-
-def main():
-    """Main entry point for the CLI."""
-    global _memory
-
-    parser = argparse.ArgumentParser(
-        prog="hindsight-api",
-        description="Hindsight API Server",
-    )
-    parser.add_argument(
-        "--host", default="0.0.0.0",
-        help="Host to bind to (default: 0.0.0.0)"
-    )
-    parser.add_argument(
-        "--port", type=int, default=8888,
-        help="Port to bind to (default: 8888)"
-    )
-    parser.add_argument(
-        "--log-level", default="info",
-        choices=["critical", "error", "warning", "info", "debug", "trace"],
-        help="Log level (default: info)"
-    )
-    parser.add_argument(
-        "--access-log", action="store_true",
-        help="Enable access log"
-    )
-
-    args = parser.parse_args()
-
-    # Register cleanup handlers
-    atexit.register(_cleanup)
-    signal.signal(signal.SIGINT, _signal_handler)
-    signal.signal(signal.SIGTERM, _signal_handler)
-
-    # Get configuration from environment variables
-    db_url = os.getenv("HINDSIGHT_API_DATABASE_URL", "pg0")
-    llm_provider = os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq")
-    llm_api_key = os.getenv("HINDSIGHT_API_LLM_API_KEY", "")
-    llm_model = os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-20b")
-    llm_base_url = os.getenv("HINDSIGHT_API_LLM_BASE_URL") or None
-
-    # Create MemoryEngine
-    _memory = MemoryEngine(
-        db_url=db_url,
-        memory_llm_provider=llm_provider,
-        memory_llm_api_key=llm_api_key,
-        memory_llm_model=llm_model,
-        memory_llm_base_url=llm_base_url,
-    )
-
-    # Create FastAPI app
-    app = create_app(
-        memory=_memory,
-        http_api_enabled=True,
-        mcp_api_enabled=True,
-        mcp_mount_path="/mcp",
-        run_migrations=True,
-        initialize_memory=True,
-    )
-
-    # Prepare uvicorn config
-    uvicorn_config = {
-        "app": app,
-        "host": args.host,
-        "port": args.port,
-        "log_level": args.log_level,
-        "access_log": args.access_log,
-    }
-
-    print(f"\nStarting Hindsight API...")
-    print(f"  URL: http://{args.host}:{args.port}")
-    print(f"  Database: {db_url}")
-    print(f"  LLM Provider: {llm_provider}")
-    print()
-
-    uvicorn.run(**uvicorn_config)
-
-
-if __name__ == "__main__":
-    main()

hindsight-api/hindsight_api/config.py

@@ -0,0 +1,225 @@
+"""
+Centralized configuration for Hindsight API.
+
+All environment variables and their defaults are defined here.
+"""
+
+import logging
+import os
+from dataclasses import dataclass
+
+logger = logging.getLogger(__name__)
+
+# Environment variable names
+ENV_DATABASE_URL = "HINDSIGHT_API_DATABASE_URL"
+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_LLM_GROQ_SERVICE_TIER = "HINDSIGHT_API_LLM_GROQ_SERVICE_TIER"
+
+ENV_EMBEDDINGS_PROVIDER = "HINDSIGHT_API_EMBEDDINGS_PROVIDER"
+ENV_EMBEDDINGS_LOCAL_MODEL = "HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL"
+ENV_EMBEDDINGS_TEI_URL = "HINDSIGHT_API_EMBEDDINGS_TEI_URL"
+ENV_EMBEDDINGS_OPENAI_API_KEY = "HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY"
+ENV_EMBEDDINGS_OPENAI_MODEL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_MODEL"
+
+ENV_RERANKER_PROVIDER = "HINDSIGHT_API_RERANKER_PROVIDER"
+ENV_RERANKER_LOCAL_MODEL = "HINDSIGHT_API_RERANKER_LOCAL_MODEL"
+ENV_RERANKER_TEI_URL = "HINDSIGHT_API_RERANKER_TEI_URL"
+
+ENV_HOST = "HINDSIGHT_API_HOST"
+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"
+DEFAULT_EMBEDDINGS_OPENAI_MODEL = "text-embedding-3-small"
+DEFAULT_EMBEDDING_DIMENSION = 384
+
+DEFAULT_RERANKER_PROVIDER = "local"
+DEFAULT_RERANKER_LOCAL_MODEL = "cross-encoder/ms-marco-MiniLM-L-6-v2"
+
+DEFAULT_HOST = "0.0.0.0"
+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"""
+
+# Default embedding dimension (used by initial migration, adjusted at runtime)
+EMBEDDING_DIMENSION = DEFAULT_EMBEDDING_DIMENSION
+
+
+@dataclass
+class HindsightConfig:
+    """Configuration container for Hindsight API."""
+
+    # Database
+    database_url: str
+
+    # LLM
+    llm_provider: str
+    llm_api_key: str | None
+    llm_model: str
+    llm_base_url: str | None
+    llm_max_concurrent: int
+    llm_timeout: float
+
+    # Embeddings
+    embeddings_provider: str
+    embeddings_local_model: str
+    embeddings_tei_url: str | None
+
+    # Reranker
+    reranker_provider: str
+    reranker_local_model: str
+    reranker_tei_url: str | None
+
+    # Server
+    host: str
+    port: int
+    log_level: str
+    mcp_enabled: bool
+
+    # 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:
+        """Get the LLM base URL, with provider-specific defaults."""
+        if self.llm_base_url:
+            return self.llm_base_url
+
+        provider = self.llm_provider.lower()
+        if provider == "groq":
+            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 ""
+
+    def get_python_log_level(self) -> int:
+        """Get the Python logging level from the configured log level string."""
+        log_level_map = {
+            "critical": logging.CRITICAL,
+            "error": logging.ERROR,
+            "warning": logging.WARNING,
+            "info": logging.INFO,
+            "debug": logging.DEBUG,
+            "trace": logging.DEBUG,  # Python doesn't have TRACE, use DEBUG
+        }
+        return log_level_map.get(self.log_level.lower(), logging.INFO)
+
+    def configure_logging(self) -> None:
+        """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",
+            force=True,  # Override any existing configuration
+        )
+
+    def log_config(self) -> None:
+        """Log the current configuration (without sensitive values)."""
+        logger.info(f"Database: {self.database_url}")
+        logger.info(f"LLM: provider={self.llm_provider}, model={self.llm_model}")
+        logger.info(f"Embeddings: provider={self.embeddings_provider}")
+        logger.info(f"Reranker: provider={self.reranker_provider}")
+        logger.info(f"Graph retriever: {self.graph_retriever}")
+
+
+def get_config() -> HindsightConfig:
+    """Get the current configuration from environment variables."""
+    return HindsightConfig.from_env()

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

@@ -7,29 +7,40 @@ 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, SentenceTransformersEmbeddings
+from .embeddings import Embeddings, LocalSTEmbeddings, RemoteTEIEmbeddings
+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",
     "acquire_with_retry",
     "Embeddings",
-    "SentenceTransformersEmbeddings",
+    "LocalSTEmbeddings",
+    "RemoteTEIEmbeddings",
+    "CrossEncoderModel",
+    "LocalSTCrossEncoder",
+    "RemoteTEICrossEncoder",
     "SearchTrace",
     "SearchTracer",
     "QueryInfo",
@@ -44,4 +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

@@ -2,10 +2,23 @@
 Cross-encoder abstraction for reranking.
 
 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
+
 import logging
+import os
+from abc import ABC, abstractmethod
+
+import httpx
+
+from ..config import (
+    DEFAULT_RERANKER_LOCAL_MODEL,
+    DEFAULT_RERANKER_PROVIDER,
+    ENV_RERANKER_LOCAL_MODEL,
+    ENV_RERANKER_PROVIDER,
+    ENV_RERANKER_TEI_URL,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -17,18 +30,24 @@ class CrossEncoderModel(ABC):
     Cross-encoders take query-document pairs and return relevance scores.
     """
 
+    @property
     @abstractmethod
-    def load(self) -> None:
-        """
-        Load the cross-encoder model.
+    def provider_name(self) -> str:
+        """Return a human-readable name for this provider (e.g., 'local', 'tei')."""
+        pass
 
-        This should be called during initialization to load the model
+    @abstractmethod
+    async def initialize(self) -> None:
+        """
+        Initialize the cross-encoder model asynchronously.
+
+        This should be called during startup to load/connect to the model
         and avoid cold start latency on first predict() call.
         """
         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.
 
@@ -41,11 +60,11 @@ class CrossEncoderModel(ABC):
         pass
 
 
-class SentenceTransformersCrossEncoder(CrossEncoderModel):
+class LocalSTCrossEncoder(CrossEncoderModel):
     """
-    Cross-encoder implementation using SentenceTransformers.
+    Local cross-encoder implementation using SentenceTransformers.
 
-    Call load() during initialization to load the model and avoid cold starts.
+    Call initialize() during startup to load the model and avoid cold starts.
 
     Default model is cross-encoder/ms-marco-MiniLM-L-6-v2:
     - Fast inference (~80ms for 100 pairs on CPU)
@@ -53,18 +72,22 @@ class SentenceTransformersCrossEncoder(CrossEncoderModel):
     - Trained for passage re-ranking
     """
 
-    def __init__(self, model_name: str = "cross-encoder/ms-marco-MiniLM-L-6-v2"):
+    def __init__(self, model_name: str | None = None):
         """
-        Initialize SentenceTransformers cross-encoder.
+        Initialize local SentenceTransformers cross-encoder.
 
         Args:
             model_name: Name of the CrossEncoder model to use.
                        Default: cross-encoder/ms-marco-MiniLM-L-6-v2
         """
-        self.model_name = model_name
+        self.model_name = model_name or DEFAULT_RERANKER_LOCAL_MODEL
         self._model = None
 
-    def load(self) -> None:
+    @property
+    def provider_name(self) -> str:
+        return "local"
+
+    async def initialize(self) -> None:
         """Load the cross-encoder model."""
         if self._model is not None:
             return
@@ -73,15 +96,15 @@ class SentenceTransformersCrossEncoder(CrossEncoderModel):
             from sentence_transformers import CrossEncoder
         except ImportError:
             raise ImportError(
-                "sentence-transformers is required for SentenceTransformersCrossEncoder. "
+                "sentence-transformers is required for LocalSTCrossEncoder. "
                 "Install it with: pip install sentence-transformers"
             )
 
-        logger.info(f"Loading cross-encoder model: {self.model_name}...")
+        logger.info(f"Reranker: initializing local provider with model {self.model_name}")
         self._model = CrossEncoder(self.model_name)
-        logger.info("Cross-encoder model loaded")
+        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.
 
@@ -92,6 +115,188 @@ class SentenceTransformersCrossEncoder(CrossEncoderModel):
             List of relevance scores (raw logits from the model)
         """
         if self._model is None:
-            self.load()
+            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):
+    """
+    Remote cross-encoder implementation using HuggingFace Text Embeddings Inference (TEI) HTTP API.
+
+    TEI supports reranking via the /rerank endpoint.
+    See: https://github.com/huggingface/text-embeddings-inference
+
+    Note: The TEI server must be running a cross-encoder/reranker model.
+    """
+
+    def __init__(
+        self,
+        base_url: str,
+        timeout: float = 30.0,
+        batch_size: int = 32,
+        max_retries: int = 3,
+        retry_delay: float = 0.5,
+    ):
+        """
+        Initialize remote TEI cross-encoder client.
+
+        Args:
+            base_url: Base URL of the TEI server (e.g., "http://localhost:8080")
+            timeout: Request timeout in seconds (default: 30.0)
+            batch_size: Maximum batch size for rerank requests (default: 32)
+            max_retries: Maximum number of retries for failed requests (default: 3)
+            retry_delay: Initial delay between retries in seconds, doubles each retry (default: 0.5)
+        """
+        self.base_url = base_url.rstrip("/")
+        self.timeout = timeout
+        self.batch_size = batch_size
+        self.max_retries = max_retries
+        self.retry_delay = retry_delay
+        self._client: httpx.Client | None = None
+        self._model_id: str | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "tei"
+
+    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
+
+        for attempt in range(self.max_retries + 1):
+            try:
+                if method == "GET":
+                    response = self._client.get(url, **kwargs)
+                else:
+                    response = self._client.post(url, **kwargs)
+                response.raise_for_status()
+                return response
+            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..."
+                    )
+                    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..."
+                    )
+                    time.sleep(delay)
+                    delay *= 2
+                else:
+                    raise
+
+        raise last_error
+
+    async def initialize(self) -> None:
+        """Initialize the HTTP client and verify server connectivity."""
+        if self._client is not None:
+            return
+
+        logger.info(f"Reranker: initializing TEI provider at {self.base_url}")
+        self._client = httpx.Client(timeout=self.timeout)
+
+        # Verify server is reachable and get model info
+        try:
+            response = self._request_with_retry("GET", f"{self.base_url}/info")
+            info = response.json()
+            self._model_id = info.get("model_id", "unknown")
+            logger.info(f"Reranker: TEI provider initialized (model: {self._model_id})")
+        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]:
+        """
+        Score query-document pairs using the remote TEI reranker.
+
+        Args:
+            pairs: List of (query, document) tuples to score
+
+        Returns:
+            List of relevance scores
+        """
+        if self._client is None:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        if not pairs:
+            return []
+
+        all_scores = []
+
+        # Process in batches
+        for i in range(0, len(pairs), 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
+            # but we handle mixed queries by making separate requests per unique query
+            query_groups: dict[str, list[tuple[int, str]]] = {}
+            for idx, (query, text) in enumerate(batch):
+                if query not in query_groups:
+                    query_groups[query] = []
+                query_groups[query].append((idx, text))
+
+            batch_scores = [0.0] * len(batch)
+
+            for query, indexed_texts in query_groups.items():
+                texts = [text for _, text in indexed_texts]
+                indices = [idx for idx, _ in indexed_texts]
+
+                try:
+                    response = self._request_with_retry(
+                        "POST",
+                        f"{self.base_url}/rerank",
+                        json={
+                            "query": query,
+                            "texts": texts,
+                            "return_text": False,
+                        },
+                    )
+                    results = response.json()
+
+                    # TEI returns results sorted by score descending, with original index
+                    for result in results:
+                        original_idx = result["index"]
+                        score = result["score"]
+                        # Map back to batch position
+                        batch_scores[indices[original_idx]] = score
+
+                except httpx.HTTPError as e:
+                    raise RuntimeError(f"TEI rerank request failed: {e}")
+
+            all_scores.extend(batch_scores)
+
+        return all_scores
+
+
+def create_cross_encoder_from_env() -> CrossEncoderModel:
+    """
+    Create a CrossEncoderModel instance based on environment variables.
+
+    See hindsight_api.config for environment variable names and defaults.
+
+    Returns:
+        Configured CrossEncoderModel instance
+    """
+    provider = os.environ.get(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER).lower()
+
+    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'")
+        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'")

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

@@ -3,74 +3,108 @@ Embeddings abstraction for the memory system.
 
 Provides an interface for generating embeddings with different backends.
 
-IMPORTANT: All embeddings must produce 384-dimensional vectors to match
-the database schema (pgvector column defined as vector(384)).
+The embedding dimension is auto-detected from the model at initialization.
+The database schema is automatically adjusted to match the model's dimension.
+
+Configuration via environment variables - see hindsight_api.config for all env var names.
 """
-from abc import ABC, abstractmethod
-from typing import List
+
 import logging
+import os
+from abc import ABC, abstractmethod
+
+import httpx
+
+from ..config import (
+    DEFAULT_EMBEDDINGS_LOCAL_MODEL,
+    DEFAULT_EMBEDDINGS_OPENAI_MODEL,
+    DEFAULT_EMBEDDINGS_PROVIDER,
+    ENV_EMBEDDINGS_LOCAL_MODEL,
+    ENV_EMBEDDINGS_OPENAI_API_KEY,
+    ENV_EMBEDDINGS_OPENAI_MODEL,
+    ENV_EMBEDDINGS_PROVIDER,
+    ENV_EMBEDDINGS_TEI_URL,
+    ENV_LLM_API_KEY,
+)
 
 logger = logging.getLogger(__name__)
 
-# Fixed embedding dimension required by database schema
-EMBEDDING_DIMENSION = 384
-
 
 class Embeddings(ABC):
     """
     Abstract base class for embedding generation.
 
-    All implementations MUST generate 384-dimensional embeddings to match
-    the database schema.
+    The embedding dimension is determined by the model and detected at initialization.
+    The database schema is automatically adjusted to match the model's dimension.
     """
 
+    @property
     @abstractmethod
-    def load(self) -> None:
-        """
-        Load the embedding model.
+    def provider_name(self) -> str:
+        """Return a human-readable name for this provider (e.g., 'local', 'tei')."""
+        pass
 
-        This should be called during initialization to load the model
+    @property
+    @abstractmethod
+    def dimension(self) -> int:
+        """Return the embedding dimension produced by this model."""
+        pass
+
+    @abstractmethod
+    async def initialize(self) -> None:
+        """
+        Initialize the embedding model asynchronously.
+
+        This should be called during startup to load/connect to the model
         and avoid cold start latency on first encode() call.
         """
         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.
+        Generate embeddings for a list of texts.
 
         Args:
             texts: List of text strings to encode
 
         Returns:
-            List of 384-dimensional embedding vectors (each is a list of floats)
+            List of embedding vectors (each is a list of floats)
         """
         pass
 
 
-class SentenceTransformersEmbeddings(Embeddings):
+class LocalSTEmbeddings(Embeddings):
     """
-    Embeddings implementation using SentenceTransformers.
+    Local embeddings implementation using SentenceTransformers.
 
-    Call load() during initialization to load the model and avoid cold starts.
-
-    Default model is BAAI/bge-small-en-v1.5 which produces 384-dimensional
-    embeddings matching the database schema.
+    Call initialize() during startup to load the model and avoid cold starts.
+    The embedding dimension is auto-detected from the model.
     """
 
-    def __init__(self, model_name: str = "BAAI/bge-small-en-v1.5"):
+    def __init__(self, model_name: str | None = None):
         """
-        Initialize SentenceTransformers embeddings.
+        Initialize local SentenceTransformers embeddings.
 
         Args:
             model_name: Name of the SentenceTransformer model to use.
-                       Must produce 384-dimensional embeddings.
                        Default: BAAI/bge-small-en-v1.5
         """
-        self.model_name = model_name
+        self.model_name = model_name or DEFAULT_EMBEDDINGS_LOCAL_MODEL
         self._model = None
+        self._dimension: int | None = None
 
-    def load(self) -> None:
+    @property
+    def provider_name(self) -> str:
+        return "local"
+
+    @property
+    def dimension(self) -> int:
+        if self._dimension is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+        return self._dimension
+
+    async def initialize(self) -> None:
         """Load the embedding model."""
         if self._model is not None:
             return
@@ -79,35 +113,331 @@ class SentenceTransformersEmbeddings(Embeddings):
             from sentence_transformers import SentenceTransformer
         except ImportError:
             raise ImportError(
-                "sentence-transformers is required for SentenceTransformersEmbeddings. "
+                "sentence-transformers is required for LocalSTEmbeddings. "
                 "Install it with: pip install sentence-transformers"
             )
 
-        logger.info(f"Loading embedding model: {self.model_name}...")
-        self._model = SentenceTransformer(self.model_name)
+        logger.info(f"Embeddings: initializing local provider with model {self.model_name}")
+        # Disable lazy loading (meta tensors) which causes issues with newer transformers/accelerate
+        # Setting low_cpu_mem_usage=False and device_map=None ensures tensors are fully materialized
+        self._model = SentenceTransformer(
+            self.model_name,
+            model_kwargs={"low_cpu_mem_usage": False, "device_map": None},
+        )
 
-        # Validate dimension matches database schema
-        model_dim = self._model.get_sentence_embedding_dimension()
-        if model_dim != EMBEDDING_DIMENSION:
-            raise ValueError(
-                f"Model {self.model_name} produces {model_dim}-dimensional embeddings, "
-                f"but database schema requires {EMBEDDING_DIMENSION} dimensions. "
-                f"Use a model that produces {EMBEDDING_DIMENSION}-dimensional embeddings."
-            )
+        self._dimension = self._model.get_sentence_embedding_dimension()
+        logger.info(f"Embeddings: local provider initialized (dim: {self._dimension})")
 
-        logger.info(f"Model loaded (embedding 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.
+        Generate embeddings for a list of texts.
 
         Args:
             texts: List of text strings to encode
 
         Returns:
-            List of 384-dimensional embedding vectors
+            List of embedding vectors
         """
         if self._model is None:
-            self.load()
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
         embeddings = self._model.encode(texts, convert_to_numpy=True, show_progress_bar=False)
         return [emb.tolist() for emb in embeddings]
+
+
+class RemoteTEIEmbeddings(Embeddings):
+    """
+    Remote embeddings implementation using HuggingFace Text Embeddings Inference (TEI) HTTP API.
+
+    TEI provides a high-performance inference server for embedding models.
+    See: https://github.com/huggingface/text-embeddings-inference
+
+    The embedding dimension is auto-detected from the server at initialization.
+    """
+
+    def __init__(
+        self,
+        base_url: str,
+        timeout: float = 30.0,
+        batch_size: int = 32,
+        max_retries: int = 3,
+        retry_delay: float = 0.5,
+    ):
+        """
+        Initialize remote TEI embeddings client.
+
+        Args:
+            base_url: Base URL of the TEI server (e.g., "http://localhost:8080")
+            timeout: Request timeout in seconds (default: 30.0)
+            batch_size: Maximum batch size for embedding requests (default: 32)
+            max_retries: Maximum number of retries for failed requests (default: 3)
+            retry_delay: Initial delay between retries in seconds, doubles each retry (default: 0.5)
+        """
+        self.base_url = base_url.rstrip("/")
+        self.timeout = timeout
+        self.batch_size = batch_size
+        self.max_retries = max_retries
+        self.retry_delay = retry_delay
+        self._client: httpx.Client | None = None
+        self._model_id: str | None = None
+        self._dimension: int | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "tei"
+
+    @property
+    def dimension(self) -> int:
+        if self._dimension is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+        return self._dimension
+
+    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
+
+        for attempt in range(self.max_retries + 1):
+            try:
+                if method == "GET":
+                    response = self._client.get(url, **kwargs)
+                else:
+                    response = self._client.post(url, **kwargs)
+                response.raise_for_status()
+                return response
+            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..."
+                    )
+                    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..."
+                    )
+                    time.sleep(delay)
+                    delay *= 2
+                else:
+                    raise
+
+        raise last_error
+
+    async def initialize(self) -> None:
+        """Initialize the HTTP client and verify server connectivity."""
+        if self._client is not None:
+            return
+
+        logger.info(f"Embeddings: initializing TEI provider at {self.base_url}")
+        self._client = httpx.Client(timeout=self.timeout)
+
+        # Verify server is reachable and get model info
+        try:
+            response = self._request_with_retry("GET", f"{self.base_url}/info")
+            info = response.json()
+            self._model_id = info.get("model_id", "unknown")
+
+            # Get dimension from server info or by doing a test embedding
+            if "max_input_length" in info and "model_dtype" in info:
+                # Try to get dimension from info endpoint (some TEI versions expose it)
+                # If not available, do a test embedding
+                pass
+
+            # Do a test embedding to detect dimension
+            test_response = self._request_with_retry(
+                "POST",
+                f"{self.base_url}/embed",
+                json={"inputs": ["test"]},
+            )
+            test_embeddings = test_response.json()
+            if test_embeddings and len(test_embeddings) > 0:
+                self._dimension = len(test_embeddings[0])
+
+            logger.info(f"Embeddings: TEI provider initialized (model: {self._model_id}, dim: {self._dimension})")
+        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]]:
+        """
+        Generate embeddings using the remote TEI server.
+
+        Args:
+            texts: List of text strings to encode
+
+        Returns:
+            List of embedding vectors
+        """
+        if self._client is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+
+        if not texts:
+            return []
+
+        all_embeddings = []
+
+        # Process in batches
+        for i in range(0, len(texts), self.batch_size):
+            batch = texts[i : i + self.batch_size]
+
+            try:
+                response = self._request_with_retry(
+                    "POST",
+                    f"{self.base_url}/embed",
+                    json={"inputs": batch},
+                )
+                batch_embeddings = response.json()
+                all_embeddings.extend(batch_embeddings)
+            except httpx.HTTPError as e:
+                raise RuntimeError(f"TEI embedding request failed: {e}")
+
+        return all_embeddings
+
+
+class OpenAIEmbeddings(Embeddings):
+    """
+    OpenAI embeddings implementation using the OpenAI API.
+
+    Supports text-embedding-3-small (1536 dims), text-embedding-3-large (3072 dims),
+    and text-embedding-ada-002 (1536 dims, legacy).
+
+    The embedding dimension is auto-detected from the model at initialization.
+    """
+
+    # Known dimensions for OpenAI embedding models
+    MODEL_DIMENSIONS = {
+        "text-embedding-3-small": 1536,
+        "text-embedding-3-large": 3072,
+        "text-embedding-ada-002": 1536,
+    }
+
+    def __init__(
+        self,
+        api_key: str,
+        model: str = DEFAULT_EMBEDDINGS_OPENAI_MODEL,
+        batch_size: int = 100,
+        max_retries: int = 3,
+    ):
+        """
+        Initialize OpenAI embeddings client.
+
+        Args:
+            api_key: OpenAI API key
+            model: OpenAI embedding model name (default: text-embedding-3-small)
+            batch_size: Maximum batch size for embedding requests (default: 100)
+            max_retries: Maximum number of retries for failed requests (default: 3)
+        """
+        self.api_key = api_key
+        self.model = model
+        self.batch_size = batch_size
+        self.max_retries = max_retries
+        self._client = None
+        self._dimension: int | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "openai"
+
+    @property
+    def dimension(self) -> int:
+        if self._dimension is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+        return self._dimension
+
+    async def initialize(self) -> None:
+        """Initialize the OpenAI client and detect dimension."""
+        if self._client is not None:
+            return
+
+        try:
+            from openai import OpenAI
+        except ImportError:
+            raise ImportError("openai is required for OpenAIEmbeddings. Install it with: pip install openai")
+
+        logger.info(f"Embeddings: initializing OpenAI provider with model {self.model}")
+        self._client = OpenAI(api_key=self.api_key, max_retries=self.max_retries)
+
+        # Try to get dimension from known models, otherwise do a test embedding
+        if self.model in self.MODEL_DIMENSIONS:
+            self._dimension = self.MODEL_DIMENSIONS[self.model]
+        else:
+            # Do a test embedding to detect dimension
+            response = self._client.embeddings.create(
+                model=self.model,
+                input=["test"],
+            )
+            if response.data:
+                self._dimension = len(response.data[0].embedding)
+
+        logger.info(f"Embeddings: OpenAI provider initialized (model: {self.model}, dim: {self._dimension})")
+
+    def encode(self, texts: list[str]) -> list[list[float]]:
+        """
+        Generate embeddings using the OpenAI API.
+
+        Args:
+            texts: List of text strings to encode
+
+        Returns:
+            List of embedding vectors
+        """
+        if self._client is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+
+        if not texts:
+            return []
+
+        all_embeddings = []
+
+        # Process in batches
+        for i in range(0, len(texts), self.batch_size):
+            batch = texts[i : i + self.batch_size]
+
+            response = self._client.embeddings.create(
+                model=self.model,
+                input=batch,
+            )
+
+            # Sort by index to ensure correct order
+            batch_embeddings = sorted(response.data, key=lambda x: x.index)
+            all_embeddings.extend([e.embedding for e in batch_embeddings])
+
+        return all_embeddings
+
+
+def create_embeddings_from_env() -> Embeddings:
+    """
+    Create an Embeddings instance based on environment variables.
+
+    See hindsight_api.config for environment variable names and defaults.
+
+    Returns:
+        Configured Embeddings instance
+    """
+    provider = os.environ.get(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER).lower()
+
+    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'")
+        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)
+    elif provider == "openai":
+        # Use dedicated embeddings API key, or fall back to LLM API key
+        api_key = os.environ.get(ENV_EMBEDDINGS_OPENAI_API_KEY) or os.environ.get(ENV_LLM_API_KEY)
+        if not api_key:
+            raise ValueError(
+                f"{ENV_EMBEDDINGS_OPENAI_API_KEY} or {ENV_LLM_API_KEY} is required "
+                f"when {ENV_EMBEDDINGS_PROVIDER} is 'openai'"
+            )
+        model = os.environ.get(ENV_EMBEDDINGS_OPENAI_MODEL, DEFAULT_EMBEDDINGS_OPENAI_MODEL)
+        return OpenAIEmbeddings(api_key=api_key, model=model)
+    else:
+        raise ValueError(f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai'")

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/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):
@@ -43,41 +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)")
-
-    # Internal metrics (used by system but may not be exposed in API)
-    activation: Optional[float] = Field(None, description="Internal activation score")
+    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")
@@ -90,35 +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}'"
     )
 
 
@@ -127,38 +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.",
     )
 
 
@@ -169,12 +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")
@@ -187,15 +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):
@@ -204,20 +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_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_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

@@ -3,12 +3,11 @@ 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__)
 
@@ -17,18 +16,20 @@ 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]:
+    unit_ids: list[str],
+    facts: list[ProcessedFact],
+    log_buffer: list[str] = None,
+    user_entities_per_content: dict[int, list[dict]] = None,
+) -> list[EntityLink]:
     """
     Process entities for all facts and create entity links.
 
     This function:
     1. Extracts entity mentions from fact texts
-    2. Resolves entity names to canonical entities
-    3. Creates entity records in the database
-    4. Returns entity links ready for insertion
+    2. Merges user-provided entities with LLM-extracted entities
+    3. Resolves entity names to canonical entities
+    4. Creates entity records in the database
+    5. Returns entity links ready for insertion
 
     Args:
         entity_resolver: EntityResolver instance for entity resolution
@@ -37,6 +38,7 @@ async def process_entities_batch(
         unit_ids: List of unit IDs (same length as facts)
         facts: List of ProcessedFact objects
         log_buffer: Optional buffer for detailed logging
+        user_entities_per_content: Dict mapping content_index to list of user-provided entities
 
     Returns:
         List of EntityLink objects for batch insertion
@@ -47,15 +49,35 @@ async def process_entities_batch(
     if len(unit_ids) != len(facts):
         raise ValueError(f"Mismatch between unit_ids ({len(unit_ids)}) and facts ({len(facts)})")
 
+    user_entities_per_content = user_entities_per_content or {}
+
     # Extract data for link_utils function
     fact_texts = [fact.fact_text for fact in facts]
     # Use occurred_start if available, otherwise use mentioned_at for entity timestamps
     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
-    ]
+
+    # Convert EntityRef objects to dict format and merge with user-provided entities
+    entities_per_fact = []
+    for fact in facts:
+        # Start with LLM-extracted entities
+        llm_entities = [{"text": entity.name, "type": "CONCEPT"} for entity in (fact.entities or [])]
+
+        # Get user entities for this content (use content_index from fact)
+        user_entities = user_entities_per_content.get(fact.content_index, [])
+
+        # Merge with case-insensitive deduplication
+        seen_texts = {e["text"].lower() for e in llm_entities}
+        for user_entity in user_entities:
+            if user_entity["text"].lower() not in seen_texts:
+                llm_entities.append(
+                    {
+                        "text": user_entity["text"],
+                        "type": user_entity.get("type", "CONCEPT"),
+                    }
+                )
+                seen_texts.add(user_entity["text"].lower())
+
+        entities_per_fact.append(llm_entities)
 
     # Use existing link_utils function for entity processing
     entity_links = await link_utils.extract_entities_batch_optimized(
@@ -67,16 +89,13 @@ 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 +4,55 @@ 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 _infer_temporal_date(fact_text: str, event_date: datetime) -> str | None:
+    """
+    Infer a temporal date from fact text when LLM didn't provide occurred_start.
+
+    This is a fallback for when the LLM fails to extract temporal information
+    from relative time expressions like "last night", "yesterday", etc.
+    """
+    import re
+
+    fact_lower = fact_text.lower()
+
+    # Map relative time expressions to day offsets
+    temporal_patterns = {
+        r"\blast night\b": -1,
+        r"\byesterday\b": -1,
+        r"\btoday\b": 0,
+        r"\bthis morning\b": 0,
+        r"\bthis afternoon\b": 0,
+        r"\bthis evening\b": 0,
+        r"\btonigh?t\b": 0,
+        r"\btomorrow\b": 1,
+        r"\blast week\b": -7,
+        r"\bthis week\b": 0,
+        r"\bnext week\b": 7,
+        r"\blast month\b": -30,
+        r"\bthis month\b": 0,
+        r"\bnext month\b": 30,
+    }
+
+    for pattern, offset_days in temporal_patterns.items():
+        if re.search(pattern, fact_lower):
+            target_date = event_date + timedelta(days=offset_days)
+            return target_date.replace(hour=0, minute=0, second=0, microsecond=0).isoformat()
+
+    # If no relative time expression found, return None
+    return None
 
 
 def _sanitize_text(text: str) -> str:
@@ -31,11 +70,12 @@ 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 +88,46 @@ 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 +136,7 @@ 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 +145,43 @@ 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 +190,17 @@ 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 +210,15 @@ 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 +226,7 @@ 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 +239,11 @@ 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 +254,11 @@ 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 +272,6 @@ 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 +295,21 @@ 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 +320,6 @@ 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 +353,10 @@ 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 +371,9 @@ 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.
 
@@ -382,13 +422,42 @@ WRONG output:
 - where: (missing) ← WRONG - include the location!
 
 ══════════════════════════════════════════════════════════════════════════
-TEMPORAL HANDLING
+FACT_KIND CLASSIFICATION (CRITICAL FOR TEMPORAL HANDLING)
 ══════════════════════════════════════════════════════════════════════════
 
-For EVENTS (fact_kind="event"):
-- Convert relative dates → absolute WITH DAY OF WEEK: "yesterday" on Saturday March 15 → "Friday, March 14, 2024"
+⚠️ MUST set fact_kind correctly - this determines whether occurred_start/end are set!
+
+fact_kind="event" - USE FOR:
+- Actions that happened at a specific time: "went to", "attended", "visited", "bought", "made"
+- Past events: "yesterday I...", "last week...", "in March 2020..."
+- Future plans with dates: "will go to", "scheduled for"
+- Examples: "I went to a pottery workshop" → event
+           "Alice visited Paris in February" → event
+           "I bought a new car yesterday" → event
+           "The user graduated from MIT in March 2020" → event
+
+fact_kind="conversation" - USE FOR:
+- Ongoing states: "works as", "lives in", "is married to"
+- Preferences: "loves", "prefers", "enjoys"
+- Traits/abilities: "speaks fluent French", "knows Python"
+- Examples: "I love Italian food" → conversation
+           "Alice works at Google" → conversation
+           "I prefer outdoor dining" → conversation
+
+══════════════════════════════════════════════════════════════════════════
+TEMPORAL HANDLING (CRITICAL - USE EVENT DATE AS REFERENCE)
+══════════════════════════════════════════════════════════════════════════
+
+⚠️ IMPORTANT: Use the "Event Date" provided in the input as your reference point!
+All relative dates ("yesterday", "last week", "recently") must be resolved relative to the Event Date, NOT today's date.
+
+For EVENTS (fact_kind="event") - MUST SET BOTH occurred_start AND occurred_end:
+- Convert relative dates → absolute using Event Date as reference
+- If Event Date is "Saturday, March 15, 2020", then "yesterday" = Friday, March 14, 2020
+- Dates mentioned in text (e.g., "in March 2020") should use THAT year, not current year
 - Always include the day name (Monday, Tuesday, etc.) in the 'when' field
-- Set occurred_start/occurred_end to WHEN IT HAPPENED (not when mentioned)
+- Set occurred_start AND occurred_end to WHEN IT HAPPENED (not when mentioned)
+- For single-day/point events: set occurred_end = occurred_start (same timestamp)
 
 For CONVERSATIONS (fact_kind="conversation"):
 - General info, preferences, ongoing states → NO occurred dates
@@ -440,7 +509,7 @@ Extract entities that help link related facts together. Include:
 EXAMPLES
 ══════════════════════════════════════════════════════════════════════════
 
-Example 1 - World Facts (Context: June 10, 2024):
+Example 1 - World Facts (Event Date: Tuesday, June 10, 2024):
 Input: "I'm planning my wedding and want a small outdoor ceremony. I just got back from my college roommate Emily's wedding - she married Sarah at a rooftop garden, it was so romantic!"
 
 Output facts:
@@ -459,12 +528,13 @@ Output facts:
    - fact_type: "world", fact_kind: "conversation"
    - entities: ["user", "wedding"]
 
-3. Emily's wedding (THE EVENT)
+3. Emily's wedding (THE EVENT - note occurred_start AND occurred_end both set)
    - what: "Emily got married to Sarah at a rooftop garden ceremony in the city"
    - who: "Emily (user's college roommate), Sarah (Emily's partner)"
    - why: "User found it romantic and beautiful"
    - fact_type: "world", fact_kind: "event"
-   - occurred_start: "2024-06-09T00:00:00Z" (recently, user "just got back")
+   - occurred_start: "2024-06-09T00:00:00Z" (recently, user "just got back" - relative to Event Date June 10, 2024)
+   - occurred_end: "2024-06-09T23:59:59Z" (same day - point event)
    - entities: ["user", "Emily", "Sarah", "wedding", "rooftop garden"]
 
 Example 2 - Assistant Facts (Context: March 5, 2024):
@@ -479,16 +549,17 @@ Output fact:
    - fact_type: "assistant", fact_kind: "conversation"
    - entities: ["user", "API", "Redis"]
 
-Example 3 - Kitchen Items with Concept Inference (Context: May 30, 2024):
+Example 3 - Kitchen Items with Concept Inference (Event Date: Thursday, May 30, 2024):
 Input: "I finally donated my old coffee maker to Goodwill. I upgraded to that new espresso machine last month and the old one was just taking up counter space."
 
 Output fact:
    - what: "User donated their old coffee maker to Goodwill after upgrading to a new espresso machine"
-   - when: "May 30, 2024"
+   - when: "Thursday, May 30, 2024"
    - who: "user"
    - why: "The old coffee maker was taking up counter space after the upgrade"
    - fact_type: "world", fact_kind: "event"
-   - occurred_start: "2024-05-30T00:00:00Z"
+   - occurred_start: "2024-05-30T00:00:00Z" (uses Event Date year)
+   - occurred_end: "2024-05-30T23:59:59Z" (same day - point event)
    - entities: ["user", "coffee maker", "Goodwill", "espresso machine", "kitchen"]
 
 Note: "kitchen" is inferred as a concept because coffee makers and espresso machines are kitchen appliances.
@@ -503,10 +574,8 @@ 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__)
@@ -517,11 +586,11 @@ 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}
 
@@ -535,20 +604,11 @@ 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,
-                max_tokens=65000,
+                max_completion_tokens=65000,
                 skip_validation=True,  # Get raw JSON, we'll validate leniently
             )
 
@@ -570,7 +630,7 @@ 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. "
@@ -591,48 +651,48 @@ 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 = {}
@@ -651,17 +711,25 @@ 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 occurred_start:
-                        fact_data['occurred_start'] = occurred_start
+                if fact_kind == "event":
+                    occurred_start = get_value("occurred_start")
+                    occurred_end = get_value("occurred_end")
+
+                    # If LLM didn't set temporal fields, try to extract them from the fact text
+                    if not occurred_start:
+                        fact_data["occurred_start"] = _infer_temporal_date(combined_text, event_date)
+                    else:
+                        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
+                    elif fact_data.get("occurred_start"):
+                        fact_data["occurred_end"] = fact_data["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 = []
@@ -669,38 +737,34 @@ 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}")
@@ -719,7 +783,9 @@ 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
@@ -738,8 +804,8 @@ 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.
 
@@ -760,6 +826,7 @@ 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:
@@ -772,9 +839,9 @@ 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} "
@@ -790,7 +857,7 @@ 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:
@@ -804,8 +871,7 @@ 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)
@@ -818,7 +884,7 @@ 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,
@@ -828,8 +894,8 @@ 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)
@@ -839,9 +905,7 @@ 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
 
@@ -853,7 +917,7 @@ 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.
 
@@ -886,7 +950,7 @@ 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)
     ]
@@ -904,8 +968,10 @@ 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__)
 
@@ -914,11 +980,8 @@ 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.
 
@@ -951,7 +1014,7 @@ 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)
 
@@ -959,8 +1022,8 @@ 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
@@ -974,7 +1037,7 @@ 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
@@ -995,18 +1058,21 @@ 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)
@@ -1022,13 +1088,14 @@ 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.
 
@@ -1039,13 +1106,13 @@ 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)
@@ -383,24 +444,48 @@ 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()
         links = compute_temporal_links(new_units, all_candidates, time_window_hours)
+
+        # Also compute temporal links WITHIN the new batch (new units to each other)
+        if len(new_units) > 1:
+            # Convert new_units dict to candidate format for within-batch linking
+            new_unit_items = list(new_units.items())
+            for i, (unit_id, event_date) in enumerate(new_unit_items):
+                unit_event_date_norm = _normalize_datetime(event_date)
+
+                # Compare with other new units (only those after this one to avoid duplicates)
+                for j in range(i + 1, len(new_unit_items)):
+                    other_id, other_event_date = new_unit_items[j]
+                    other_event_date_norm = _normalize_datetime(other_event_date)
+
+                    # Check if within time window
+                    time_diff_hours = abs((unit_event_date_norm - other_event_date_norm).total_seconds() / 3600)
+                    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))
+
         _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")
 
@@ -409,6 +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
 
@@ -416,11 +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.
@@ -444,22 +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()
@@ -467,15 +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)
@@ -514,33 +605,68 @@ async def create_semantic_links_batch(
 
                     for idx in sorted_indices:
                         similar_id = existing_ids[idx]
-                        similarity = float(similarities[idx])
-                        all_links.append((unit_id, similar_id, 'semantic', similarity, None))
+                        # 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))
 
-        _log(log_buffer, f"      [8.2] Compute similarities & generate {len(all_links)} semantic links: {time_mod.time() - compute_start:.3f}s")
+        # 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
+        if len(unit_ids) > 1:
+            new_embeddings_matrix = np.array(embeddings)
+
+            for i, unit_id in enumerate(unit_ids):
+                # Compute similarities with all OTHER new units
+                other_indices = [j for j in range(len(unit_ids)) if j != i]
+                if not other_indices:
+                    continue
+
+                other_embeddings = new_embeddings_matrix[other_indices]
+                similarities = np.dot(other_embeddings, new_embeddings_matrix[i])
+
+                # Find top-k above threshold (same logic as existing units)
+                above_threshold = np.where(similarities >= threshold)[0]
+
+                if len(above_threshold) > 0:
+                    # Sort by similarity (descending) and take top-k
+                    sorted_local_indices = above_threshold[np.argsort(-similarities[above_threshold])][:top_k]
+
+                    for local_idx in sorted_local_indices:
+                        other_idx = other_indices[local_idx]
+                        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))
+
+        _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.
 
@@ -556,7 +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()
@@ -583,28 +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
@@ -615,8 +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.
@@ -644,6 +763,7 @@ async def create_causal_links_batch(
 
     try:
         import time as time_mod
+
         create_start = time_mod.time()
 
         # Build links list
@@ -655,12 +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__}) "
@@ -685,24 +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)
@@ -710,5 +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 +90,80 @@ 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", {}),
+            entities=item.get("entities", []),
         )
         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 +188,7 @@ 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 +214,11 @@ 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 +258,11 @@ 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 +272,9 @@ 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 +299,9 @@ 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 +336,9 @@ 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)
@@ -280,8 +353,18 @@ async def retain_batch(
 
             # Process entities
             step_start = time.time()
+            # Build map of content_index -> user entities for merging
+            user_entities_per_content = {
+                idx: content.entities for idx, content in enumerate(contents) if content.entities
+            }
             entity_links = await entity_processing.process_entities_batch(
-                entity_resolver, conn, bank_id, unit_ids, non_duplicate_facts, log_buffer
+                entity_resolver,
+                conn,
+                bank_id,
+                unit_ids,
+                non_duplicate_facts,
+                log_buffer,
+                user_entities_per_content=user_entities_per_content,
             )
             log_buffer.append(f"[6] Process entities: {len(entity_links)} links in {time.time() - step_start:.3f}s")
 
@@ -293,14 +376,18 @@ 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 +396,22 @@ 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 +419,11 @@ 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 +451,19 @@ 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 +6,36 @@ 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)
+        entities: User-provided entities to merge with extracted entities (optional)
+    """
+
+    content: str  # Required
+    context: str
+    event_date: datetime
+    metadata: dict[str, str]
+    document_id: str
+    entities: list[dict[str, str]]  # [{"text": "...", "type": "..."}]
+
+
+def _now_utc() -> datetime:
+    """Factory function for default event_date."""
+    return datetime.now(UTC)
+
+
 @dataclass
 class RetainContent:
     """
@@ -18,16 +43,12 @@ 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)
+    entities: list[dict[str, str]] = field(default_factory=list)  # User-provided entities
 
 
 @dataclass
@@ -37,6 +58,7 @@ 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 +72,10 @@ 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 +85,7 @@ 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 +98,21 @@ 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 +122,41 @@ 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
+
+    # Track which content this fact came from (for user entity merging)
+    content_index: int = 0
 
     @property
     def is_duplicate(self) -> bool:
@@ -136,10 +165,8 @@ 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 +178,12 @@ 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 +199,8 @@ class ProcessedFact:
             metadata=extracted_fact.metadata,
             entities=entities,
             causal_relations=extracted_fact.causal_relations,
-            chunk_id=chunk_id
+            chunk_id=chunk_id,
+            content_index=extracted_fact.content_index,
         )
 
 
@@ -183,10 +211,11 @@ 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 +226,25 @@ 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

@@ -3,13 +3,27 @@ Search module for memory retrieval.
 
 Provides modular search architecture:
 - Retrieval: 4-way parallel (semantic + BM25 + graph + temporal)
+- Graph retrieval: Pluggable strategies (BFS, PPR)
 - Reranking: Pluggable strategies (heuristic, cross-encoder)
 """
 
-from .retrieval import retrieve_parallel
+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",
+    "get_default_graph_retriever",
+    "set_default_graph_retriever",
+    "ParallelRetrievalResult",
+    "GraphRetriever",
+    "BFSGraphRetriever",
+    "MPFPGraphRetriever",
     "CrossEncoderReranker",
 ]

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

@@ -0,0 +1,235 @@
+"""
+Graph retrieval strategies for memory recall.
+
+This module provides an abstraction for graph-based memory retrieval,
+allowing different algorithms (BFS spreading activation, PPR, etc.) to be
+swapped without changing the rest of the recall pipeline.
+"""
+
+import logging
+from abc import ABC, abstractmethod
+
+from ..db_utils import acquire_with_retry
+from ..memory_engine import fq_table
+from .types import RetrievalResult
+
+logger = logging.getLogger(__name__)
+
+
+class GraphRetriever(ABC):
+    """
+    Abstract base class for graph-based memory retrieval.
+
+    Implementations traverse the memory graph (entity links, temporal links,
+    causal links) to find relevant facts that might not be found by
+    semantic or keyword search alone.
+    """
+
+    @property
+    @abstractmethod
+    def name(self) -> str:
+        """Return identifier for this retrieval strategy (e.g., 'bfs', 'mpfp')."""
+        pass
+
+    @abstractmethod
+    async def retrieve(
+        self,
+        pool,
+        query_embedding_str: str,
+        bank_id: str,
+        fact_type: str,
+        budget: int,
+        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.
+
+        Args:
+            pool: Database connection pool
+            query_embedding_str: Query embedding as string (for finding entry points)
+            bank_id: Memory bank identifier
+            fact_type: Fact type to filter ('world', 'experience', 'opinion', 'observation')
+            budget: Maximum number of nodes to explore/return
+            query_text: Original query text (optional, for some strategies)
+            semantic_seeds: Pre-computed semantic entry points (from semantic retrieval)
+            temporal_seeds: Pre-computed temporal entry points (from temporal retrieval)
+
+        Returns:
+            List of RetrievalResult objects with activation scores set
+        """
+        pass
+
+
+class BFSGraphRetriever(GraphRetriever):
+    """
+    Graph retrieval using BFS-style spreading activation.
+
+    Starting from semantic entry points, spreads activation through
+    the memory graph (entity, temporal, causal links) using breadth-first
+    traversal with decaying activation.
+
+    This is the original Hindsight graph retrieval algorithm.
+    """
+
+    def __init__(
+        self,
+        entry_point_limit: int = 5,
+        entry_point_threshold: float = 0.5,
+        activation_decay: float = 0.8,
+        min_activation: float = 0.1,
+        batch_size: int = 20,
+    ):
+        """
+        Initialize BFS graph retriever.
+
+        Args:
+            entry_point_limit: Maximum number of entry points to start from
+            entry_point_threshold: Minimum semantic similarity for entry points
+            activation_decay: Decay factor per hop (activation *= decay)
+            min_activation: Minimum activation to continue spreading
+            batch_size: Number of nodes to process per batch (for neighbor fetching)
+        """
+        self.entry_point_limit = entry_point_limit
+        self.entry_point_threshold = entry_point_threshold
+        self.activation_decay = activation_decay
+        self.min_activation = min_activation
+        self.batch_size = batch_size
+
+    @property
+    def name(self) -> str:
+        return "bfs"
+
+    async def retrieve(
+        self,
+        pool,
+        query_embedding_str: str,
+        bank_id: str,
+        fact_type: str,
+        budget: int,
+        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.
+
+        Algorithm:
+        1. Find entry points (top semantic matches above threshold)
+        2. BFS traversal: visit neighbors, propagate decaying activation
+        3. Boost causal links (causes, enables, prevents)
+        4. Return visited nodes up to budget
+
+        Note: BFS finds its own entry points via embedding search.
+        The semantic_seeds and temporal_seeds parameters are accepted
+        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)
+
+    async def _retrieve_with_conn(
+        self,
+        conn,
+        query_embedding_str: str,
+        bank_id: str,
+        fact_type: str,
+        budget: int,
+    ) -> 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 {fq_table("memory_units")}
+            WHERE bank_id = $2
+              AND embedding IS NOT NULL
+              AND fact_type = $3
+              AND (1 - (embedding <=> $1::vector)) >= $4
+            ORDER BY embedding <=> $1::vector
+            LIMIT $5
+            """,
+            query_embedding_str,
+            bank_id,
+            fact_type,
+            self.entry_point_threshold,
+            self.entry_point_limit,
+        )
+
+        if not entry_points:
+            return []
+
+        # Step 2: BFS spreading activation
+        visited = set()
+        results = []
+        queue = [(RetrievalResult.from_db_row(dict(r)), r["similarity"]) for r in entry_points]
+        budget_remaining = budget
+
+        while queue and budget_remaining > 0:
+            # Collect a batch of nodes to process
+            batch_nodes = []
+            batch_activations = {}
+
+            while queue and len(batch_nodes) < self.batch_size and budget_remaining > 0:
+                current, activation = queue.pop(0)
+                unit_id = current.id
+
+                if unit_id not in visited:
+                    visited.add(unit_id)
+                    budget_remaining -= 1
+                    current.activation = activation
+                    results.append(current)
+                    batch_nodes.append(current.id)
+                    batch_activations[unit_id] = activation
+
+            # Batch fetch neighbors
+            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 {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,
+                )
+
+                for n in neighbors:
+                    neighbor_id = str(n["id"])
+                    if neighbor_id not in visited:
+                        parent_id = str(n["from_unit_id"])
+                        parent_activation = batch_activations.get(parent_id, 0.5)
+
+                        # Boost causal links
+                        link_type = n["link_type"]
+                        base_weight = n["weight"]
+
+                        if link_type in ("causes", "caused_by"):
+                            causal_boost = 2.0
+                        elif link_type in ("enables", "prevents"):
+                            causal_boost = 1.5
+                        else:
+                            causal_boost = 1.0
+
+                        effective_weight = base_weight * causal_boost
+                        new_activation = parent_activation * effective_weight * self.activation_decay
+
+                        if new_activation > self.min_activation:
+                            neighbor_result = RetrievalResult.from_db_row(dict(n))
+                            queue.append((neighbor_result, new_activation))
+
+        return results

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

@@ -0,0 +1,439 @@
+"""
+Meta-Path Forward Push (MPFP) graph retrieval.
+
+A sublinear graph traversal algorithm for memory retrieval over heterogeneous
+graphs with multiple edge types (semantic, temporal, causal, entity).
+
+Combines meta-path patterns from HIN literature with Forward Push local
+propagation from Approximate PPR.
+
+Key properties:
+- Sublinear in graph size (threshold pruning bounds active nodes)
+- Predefined patterns capture different retrieval intents
+- All patterns run in parallel, results fused via RRF
+- No LLM in the loop during traversal
+"""
+
+import asyncio
+import logging
+from collections import defaultdict
+from dataclasses import dataclass, field
+
+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__)
+
+
+# -----------------------------------------------------------------------------
+# Data Classes
+# -----------------------------------------------------------------------------
+
+
+@dataclass
+class EdgeTarget:
+    """A neighbor node with its edge weight."""
+
+    node_id: str
+    weight: float
+
+
+@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]:
+        """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]:
+        """Get top-k neighbors with weights normalized to sum to 1."""
+        neighbors = self.get_neighbors(edge_type, node_id)[:top_k]
+        if not neighbors:
+            return []
+
+        total = sum(n.weight for n in neighbors)
+        if total == 0:
+            return []
+
+        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
+
+
+@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
+
+    # 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 from temporal seeds
+    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)
+
+
+# -----------------------------------------------------------------------------
+# Core Algorithm
+# -----------------------------------------------------------------------------
+
+
+def mpfp_traverse(
+    seeds: list[SeedNode],
+    pattern: list[str],
+    adjacency: TypedAdjacency,
+    config: MPFPConfig,
+) -> PatternResult:
+    """
+    Forward Push traversal following a meta-path pattern.
+
+    Args:
+        seeds: Entry point nodes with initial scores
+        pattern: Sequence of edge types to follow
+        adjacency: Typed adjacency structure
+        config: Algorithm parameters
+
+    Returns:
+        PatternResult with accumulated scores per node
+    """
+    if not seeds:
+        return PatternResult(pattern=pattern, scores={})
+
+    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}
+
+    # Follow pattern hop by hop
+    for edge_type in pattern:
+        next_frontier: dict[str, float] = {}
+
+        for node_id, mass in frontier.items():
+            if mass < config.threshold:
+                continue
+
+            # Keep α portion for this node
+            scores[node_id] = scores.get(node_id, 0) + config.alpha * mass
+
+            # Push (1-α) to neighbors
+            push_mass = (1 - config.alpha) * mass
+            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
+
+        frontier = next_frontier
+
+    # Final frontier nodes get their remaining mass
+    for node_id, mass in frontier.items():
+        if mass >= config.threshold:
+            scores[node_id] = scores.get(node_id, 0) + mass
+
+    return PatternResult(pattern=pattern, scores=scores)
+
+
+def rrf_fusion(
+    results: list[PatternResult],
+    k: int = 60,
+    top_k: int = 50,
+) -> list[tuple[str, float]]:
+    """
+    Reciprocal Rank Fusion to combine pattern results.
+
+    Args:
+        results: List of pattern results
+        k: RRF constant (higher = more uniform weighting)
+        top_k: Number of results to return
+
+    Returns:
+        List of (node_id, fused_score) tuples, sorted by score descending
+    """
+    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)
+
+        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)
+
+    return sorted_results[:top_k]
+
+
+# -----------------------------------------------------------------------------
+# Database Loading
+# -----------------------------------------------------------------------------
+
+
+async def load_typed_adjacency(pool, bank_id: str) -> TypedAdjacency:
+    """
+    Load all edges for a bank, split by edge type.
+
+    Single query, then organize in-memory for fast traversal.
+    """
+    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 {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,
+        )
+
+    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"]
+
+        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],
+    fact_type: str,
+) -> 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 {fq_table("memory_units")}
+            WHERE id = ANY($1::uuid[])
+              AND fact_type = $2
+            """,
+            node_ids,
+            fact_type,
+        )
+
+    return [RetrievalResult.from_db_row(dict(r)) for r in rows]
+
+
+# -----------------------------------------------------------------------------
+# Graph Retriever Implementation
+# -----------------------------------------------------------------------------
+
+
+class MPFPGraphRetriever(GraphRetriever):
+    """
+    Graph retrieval using Meta-Path Forward Push.
+
+    Runs predefined patterns in parallel from semantic and temporal seeds,
+    then fuses results via RRF.
+    """
+
+    def __init__(self, config: MPFPConfig | None = None):
+        """
+        Initialize MPFP retriever.
+
+        Args:
+            config: Algorithm configuration (uses defaults if None)
+        """
+        self.config = config or MPFPConfig()
+        self._adjacency_cache: dict[str, TypedAdjacency] = {}
+
+    @property
+    def name(self) -> str:
+        return "mpfp"
+
+    async def retrieve(
+        self,
+        pool,
+        query_embedding_str: str,
+        bank_id: str,
+        fact_type: str,
+        budget: int,
+        query_text: str | None = None,
+        semantic_seeds: list[RetrievalResult] | None = None,
+        temporal_seeds: list[RetrievalResult] | None = None,
+    ) -> list[RetrievalResult]:
+        """
+        Retrieve facts using MPFP algorithm.
+
+        Args:
+            pool: Database connection pool
+            query_embedding_str: Query embedding (used for fallback seed finding)
+            bank_id: Memory bank ID
+            fact_type: Fact type to filter
+            budget: Maximum results to return
+            query_text: Original query text (optional)
+            semantic_seeds: Pre-computed semantic entry points
+            temporal_seeds: Pre-computed temporal entry points
+
+        Returns:
+            List of RetrievalResult with activation scores
+        """
+        # Load typed adjacency (could cache per bank_id with TTL)
+        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")
+
+        # 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)
+
+        # Run all patterns in parallel
+        tasks = []
+
+        # Patterns from semantic seeds
+        for pattern in self.config.patterns_semantic:
+            if semantic_seed_nodes:
+                tasks.append(
+                    asyncio.to_thread(
+                        mpfp_traverse,
+                        semantic_seed_nodes,
+                        pattern,
+                        adjacency,
+                        self.config,
+                    )
+                )
+
+        # Patterns from temporal seeds
+        for pattern in self.config.patterns_temporal:
+            if temporal_seed_nodes:
+                tasks.append(
+                    asyncio.to_thread(
+                        mpfp_traverse,
+                        temporal_seed_nodes,
+                        pattern,
+                        adjacency,
+                        self.config,
+                    )
+                )
+
+        if not tasks:
+            return []
+
+        # Gather pattern results
+        pattern_results = await asyncio.gather(*tasks)
+
+        # Fuse results
+        fused = rrf_fusion(pattern_results, top_k=budget)
+
+        if not fused:
+            return []
+
+        # Get top result IDs (don't exclude seeds - they may be highly relevant)
+        result_ids = [node_id for node_id, score in fused][:budget]
+
+        # Fetch full details
+        results = await fetch_memory_units_by_ids(pool, result_ids, fact_type)
+
+        # Add activation scores from fusion
+        score_map = {node_id: score for node_id, score in fused}
+        for result in results:
+            result.activation = score_map.get(result.id, 0.0)
+
+        # Sort by activation
+        results.sort(key=lambda r: r.activation or 0, reverse=True)
+
+        return results
+
+    def _convert_seeds(
+        self,
+        seeds: list[RetrievalResult] | None,
+        score_attr: str,
+    ) -> list[SeedNode]:
+        """Convert RetrievalResult seeds to SeedNode format."""
+        if not seeds:
+            return []
+
+        result = []
+        for seed in seeds:
+            score = getattr(seed, score_attr, None)
+            if score is None:
+                score = seed.activation or seed.similarity or 1.0
+            result.append(SeedNode(node_id=seed.id, score=score))
+
+        return result
+
+    async def _find_semantic_seeds(
+        self,
+        pool,
+        query_embedding_str: str,
+        bank_id: str,
+        fact_type: str,
+        limit: int = 20,
+        threshold: float = 0.3,
+    ) -> 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 {fq_table("memory_units")}
+                WHERE bank_id = $2
+                  AND embedding IS NOT NULL
+                  AND fact_type = $3
+                  AND (1 - (embedding <=> $1::vector)) >= $4
+                ORDER BY embedding <=> $1::vector
+                LIMIT $5
+                """,
+                query_embedding_str,
+                bank_id,
+                fact_type,
+                threshold,
+                limit,
+            )
+
+        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
 
 
@@ -10,10 +9,8 @@ class CrossEncoderReranker:
     """
     Neural reranking using a cross-encoder model.
 
-    Uses cross-encoder/ms-marco-MiniLM-L-6-v2 by default:
-    - Fast inference (~80ms for 100 pairs on CPU)
-    - Small model (80MB)
-    - Trained for passage re-ranking
+    Configured via environment variables (see cross_encoder.py).
+    Default local model is cross-encoder/ms-marco-MiniLM-L-6-v2.
     """
 
     def __init__(self, cross_encoder=None):
@@ -21,21 +18,33 @@ class CrossEncoderReranker:
         Initialize cross-encoder reranker.
 
         Args:
-            cross_encoder: CrossEncoderReranker instance. If None, uses default
-                          SentenceTransformersCrossEncoder with ms-marco-MiniLM-L-6-v2
-                          (loaded lazily for faster startup)
+            cross_encoder: CrossEncoderModel instance. If None, creates one from
+                          environment variables (defaults to local provider)
         """
         if cross_encoder is None:
-            from hindsight_api.engine.cross_encoder import SentenceTransformersCrossEncoder
-            # Model is loaded lazily - call ensure_loaded() during initialize()
-            cross_encoder = SentenceTransformersCrossEncoder()
-        self.cross_encoder = cross_encoder
+            from hindsight_api.engine.cross_encoder import create_cross_encoder_from_env
 
-    def rerank(
-        self,
-        query: str,
-        candidates: List[MergedCandidate]
-    ) -> List[ScoredResult]:
+            cross_encoder = create_cross_encoder_from_env()
+        self.cross_encoder = cross_encoder
+        self._initialized = False
+
+    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.
 
@@ -81,6 +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))
 
@@ -93,7 +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

@@ -4,24 +4,69 @@ Retrieval module for 4-way parallel search.
 Implements:
 1. Semantic retrieval (vector similarity)
 2. BM25 retrieval (keyword/full-text search)
-3. Graph retrieval (spreading activation)
+3. Graph retrieval (via pluggable GraphRetriever interface)
 4. Temporal retrieval (time-aware search with spreading)
 """
 
-from typing import List, Dict, Any, Tuple, Optional
-from datetime import datetime
 import asyncio
+import logging
+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__)
+
+
+@dataclass
+class ParallelRetrievalResult:
+    """Result from parallel retrieval across all methods."""
+
+    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: GraphRetriever | None = None
+
+
+def get_default_graph_retriever() -> GraphRetriever:
+    """Get or create the default graph retriever based on config."""
+    global _default_graph_retriever
+    if _default_graph_retriever is None:
+        config = get_config()
+        retriever_type = config.graph_retriever.lower()
+        if retriever_type == "mpfp":
+            _default_graph_retriever = MPFPGraphRetriever()
+            logger.info("Using MPFP graph retriever")
+        elif retriever_type == "bfs":
+            _default_graph_retriever = BFSGraphRetriever()
+            logger.info("Using BFS graph retriever")
+        else:
+            logger.warning(f"Unknown graph retriever '{retriever_type}', falling back to MPFP")
+            _default_graph_retriever = MPFPGraphRetriever()
+    return _default_graph_retriever
+
+
+def set_default_graph_retriever(retriever: GraphRetriever) -> None:
+    """Set the default graph retriever (for configuration/testing)."""
+    global _default_graph_retriever
+    _default_graph_retriever = retriever
+
 
 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.
 
@@ -36,10 +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
@@ -47,18 +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.
 
@@ -76,7 +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]
@@ -90,136 +132,24 @@ 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]
 
 
-async def retrieve_graph(
-    conn,
-    query_emb_str: str,
-    bank_id: str,
-    fact_type: str,
-    budget: int
-) -> List[RetrievalResult]:
-    """
-    Graph retrieval via spreading activation.
-
-    Args:
-        conn: Database connection
-        query_emb_str: Query embedding as string
-        agent_id: bank ID
-        fact_type: Fact type to filter
-        budget: Node budget for graph traversal
-
-    Returns:
-        List of RetrievalResult objects
-    """
-    # Find entry points
-    entry_points = await conn.fetch(
-        """
-        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
-        WHERE bank_id = $2
-          AND embedding IS NOT NULL
-          AND fact_type = $3
-          AND (1 - (embedding <=> $1::vector)) >= 0.5
-        ORDER BY embedding <=> $1::vector
-        LIMIT 5
-        """,
-        query_emb_str, bank_id, fact_type
-    )
-
-    if not entry_points:
-        return []
-
-    # BFS-style spreading activation with batched neighbor fetching
-    visited = set()
-    results = []
-    queue = [(RetrievalResult.from_db_row(dict(r)), r["similarity"]) for r in entry_points]
-    budget_remaining = budget
-
-    # Process nodes in batches to reduce DB roundtrips
-    batch_size = 20  # Fetch neighbors for up to 20 nodes at once
-
-    while queue and budget_remaining > 0:
-        # Collect a batch of nodes to process
-        batch_nodes = []
-        batch_activations = {}
-
-        while queue and len(batch_nodes) < batch_size and budget_remaining > 0:
-            current, activation = queue.pop(0)
-            unit_id = current.id
-
-            if unit_id not in visited:
-                visited.add(unit_id)
-                budget_remaining -= 1
-                results.append(current)
-                batch_nodes.append(current.id)
-                batch_activations[unit_id] = activation
-
-        # Batch fetch neighbors for all nodes in this batch
-        # Fetch top weighted neighbors (batch_size * 20 = ~400 for good distribution)
-        if batch_nodes and budget_remaining > 0:
-            max_neighbors = len(batch_nodes) * 20
-            neighbors = await conn.fetch(
-                """
-                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
-                WHERE ml.from_unit_id = ANY($1::uuid[])
-                  AND ml.weight >= 0.1
-                  AND mu.fact_type = $2
-                ORDER BY ml.weight DESC
-                LIMIT $3
-                """,
-                batch_nodes, fact_type, max_neighbors
-            )
-
-            for n in neighbors:
-                neighbor_id = str(n["id"])
-                if neighbor_id not in visited:
-                    # Get parent activation
-                    parent_id = str(n["from_unit_id"])
-                    activation = batch_activations.get(parent_id, 0.5)
-
-                    # Boost activation for causal links (they're high-value relationships)
-                    link_type = n["link_type"]
-                    base_weight = n["weight"]
-
-                    # Causal links get 1.5-2.0x boost depending on type
-                    if link_type in ("causes", "caused_by"):
-                        # Direct causation - very strong relationship
-                        causal_boost = 2.0
-                    elif link_type in ("enables", "prevents"):
-                        # Conditional causation - strong but not as direct
-                        causal_boost = 1.5
-                    else:
-                        # Temporal, semantic, entity links - standard weight
-                        causal_boost = 1.0
-
-                    effective_weight = base_weight * causal_boost
-                    new_activation = activation * effective_weight * 0.8
-                    if new_activation > 0.1:
-                        neighbor_result = RetrievalResult.from_db_row(dict(n))
-                        queue.append((neighbor_result, new_activation))
-
-    return results
-
-
 async def retrieve_temporal(
     conn,
     query_emb_str: str,
@@ -228,8 +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.
 
@@ -251,19 +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
@@ -284,7 +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:
@@ -327,7 +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:
@@ -337,12 +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
@@ -352,7 +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:
@@ -376,7 +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
 
@@ -418,9 +359,10 @@ async def retrieve_parallel(
     bank_id: str,
     fact_type: str,
     thinking_budget: int,
-    question_date: Optional[datetime] = None,
-    query_analyzer: Optional["QueryAnalyzer"] = None
-) -> Tuple[List[RetrievalResult], List[RetrievalResult], List[RetrievalResult], Optional[List[RetrievalResult]], Dict[str, float], Optional[Tuple[datetime, datetime]]]:
+    question_date: datetime | None = None,
+    query_analyzer: Optional["QueryAnalyzer"] = None,
+    graph_retriever: GraphRetriever | None = None,
+) -> ParallelRetrievalResult:
     """
     Run 3-way or 4-way parallel retrieval (adds temporal if detected).
 
@@ -428,76 +370,330 @@ async def retrieve_parallel(
         pool: Database connection pool
         query_text: Query text
         query_embedding_str: Query embedding as string
-        agent_id: bank ID
+        bank_id: Bank ID
         fact_type: Fact type to filter
         thinking_budget: Budget for graph traversal and retrieval limits
         question_date: Optional date when question was asked (for temporal filtering)
         query_analyzer: Query analyzer to use (defaults to TransformerQueryAnalyzer)
+        graph_retriever: Graph retrieval strategy (defaults to configured retriever)
 
     Returns:
-        Tuple of (semantic_results, bm25_results, graph_results, temporal_results, timings, temporal_constraint)
-        Each results list contains RetrievalResult objects
-        temporal_results is None if no temporal constraint detected
-        timings is a dict with per-method latencies in seconds
-        temporal_constraint is the (start_date, end_date) tuple if detected, else None
+        ParallelRetrievalResult with semantic, bm25, graph, temporal results and timings
     """
-    # Detect temporal constraint
     from .temporal_extraction import extract_temporal_constraint
+
+    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
+        )
+    else:
+        return await _retrieve_parallel_bfs(
+            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_time: float
+    graph_time: float
+
+
+@dataclass
+class _TimedResult:
+    """Internal result with timing."""
+
+    results: list[RetrievalResult]
+    time: float
+
+
+async def _retrieve_parallel_mpfp(
+    pool,
+    query_text: str,
+    query_embedding_str: str,
+    bank_id: str,
+    fact_type: str,
+    thinking_budget: int,
+    temporal_constraint: tuple | None,
+    retriever: GraphRetriever,
+) -> ParallelRetrievalResult:
+    """
+    MPFP retrieval with optimized parallelization.
+
+    Runs 2-3 parallel task chains:
+    - Task 1: Semantic → Graph (chained, graph uses semantic seeds)
+    - Task 2: BM25 (independent)
+    - Task 3: Temporal (if constraint detected)
+    """
     import time
 
-    temporal_constraint = extract_temporal_constraint(
-        query_text, reference_date=question_date, analyzer=query_analyzer
+    async def run_semantic_then_graph() -> _SemanticGraphResult:
+        """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_time = time.time() - start
+
+        # Get temporal seeds if needed (quick query, part of this chain)
+        temporal_seeds = None
+        if temporal_constraint:
+            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
+                )
+
+        # Run graph with seeds
+        start = time.time()
+        graph = await retriever.retrieve(
+            pool=pool,
+            query_embedding_str=query_embedding_str,
+            bank_id=bank_id,
+            fact_type=fact_type,
+            budget=thinking_budget,
+            query_text=query_text,
+            semantic_seeds=semantic,
+            temporal_seeds=temporal_seeds,
+        )
+        graph_time = time.time() - start
+
+        return _SemanticGraphResult(semantic, graph, semantic_time, graph_time)
+
+    async def run_bm25() -> _TimedResult:
+        """Independent BM25 retrieval."""
+        start = time.time()
+        async with acquire_with_retry(pool) as conn:
+            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget)
+        return _TimedResult(results, time.time() - start)
+
+    async def run_temporal(tc_start, tc_end) -> _TimedResult:
+        """Temporal retrieval (uses its own entry point finding)."""
+        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,
+            )
+        return _TimedResult(results, time.time() - start)
+
+    # Run parallel task chains
+    if temporal_constraint:
+        tc_start, tc_end = temporal_constraint
+        sg_result, bm25_result, temporal_result = await asyncio.gather(
+            run_semantic_then_graph(),
+            run_bm25(),
+            run_temporal(tc_start, tc_end),
+        )
+        return ParallelRetrievalResult(
+            semantic=sg_result.semantic,
+            bm25=bm25_result.results,
+            graph=sg_result.graph,
+            temporal=temporal_result.results,
+            timings={
+                "semantic": sg_result.semantic_time,
+                "graph": sg_result.graph_time,
+                "bm25": bm25_result.time,
+                "temporal": temporal_result.time,
+            },
+            temporal_constraint=temporal_constraint,
+        )
+    else:
+        sg_result, bm25_result = await asyncio.gather(
+            run_semantic_then_graph(),
+            run_bm25(),
+        )
+        return ParallelRetrievalResult(
+            semantic=sg_result.semantic,
+            bm25=bm25_result.results,
+            graph=sg_result.graph,
+            temporal=None,
+            timings={
+                "semantic": sg_result.semantic_time,
+                "graph": sg_result.graph_time,
+                "bm25": bm25_result.time,
+            },
+            temporal_constraint=None,
+        )
+
+
+async def _get_temporal_entry_points(
+    conn,
+    query_embedding_str: str,
+    bank_id: str,
+    fact_type: str,
+    start_date: datetime,
+    end_date: datetime,
+    limit: int = 20,
+    semantic_threshold: float = 0.1,
+) -> list[RetrievalResult]:
+    """Get temporal entry points (facts in date range with semantic relevance)."""
+
+    if start_date.tzinfo is None:
+        start_date = start_date.replace(tzinfo=UTC)
+    if end_date.tzinfo is None:
+        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 {fq_table("memory_units")}
+        WHERE bank_id = $2
+          AND fact_type = $3
+          AND embedding IS NOT NULL
+          AND (
+              (occurred_start IS NOT NULL AND occurred_end IS NOT NULL
+               AND occurred_start <= $5 AND occurred_end >= $4)
+              OR (mentioned_at IS NOT NULL AND mentioned_at BETWEEN $4 AND $5)
+              OR (occurred_start IS NOT NULL AND occurred_start BETWEEN $4 AND $5)
+              OR (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
+          )
+          AND (1 - (embedding <=> $1::vector)) >= $6
+        ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC,
+                 (embedding <=> $1::vector) ASC
+        LIMIT $7
+        """,
+        query_embedding_str,
+        bank_id,
+        fact_type,
+        start_date,
+        end_date,
+        semantic_threshold,
+        limit,
     )
 
-    # Wrapper to track timing for each retrieval method
-    async def timed_retrieval(name: str, coro):
+    results = []
+    total_days = max((end_date - start_date).total_seconds() / 86400, 1)
+    mid_date = start_date + (end_date - start_date) / 2
+
+    for row in rows:
+        result = RetrievalResult.from_db_row(dict(row))
+
+        # Calculate temporal proximity score
+        best_date = None
+        if row["occurred_start"] and row["occurred_end"]:
+            best_date = row["occurred_start"] + (row["occurred_end"] - row["occurred_start"]) / 2
+        elif row["occurred_start"]:
+            best_date = row["occurred_start"]
+        elif row["occurred_end"]:
+            best_date = row["occurred_end"]
+        elif row["mentioned_at"]:
+            best_date = row["mentioned_at"]
+
+        if best_date:
+            days_from_mid = abs((best_date - mid_date).total_seconds() / 86400)
+            result.temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0)
+        else:
+            result.temporal_proximity = 0.5
+
+        result.temporal_score = result.temporal_proximity
+        results.append(result)
+
+    return results
+
+
+async def _retrieve_parallel_bfs(
+    pool,
+    query_text: str,
+    query_embedding_str: str,
+    bank_id: str,
+    fact_type: str,
+    thinking_budget: int,
+    temporal_constraint: tuple | None,
+    retriever: GraphRetriever,
+) -> ParallelRetrievalResult:
+    """BFS retrieval: all methods run in parallel (original behavior)."""
+    import time
+
+    async def run_semantic() -> _TimedResult:
         start = time.time()
-        result = await coro
-        duration = time.time() - start
-        return result, name, duration
-
-    async def run_semantic():
         async with acquire_with_retry(pool) as conn:
-            return await retrieve_semantic(conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget)
+            results = await retrieve_semantic(conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget)
+        return _TimedResult(results, time.time() - start)
 
-    async def run_bm25():
+    async def run_bm25() -> _TimedResult:
+        start = time.time()
         async with acquire_with_retry(pool) as conn:
-            return await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget)
+            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget)
+        return _TimedResult(results, time.time() - start)
 
-    async def run_graph():
-        async with acquire_with_retry(pool) as conn:
-            return await retrieve_graph(conn, query_embedding_str, bank_id, fact_type, budget=thinking_budget)
+    async def run_graph() -> _TimedResult:
+        start = time.time()
+        results = await retriever.retrieve(
+            pool=pool,
+            query_embedding_str=query_embedding_str,
+            bank_id=bank_id,
+            fact_type=fact_type,
+            budget=thinking_budget,
+            query_text=query_text,
+        )
+        return _TimedResult(results, time.time() - start)
 
-    async def run_temporal(start_date, end_date):
+    async def run_temporal(tc_start, tc_end) -> _TimedResult:
+        start = time.time()
         async with acquire_with_retry(pool) as conn:
-            return await retrieve_temporal(
-                conn, query_embedding_str, bank_id, fact_type,
-                start_date, end_date, budget=thinking_budget, semantic_threshold=0.1
+            results = await retrieve_temporal(
+                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)
 
-    # Run retrievals in parallel with timing
-    timings = {}
     if temporal_constraint:
-        start_date, end_date = temporal_constraint
-        results = await asyncio.gather(
-            timed_retrieval("semantic", run_semantic()),
-            timed_retrieval("bm25", run_bm25()),
-            timed_retrieval("graph", run_graph()),
-            timed_retrieval("temporal", run_temporal(start_date, end_date))
+        tc_start, tc_end = temporal_constraint
+        semantic_r, bm25_r, graph_r, temporal_r = await asyncio.gather(
+            run_semantic(),
+            run_bm25(),
+            run_graph(),
+            run_temporal(tc_start, tc_end),
+        )
+        return ParallelRetrievalResult(
+            semantic=semantic_r.results,
+            bm25=bm25_r.results,
+            graph=graph_r.results,
+            temporal=temporal_r.results,
+            timings={
+                "semantic": semantic_r.time,
+                "bm25": bm25_r.time,
+                "graph": graph_r.time,
+                "temporal": temporal_r.time,
+            },
+            temporal_constraint=temporal_constraint,
         )
-        semantic_results, _, timings["semantic"] = results[0]
-        bm25_results, _, timings["bm25"] = results[1]
-        graph_results, _, timings["graph"] = results[2]
-        temporal_results, _, timings["temporal"] = results[3]
     else:
-        results = await asyncio.gather(
-            timed_retrieval("semantic", run_semantic()),
-            timed_retrieval("bm25", run_bm25()),
-            timed_retrieval("graph", run_graph())
+        semantic_r, bm25_r, graph_r = await asyncio.gather(
+            run_semantic(),
+            run_bm25(),
+            run_graph(),
+        )
+        return ParallelRetrievalResult(
+            semantic=semantic_r.results,
+            bm25=bm25_r.results,
+            graph=graph_r.results,
+            temporal=None,
+            timings={
+                "semantic": semantic_r.time,
+                "bm25": bm25_r.time,
+                "graph": graph_r.time,
+            },
+            temporal_constraint=None,
         )
-        semantic_results, _, timings["semantic"] = results[0]
-        bm25_results, _, timings["bm25"] = results[1]
-        graph_results, _, timings["graph"] = results[2]
-        temporal_results = None
-
-    return semantic_results, bm25_results, graph_results, temporal_results, timings, temporal_constraint

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:
@@ -94,11 +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')
-
-        # Add activation if available
-        if fact.activation is not None:
-            fact_obj["score"] = fact.activation
+                fact_obj["occurred_start"] = occurred_start.strftime("%Y-%m-%d %H:%M:%S")
 
         formatted.append(fact_obj)
 
@@ -113,7 +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)
@@ -180,16 +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.
 
@@ -242,11 +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
@@ -257,14 +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
@@ -272,17 +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)
@@ -43,98 +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")
-    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")
@@ -149,33 +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."""
@@ -185,14 +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)
@@ -206,10 +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.
 
@@ -286,10 +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
+        metadata: dict[str, Any] | None = None,
+        fact_type: str | None = None,
     ):
         """
         Record results from a single retrieval method.
@@ -300,6 +302,7 @@ class SearchTracer:
             duration_seconds: Time taken for this retrieval
             score_field: Field name containing the score in data dict
             metadata: Optional metadata about this retrieval method
+            fact_type: Fact type this retrieval was for (world, experience, opinion)
         """
         retrieval_results = []
         for rank, (doc_id, data) in enumerate(results, start=1):
@@ -313,7 +316,7 @@ class SearchTracer:
                     text=data.get("text", ""),
                     context=data.get("context", ""),
                     event_date=data.get("event_date"),
-                    fact_type=data.get("fact_type"),
+                    fact_type=data.get("fact_type") or fact_type,
                     score=score,
                     score_name=score_field,
                 )
@@ -322,13 +325,14 @@ class SearchTracer:
         self.retrieval_results.append(
             RetrievalMethodResults(
                 method_name=method_name,
+                fact_type=fact_type,
                 results=retrieval_results,
                 duration_seconds=duration_seconds,
                 metadata=metadata or {},
             )
         )
 
-    def add_rrf_merged(self, merged_results: List[tuple]):
+    def add_rrf_merged(self, merged_results: list[tuple]):
         """
         Record RRF merged results.
 
@@ -347,7 +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.
 
@@ -367,8 +371,18 @@ class SearchTracer:
             rank_change = rrf_rank - rank  # Positive = moved up
 
             # Extract score components (only include non-None values)
+            # Keys from ScoredResult.to_dict(): cross_encoder_score, cross_encoder_score_normalized,
+            # rrf_normalized, temporal, recency, combined_score, weight
             score_components = {}
-            for key in ["semantic_similarity", "bm25_score", "rrf_score", "recency_normalized", "frequency_normalized"]:
+            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]
 
@@ -384,7 +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.
 
@@ -411,7 +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
@@ -17,27 +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/graph retrieval
-    bm25_score: Optional[float] = None  # BM25 retrieval
-    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"]),
@@ -54,6 +56,7 @@ class RetrievalResult:
             embedding=row.get("embedding"),
             similarity=row.get("similarity"),
             bm25_score=row.get("bm25_score"),
+            activation=row.get("activation"),
             temporal_score=row.get("temporal_score"),
             temporal_proximity=row.get("temporal_proximity"),
         )
@@ -66,13 +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:
@@ -87,6 +91,7 @@ class ScoredResult:
 
     Contains all retrieval/merge data plus reranking scores and combined score.
     """
+
     # Original merged candidate
     candidate: MergedCandidate
 
@@ -113,7 +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.
 
@@ -152,6 +157,7 @@ class ScoredResult:
         result["cross_encoder_score"] = self.cross_encoder_score
         result["cross_encoder_score_normalized"] = self.cross_encoder_score_normalized
         result["rrf_normalized"] = self.rrf_normalized
+        result["temporal"] = self.temporal
         result["recency"] = self.recency
         result["combined_score"] = self.combined_score
         result["weight"] = self.weight

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.
+        """
+        ...