refactor(claude-code): remove recallTopK setting

Unused client-side cap — Hindsight server already controls result
count via recallBudget and recallMaxTokens.

.claude-plugin/marketplace.json

@@ -0,0 +1,15 @@
+{
+  "$schema": "https://anthropic.com/claude-code/marketplace.schema.json",
+  "name": "hindsight",
+  "description": "Official Hindsight integrations for Claude Code",
+  "owner": {
+    "name": "vectorize-io"
+  },
+  "plugins": [
+    {
+      "name": "hindsight-memory",
+      "description": "Automatic long-term memory for Claude Code via Hindsight",
+      "source": "./hindsight-integrations/claude-code"
+    }
+  ]
+}

.env.example

@@ -2,7 +2,7 @@
 # Copy this file to .env and fill in your values
 
 # LLM Configuration (Required)
-# Supported providers: openai, groq, ollama, gemini, anthropic, lmstudio, vertexai
+# Supported providers: openai, groq, ollama, gemini, anthropic, lmstudio, vertexai, minimax
 HINDSIGHT_API_LLM_PROVIDER=openai
 HINDSIGHT_API_LLM_API_KEY=your-api-key-here
 HINDSIGHT_API_LLM_MODEL=gpt-4o-mini
@@ -20,6 +20,11 @@ HINDSIGHT_API_LLM_BASE_URL=https://api.openai.com/v1
 # HINDSIGHT_API_LLM_VERTEXAI_REGION=us-central1
 # HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/path/to/service-account-key.json  # Optional, uses ADC if not set
 
+# Example: MiniMax configuration (1M context window)
+# HINDSIGHT_API_LLM_PROVIDER=minimax
+# HINDSIGHT_API_LLM_API_KEY=your-minimax-api-key
+# HINDSIGHT_API_LLM_MODEL=MiniMax-M2.7
+
 # Example: LM Studio local configuration (Qwen 2.5 32B recommended)
 # HINDSIGHT_API_LLM_PROVIDER=lmstudio
 # HINDSIGHT_API_LLM_API_KEY=lmstudio

.github/dependabot.yml

@@ -0,0 +1,6 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"

.github/workflows/deploy-docs.yml

@@ -21,20 +21,20 @@ jobs:
   build:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-node@v4
+      - uses: actions/checkout@v6
+      - uses: actions/setup-node@v6
         with:
           node-version: 20
           cache: npm
           cache-dependency-path: package-lock.json
-      - uses: astral-sh/setup-uv@v4
+      - uses: astral-sh/setup-uv@v7
       - run: npm ci --workspace=hindsight-docs
       - run: uv run generate-llms-full
       - run: npm run build --workspace=hindsight-docs
         env:
           UMAMI_URL: https://analytics.hindsight.vectorize.io
           UMAMI_WEBSITE_ID: ${{ secrets.UMAMI_WEBSITE_ID }}
-      - uses: actions/upload-pages-artifact@v3
+      - uses: actions/upload-pages-artifact@v4
         with:
           path: hindsight-docs/build
   deploy:

.github/workflows/release-integration.yml

@@ -0,0 +1,111 @@
+name: Release Integration
+
+on:
+  push:
+    tags:
+      - 'integrations/**'
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write  # for PyPI trusted publishing
+
+    steps:
+    - uses: actions/checkout@v6
+
+    - name: Extract integration info
+      id: info
+      run: |
+        # refs/tags/integrations/litellm/v0.1.0 → integration=litellm, version=0.1.0
+        TAG="${GITHUB_REF#refs/tags/}"
+        INTEGRATION=$(echo "$TAG" | cut -d'/' -f2)
+        VERSION=$(echo "$TAG" | cut -d'/' -f3 | sed 's/^v//')
+        echo "integration=$INTEGRATION" >> $GITHUB_OUTPUT
+        echo "version=$VERSION" >> $GITHUB_OUTPUT
+        echo "tag=$TAG" >> $GITHUB_OUTPUT
+        echo "Integration: $INTEGRATION, Version: $VERSION"
+
+    - name: Detect integration type
+      id: type
+      run: |
+        if [ -f "hindsight-integrations/${{ steps.info.outputs.integration }}/pyproject.toml" ]; then
+          echo "type=python" >> $GITHUB_OUTPUT
+        elif [ -f "hindsight-integrations/${{ steps.info.outputs.integration }}/package.json" ]; then
+          echo "type=typescript" >> $GITHUB_OUTPUT
+        else
+          echo "type=plugin" >> $GITHUB_OUTPUT
+        fi
+
+    # ── Python integrations (litellm, pydantic-ai, crewai) ──────────────────
+
+    - name: Install uv
+      if: steps.type.outputs.type == 'python'
+      uses: astral-sh/setup-uv@v7
+      with:
+        enable-cache: true
+
+    - name: Set up Python
+      if: steps.type.outputs.type == 'python'
+      uses: actions/setup-python@v6
+      with:
+        python-version-file: ".python-version"
+
+    - name: Build Python package
+      if: steps.type.outputs.type == 'python'
+      working-directory: ./hindsight-integrations/${{ steps.info.outputs.integration }}
+      run: uv build --out-dir dist
+
+    - name: Publish Python package to PyPI
+      if: steps.type.outputs.type == 'python'
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        packages-dir: ./hindsight-integrations/${{ steps.info.outputs.integration }}/dist
+        skip-existing: true
+
+    # ── TypeScript integrations (ai-sdk, chat, openclaw) ────────────────────
+
+    # ── Plugin integrations (claude-code) — no package to publish ───────────
+
+    - name: Plugin release
+      if: steps.type.outputs.type == 'plugin'
+      run: |
+        echo "Plugin integration ${{ steps.info.outputs.integration }} v${{ steps.info.outputs.version }} — no package to publish."
+        echo "Users install via: claude plugin marketplace add vectorize-io/hindsight --sparse hindsight-integrations"
+
+    # ── TypeScript integrations (ai-sdk, chat, openclaw) ────────────────────
+
+    - name: Set up Node.js
+      if: steps.type.outputs.type == 'typescript'
+      uses: actions/setup-node@v6
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+
+    - name: Install dependencies
+      if: steps.type.outputs.type == 'typescript'
+      working-directory: ./hindsight-integrations/${{ steps.info.outputs.integration }}
+      run: npm ci
+
+    - name: Build TypeScript package
+      if: steps.type.outputs.type == 'typescript'
+      working-directory: ./hindsight-integrations/${{ steps.info.outputs.integration }}
+      run: npm run build
+
+    - name: Publish TypeScript package to npm
+      if: steps.type.outputs.type == 'typescript'
+      working-directory: ./hindsight-integrations/${{ steps.info.outputs.integration }}
+      run: |
+        set +e
+        OUTPUT=$(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 }}

.github/workflows/release.yml

@@ -13,15 +13,15 @@ jobs:
       id-token: write
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
 
     - name: Install uv
-      uses: astral-sh/setup-uv@v5
+      uses: astral-sh/setup-uv@v7
       with:
         enable-cache: true
 
     - name: Set up Python
-      uses: actions/setup-python@v5
+      uses: actions/setup-python@v6
       with:
         python-version-file: ".python-version"
 
@@ -30,37 +30,39 @@ jobs:
       working-directory: ./hindsight-clients/python
       run: uv build --out-dir dist
 
+    - name: Build hindsight-api-slim
+      working-directory: ./hindsight-api-slim
+      run: uv build --out-dir dist
+
     - name: Build hindsight-api
       working-directory: ./hindsight-api
       run: uv build --out-dir dist
 
     - name: Build hindsight-all
-      working-directory: ./hindsight
+      working-directory: ./hindsight-all
       run: uv build --out-dir dist
 
-    - name: Build hindsight-litellm
-      working-directory: ./hindsight-integrations/litellm
+    - name: Build hindsight-all-slim
+      working-directory: ./hindsight-all-slim
       run: uv build --out-dir dist
 
     - name: Build hindsight-embed
       working-directory: ./hindsight-embed
       run: uv build --out-dir dist
 
-    - name: Build hindsight-crewai
-      working-directory: ./hindsight-integrations/crewai
-      run: uv build --out-dir dist
-
-    - name: Build hindsight-pydantic-ai
-      working-directory: ./hindsight-integrations/pydantic-ai
-      run: uv build --out-dir dist
-
-    # Publish in order (client and api first, then hindsight-all which depends on them)
+    # Publish in order (client and api-slim first, then api/all wrappers which depend on them)
     - name: Publish hindsight-client to PyPI
       uses: pypa/gh-action-pypi-publish@release/v1
       with:
         packages-dir: ./hindsight-clients/python/dist
         skip-existing: true
 
+    - name: Publish hindsight-api-slim to PyPI
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        packages-dir: ./hindsight-api-slim/dist
+        skip-existing: true
+
     - name: Publish hindsight-api to PyPI
       uses: pypa/gh-action-pypi-publish@release/v1
       with:
@@ -70,13 +72,13 @@ jobs:
     - name: Publish hindsight-all to PyPI
       uses: pypa/gh-action-pypi-publish@release/v1
       with:
-        packages-dir: ./hindsight/dist
+        packages-dir: ./hindsight-all/dist
         skip-existing: true
 
-    - name: Publish hindsight-litellm to PyPI
+    - name: Publish hindsight-all-slim to PyPI
       uses: pypa/gh-action-pypi-publish@release/v1
       with:
-        packages-dir: ./hindsight-integrations/litellm/dist
+        packages-dir: ./hindsight-all-slim/dist
         skip-existing: true
 
     - name: Publish hindsight-embed to PyPI
@@ -85,31 +87,18 @@ jobs:
         packages-dir: ./hindsight-embed/dist
         skip-existing: true
 
-    - name: Publish hindsight-crewai to PyPI
-      uses: pypa/gh-action-pypi-publish@release/v1
-      with:
-        packages-dir: ./hindsight-integrations/crewai/dist
-        skip-existing: true
-
-    - name: Publish hindsight-pydantic-ai to PyPI
-      uses: pypa/gh-action-pypi-publish@release/v1
-      with:
-        packages-dir: ./hindsight-integrations/pydantic-ai/dist
-        skip-existing: true
-
     # Upload artifacts for GitHub release
     - name: Upload artifacts
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
       with:
         name: python-packages
         path: |
           hindsight-clients/python/dist/*
+          hindsight-api-slim/dist/*
           hindsight-api/dist/*
-          hindsight/dist/*
-          hindsight-integrations/litellm/dist/*
+          hindsight-all/dist/*
+          hindsight-all-slim/dist/*
           hindsight-embed/dist/*
-          hindsight-integrations/crewai/dist/*
-          hindsight-integrations/pydantic-ai/dist/*
         retention-days: 1
 
   release-typescript-client:
@@ -117,10 +106,10 @@ jobs:
     environment: npm
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
 
     - name: Set up Node.js
-      uses: actions/setup-node@v4
+      uses: actions/setup-node@v6
       with:
         node-version: '20'
         registry-url: 'https://registry.npmjs.org'
@@ -155,168 +144,21 @@ jobs:
       run: npm pack
 
     - name: Upload artifacts
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
       with:
         name: typescript-client
         path: hindsight-clients/typescript/*.tgz
         retention-days: 1
 
-  release-openclaw-integration:
-    runs-on: ubuntu-latest
-    environment: npm
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Set up Node.js
-      uses: actions/setup-node@v4
-      with:
-        node-version: '22'
-        registry-url: 'https://registry.npmjs.org'
-
-    - name: Install dependencies
-      working-directory: ./hindsight-integrations/openclaw
-      run: npm ci
-
-    - name: Build
-      working-directory: ./hindsight-integrations/openclaw
-      run: npm run build
-
-    - name: Publish to npm
-      working-directory: ./hindsight-integrations/openclaw
-      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: ${{ secrets.NPM_TOKEN }}
-
-    - name: Pack for GitHub release
-      working-directory: ./hindsight-integrations/openclaw
-      run: npm pack
-
-    - name: Upload artifacts
-      uses: actions/upload-artifact@v4
-      with:
-        name: openclaw-integration
-        path: hindsight-integrations/openclaw/*.tgz
-        retention-days: 1
-
-  release-ai-sdk-integration:
-    runs-on: ubuntu-latest
-    environment: npm
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Set up Node.js
-      uses: actions/setup-node@v4
-      with:
-        node-version: '22'
-        registry-url: 'https://registry.npmjs.org'
-
-    - name: Install dependencies
-      working-directory: ./hindsight-integrations/ai-sdk
-      run: npm ci
-
-    - name: Build
-      working-directory: ./hindsight-integrations/ai-sdk
-      run: npm run build
-
-    - name: Publish to npm
-      working-directory: ./hindsight-integrations/ai-sdk
-      run: |
-        set +e
-        OUTPUT=$(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: ${{ secrets.NPM_TOKEN }}
-
-    - name: Pack for GitHub release
-      working-directory: ./hindsight-integrations/ai-sdk
-      run: npm pack
-
-    - name: Upload artifacts
-      uses: actions/upload-artifact@v4
-      with:
-        name: ai-sdk-integration
-        path: hindsight-integrations/ai-sdk/*.tgz
-        retention-days: 1
-
-  release-chat-integration:
-    runs-on: ubuntu-latest
-    environment: npm
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Set up Node.js
-      uses: actions/setup-node@v4
-      with:
-        node-version: '22'
-        registry-url: 'https://registry.npmjs.org'
-
-    - name: Install dependencies
-      working-directory: ./hindsight-integrations/chat
-      run: npm ci
-
-    - name: Build
-      working-directory: ./hindsight-integrations/chat
-      run: npm run build
-
-    - name: Publish to npm
-      working-directory: ./hindsight-integrations/chat
-      run: |
-        set +e
-        OUTPUT=$(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-integrations/chat
-      run: npm pack
-
-    - name: Upload artifacts
-      uses: actions/upload-artifact@v4
-      with:
-        name: chat-integration
-        path: hindsight-integrations/chat/*.tgz
-        retention-days: 1
-
   release-control-plane:
     runs-on: ubuntu-latest
     environment: npm
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
 
     - name: Set up Node.js
-      uses: actions/setup-node@v4
+      uses: actions/setup-node@v6
       with:
         node-version: '20'
         registry-url: 'https://registry.npmjs.org'
@@ -364,7 +206,7 @@ jobs:
       run: npm pack
 
     - name: Upload artifacts
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
       with:
         name: control-plane
         path: hindsight-control-plane/*.tgz
@@ -393,7 +235,7 @@ jobs:
             asset_name: hindsight-linux-arm64
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
 
     - name: Install Rust
       uses: dtolnay/rust-toolchain@stable
@@ -411,7 +253,7 @@ jobs:
         chmod +x artifacts/${{ matrix.asset_name }}
 
     - name: Upload artifacts
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
       with:
         name: rust-cli-${{ matrix.asset_name }}
         path: artifacts/${{ matrix.asset_name }}
@@ -452,7 +294,7 @@ jobs:
               PRELOAD_ML_MODELS=false
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
 
     - name: Free Disk Space
       uses: jlumbroso/free-disk-space@main
@@ -466,13 +308,13 @@ jobs:
         swap-storage: true
 
     - name: Set up QEMU
-      uses: docker/setup-qemu-action@v3
+      uses: docker/setup-qemu-action@v4
 
     - name: Set up Docker Buildx
-      uses: docker/setup-buildx-action@v3
+      uses: docker/setup-buildx-action@v4
 
     - name: Log in to GitHub Container Registry
-      uses: docker/login-action@v3
+      uses: docker/login-action@v4
       with:
         registry: ghcr.io
         username: ${{ github.actor }}
@@ -484,7 +326,7 @@ jobs:
 
     - name: Extract metadata for release tags
       id: meta
-      uses: docker/metadata-action@v5
+      uses: docker/metadata-action@v6
       with:
         images: ghcr.io/${{ github.repository_owner }}/${{ matrix.image_name }}
         flavor: |
@@ -500,7 +342,7 @@ jobs:
     # # 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
+    #   uses: docker/build-push-action@v7
     #   with:
     #     context: .
     #     file: docker/standalone/Dockerfile
@@ -519,7 +361,7 @@ jobs:
 
     # Build multi-platform and push to release tags
     - name: Build and push release images
-      uses: docker/build-push-action@v6
+      uses: docker/build-push-action@v7
       with:
         context: .
         file: docker/standalone/Dockerfile
@@ -537,7 +379,7 @@ jobs:
       packages: write
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
 
     - name: Install Helm
       uses: azure/setup-helm@v4
@@ -557,7 +399,7 @@ jobs:
       run: helm push helm-packages/*.tgz oci://ghcr.io/${{ github.repository_owner }}/charts
 
     - name: Upload artifacts
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
       with:
         name: helm-chart
         path: helm-packages/*.tgz
@@ -565,73 +407,55 @@ jobs:
 
   create-github-release:
     runs-on: ubuntu-latest
-    needs: [release-python-packages, release-typescript-client, release-openclaw-integration, release-ai-sdk-integration, release-chat-integration, release-control-plane, 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
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
 
     - name: Extract version from tag
       id: get_version
       run: echo "VERSION=${GITHUB_REF#refs/tags/v}" >> $GITHUB_OUTPUT
 
     - name: Download Python packages
-      uses: actions/download-artifact@v4
+      uses: actions/download-artifact@v8
       with:
         name: python-packages
         path: ./artifacts/python-packages
 
     - name: Download TypeScript client
-      uses: actions/download-artifact@v4
+      uses: actions/download-artifact@v8
       with:
         name: typescript-client
         path: ./artifacts/typescript-client
 
-    - name: Download OpenClaw Integration
-      uses: actions/download-artifact@v4
-      with:
-        name: openclaw-integration
-        path: ./artifacts/openclaw-integration
-
-    - name: Download AI SDK Integration
-      uses: actions/download-artifact@v4
-      with:
-        name: ai-sdk-integration
-        path: ./artifacts/ai-sdk-integration
-
-    - name: Download Chat Integration
-      uses: actions/download-artifact@v4
-      with:
-        name: chat-integration
-        path: ./artifacts/chat-integration
-
     - name: Download Control Plane
-      uses: actions/download-artifact@v4
+      uses: actions/download-artifact@v8
       with:
         name: control-plane
         path: ./artifacts/control-plane
 
     - name: Download Rust CLI (Linux)
-      uses: actions/download-artifact@v4
+      uses: actions/download-artifact@v8
       with:
         name: rust-cli-hindsight-linux-amd64
         path: ./artifacts/rust-cli-linux
 
     - name: Download Rust CLI (macOS Intel)
-      uses: actions/download-artifact@v4
+      uses: actions/download-artifact@v8
       with:
         name: rust-cli-hindsight-darwin-amd64
         path: ./artifacts/rust-cli-darwin-amd64
 
     - name: Download Rust CLI (macOS ARM)
-      uses: actions/download-artifact@v4
+      uses: actions/download-artifact@v8
       with:
         name: rust-cli-hindsight-darwin-arm64
         path: ./artifacts/rust-cli-darwin-arm64
 
     - name: Download Helm chart
-      uses: actions/download-artifact@v4
+      uses: actions/download-artifact@v8
       with:
         name: helm-chart
         path: ./artifacts/helm-chart
@@ -641,19 +465,13 @@ jobs:
         mkdir -p release-assets
         # Python packages
         cp artifacts/python-packages/hindsight-clients/python/dist/* release-assets/ || true
+        cp artifacts/python-packages/hindsight-api-slim/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-integrations/pydantic-ai/dist/* release-assets/ || true
+        cp artifacts/python-packages/hindsight-all/dist/* release-assets/ || true
+        cp artifacts/python-packages/hindsight-all-slim/dist/* release-assets/ || true
         cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
         # TypeScript client
         cp artifacts/typescript-client/*.tgz release-assets/ || true
-        # OpenClaw Integration
-        cp artifacts/openclaw-integration/*.tgz release-assets/ || true
-        # AI SDK Integration
-        cp artifacts/ai-sdk-integration/*.tgz release-assets/ || true
-        # Chat Integration
-        cp artifacts/chat-integration/*.tgz release-assets/ || true
         # Control Plane
         cp artifacts/control-plane/*.tgz release-assets/ || true
         # Rust CLI binaries

CLAUDE.md

@@ -17,20 +17,20 @@ Hindsight is an agent memory system that provides long-term memory for AI agents
 ./scripts/dev/start-api.sh
 
 # Run all tests (parallelized with pytest-xdist)
-cd hindsight-api && uv run pytest tests/
+cd hindsight-api-slim && uv run pytest tests/
 
 # Run specific test file
-cd hindsight-api && uv run pytest tests/test_http_api_integration.py -v
+cd hindsight-api-slim && uv run pytest tests/test_http_api_integration.py -v
 
 # Run single test function
-cd hindsight-api && uv run pytest tests/test_retain.py::test_retain_simple -v
+cd hindsight-api-slim && uv run pytest tests/test_retain.py::test_retain_simple -v
 
 # Lint and format
-cd hindsight-api && uv run ruff check .
-cd hindsight-api && uv run ruff format .
+cd hindsight-api-slim && uv run ruff check .
+cd hindsight-api-slim && uv run ruff format .
 
 # Type checking (uses ty - extremely fast type checker from Astral)
-cd hindsight-api && uv run ty check hindsight_api/
+cd hindsight-api-slim && uv run ty check hindsight_api/
 ```
 
 ### Control Plane (Next.js)
@@ -72,7 +72,7 @@ cd hindsight-control-plane && npm run dev
 ## Architecture
 
 ### Monorepo Structure
-- **hindsight-api/**: Core FastAPI server with memory engine (Python, uv)
+- **hindsight-api-slim/**: 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, uses progenitor for API client)
@@ -81,9 +81,9 @@ cd hindsight-control-plane && npm run dev
 - **hindsight-integrations/**: Framework integrations (LiteLLM, OpenAI)
 - **hindsight-dev/**: Development tools and benchmarks
 
-### Core Engine (hindsight-api/hindsight_api/engine/)
+### Core Engine (hindsight-api-slim/hindsight_api/engine/)
 - `memory_engine.py`: Main orchestrator (~170KB) for retain/recall/reflect operations
-- `llm_wrapper.py`: LLM abstraction supporting OpenAI, Anthropic, Gemini, Groq, Ollama, LM Studio
+- `llm_wrapper.py`: LLM abstraction supporting OpenAI, Anthropic, Gemini, Groq, MiniMax, Ollama, LM Studio
 - `embeddings.py`: Embedding generation (local sentence-transformers or TEI)
 - `cross_encoder.py`: Reranking (local or TEI)
 - `entity_resolver.py`: Entity extraction and normalization
@@ -101,7 +101,7 @@ cd hindsight-control-plane && npm run dev
 - `fusion.py`: Reciprocal rank fusion for combining results
 - `reranking.py`: Cross-encoder reranking
 
-### API Layer (hindsight-api/hindsight_api/api/)
+### API Layer (hindsight-api-slim/hindsight_api/api/)
 - `http.py`: FastAPI HTTP routers (~80KB) for all REST endpoints
 - `mcp.py`: Model Context Protocol server implementation
 
@@ -111,13 +111,13 @@ Main operations:
 - **Reflect**: Disposition-aware reasoning using memories and mental models.
 
 ### Database
-PostgreSQL with pgvector. Schema managed via Alembic migrations in `hindsight-api/hindsight_api/alembic/`. Migrations run automatically on API startup.
+PostgreSQL with pgvector. Schema managed via Alembic migrations in `hindsight-api-slim/hindsight_api/alembic/`. Migrations run automatically on API startup.
 
 Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
 
 ### Adding Database Migrations
 
-1. **Create a new migration file** in `hindsight-api/hindsight_api/alembic/versions/`:
+1. **Create a new migration file** in `hindsight-api-slim/hindsight_api/alembic/versions/`:
    - File name format: `<revision_id>_<description>.py` (e.g., `f1a2b3c4d5e6_add_new_index.py`)
    - Use a unique hex revision ID (12 chars)
    - Set `down_revision` to the previous migration's revision ID
@@ -154,7 +154,7 @@ Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
 
 3. **Run migrations locally**:
    ```bash
-   # Set database URL and run migrations
+   # Set database URL and run migrations for the base schema plus all tenants
    uv run hindsight-admin run-db-migration
 
    # Run on a specific tenant schema
@@ -251,7 +251,7 @@ Fields must be categorized as either **hierarchical** (can be overridden per-ten
 
 #### Adding a New Configuration Field
 
-1. **config.py** (`hindsight-api/hindsight_api/config.py`):
+1. **config.py** (`hindsight-api-slim/hindsight_api/config.py`):
    - Add `ENV_*` constant for the environment variable name (e.g., `ENV_MY_SETTING = "HINDSIGHT_API_MY_SETTING"`)
    - Add `DEFAULT_*` constant for the default value
    - Add field to `HindsightConfig` dataclass with type annotation
@@ -268,7 +268,7 @@ Fields must be categorized as either **hierarchical** (can be overridden per-ten
    # Static field - just don't add to _HIERARCHICAL_FIELDS
    ```
 
-2. **main.py** (`hindsight-api/hindsight_api/main.py`):
+2. **main.py** (`hindsight-api-slim/hindsight_api/main.py`):
    - Add field to the manual `HindsightConfig()` constructor call (search for "CLI override")
 
 3. **Use hierarchical config in MemoryEngine**:
@@ -308,14 +308,14 @@ cp .env.example .env
 # Edit .env with LLM API key
 
 # Python deps
-uv sync --directory hindsight-api/
+uv sync --directory hindsight-api-slim/
 
 # Node deps (uses npm workspaces)
 npm install
 ```
 
 Required env vars:
-- `HINDSIGHT_API_LLM_PROVIDER`: openai, anthropic, gemini, groq, ollama, lmstudio
+- `HINDSIGHT_API_LLM_PROVIDER`: openai, anthropic, gemini, groq, minimax, ollama, lmstudio
 - `HINDSIGHT_API_LLM_API_KEY`: Your API key
 - `HINDSIGHT_API_LLM_MODEL`: Model name (e.g., gpt-4o-mini, claude-sonnet-4-20250514)
 

README.md

@@ -7,10 +7,12 @@
 [![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-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+[![gitcgr](https://gitcgr.com/badge/vectorize-io/hindsight.svg)](https://gitcgr.com/vectorize-io/hindsight)
 ![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)
+<br/>
 
-
+<a href="https://trendshift.io/repositories/15603" target="_blank"><img src="https://trendshift.io/api/badge/repositories/15603" alt="vectorize-io%2Fhindsight | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
 </div>
 
 ---
@@ -69,7 +71,7 @@ docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
 >API: http://localhost:8888
 >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).
+You can modify the LLM provider by setting `HINDSIGHT_API_LLM_PROVIDER`. Valid options are `openai`, `anthropic`, `gemini`, `groq`, `ollama`, `lmstudio`, and `minimax`. The documentation provides more details on [supported models](https://hindsight.vectorize.io/developer/models).
 
 
 

deno.lock

@@ -0,0 +1,139 @@
+{
+  "version": "5",
+  "specifiers": {
+    "jsr:@std/assert@^1.0.17": "1.0.19",
+    "jsr:@std/assert@^1.0.19": "1.0.19",
+    "jsr:@std/expect@*": "1.0.18",
+    "jsr:@std/internal@^1.0.12": "1.0.12",
+    "jsr:@std/path@^1.1.4": "1.1.4",
+    "jsr:@std/testing@*": "1.0.17"
+  },
+  "jsr": {
+    "@std/[email protected]": {
+      "integrity": "eaada96ee120cb980bc47e040f82814d786fe8162ecc53c91d8df60b8755991e",
+      "dependencies": [
+        "jsr:@std/internal"
+      ]
+    },
+    "@std/[email protected]": {
+      "integrity": "8566eab35200466f8609eb7e7aed062ed0db314e9a258d5d201b1b8997ce801a",
+      "dependencies": [
+        "jsr:@std/assert@^1.0.19",
+        "jsr:@std/internal",
+        "jsr:@std/path"
+      ]
+    },
+    "@std/[email protected]": {
+      "integrity": "972a634fd5bc34b242024402972cd5143eac68d8dffaca5eaa4dba30ce17b027"
+    },
+    "@std/[email protected]": {
+      "integrity": "1d2d43f39efb1b42f0b1882a25486647cb851481862dc7313390b2bb044314b5",
+      "dependencies": [
+        "jsr:@std/internal"
+      ]
+    },
+    "@std/[email protected]": {
+      "integrity": "87bdc2700fa98249d48a17cd72413352d3d3680dcfbdb64947fd0982d6bbf681",
+      "dependencies": [
+        "jsr:@std/assert@^1.0.17",
+        "jsr:@std/internal"
+      ]
+    }
+  },
+  "workspace": {
+    "members": {
+      "hindsight-clients/typescript": {
+        "packageJson": {
+          "dependencies": [
+            "npm:@hey-api/[email protected]",
+            "npm:@types/jest@29",
+            "npm:@types/node@20",
+            "npm:jest@29",
+            "npm:ts-jest@29",
+            "npm:tsup@^8.5.1",
+            "npm:typescript@5"
+          ]
+        }
+      },
+      "hindsight-control-plane": {
+        "packageJson": {
+          "dependencies": [
+            "npm:@eslint/eslintrc@^3.3.3",
+            "npm:@eslint/js@^9.39.2",
+            "npm:@radix-ui/react-alert-dialog@^1.1.15",
+            "npm:@radix-ui/react-checkbox@^1.3.3",
+            "npm:@radix-ui/react-dialog@^1.1.15",
+            "npm:@radix-ui/react-dropdown-menu@^2.1.16",
+            "npm:@radix-ui/react-label@^2.1.8",
+            "npm:@radix-ui/react-popover@^1.1.15",
+            "npm:@radix-ui/react-radio-group@^1.3.8",
+            "npm:@radix-ui/react-select@^2.2.6",
+            "npm:@radix-ui/react-slider@^1.3.6",
+            "npm:@radix-ui/react-slot@^1.2.4",
+            "npm:@radix-ui/react-switch@^1.2.6",
+            "npm:@radix-ui/react-tabs@^1.1.13",
+            "npm:@radix-ui/react-tooltip@^1.2.8",
+            "npm:@tailwindcss/postcss@^4.1.17",
+            "npm:@tailwindcss/typography@~0.5.19",
+            "npm:@types/cytoscape@^3.21.9",
+            "npm:@types/node@^24.10.0",
+            "npm:@types/react-dom@^19.2.2",
+            "npm:@types/react@^19.2.2",
+            "npm:autoprefixer@^10.4.21",
+            "npm:class-variance-authority@~0.7.1",
+            "npm:clsx@^2.1.1",
+            "npm:cmdk@^1.1.1",
+            "npm:cytoscape-fcose@^2.2.0",
+            "npm:cytoscape@^3.33.1",
+            "npm:eslint-config-next@^16.0.1",
+            "npm:eslint-plugin-react-hooks@^7.0.1",
+            "npm:eslint-plugin-react@^7.37.5",
+            "npm:eslint@^9.39.1",
+            "npm:[email protected]",
+            "npm:next-themes@~0.4.6",
+            "npm:next@^16.1.6",
+            "npm:postcss@^8.5.6",
+            "npm:prettier@^3.7.4",
+            "npm:react-chrono@^2.9.1",
+            "npm:react-dom@^19.2.0",
+            "npm:react-markdown@^10.1.0",
+            "npm:react18-json-view@~0.2.9",
+            "npm:react@^19.2.0",
+            "npm:recharts@^3.5.1",
+            "npm:remark-gfm@^4.0.1",
+            "npm:sonner@^2.0.7",
+            "npm:tailwind-merge@^3.4.0",
+            "npm:tailwindcss-animate@^1.0.7",
+            "npm:tailwindcss@^4.1.17",
+            "npm:[email protected]",
+            "npm:typescript-eslint@^8.50.0",
+            "npm:typescript@^5.9.3"
+          ]
+        }
+      },
+      "hindsight-docs": {
+        "packageJson": {
+          "dependencies": [
+            "npm:@docusaurus/[email protected]",
+            "npm:@docusaurus/[email protected]",
+            "npm:@docusaurus/[email protected]",
+            "npm:@docusaurus/theme-common@^3.9.2",
+            "npm:@docusaurus/theme-mermaid@^3.9.2",
+            "npm:@docusaurus/[email protected]",
+            "npm:@docusaurus/[email protected]",
+            "npm:@easyops-cn/docusaurus-search-local@~0.52.2",
+            "npm:@mdx-js/react@3",
+            "npm:clsx@2",
+            "npm:prism-react-renderer@^2.3.0",
+            "npm:raw-loader@^4.0.2",
+            "npm:react-dom@19",
+            "npm:react-icons@^5.6.0",
+            "npm:react@19",
+            "npm:redocusaurus@^2.5.0",
+            "npm:typescript@~5.6.2"
+          ]
+        }
+      }
+    }
+  }
+}

docker/standalone/Dockerfile

@@ -42,25 +42,22 @@ RUN apt-get update && apt-get install -y \
     && pip install --no-cache-dir uv
 
 # Copy dependency files and README (required by pyproject.toml)
-COPY hindsight-api/pyproject.toml ./api/
-COPY hindsight-api/README.md ./api/
+COPY hindsight-api-slim/pyproject.toml ./api/
+COPY hindsight-api-slim/README.md ./api/
 
 WORKDIR /app/api
 
-# Remove local ML model dependencies if INCLUDE_LOCAL_MODELS=false
-# This creates a smaller image when using external providers (TEI, OpenAI, Cohere)
-RUN if [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then \
-        echo "Removing local-models dependencies (sentence-transformers, torch, transformers)..." && \
-        sed -i '/"sentence-transformers/d' pyproject.toml && \
-        sed -i '/"transformers/d' pyproject.toml && \
-        sed -i '/"torch/d' pyproject.toml; \
+# Sync dependencies using appropriate extras based on INCLUDE_LOCAL_MODELS
+# local-ml: torch, sentence-transformers, transformers, einops, flashrank, mlx (optional)
+# embedded-db: pg0-embedded (always included for embedded PostgreSQL support)
+RUN if [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
+        uv sync --extra local-ml --extra embedded-db; \
+    else \
+        uv sync --extra embedded-db; \
     fi
 
-# Sync dependencies (will create lock file if needed)
-RUN uv sync
-
 # Copy source code (alembic migrations are inside hindsight_api/)
-COPY hindsight-api/hindsight_api ./hindsight_api
+COPY hindsight-api-slim/hindsight_api ./hindsight_api
 
 # Install the local package (uv sync only installed dependencies, not the package itself)
 RUN uv pip install -e .

docker/standalone/start-all.sh

@@ -77,18 +77,32 @@ PIDS=()
 # Start API if enabled
 if [ "$ENABLE_API" = "true" ]; then
     cd /app/api
+    API_HEALTH_URL="${HINDSIGHT_API_HEALTH_URL:-http://localhost:8888/health}"
+    API_STARTUP_WAIT_SECONDS="${HINDSIGHT_API_STARTUP_WAIT_SECONDS:-300}"
+
     # Run API directly - Python's PYTHONUNBUFFERED=1 handles output buffering
     hindsight-api &
     API_PID=$!
     PIDS+=($API_PID)
 
     # Wait for API to be ready
-    for i in {1..60}; do
-        if curl -sf http://localhost:8888/health &>/dev/null; then
+    api_ready=false
+    for ((i=1; i<=API_STARTUP_WAIT_SECONDS; i++)); do
+        if ! kill -0 "$API_PID" 2>/dev/null; then
+            wait "$API_PID"
+            exit $?
+        fi
+        if curl -sf "$API_HEALTH_URL" &>/dev/null; then
+            api_ready=true
             break
         fi
         sleep 1
     done
+
+    if [ "$api_ready" != "true" ]; then
+        echo "❌ API did not become healthy within ${API_STARTUP_WAIT_SECONDS}s"
+        exit 1
+    fi
 else
     echo "API disabled (HINDSIGHT_ENABLE_API=false)"
 fi
@@ -97,6 +111,7 @@ fi
 if [ "$ENABLE_CP" = "true" ]; then
     echo "🎛️  Starting Control Plane..."
     cd /app/control-plane
+    export HOSTNAME="${HINDSIGHT_CP_HOSTNAME:-0.0.0.0}"
     PORT="${HINDSIGHT_CP_PORT:-9999}" node server.js &
     CP_PID=$!
     PIDS+=($CP_PID)

docker/test-image.sh

@@ -49,6 +49,9 @@
 
 set -euo pipefail
 
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(dirname "$SCRIPT_DIR")"
+
 # Colors for output
 RED='\033[0;31m'
 GREEN='\033[0;32m'
@@ -178,6 +181,21 @@ for i in $(seq 1 "$TIMEOUT"); do
         echo "=== Health Response ==="
         curl -s "http://localhost:${HEALTH_PORT}${HEALTH_PATH}" | python3 -m json.tool 2>/dev/null || curl -s "http://localhost:${HEALTH_PORT}${HEALTH_PATH}"
         echo ""
+
+        # Run retain/recall smoke test for API targets
+        if [ "$TARGET" != "cp-only" ]; then
+            echo ""
+            echo "=== Retain/Recall Smoke Test ==="
+            if ! "$REPO_ROOT/scripts/smoke-test-slim.sh" "http://localhost:${HEALTH_PORT}"; then
+                echo ""
+                echo "=== Container Logs (last 50 lines) ==="
+                docker logs "$CONTAINER_NAME" 2>&1 | tail -50
+                echo ""
+                echo -e "${RED}Smoke test FAILED${NC}"
+                exit 1
+            fi
+        fi
+
         echo ""
         echo "=== Container Logs (last 50 lines) ==="
         docker logs "$CONTAINER_NAME" 2>&1 | tail -50

helm/hindsight/Chart.yaml

@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.4.16
-appVersion: "0.4.16"
+version: 0.4.20
+appVersion: "0.4.20"
 keywords:
   - ai
   - memory

hindsight-all-slim/pyproject.toml

@@ -0,0 +1,33 @@
+[build-system]
+requires = ["setuptools>=61"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "hindsight-all-slim"
+version = "0.4.20"
+description = "Hindsight: Agent Memory That Works Like Human Memory - Slim All-in-One Bundle"
+readme = "README.md"
+requires-python = ">=3.11"
+dependencies = [
+    "hindsight-api-slim>=0.4.17",
+    "hindsight-client>=0.0.7",
+    "hindsight-embed>=0.1.0",
+]
+
+[tool.uv.sources]
+hindsight-api-slim = { workspace = true }
+hindsight-client = { workspace = true }
+hindsight-embed = { workspace = true }
+
+[project.optional-dependencies]
+test = [
+    "pytest>=7.0.0",
+    "pytest-asyncio>=0.21.0",
+]
+
+[tool.setuptools]
+packages = []
+
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+asyncio_default_fixture_loop_scope = "function"

hindsight-all/README.md

@@ -0,0 +1,48 @@
+# hindsight-all
+
+All-in-one package for Hindsight - Agent Memory That Works Like Human Memory
+
+## Quick Start
+
+```python
+from hindsight import start_server, HindsightClient
+
+# Start server with embedded PostgreSQL
+server = start_server(
+    llm_provider="groq",
+    llm_api_key="your-api-key",
+    llm_model="openai/gpt-oss-120b"
+)
+
+# Create client
+client = HindsightClient(base_url=server.url)
+
+# Store memories
+client.put(agent_id="assistant", content="User prefers Python for data analysis")
+
+# Search memories
+results = client.search(agent_id="assistant", query="programming preferences")
+
+# Generate contextual response
+response = client.think(agent_id="assistant", query="What languages should I recommend?")
+
+# Stop server when done
+server.stop()
+```
+
+## Using Context Manager
+
+```python
+from hindsight import HindsightServer, HindsightClient
+
+with HindsightServer(llm_provider="groq", llm_api_key="...") as server:
+    client = HindsightClient(base_url=server.url)
+    # ... use client ...
+# Server automatically stops
+```
+
+## Installation
+
+```bash
+pip install hindsight-all
+```

hindsight-all/hindsight/client_wrapper.py

@@ -0,0 +1,423 @@
+"""
+Wrapper for Hindsight client that adds API namespaces.
+
+Provides organized access to different parts of the Hindsight API through
+namespaces like .banks, .mental_models, etc.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from hindsight_client import Hindsight
+
+
+class BanksAPI:
+    """Namespace for bank-related operations.
+    
+    Provides methods to create, delete, and manage memory banks.
+    """
+
+    def __init__(self, client: Hindsight):
+        self._client = client
+
+    def create(
+        self,
+        bank_id: str,
+        name: str | None = None,
+        mission: str | None = None,
+        disposition: dict[str, Any] | None = None,
+    ) -> Any:
+        """Create a new bank.
+        
+        Args:
+            bank_id: Unique identifier for the bank.
+            name: Optional display name for the bank.
+            mission: Optional mission statement for the bank.
+            disposition: Optional disposition configuration dict.
+            
+        Returns:
+            Bank creation response from the API.
+        """
+        return self._client.create_bank(
+            bank_id=bank_id,
+            name=name,
+            mission=mission,
+            disposition=disposition,
+        )
+
+    def delete(self, bank_id: str) -> Any:
+        """Delete a bank.
+        
+        Args:
+            bank_id: The ID of the bank to delete.
+            
+        Returns:
+            Deletion response from the API.
+        """
+        return self._client.delete_bank(bank_id=bank_id)
+
+    def set_mission(self, bank_id: str, mission: str) -> Any:
+        """Set or update the mission for a bank.
+        
+        Args:
+            bank_id: The ID of the bank.
+            mission: The mission statement to set.
+            
+        Returns:
+            API response confirming the update.
+        """
+        return self._client.set_mission(bank_id=bank_id, mission=mission)
+
+    def set_disposition(self, bank_id: str, disposition: dict[str, Any]) -> Any:
+        """Set or update the disposition for a bank.
+        
+        Args:
+            bank_id: The ID of the bank.
+            disposition: The disposition configuration dict.
+            
+        Returns:
+            API response confirming the update.
+        """
+        return self._client.set_disposition(bank_id=bank_id, disposition=disposition)
+
+    def list(self) -> Any:
+        """List all banks.
+        
+        Returns:
+            List of banks from the API.
+        """
+        from hindsight_client.hindsight_client import _run_async
+
+        return _run_async(self._client._banks_api.list_banks())
+
+
+class MentalModelsAPI:
+    """Namespace for mental model operations.
+    
+    Mental models are reusable knowledge structures that guide agent behavior.
+    """
+
+    def __init__(self, client: Hindsight):
+        self._client = client
+
+    def create(
+        self,
+        bank_id: str,
+        name: str,
+        content: str,
+        tags: list[str] | None = None,
+    ) -> Any:
+        """Create a new mental model.
+        
+        Args:
+            bank_id: The ID of the bank to add the model to.
+            name: Name for the mental model.
+            content: The content/instructions for the mental model.
+            tags: Optional list of tags for categorization.
+            
+        Returns:
+            Creation response from the API.
+        """
+        return self._client.create_mental_model(
+            bank_id=bank_id,
+            name=name,
+            content=content,
+            tags=tags,
+        )
+
+    def list(self, bank_id: str, tags: list[str] | None = None) -> Any:
+        """List all mental models for a bank.
+        
+        Args:
+            bank_id: The ID of the bank.
+            tags: Optional filter by tags.
+            
+        Returns:
+            List of mental models.
+        """
+        return self._client.list_mental_models(bank_id=bank_id, tags=tags)
+
+    def get(self, bank_id: str, mental_model_id: str) -> Any:
+        """Get a specific mental model.
+        
+        Args:
+            bank_id: The ID of the bank.
+            mental_model_id: The ID of the mental model.
+            
+        Returns:
+            The mental model details.
+        """
+        return self._client.get_mental_model(bank_id=bank_id, mental_model_id=mental_model_id)
+
+    def refresh(self, bank_id: str, mental_model_id: str) -> Any:
+        """Refresh a mental model.
+        
+        Args:
+            bank_id: The ID of the bank.
+            mental_model_id: The ID of the mental model to refresh.
+            
+        Returns:
+            Refresh response from the API.
+        """
+        return self._client.refresh_mental_model(bank_id=bank_id, mental_model_id=mental_model_id)
+
+    def update(
+        self,
+        bank_id: str,
+        mental_model_id: str,
+        name: str | None = None,
+        content: str | None = None,
+        tags: list[str] | None = None,
+    ) -> Any:
+        """Update a mental model.
+        
+        Args:
+            bank_id: The ID of the bank.
+            mental_model_id: The ID of the mental model to update.
+            name: Optional new name.
+            content: Optional new content.
+            tags: Optional new tags list.
+            
+        Returns:
+            Update response from the API.
+        """
+        return self._client.update_mental_model(
+            bank_id=bank_id,
+            mental_model_id=mental_model_id,
+            name=name,
+            content=content,
+            tags=tags,
+        )
+
+    def delete(self, bank_id: str, mental_model_id: str) -> Any:
+        """Delete a mental model.
+        
+        Args:
+            bank_id: The ID of the bank.
+            mental_model_id: The ID of the mental model to delete.
+            
+        Returns:
+            Deletion response from the API.
+        """
+        return self._client.delete_mental_model(bank_id=bank_id, mental_model_id=mental_model_id)
+
+
+class DirectivesAPI:
+    """Namespace for directive operations.
+    
+    Directives are explicit instructions that guide agent behavior.
+    """
+
+    def __init__(self, client: Hindsight):
+        self._client = client
+
+    def create(
+        self,
+        bank_id: str,
+        name: str,
+        content: str,
+        tags: list[str] | None = None,
+    ) -> Any:
+        """Create a new directive.
+        
+        Args:
+            bank_id: The ID of the bank to add the directive to.
+            name: Name for the directive.
+            content: The directive content/instructions.
+            tags: Optional list of tags for categorization.
+            
+        Returns:
+            Creation response from the API.
+        """
+        return self._client.create_directive(
+            bank_id=bank_id,
+            name=name,
+            content=content,
+            tags=tags,
+        )
+
+    def list(self, bank_id: str, tags: list[str] | None = None) -> Any:
+        """List all directives for a bank.
+        
+        Args:
+            bank_id: The ID of the bank.
+            tags: Optional filter by tags.
+            
+        Returns:
+            List of directives.
+        """
+        return self._client.list_directives(bank_id=bank_id, tags=tags)
+
+    def get(self, bank_id: str, directive_id: str) -> Any:
+        """Get a specific directive.
+        
+        Args:
+            bank_id: The ID of the bank.
+            directive_id: The ID of the directive.
+            
+        Returns:
+            The directive details.
+        """
+        return self._client.get_directive(bank_id=bank_id, directive_id=directive_id)
+
+    def update(
+        self,
+        bank_id: str,
+        directive_id: str,
+        name: str | None = None,
+        content: str | None = None,
+        tags: list[str] | None = None,
+    ) -> Any:
+        """Update a directive.
+        
+        Args:
+            bank_id: The ID of the bank.
+            directive_id: The ID of the directive to update.
+            name: Optional new name.
+            content: Optional new content.
+            tags: Optional new tags list.
+            
+        Returns:
+            Update response from the API.
+        """
+        return self._client.update_directive(
+            bank_id=bank_id,
+            directive_id=directive_id,
+            name=name,
+            content=content,
+            tags=tags,
+        )
+
+    def delete(self, bank_id: str, directive_id: str) -> Any:
+        """Delete a directive.
+        
+        Args:
+            bank_id: The ID of the bank.
+            directive_id: The ID of the directive to delete.
+            
+        Returns:
+            Deletion response from the API.
+        """
+        return self._client.delete_directive(bank_id=bank_id, directive_id=directive_id)
+
+
+class MemoriesAPI:
+    """Namespace for memory operations.
+    
+    Provides methods to query and retrieve stored memories.
+    """
+
+    def __init__(self, client: Hindsight):
+        self._client = client
+
+    def list(
+        self,
+        bank_id: str,
+        type: str | None = None,
+        search_query: str | None = None,
+        limit: int = 100,
+        offset: int = 0,
+    ) -> Any:
+        """List memories in a bank.
+        
+        Args:
+            bank_id: The ID of the bank to query.
+            type: Optional filter by memory type.
+            search_query: Optional search query for filtering.
+            limit: Maximum number of results to return (default: 100).
+            offset: Number of results to skip for pagination (default: 0).
+            
+        Returns:
+            List of memories matching the criteria.
+        """
+        return self._client.list_memories(
+            bank_id=bank_id,
+            type=type,
+            search_query=search_query,
+            limit=limit,
+            offset=offset,
+        )
+
+
+class HindsightClient(Hindsight):
+    """
+    Enhanced Hindsight client with organized API namespaces.
+
+    This wrapper extends the auto-generated Hindsight client with organized
+    access to different parts of the API through namespaces.
+
+    Example:
+        ```python
+        from hindsight import HindsightClient
+
+        client = HindsightClient(base_url="http://localhost:8888")
+
+        # Core operations (inherited from Hindsight)
+        client.retain(bank_id="test", content="Hello")
+        results = client.recall(bank_id="test", query="Hello")
+
+        # Organized API access through namespaces
+        client.banks.create(bank_id="test", name="Test Bank")
+        models = client.mental_models.list(bank_id="test")
+        directives = client.directives.list(bank_id="test")
+        memories = client.memories.list(bank_id="test")
+        ```
+        
+    Attributes:
+        banks: Namespace for bank management operations.
+        mental_models: Namespace for mental model operations.
+        directives: Namespace for directive operations.
+        memories: Namespace for memory listing operations.
+    """
+
+    def __init__(self, *args: Any, **kwargs: Any) -> None:
+        super().__init__(*args, **kwargs)
+        self._banks_namespace: BanksAPI | None = None
+        self._mental_models_namespace: MentalModelsAPI | None = None
+        self._directives_namespace: DirectivesAPI | None = None
+        self._memories_namespace: MemoriesAPI | None = None
+
+    @property
+    def banks(self) -> BanksAPI:
+        """Access bank management operations.
+        
+        Returns:
+            BanksAPI instance for bank operations.
+        """
+        if self._banks_namespace is None:
+            self._banks_namespace = BanksAPI(self)
+        return self._banks_namespace
+
+    @property
+    def mental_models(self) -> MentalModelsAPI:
+        """Access mental model operations.
+        
+        Returns:
+            MentalModelsAPI instance for mental model operations.
+        """
+        if self._mental_models_namespace is None:
+            self._mental_models_namespace = MentalModelsAPI(self)
+        return self._mental_models_namespace
+
+    @property
+    def directives(self) -> DirectivesAPI:
+        """Access directive operations.
+        
+        Returns:
+            DirectivesAPI instance for directive operations.
+        """
+        if self._directives_namespace is None:
+            self._directives_namespace = DirectivesAPI(self)
+        return self._directives_namespace
+
+    @property
+    def memories(self) -> MemoriesAPI:
+        """Access memory listing operations.
+        
+        Returns:
+            MemoriesAPI instance for memory operations.
+        """
+        if self._memories_namespace is None:
+            self._memories_namespace = MemoriesAPI(self)
+        return self._memories_namespace

hindsight-all/hindsight/embedded.py

@@ -34,7 +34,6 @@ Using context manager:
 """
 
 import logging
-import os
 import threading
 from typing import Optional
 
@@ -140,13 +139,17 @@ class HindsightEmbedded:
                 return
 
             if self._closed:
-                raise RuntimeError("Cannot use HindsightEmbedded after it has been closed")
+                raise RuntimeError(
+                    "Cannot use HindsightEmbedded after it has been closed"
+                )
 
             # Use embed manager interface for daemon management
             logger.info(f"Ensuring daemon is running for profile '{self.profile}'...")
             success = self._manager.ensure_running(self.config, self.profile)
             if not success:
-                raise RuntimeError(f"Failed to start daemon for profile '{self.profile}'")
+                raise RuntimeError(
+                    f"Failed to start daemon for profile '{self.profile}'"
+                )
 
             # Get daemon URL and create client
             daemon_url = self._manager.get_url(self.profile)
@@ -375,3 +378,45 @@ class HindsightEmbedded:
     def is_running(self) -> bool:
         """Check if the client is initialized."""
         return self._started and not self._closed and self._client is not None
+
+    def start_ui(self, ui_port: int | None = None, hostname: str = "0.0.0.0") -> bool:
+        """Start the control plane web UI.
+
+        The daemon is started automatically if not already running.
+
+        Args:
+            ui_port: Port for the UI. Defaults to daemon_port + 10000.
+            hostname: Hostname to bind to. Defaults to 0.0.0.0.
+
+        Returns:
+            True if UI started successfully.
+        """
+        self._ensure_started()
+        return self._manager.start_ui(self.profile, ui_port, hostname)
+
+    def stop_ui(self, ui_port: int | None = None) -> bool:
+        """Stop the control plane web UI.
+
+        Args:
+            ui_port: Port the UI is running on. Defaults to daemon_port + 10000.
+
+        Returns:
+            True if stopped successfully.
+        """
+        return self._manager.stop_ui(self.profile, ui_port)
+
+    def is_ui_running(self, ui_port: int | None = None) -> bool:
+        """Check if the control plane web UI is running.
+
+        Args:
+            ui_port: Port to check. Defaults to daemon_port + 10000.
+
+        Returns:
+            True if UI is running and responsive.
+        """
+        return self._manager.is_ui_running(self.profile, ui_port)
+
+    @property
+    def ui_url(self) -> str:
+        """Get the UI URL for this profile."""
+        return self._manager.get_ui_url(self.profile)

hindsight-all/pyproject.toml

@@ -4,18 +4,18 @@ build-backend = "hatchling.build"
 
 [project]
 name = "hindsight-all"
-version = "0.4.16"
+version = "0.4.20"
 description = "Hindsight: Agent Memory That Works Like Human Memory - All-in-One Bundle"
 readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
-    "hindsight-api>=0.0.7",
+    "hindsight-api-slim[all]>=0.4.17",
     "hindsight-client>=0.0.7",
     "hindsight-embed>=0.1.0",
 ]
 
 [tool.uv.sources]
-hindsight-api = { workspace = true }
+hindsight-api-slim = { workspace = true }
 hindsight-client = { workspace = true }
 hindsight-embed = { workspace = true }
 

hindsight-api-slim/README.md

@@ -0,0 +1,137 @@
+# 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-slim/hindsight_api/__init__.py

@@ -46,4 +46,4 @@ __all__ = [
     "RemoteTEICrossEncoder",
     "LLMConfig",
 ]
-__version__ = "0.4.16"
+__version__ = "0.4.20"

hindsight-api-slim/hindsight_api/admin/cli.py

@@ -14,7 +14,8 @@ from typing import Any
 import asyncpg
 import typer
 
-from ..config import HindsightConfig
+from ..config import DEFAULT_DATABASE_SCHEMA, HindsightConfig
+from ..extensions import TenantExtension, load_extension
 from ..pg0 import parse_pg0_url, resolve_database_url
 
 
@@ -214,20 +215,81 @@ def restore(
     typer.echo("Restore complete")
 
 
-async def _run_migration(db_url: str, schema: str = "public") -> None:
-    """Resolve database URL and run migrations."""
-    from ..migrations import run_migrations
+async def _run_migration(
+    db_url: str,
+    schema: str | None = None,
+    base_schema: str = DEFAULT_DATABASE_SCHEMA,
+    embedding_dimension: int | None = None,
+) -> list[str]:
+    """Resolve database URL and run migrations for one schema or all discovered schemas."""
+    from ..migrations import (
+        ensure_embedding_dimension,
+        ensure_text_search_extension,
+        ensure_vector_extension,
+        run_migrations,
+    )
 
     is_pg0, instance_name, _ = parse_pg0_url(db_url)
     if is_pg0:
         typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
     resolved_url = await resolve_database_url(db_url)
-    run_migrations(resolved_url, schema=schema)
+
+    config = HindsightConfig.from_env()
+    if schema:
+        schemas = [schema]
+    else:
+        tenant_extension = load_extension("TENANT", TenantExtension)
+
+        schemas = [base_schema or DEFAULT_DATABASE_SCHEMA]
+        if tenant_extension:
+            tenants = await tenant_extension.list_tenants()
+            schemas.extend(tenant.schema for tenant in tenants if tenant.schema)
+
+        # Preserve order while removing duplicates.
+        schemas = list(dict.fromkeys(schemas))
+
+    for schema in schemas:
+        run_migrations(resolved_url, schema=schema)
+
+    if embedding_dimension is not None:
+        for schema in schemas:
+            ensure_embedding_dimension(
+                resolved_url,
+                embedding_dimension,
+                schema=schema,
+                vector_extension=config.vector_extension,
+            )
+
+    for schema in schemas:
+        ensure_vector_extension(
+            resolved_url,
+            vector_extension=config.vector_extension,
+            schema=schema,
+        )
+
+    for schema in schemas:
+        ensure_text_search_extension(
+            resolved_url,
+            text_search_extension=config.text_search_extension,
+            schema=schema,
+        )
+
+    return schemas
 
 
 @app.command(name="run-db-migration")
 def run_db_migration(
-    schema: str = typer.Option("public", "--schema", "-s", help="Database schema to run migrations on"),
+    schema: str | None = typer.Option(
+        None,
+        "--schema",
+        "-s",
+        help="Database schema to run migrations on. If omitted, migrate the base schema and all discovered tenant schemas.",
+    ),
+    embedding_dimension: int | None = typer.Option(
+        None,
+        "--embedding-dimension",
+        help="Expected embedding dimension to enforce after migrations. Omit to skip dimension sync.",
+    ),
 ):
     """Run database migrations to the latest version."""
     config = HindsightConfig.from_env()
@@ -237,11 +299,21 @@ def run_db_migration(
         typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
         raise typer.Exit(1)
 
-    typer.echo(f"Running database migrations (schema: {schema})...")
+    if schema:
+        typer.echo(f"Running database migrations for schema: {schema}...")
+    else:
+        typer.echo("Running database migrations for base schema and all discovered tenant schemas...")
 
-    asyncio.run(_run_migration(config.database_url, schema))
+    schemas = asyncio.run(
+        _run_migration(
+            config.database_url,
+            schema=schema,
+            base_schema=config.database_schema,
+            embedding_dimension=embedding_dimension,
+        )
+    )
 
-    typer.echo("Database migrations completed successfully")
+    typer.echo(f"Database migrations completed successfully for {len(schemas)} schema(s)")
 
 
 async def _decommission_worker(db_url: str, worker_id: str, schema: str = "public") -> int:

hindsight-api-slim/hindsight_api/alembic/versions/a1b2c3d4e5f6_add_file_storage_table.py

@@ -34,7 +34,7 @@ def upgrade() -> None:
     # Create file_storage table (minimal: just key + data)
     op.execute(
         f"""
-        CREATE TABLE {schema}file_storage (
+        CREATE TABLE IF NOT EXISTS {schema}file_storage (
             storage_key TEXT PRIMARY KEY,
             data BYTEA NOT NULL
         )

hindsight-api-slim/hindsight_api/alembic/versions/a3b4c5d6e7f8_add_consolidation_failed_at_to_memory_units.py

@@ -0,0 +1,52 @@
+"""Add consolidation_failed_at column to memory_units for tracking persistent LLM failures.
+
+When all LLM retries are exhausted on a single-memory batch, the memory is marked
+with consolidation_failed_at instead of consolidated_at, so it is not silently lost
+and can be retried later via the API.
+
+Revision ID: a3b4c5d6e7f8
+Revises: g7h8i9j0k1l2
+Create Date: 2026-03-17
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "a3b4c5d6e7f8"
+down_revision: str | Sequence[str] | None = "g7h8i9j0k1l2"
+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 (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+
+    op.execute(
+        f"""
+        ALTER TABLE {schema}memory_units
+        ADD COLUMN IF NOT EXISTS consolidation_failed_at TIMESTAMPTZ DEFAULT NULL
+        """
+    )
+
+    # Index to efficiently query memories that failed consolidation for a given bank
+    op.execute(
+        f"""
+        CREATE INDEX IF NOT EXISTS idx_memory_units_consolidation_failed
+        ON {schema}memory_units (bank_id, consolidation_failed_at)
+        WHERE consolidation_failed_at IS NOT NULL AND fact_type IN ('experience', 'world')
+        """
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_consolidation_failed")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS consolidation_failed_at")

hindsight-api-slim/hindsight_api/alembic/versions/c1a2b3d4e5f6_enable_pg_trgm_and_entities_trgm_index.py

@@ -11,6 +11,7 @@ block; see migrations.py for how this is handled safely.
 
 from collections.abc import Sequence
 
+import sqlalchemy as sa
 from alembic import context, op
 
 revision: str = "c1a2b3d4e5f6"
@@ -25,9 +26,21 @@ def _get_schema_prefix() -> str:
 
 
 def upgrade() -> None:
-    # pg_trgm ships with every standard PostgreSQL installation as a contrib module.
+    # pg_trgm ships with most PostgreSQL installations as a contrib module.
     # It enables fast similarity lookups via GIN indexes, used for entity name matching.
-    op.execute("CREATE EXTENSION IF NOT EXISTS pg_trgm")
+    # On managed services (e.g. Azure Flexible Server), the extension may not be
+    # available or may require manual enablement.  We gracefully skip the index
+    # creation if the extension cannot be loaded — the entity resolver will
+    # auto-detect and fall back to the "full" lookup strategy at runtime.  See #626.
+    conn = op.get_bind()
+    try:
+        conn.execute(sa.text("CREATE EXTENSION IF NOT EXISTS pg_trgm"))
+    except Exception:
+        # Extension not available (managed Postgres, insufficient privileges, etc.)
+        # Roll back the failed statement and skip index creation.
+        conn.execute(sa.text("ROLLBACK"))
+        conn.execute(sa.text("BEGIN"))
+        return
 
     schema = _get_schema_prefix()
     # GIN index on canonical_name enables sub-millisecond trigram similarity queries

hindsight-api-slim/hindsight_api/alembic/versions/c3d4e5f6g7h8_add_history_to_mental_models.py

@@ -0,0 +1,30 @@
+"""Add history column to mental_models
+
+Revision ID: c3d4e5f6g7h8
+Revises: a2b3c4d5e6f7, a2b3c4d5e6f8
+Create Date: 2026-03-06
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "c3d4e5f6g7h8"
+down_revision: str | Sequence[str] | None = ("a2b3c4d5e6f7", "a2b3c4d5e6f8")
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}mental_models ADD COLUMN IF NOT EXISTS history JSONB DEFAULT '[]'::jsonb")
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS history")

hindsight-api-slim/hindsight_api/alembic/versions/d4e5f6g7h8i9_gin_source_memory_ids_fastupdate_off.py

@@ -0,0 +1,53 @@
+"""Recreate idx_memory_units_source_memory_ids GIN index with fastupdate=off
+
+GIN indexes use a "fastupdate" pending list by default: small writes are
+buffered there and flushed to the main GIN tree in bulk. Flushing requires
+AccessExclusiveLock on the index. Under high insert concurrency (e.g. 8
+parallel pytest-xdist workers all calling retain_async) two transactions can
+each trigger a flush simultaneously and deadlock.
+
+Disabling fastupdate makes every insert write directly to the GIN tree
+(slightly slower per insert, but no pending-list lock cycles).
+
+Revision ID: d4e5f6g7h8i9
+Revises: d5e6f7a8b9c0
+Create Date: 2026-03-11
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "d4e5f6g7h8i9"
+down_revision: str | Sequence[str] | None = "d5e6f7a8b9c0"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    # DROP + CREATE CONCURRENTLY must run outside a transaction block.
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_source_memory_ids")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_source_memory_ids "
+        f"ON {schema}memory_units USING GIN (source_memory_ids) "
+        f"WITH (fastupdate=off) "
+        f"WHERE source_memory_ids IS NOT NULL"
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_source_memory_ids")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_source_memory_ids "
+        f"ON {schema}memory_units USING GIN (source_memory_ids) "
+        f"WHERE source_memory_ids IS NOT NULL"
+    )

hindsight-api-slim/hindsight_api/alembic/versions/d5e6f7a8b9c0_add_bank_internal_id_and_per_bank_hnsw.py

@@ -0,0 +1,131 @@
+"""Add internal_id to banks and per-(bank, fact_type) partial HNSW indexes
+
+Revision ID: d5e6f7a8b9c0
+Revises: a3b4c5d6e7f8
+Create Date: 2026-03-11
+
+This migration:
+1. Adds internal_id UUID column to banks (stable identifier for index naming)
+2. Drops the global HNSW index (competes with per-bank partial indexes)
+3. Creates per-(bank_id, fact_type) partial HNSW indexes for all existing banks
+   (new banks get indexes created at bank-creation time via bank_utils.create_bank_hnsw_indexes)
+
+Why per-(bank, fact_type) indexes:
+- fact_type-only partial indexes are never chosen by the planner when bank_id is in the WHERE
+  clause, because the idx_memory_units_bank_id B-tree index always wins at planning time.
+- Per-(bank, fact_type) partial indexes have both predicates matching → planner selects them.
+- The global HNSW index competes for larger partitions (world, observation) and must be dropped.
+
+For large deployments, create indexes CONCURRENTLY before running this migration:
+    SELECT internal_id, bank_id FROM banks;
+    -- for each bank and each fact_type in (world, experience, observation):
+    CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_mu_emb_{ft}_{uid16}
+        ON memory_units USING hnsw (embedding vector_cosine_ops)
+        WHERE fact_type = '{ft}' AND bank_id = '{bank_id}';
+    DROP INDEX CONCURRENTLY IF EXISTS idx_memory_units_embedding;
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+from sqlalchemy import text
+
+revision: str = "d5e6f7a8b9c0"
+down_revision: str | Sequence[str] | None = "c3d4e5f6g7h8"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+_HNSW_FACT_TYPES: dict[str, str] = {
+    "world": "worl",
+    "experience": "expr",
+    "observation": "obsv",
+}
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+
+    # 1. Add internal_id column to banks
+    op.execute(
+        f"ALTER TABLE {schema}banks ADD COLUMN IF NOT EXISTS internal_id UUID DEFAULT gen_random_uuid() NOT NULL"
+    )
+    op.execute(f"ALTER TABLE {schema}banks ADD CONSTRAINT banks_internal_id_unique UNIQUE (internal_id)")
+
+    # 2. Drop any fact_type-only partial HNSW indexes that may exist from prior migrations
+    #    (bank_id B-tree always wins over them when bank_id is in the WHERE clause)
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_mu_emb_world")
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_mu_emb_observation")
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_mu_emb_experience")
+
+    # 4. Drop global HNSW index (competes with per-bank partial indexes)
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_embedding")
+
+    # 5. Create per-(bank, fact_type) partial HNSW indexes for all existing banks
+    bind = op.get_bind()
+    schema_name = context.config.get_main_option("target_schema")
+    table_ref = f'"{schema_name}".memory_units' if schema_name else "memory_units"
+    banks_ref = f'"{schema_name}".banks' if schema_name else "banks"
+
+    rows = bind.execute(text(f"SELECT bank_id, internal_id FROM {banks_ref}")).fetchall()  # noqa: S608
+    for row in rows:
+        bank_id = row[0]
+        internal_id = str(row[1]).replace("-", "")[:16]
+        escaped_bank_id = bank_id.replace("'", "''")
+        for ft, ft_short in _HNSW_FACT_TYPES.items():
+            idx_name = f"idx_mu_emb_{ft_short}_{internal_id}"
+            # Index name is schema-unqualified (indexes live in the schema of their table)
+            bind.execute(
+                text(
+                    f"CREATE INDEX IF NOT EXISTS {idx_name} "
+                    f"ON {table_ref} USING hnsw (embedding vector_cosine_ops) "
+                    f"WHERE fact_type = '{ft}' AND bank_id = '{escaped_bank_id}'"
+                )
+            )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+
+    # Drop per-bank HNSW indexes (iterate existing banks)
+    bind = op.get_bind()
+    schema_name = context.config.get_main_option("target_schema")
+    banks_ref = f'"{schema_name}".banks' if schema_name else "banks"
+
+    rows = bind.execute(text(f"SELECT internal_id FROM {banks_ref}")).fetchall()  # noqa: S608
+    for row in rows:
+        internal_id = str(row[0]).replace("-", "")[:16]
+        for ft_short in _HNSW_FACT_TYPES.values():
+            idx_name = f"idx_mu_emb_{ft_short}_{internal_id}"
+            bind.execute(text(f"DROP INDEX IF EXISTS {schema}{idx_name}"))
+
+    # Restore the global HNSW index
+    table_ref = f'"{schema_name}".memory_units' if schema_name else "memory_units"
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_memory_units_embedding ON {table_ref} USING hnsw (embedding vector_cosine_ops)"
+    )
+
+    # Restore old fact_type-only partial indexes
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_mu_emb_world "
+        f"ON {table_ref} USING hnsw (embedding vector_cosine_ops) "
+        f"WHERE fact_type = 'world'"
+    )
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_mu_emb_observation "
+        f"ON {table_ref} USING hnsw (embedding vector_cosine_ops) "
+        f"WHERE fact_type = 'observation'"
+    )
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_mu_emb_experience "
+        f"ON {table_ref} USING hnsw (embedding vector_cosine_ops) "
+        f"WHERE fact_type = 'experience'"
+    )
+
+    # Drop internal_id column
+    op.execute(f"ALTER TABLE {schema}banks DROP CONSTRAINT IF EXISTS banks_internal_id_unique")
+    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS internal_id")

hindsight-api-slim/hindsight_api/alembic/versions/e5f6g7h8i9j0_cascade_delete_ops_on_bank_delete.py

@@ -0,0 +1,73 @@
+"""Add CASCADE DELETE FK from async_operations and webhooks to banks.
+
+When a bank is deleted, all its async_operations and webhooks rows are
+automatically deleted by the database. This ensures that any in-flight
+worker tasks detect the deletion via _check_op_alive() and abort early.
+
+Revision ID: e5f6g7h8i9j0
+Revises: d4e5f6g7h8i9
+Create Date: 2026-03-11
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "e5f6g7h8i9j0"
+down_revision: str | Sequence[str] | None = "d4e5f6g7h8i9"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+
+    # Remove orphaned async_operations rows whose bank no longer exists
+    # (can happen because there was no FK before this migration).
+    op.execute(
+        f"""
+        DELETE FROM {schema}async_operations
+        WHERE bank_id IS NOT NULL
+          AND bank_id NOT IN (SELECT bank_id FROM {schema}banks)
+        """
+    )
+
+    # Remove orphaned webhooks rows whose bank no longer exists.
+    op.execute(
+        f"""
+        DELETE FROM {schema}webhooks
+        WHERE bank_id IS NOT NULL
+          AND bank_id NOT IN (SELECT bank_id FROM {schema}banks)
+        """
+    )
+
+    # Add FK with ON DELETE CASCADE so that deleting a bank automatically
+    # cleans up all its pending/processing operations and webhook configs.
+    op.execute(
+        f"""
+        ALTER TABLE {schema}async_operations
+            ADD CONSTRAINT fk_async_operations_bank_id
+            FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id)
+            ON DELETE CASCADE
+        """
+    )
+
+    op.execute(
+        f"""
+        ALTER TABLE {schema}webhooks
+            ADD CONSTRAINT fk_webhooks_bank_id
+            FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id)
+            ON DELETE CASCADE
+        """
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}async_operations DROP CONSTRAINT IF EXISTS fk_async_operations_bank_id")
+    op.execute(f"ALTER TABLE {schema}webhooks DROP CONSTRAINT IF EXISTS fk_webhooks_bank_id")

hindsight-api-slim/hindsight_api/alembic/versions/f6g7h8i9j0k1_chunk_fk_cascade_delete.py

@@ -0,0 +1,38 @@
+"""chunk_fk_cascade_delete
+
+Revision ID: f6g7h8i9j0k1
+Revises: e5f6g7h8i9j0
+Create Date: 2026-03-16 00:00:00.000000
+
+"""
+
+from collections.abc import Sequence
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "f6g7h8i9j0k1"
+down_revision: str | Sequence[str] | None = "e5f6g7h8i9j0"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    """Change memory_units.chunk_id FK from SET NULL to CASCADE.
+
+    When a document is deleted the CASCADE reaches chunks first; with SET NULL
+    the memory_units rows survived with chunk_id = NULL, leaving ghost records.
+    Switching to CASCADE ensures they are removed together with their chunk.
+    """
+    op.drop_constraint("memory_units_chunk_fkey", "memory_units", type_="foreignkey")
+    op.create_foreign_key(
+        "memory_units_chunk_fkey", "memory_units", "chunks", ["chunk_id"], ["chunk_id"], ondelete="CASCADE"
+    )
+
+
+def downgrade() -> None:
+    """Revert to SET NULL behaviour."""
+    op.drop_constraint("memory_units_chunk_fkey", "memory_units", type_="foreignkey")
+    op.create_foreign_key(
+        "memory_units_chunk_fkey", "memory_units", "chunks", ["chunk_id"], ["chunk_id"], ondelete="SET NULL"
+    )

hindsight-api-slim/hindsight_api/alembic/versions/g7h8i9j0k1l2_backsweep_orphan_observations.py

@@ -0,0 +1,71 @@
+"""backsweep_orphan_memory_units
+
+Two-pass cleanup of memory_units rows that were never removed by earlier bugs:
+
+Pass 1 — any fact_type, bank gone:
+  memory_units whose bank_id no longer exists in banks.  These accumulate when
+  a bank is deleted without a proper cascade (no FK from memory_units to banks
+  exists in the schema).
+
+Pass 2 — observations only, all sources gone:
+  observation rows whose bank still exists but every source_memory_id points
+  to a deleted memory unit.  These were left behind before PR #580 fixed the
+  chunk FK cascade and before delete_document() called
+  _delete_stale_observations_for_memories.
+
+Revision ID: g7h8i9j0k1l2
+Revises: f6g7h8i9j0k1
+Create Date: 2026-03-16
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "g7h8i9j0k1l2"
+down_revision: str | Sequence[str] | None = "f6g7h8i9j0k1"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    mu = f"{schema}memory_units"
+    banks = f"{schema}banks"
+
+    # Pass 1: delete all memory_units (any fact_type) whose bank no longer exists.
+    # There is no FK from memory_units to banks, so these never cascade away.
+    op.execute(
+        f"""
+        DELETE FROM {mu}
+        WHERE NOT EXISTS (
+            SELECT 1 FROM {banks} b WHERE b.bank_id = {mu}.bank_id
+        )
+        """
+    )
+
+    # Pass 2: delete orphaned observations whose bank still exists but every
+    # source_memory_id refers to a now-deleted memory unit (or the array is
+    # empty).  Observations with at least one surviving source are left alone.
+    op.execute(
+        f"""
+        DELETE FROM {mu} orphan
+        WHERE orphan.fact_type = 'observation'
+          AND NOT EXISTS (
+              SELECT 1
+              FROM {mu} src
+              WHERE src.id = ANY(orphan.source_memory_ids)
+                AND src.bank_id = orphan.bank_id
+          )
+        """
+    )
+
+
+def downgrade() -> None:
+    # Deleted rows cannot be restored.
+    pass

hindsight-api-slim/hindsight_api/alembic/versions/y0t1u2v3w4x5_add_result_metadata_gin_index.py

@@ -35,7 +35,7 @@ def upgrade() -> None:
 
     # Add GIN index for JSONB containment queries (@> operator)
     op.execute(f"""
-        CREATE INDEX idx_async_operations_result_metadata
+        CREATE INDEX IF NOT EXISTS idx_async_operations_result_metadata
         ON {schema}async_operations
         USING gin(result_metadata)
     """)

hindsight-api-slim/hindsight_api/api/http.py

@@ -10,7 +10,7 @@ import json
 import logging
 import uuid
 from contextlib import asynccontextmanager
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Any, Literal
 
 from fastapi import Depends, FastAPI, File, Form, Header, HTTPException, Query, UploadFile
@@ -34,7 +34,7 @@ def _parse_metadata(metadata: Any) -> dict[str, Any]:
 
 from typing import Callable
 
-from pydantic import BaseModel, ConfigDict, Field, field_validator
+from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
 
 from hindsight_api import MemoryEngine
 
@@ -72,16 +72,15 @@ def FieldWithDefault(default_factory: Callable, **kwargs) -> Any:
 
 from hindsight_api.config import get_config
 from hindsight_api.engine.memory_engine import Budget, _current_schema, _get_tiktoken_encoding, fq_table
+from hindsight_api.engine.providers.none_llm import LLMNotAvailableError
 from hindsight_api.engine.response_models import VALID_RECALL_FACT_TYPES, MemoryFact, TokenUsage
-from hindsight_api.engine.search.tags import TagsMatch
+from hindsight_api.engine.search.tags import TagGroup, TagsMatch
 from hindsight_api.extensions import HttpExtension, OperationValidationError, load_extension
 from hindsight_api.metrics import create_metrics_collector, get_metrics_collector, initialize_metrics
 from hindsight_api.models import RequestContext
 
 logger = logging.getLogger(__name__)
 
-MAX_QUERY_TOKENS = 500  # Maximum tokens allowed in recall query
-
 
 class EntityIncludeOptions(BaseModel):
     """Options for including entity observations in recall results."""
@@ -98,7 +97,12 @@ class ChunkIncludeOptions(BaseModel):
 class SourceFactsIncludeOptions(BaseModel):
     """Options for including source facts for observation-type results."""
 
-    max_tokens: int = Field(default=4096, description="Maximum tokens for source facts")
+    max_tokens: int = Field(
+        default=4096, description="Maximum total tokens for source facts across all observations (-1 = unlimited)"
+    )
+    max_tokens_per_observation: int = Field(
+        default=-1, description="Maximum tokens of source facts per observation (-1 = unlimited)"
+    )
 
 
 class IncludeOptions(BaseModel):
@@ -160,6 +164,26 @@ class RecallRequest(BaseModel):
         description="How to match tags: 'any' (OR, includes untagged), 'all' (AND, includes untagged), "
         "'any_strict' (OR, excludes untagged), 'all_strict' (AND, excludes untagged).",
     )
+    tag_groups: list[TagGroup] | None = Field(
+        default=None,
+        description="Compound tag filter using boolean groups. Groups in the list are AND-ed. "
+        "Each group is a leaf {tags, match} or compound {and: [...]}, {or: [...]}, {not: ...}.",
+    )
+
+    @field_validator("query")
+    @classmethod
+    def validate_query_not_empty(cls, v: str) -> str:
+        from ..engine.search.retrieval import tokenize_query
+
+        if not tokenize_query(v):
+            raise ValueError("query must contain at least one word character after normalization")
+        return v
+
+    @model_validator(mode="after")
+    def validate_tags_exclusive(self) -> "RecallRequest":
+        if self.tags is not None and self.tag_groups is not None:
+            raise ValueError("'tags' and 'tag_groups' are mutually exclusive. Use 'tag_groups' for compound filtering.")
+        return self
 
 
 class RecallResult(BaseModel):
@@ -401,6 +425,27 @@ class MemoryItem(BaseModel):
         default=None,
         description="Optional tags for visibility scoping. Memories with tags can be filtered during recall.",
     )
+
+    @field_validator("tags", mode="before")
+    @classmethod
+    def coerce_tags(cls, v):
+        """Coerce JSON-string tags to list.
+
+        MCP tool bridges sometimes serialize JSON arrays as strings during
+        transport, e.g. '["a", "b"]' instead of ["a", "b"]. This validator
+        parses such strings back into lists so the retain call succeeds.
+        A plain non-JSON string is wrapped in a single-element list.
+        """
+        if isinstance(v, str):
+            try:
+                parsed = json.loads(v)
+                if isinstance(parsed, list):
+                    return parsed
+            except (json.JSONDecodeError, TypeError):
+                pass
+            return [v]
+        return v
+
     observation_scopes: Literal["per_tag", "combined", "all_combinations"] | list[list[str]] | None = Field(
         default=None,
         title="ObservationScopes",
@@ -411,6 +456,11 @@ class MemoryItem(BaseModel):
             "A list of tag lists runs one pass per inner list, giving full control over which combinations to use."
         ),
     )
+    strategy: str | None = Field(
+        default=None,
+        description="Named retain strategy for this item. Overrides the bank's default strategy for this item only. "
+        "Strategies are defined in the bank config under 'retain_strategies'.",
+    )
 
     @field_validator("timestamp", mode="before")
     @classmethod
@@ -472,6 +522,16 @@ class FileRetainMetadata(BaseModel):
     metadata: dict[str, Any] | None = Field(default=None, description="Additional metadata")
     tags: list[str] | None = Field(default=None, description="Tags for this file")
     timestamp: str | None = Field(default=None, description="ISO timestamp")
+    parser: str | list[str] | None = Field(
+        default=None,
+        description="Parser or ordered fallback chain for this file (overrides request-level parser). "
+        "E.g. 'iris' or ['iris', 'markitdown'].",
+    )
+    strategy: str | None = Field(
+        default=None,
+        description="Named retain strategy for this file. Overrides the bank's default strategy. "
+        "Strategies are defined in the bank config under 'retain_strategies'.",
+    )
 
 
 class FileRetainRequest(BaseModel):
@@ -480,14 +540,21 @@ class FileRetainRequest(BaseModel):
     model_config = ConfigDict(
         json_schema_extra={
             "example": {
+                "parser": "iris",
                 "files_metadata": [
                     {"document_id": "report_2024", "tags": ["quarterly"]},
-                    {"context": "meeting notes"},
+                    {"context": "meeting notes", "parser": ["iris", "markitdown"]},
                 ],
             }
         }
     )
 
+    parser: str | list[str] | None = Field(
+        default=None,
+        description="Default parser or ordered fallback chain for all files in this request. "
+        "E.g. 'markitdown' or ['iris', 'markitdown']. Falls back to server default if not set. "
+        "Per-file 'parser' in files_metadata takes precedence over this value.",
+    )
     files_metadata: list[FileRetainMetadata] | None = Field(
         default=None,
         description="Metadata for each file (optional, must match number of files if provided)",
@@ -518,7 +585,11 @@ class RetainResponse(BaseModel):
     )
     operation_id: str | None = Field(
         default=None,
-        description="Operation ID for tracking async operations. Use GET /v1/default/banks/{bank_id}/operations to list operations. Only present when async=true.",
+        description="Operation ID for tracking async operations. Use GET /v1/default/banks/{bank_id}/operations to list operations. Only present when async=true. When items use different per-item strategies, use operation_ids instead.",
+    )
+    operation_ids: list[str] | None = Field(
+        default=None,
+        description="Operation IDs when items were submitted as multiple strategy groups (async=true with mixed per-item strategies). operation_id is set to the first entry for backward compatibility.",
     )
     usage: TokenUsage | None = Field(
         default=None,
@@ -624,6 +695,36 @@ class ReflectRequest(BaseModel):
         description="How to match tags: 'any' (OR, includes untagged), 'all' (AND, includes untagged), "
         "'any_strict' (OR, excludes untagged), 'all_strict' (AND, excludes untagged).",
     )
+    tag_groups: list[TagGroup] | None = Field(
+        default=None,
+        description="Compound tag filter using boolean groups. Groups in the list are AND-ed. "
+        "Each group is a leaf {tags, match} or compound {and: [...]}, {or: [...]}, {not: ...}.",
+    )
+    fact_types: list[Literal["world", "experience", "observation"]] | None = Field(
+        default=None,
+        description="Filter which fact types are retrieved during reflect. None means all types (world, experience, observation).",
+    )
+    exclude_mental_models: bool = Field(
+        default=False,
+        description="If true, exclude all mental models from the reflect loop (skip search_mental_models tool).",
+    )
+    exclude_mental_model_ids: list[str] | None = Field(
+        default=None,
+        description="Exclude specific mental models by ID from the reflect loop.",
+    )
+
+    @field_validator("fact_types")
+    @classmethod
+    def validate_reflect_fact_types(cls, v: list[str] | None) -> list[str] | None:
+        if v is not None and len(v) == 0:
+            raise ValueError("fact_types must not be empty. Use null to include all fact types.")
+        return v
+
+    @model_validator(mode="after")
+    def validate_tags_exclusive(self) -> "ReflectRequest":
+        if self.tags is not None and self.tag_groups is not None:
+            raise ValueError("'tags' and 'tag_groups' are mutually exclusive. Use 'tag_groups' for compound filtering.")
+        return self
 
 
 class ReflectFact(BaseModel):
@@ -1189,6 +1290,30 @@ class DocumentResponse(BaseModel):
     tags: list[str] = FieldWithDefault(list, description="Tags associated with this document")
 
 
+class UpdateDocumentRequest(BaseModel):
+    """Request model for updating a document's mutable fields."""
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "tags": ["team-a", "team-b"],
+            }
+        }
+    )
+
+    tags: list[str] | None = Field(
+        default=None,
+        description="New tags for the document and its memory units. "
+        "Triggers observation invalidation and re-consolidation.",
+    )
+
+
+class UpdateDocumentResponse(BaseModel):
+    """Response model for update document endpoint."""
+
+    success: bool = True
+
+
 class DeleteDocumentResponse(BaseModel):
     """Response model for delete document endpoint."""
 
@@ -1253,6 +1378,14 @@ class ClearMemoryObservationsResponse(BaseModel):
     deleted_count: int
 
 
+class RecoverConsolidationResponse(BaseModel):
+    """Response model for recovering failed consolidation."""
+
+    model_config = ConfigDict(json_schema_extra={"example": {"retried_count": 42}})
+
+    retried_count: int
+
+
 class BankStatsResponse(BaseModel):
     """Response model for bank statistics endpoint."""
 
@@ -1352,6 +1485,25 @@ class MentalModelTrigger(BaseModel):
         default=False,
         description="If true, refresh this mental model after observations consolidation (real-time mode)",
     )
+    fact_types: list[Literal["world", "experience", "observation"]] | None = Field(
+        default=None,
+        description="Filter which fact types are retrieved during reflect. None means all types (world, experience, observation).",
+    )
+    exclude_mental_models: bool = Field(
+        default=False,
+        description="If true, exclude all mental models from the reflect loop (skip search_mental_models tool).",
+    )
+    exclude_mental_model_ids: list[str] | None = Field(
+        default=None,
+        description="Exclude specific mental models by ID from the reflect loop.",
+    )
+
+    @field_validator("fact_types")
+    @classmethod
+    def validate_fact_types(cls, v: list[str] | None) -> list[str] | None:
+        if v is not None and len(v) == 0:
+            raise ValueError("fact_types must not be empty. Use null to include all fact types.")
+        return v
 
 
 class MentalModelResponse(BaseModel):
@@ -1517,6 +1669,24 @@ class CancelOperationResponse(BaseModel):
     operation_id: str
 
 
+class RetryOperationResponse(BaseModel):
+    """Response model for retry operation endpoint."""
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "success": True,
+                "message": "Operation 550e8400-e29b-41d4-a716-446655440000 queued for retry",
+                "operation_id": "550e8400-e29b-41d4-a716-446655440000",
+            }
+        }
+    )
+
+    success: bool
+    message: str
+    operation_id: str
+
+
 class ChildOperationStatus(BaseModel):
     """Status of a child operation (for batch operations)."""
 
@@ -1894,6 +2064,26 @@ def create_app(
     # This is required for mounted sub-applications where lifespan may not fire
     app.state.memory = memory
 
+    # ---------------------------------------------------------------------------
+    # Patch OpenAPI schema: align ValidationError with Pydantic v2 error format
+    # ---------------------------------------------------------------------------
+    # FastAPI auto-generates ValidationError with only loc/msg/type, but Pydantic
+    # v2 actually returns additional fields: input (the rejected value), ctx (extra
+    # context dict), and url (link to error docs). Without these in the spec,
+    # generated clients using strict JSON decoding break on real 422 responses.
+    _original_openapi = app.openapi
+
+    def _patched_openapi() -> dict[str, Any]:
+        schema = _original_openapi()
+        ve = schema.get("components", {}).get("schemas", {}).get("ValidationError")
+        if ve and "input" not in ve.get("properties", {}):
+            ve["properties"]["input"] = {"title": "Input"}
+            ve["properties"]["ctx"] = {"title": "Context", "type": "object"}
+            ve["properties"]["url"] = {"title": "URL", "type": "string"}
+        return schema
+
+    app.openapi = _patched_openapi  # type: ignore[assignment]
+
     # Add HTTP metrics middleware
     @app.middleware("http")
     async def http_metrics_middleware(request, call_next):
@@ -2120,7 +2310,7 @@ def _register_routes(app: FastAPI):
     @app.get(
         "/v1/default/banks/{bank_id}/memories/{memory_id}",
         summary="Get memory unit",
-        description="Get a single memory unit by ID with all its metadata including entities and tags.",
+        description="Get a single memory unit by ID with all its metadata including entities and tags. Note: the 'history' field is deprecated and always returns an empty list - use GET /memories/{memory_id}/history instead.",
         operation_id="get_memory",
         tags=["Memory"],
     )
@@ -2150,6 +2340,39 @@ def _register_routes(app: FastAPI):
             logger.error(f"Error in /v1/default/banks/{bank_id}/memories/{memory_id}: {error_detail}")
             raise HTTPException(status_code=500, detail=str(e))
 
+    @app.get(
+        "/v1/default/banks/{bank_id}/memories/{memory_id}/history",
+        summary="Get observation history",
+        description="Get the full history of an observation, with each change's source facts resolved to their text.",
+        operation_id="get_observation_history",
+        tags=["Memory"],
+    )
+    async def api_get_observation_history(
+        bank_id: str,
+        memory_id: str,
+        request_context: RequestContext = Depends(get_request_context),
+    ):
+        """Get the history of a single observation by ID."""
+        try:
+            data = await app.state.memory.get_observation_history(
+                bank_id=bank_id,
+                memory_id=memory_id,
+                request_context=request_context,
+            )
+            if data is None:
+                raise HTTPException(status_code=404, detail=f"Memory unit '{memory_id}' not found")
+            return data
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
+        except (AuthenticationError, HTTPException):
+            raise
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(f"Error in /v1/default/banks/{bank_id}/memories/{memory_id}/history: {error_detail}")
+            raise HTTPException(status_code=500, detail=str(e))
+
     @app.post(
         "/v1/default/banks/{bank_id}/memories/recall",
         response_model=RecallResponse,
@@ -2171,12 +2394,13 @@ def _register_routes(app: FastAPI):
         metrics = get_metrics_collector()
 
         # Validate query length to prevent expensive operations on oversized queries
+        max_query_tokens = get_config().recall_max_query_tokens
         encoding = _get_tiktoken_encoding()
         query_tokens = len(encoding.encode(request.query))
-        if query_tokens > MAX_QUERY_TOKENS:
+        if query_tokens > max_query_tokens:
             raise HTTPException(
                 status_code=400,
-                detail=f"Query too long: {query_tokens} tokens exceeds maximum of {MAX_QUERY_TOKENS}. Please shorten your query.",
+                detail=f"Query too long: {query_tokens} tokens exceeds maximum of {max_query_tokens}. Please shorten your query.",
             )
 
         try:
@@ -2205,6 +2429,9 @@ def _register_routes(app: FastAPI):
             # Determine source facts inclusion settings
             include_source_facts = request.include.source_facts is not None
             max_source_facts_tokens = request.include.source_facts.max_tokens if include_source_facts else 4096
+            max_source_facts_tokens_per_observation = (
+                request.include.source_facts.max_tokens_per_observation if include_source_facts else -1
+            )
 
             pre_recall = time.time() - handler_start
             # Run recall with tracing (record metrics)
@@ -2226,9 +2453,11 @@ def _register_routes(app: FastAPI):
                     max_chunk_tokens=max_chunk_tokens,
                     include_source_facts=include_source_facts,
                     max_source_facts_tokens=max_source_facts_tokens,
+                    max_source_facts_tokens_per_observation=max_source_facts_tokens_per_observation,
                     request_context=request_context,
                     tags=request.tags,
                     tags_match=request.tags_match,
+                    tag_groups=request.tag_groups,
                 )
 
             # Convert core MemoryFact objects to API RecallResult objects (excluding internal metrics)
@@ -2364,6 +2593,10 @@ def _register_routes(app: FastAPI):
                     request_context=request_context,
                     tags=request.tags,
                     tags_match=request.tags_match,
+                    tag_groups=request.tag_groups,
+                    fact_types=request.fact_types,
+                    exclude_mental_models=request.exclude_mental_models,
+                    exclude_mental_model_ids=request.exclude_mental_model_ids,
                 )
 
             # Build based_on (memories + mental_models + directives) if facts are requested
@@ -2439,6 +2672,14 @@ def _register_routes(app: FastAPI):
             raise HTTPException(status_code=e.status_code, detail=e.reason)
         except (AuthenticationError, HTTPException):
             raise
+        except LLMNotAvailableError as e:
+            raise HTTPException(status_code=400, detail=str(e))
+        except TimeoutError as e:
+            logger.error("Timeout in /v1/default/banks/%s/reflect: %s", bank_id, e)
+            raise HTTPException(
+                status_code=504,
+                detail=str(e) or "Reflect operation timed out. Consider reducing the budget or simplifying the query.",
+            )
         except Exception as e:
             import traceback
 
@@ -2694,6 +2935,41 @@ def _register_routes(app: FastAPI):
             logger.error(f"Error in GET /v1/default/banks/{bank_id}/mental-models/{mental_model_id}: {error_detail}")
             raise HTTPException(status_code=500, detail=str(e))
 
+    @app.get(
+        "/v1/default/banks/{bank_id}/mental-models/{mental_model_id}/history",
+        summary="Get mental model history",
+        description="Get the refresh history of a mental model, showing content changes over time.",
+        operation_id="get_mental_model_history",
+        tags=["Mental Models"],
+    )
+    async def api_get_mental_model_history(
+        bank_id: str,
+        mental_model_id: str,
+        request_context: RequestContext = Depends(get_request_context),
+    ):
+        """Get the refresh history of a mental model."""
+        try:
+            data = await app.state.memory.get_mental_model_history(
+                bank_id=bank_id,
+                mental_model_id=mental_model_id,
+                request_context=request_context,
+            )
+            if data is None:
+                raise HTTPException(status_code=404, detail=f"Mental model '{mental_model_id}' not found")
+            return data
+        except (AuthenticationError, HTTPException):
+            raise
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(
+                f"Error in GET /v1/default/banks/{bank_id}/mental-models/{mental_model_id}/history: {error_detail}"
+            )
+            raise HTTPException(status_code=500, detail=str(e))
+
     @app.post(
         "/v1/default/banks/{bank_id}/mental-models",
         response_model=CreateMentalModelResponse,
@@ -2764,6 +3040,8 @@ def _register_routes(app: FastAPI):
                 request_context=request_context,
             )
             return AsyncOperationSubmitResponse(operation_id=result["operation_id"], status="queued")
+        except LLMNotAvailableError as e:
+            raise HTTPException(status_code=400, detail=str(e))
         except ValueError as e:
             raise HTTPException(status_code=404, detail=str(e))
         except (AuthenticationError, HTTPException):
@@ -3215,6 +3493,55 @@ def _register_routes(app: FastAPI):
             logger.error(f"Error in /v1/default/chunks/{chunk_id}: {error_detail}")
             raise HTTPException(status_code=500, detail=str(e))
 
+    @app.patch(
+        "/v1/default/banks/{bank_id}/documents/{document_id:path}",
+        response_model=UpdateDocumentResponse,
+        summary="Update document",
+        description="Update mutable fields on a document without re-processing its content.\n\n"
+        "**Tags** (`tags`): Propagated to all associated memory units. Observations derived from "
+        "those units are invalidated and queued for re-consolidation under the new tags. "
+        "Co-source memories from other documents that shared those observations are also reset.\n\n"
+        "At least one field must be provided.",
+        operation_id="update_document",
+        tags=["Documents"],
+    )
+    async def api_update_document(
+        bank_id: str,
+        document_id: str,
+        body: UpdateDocumentRequest,
+        request_context: RequestContext = Depends(get_request_context),
+    ):
+        """
+        Update mutable fields on a document without re-processing its content.
+
+        Args:
+            bank_id: Memory Bank ID (from path)
+            document_id: Document ID (from path)
+            body: Fields to update (tags, metadata, context)
+        """
+        if body.tags is None:
+            raise HTTPException(status_code=422, detail="At least one field (tags) must be provided")
+        try:
+            result = await app.state.memory.update_document(
+                document_id,
+                bank_id,
+                tags=body.tags,
+                request_context=request_context,
+            )
+            if not result:
+                raise HTTPException(status_code=404, detail="Document not found")
+            return UpdateDocumentResponse(success=True)
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
+        except (AuthenticationError, HTTPException):
+            raise
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(f"Error in PATCH /v1/default/banks/{bank_id}/documents/{document_id}: {error_detail}")
+            raise HTTPException(status_code=500, detail=str(e))
+
     @app.delete(
         "/v1/default/banks/{bank_id}/documents/{document_id:path}",
         response_model=DeleteDocumentResponse,
@@ -3265,13 +3592,17 @@ def _register_routes(app: FastAPI):
         "/v1/default/banks/{bank_id}/operations",
         response_model=OperationsListResponse,
         summary="List async operations",
-        description="Get a list of async operations for a specific agent, with optional filtering by status. Results are sorted by most recent first.",
+        description="Get a list of async operations for a specific agent, with optional filtering by status and operation type. Results are sorted by most recent first.",
         operation_id="list_operations",
         tags=["Operations"],
     )
     async def api_list_operations(
         bank_id: str,
         status: str | None = Query(default=None, description="Filter by status: pending, completed, or failed"),
+        type: str | None = Query(
+            default=None,
+            description="Filter by operation type: retain, consolidation, refresh_mental_model, file_convert_retain, webhook_delivery",
+        ),
         limit: int = Query(default=20, ge=1, le=100, description="Maximum number of operations to return"),
         offset: int = Query(default=0, ge=0, description="Number of operations to skip"),
         request_context: RequestContext = Depends(get_request_context),
@@ -3279,7 +3610,7 @@ def _register_routes(app: FastAPI):
         """List async operations for a memory bank with optional filtering and pagination."""
         try:
             result = await app.state.memory.list_operations(
-                bank_id, status=status, limit=limit, offset=offset, request_context=request_context
+                bank_id, status=status, task_type=type, limit=limit, offset=offset, request_context=request_context
             )
             return OperationsListResponse(
                 bank_id=bank_id,
@@ -3366,6 +3697,39 @@ def _register_routes(app: FastAPI):
             logger.error(f"Error in /v1/default/banks/{bank_id}/operations/{operation_id}: {error_detail}")
             raise HTTPException(status_code=500, detail=str(e))
 
+    @app.post(
+        "/v1/default/banks/{bank_id}/operations/{operation_id}/retry",
+        response_model=RetryOperationResponse,
+        summary="Retry a failed async operation",
+        description="Re-queue a failed async operation so the worker picks it up again",
+        operation_id="retry_operation",
+        tags=["Operations"],
+    )
+    async def api_retry_operation(
+        bank_id: str, operation_id: str, request_context: RequestContext = Depends(get_request_context)
+    ):
+        """Retry a failed async operation."""
+        try:
+            try:
+                uuid.UUID(operation_id)
+            except ValueError:
+                raise HTTPException(status_code=400, detail=f"Invalid operation_id format: {operation_id}")
+
+            result = await app.state.memory.retry_operation(bank_id, operation_id, request_context=request_context)
+            return RetryOperationResponse(**result)
+        except ValueError as e:
+            raise HTTPException(status_code=404, detail=str(e))
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
+        except (AuthenticationError, HTTPException):
+            raise
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(f"Error in POST /v1/default/banks/{bank_id}/operations/{operation_id}/retry: {error_detail}")
+            raise HTTPException(status_code=500, detail=str(e))
+
     @app.get(
         "/v1/default/banks/{bank_id}/profile",
         response_model=BankProfileResponse,
@@ -3648,6 +4012,34 @@ def _register_routes(app: FastAPI):
             logger.error(f"Error in DELETE /v1/default/banks/{bank_id}/observations: {error_detail}")
             raise HTTPException(status_code=500, detail=str(e))
 
+    @app.post(
+        "/v1/default/banks/{bank_id}/consolidation/recover",
+        response_model=RecoverConsolidationResponse,
+        summary="Recover failed consolidation",
+        description=(
+            "Reset all memories that were permanently marked as failed during consolidation "
+            "(after exhausting all LLM retries and adaptive batch splitting) so they are "
+            "picked up again on the next consolidation run. Does not delete any observations."
+        ),
+        operation_id="recover_consolidation",
+        tags=["Banks"],
+    )
+    async def api_recover_consolidation(bank_id: str, request_context: RequestContext = Depends(get_request_context)):
+        """Reset consolidation-failed memories for recovery."""
+        try:
+            result = await app.state.memory.retry_failed_consolidation(bank_id, request_context=request_context)
+            return RecoverConsolidationResponse(retried_count=result["retried_count"])
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
+        except (AuthenticationError, HTTPException):
+            raise
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(f"Error in POST /v1/default/banks/{bank_id}/consolidation/recover: {error_detail}")
+            raise HTTPException(status_code=500, detail=str(e))
+
     @app.delete(
         "/v1/default/banks/{bank_id}/memories/{memory_id}/observations",
         response_model=ClearMemoryObservationsResponse,
@@ -3875,9 +4267,13 @@ def _register_routes(app: FastAPI):
         try:
             pool = await app.state.memory._get_pool()
             from hindsight_api.engine.memory_engine import fq_table
+            from hindsight_api.engine.retain import bank_utils
+
+            # Ensure the bank row exists before inserting into webhooks (FK constraint).
+            await bank_utils.get_bank_profile(pool, bank_id)
 
             webhook_id = uuid.uuid4()
-            now = datetime.utcnow().isoformat() + "Z"
+            now = datetime.now(timezone.utc).isoformat()
             row = await pool.fetchrow(
                 f"""
                 INSERT INTO {fq_table("webhooks")}
@@ -4197,10 +4593,13 @@ def _register_routes(app: FastAPI):
         metrics = get_metrics_collector()
 
         try:
-            # Prepare contents for processing
-            contents = []
+            # Group items by strategy
+            strategy_groups: dict[str | None, list[dict]] = {}
             for item in request.items:
-                content_dict = {"content": item.content}
+                effective = item.strategy
+                if effective not in strategy_groups:
+                    strategy_groups[effective] = []
+                content_dict: dict = {"content": item.content}
                 if item.timestamp == "unset":
                     content_dict["event_date"] = None
                 elif item.timestamp:
@@ -4217,20 +4616,30 @@ def _register_routes(app: FastAPI):
                     content_dict["tags"] = item.tags
                 if item.observation_scopes is not None:
                     content_dict["observation_scopes"] = item.observation_scopes
-                contents.append(content_dict)
+                strategy_groups[effective].append(content_dict)
 
             if request.async_:
-                # Async processing: queue task and return immediately
-                result = await app.state.memory.submit_async_retain(
-                    bank_id, contents, document_tags=request.document_tags, request_context=request_context
-                )
+                # Async processing: one submit per strategy group
+                all_operation_ids = []
+                total_items_count = 0
+                for group_strategy, contents in strategy_groups.items():
+                    result = await app.state.memory.submit_async_retain(
+                        bank_id,
+                        contents,
+                        document_tags=request.document_tags,
+                        strategy=group_strategy,
+                        request_context=request_context,
+                    )
+                    all_operation_ids.append(result["operation_id"])
+                    total_items_count += result["items_count"]
                 return RetainResponse.model_validate(
                     {
                         "success": True,
                         "bank_id": bank_id,
-                        "items_count": result["items_count"],
+                        "items_count": total_items_count,
                         "async": True,
-                        "operation_id": result["operation_id"],
+                        "operation_id": all_operation_ids[0] if all_operation_ids else None,
+                        "operation_ids": all_operation_ids if len(all_operation_ids) > 1 else None,
                     }
                 )
             else:
@@ -4249,24 +4658,41 @@ def _register_routes(app: FastAPI):
                         ),
                     )
 
-                # Synchronous processing: wait for completion (record metrics)
+                # Synchronous processing: one batch per strategy group, aggregate results
+                total_items_count = 0
+                total_usage = TokenUsage(input_tokens=0, output_tokens=0, total_tokens=0)
                 with metrics.record_operation("retain", bank_id=bank_id, source="api"):
-                    result, usage = await app.state.memory.retain_batch_async(
-                        bank_id=bank_id,
-                        contents=contents,
-                        document_tags=request.document_tags,
-                        request_context=request_context,
-                        return_usage=True,
-                        outbox_callback=app.state.memory._build_retain_outbox_callback(
+                    for group_strategy, contents in strategy_groups.items():
+                        result, usage = await app.state.memory.retain_batch_async(
                             bank_id=bank_id,
                             contents=contents,
-                            operation_id=None,
-                            schema=_current_schema.get(),
-                        ),
-                    )
+                            document_tags=request.document_tags,
+                            strategy=group_strategy,
+                            request_context=request_context,
+                            return_usage=True,
+                            outbox_callback=app.state.memory._build_retain_outbox_callback(
+                                bank_id=bank_id,
+                                contents=contents,
+                                operation_id=None,
+                                schema=_current_schema.get(),
+                            ),
+                        )
+                        total_items_count += len(contents)
+                        if usage:
+                            total_usage = TokenUsage(
+                                input_tokens=total_usage.input_tokens + usage.input_tokens,
+                                output_tokens=total_usage.output_tokens + usage.output_tokens,
+                                total_tokens=total_usage.total_tokens + usage.total_tokens,
+                            )
 
                 return RetainResponse.model_validate(
-                    {"success": True, "bank_id": bank_id, "items_count": len(contents), "async": False, "usage": usage}
+                    {
+                        "success": True,
+                        "bank_id": bank_id,
+                        "items_count": total_items_count,
+                        "async": False,
+                        "usage": total_usage,
+                    }
                 )
         except OperationValidationError as e:
             raise HTTPException(status_code=e.status_code, detail=e.reason)
@@ -4312,8 +4738,14 @@ def _register_routes(app: FastAPI):
         "Use the operations endpoint to monitor progress.\n\n"
         "**Request format:** multipart/form-data with:\n"
         "- `files`: One or more files to upload\n"
-        "- `request`: JSON string with FileRetainRequest model (files_metadata)\n\n"
-        "**Note:** File parser is configured server-side via `HINDSIGHT_API_FILE_PARSER` (default: markitdown).",
+        "- `request`: JSON string with FileRetainRequest model\n\n"
+        "**Parser selection:**\n"
+        "- Set `parser` in the request body to override the server default for all files.\n"
+        "- Set `parser` inside a `files_metadata` entry for per-file control.\n"
+        "- Pass a list (e.g. `['iris', 'markitdown']`) to define an ordered fallback chain — "
+        "each parser is tried in sequence until one succeeds.\n"
+        "- Falls back to the server default (`HINDSIGHT_API_FILE_PARSER`) if not specified.\n"
+        "- Only parsers enabled on the server may be requested; others return HTTP 400.",
         operation_id="file_retain",
         tags=["Files"],
     )
@@ -4359,20 +4791,39 @@ def _register_routes(app: FastAPI):
                     detail=f"files_metadata count ({len(request_data.files_metadata)}) must match files count ({len(files)})",
                 )
 
+            # Resolve the registered parser names for allowlist validation
+            registered_parsers = app.state.memory._parser_registry.list_parsers()
+            allowlist = config.file_parser_allowlist if config.file_parser_allowlist is not None else registered_parsers
+
+            def _resolve_parser(raw: str | list[str] | None) -> list[str]:
+                """Normalize parser value to a non-empty list of names."""
+                if raw is None:
+                    return config.file_parser
+                return [raw] if isinstance(raw, str) else list(raw)
+
+            def _validate_parsers(parsers: list[str], context: str) -> None:
+                """Raise HTTP 400 if any parser name is not in the allowlist."""
+                disallowed = [p for p in parsers if p not in allowlist]
+                if disallowed:
+                    raise HTTPException(
+                        status_code=400,
+                        detail=f"Parser(s) not available ({context}): {disallowed}. Available: {allowlist}",
+                    )
+
+            # Validate request-level parser early (before reading files)
+            if request_data.parser is not None:
+                _validate_parsers(_resolve_parser(request_data.parser), "request-level parser")
+
             # Prepare file items and calculate total batch size
+            import io
+
             file_items = []
             total_batch_size = 0
 
             for i, file in enumerate(files):
                 # Read file content to check size
                 file_content = await file.read()
-                size = len(file_content)
-                total_batch_size += size
-
-                # Create a temporary file-like object from the bytes
-                import io
-
-                file_obj = io.BytesIO(file_content)
+                total_batch_size += len(file_content)
 
                 # Create a mock UploadFile with the necessary attributes
                 class FileWrapper:
@@ -4380,7 +4831,6 @@ def _register_routes(app: FastAPI):
                         self._content = content
                         self.filename = filename
                         self.content_type = content_type
-                        self._buffer = io.BytesIO(content)
 
                     async def read(self):
                         return self._content
@@ -4391,6 +4841,12 @@ def _register_routes(app: FastAPI):
                 file_meta = request_data.files_metadata[i] if request_data.files_metadata else FileRetainMetadata()
                 doc_id = file_meta.document_id or f"file_{uuid.uuid4()}"
 
+                # Resolve and validate per-file parser chain
+                # Priority: per-file > request-level > server default
+                raw_parser = file_meta.parser if file_meta.parser is not None else request_data.parser
+                parser_chain = _resolve_parser(raw_parser)
+                _validate_parsers(parser_chain, f"file '{file.filename}'")
+
                 item = {
                     "file": wrapped_file,
                     "document_id": doc_id,
@@ -4398,6 +4854,8 @@ def _register_routes(app: FastAPI):
                     "metadata": file_meta.metadata or {},
                     "tags": file_meta.tags or [],
                     "timestamp": file_meta.timestamp,
+                    "parser": parser_chain,
+                    "strategy": file_meta.strategy,
                 }
                 file_items.append(item)
 
@@ -4412,7 +4870,6 @@ def _register_routes(app: FastAPI):
             result = await app.state.memory.submit_async_file_retain(
                 bank_id=bank_id,
                 file_items=file_items,
-                parser=config.file_parser,
                 document_tags=None,
                 request_context=request_context,
             )

hindsight-api-slim/hindsight_api/api/mcp.py

@@ -83,6 +83,9 @@ _current_api_key: ContextVar[str | None] = ContextVar("current_api_key", default
 _current_tenant_id: ContextVar[str | None] = ContextVar("current_tenant_id", default=None)
 _current_api_key_id: ContextVar[str | None] = ContextVar("current_api_key_id", default=None)
 
+# Context variable for MCP pre-authentication flag (set when MCP_AUTH_TOKEN validates)
+_current_mcp_authenticated: ContextVar[bool] = ContextVar("current_mcp_authenticated", default=False)
+
 
 def get_current_bank_id() -> str | None:
     """Get the current bank_id from context."""
@@ -104,6 +107,11 @@ def get_current_api_key_id() -> str | None:
     return _current_api_key_id.get()
 
 
+def get_current_mcp_authenticated() -> bool:
+    """Get whether the request was pre-authenticated by MCP transport auth."""
+    return _current_mcp_authenticated.get()
+
+
 def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
     """
     Create and configure the Hindsight MCP server.
@@ -164,6 +172,7 @@ def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
         api_key_resolver=get_current_api_key,  # Propagate API key for tenant auth
         tenant_id_resolver=get_current_tenant_id,  # Propagate tenant_id for usage metering
         api_key_id_resolver=get_current_api_key_id,  # Propagate api_key_id for usage metering
+        mcp_authenticated_resolver=get_current_mcp_authenticated,  # Propagate MCP pre-auth flag
         include_bank_id_param=multi_bank,
         tools=base_tools,
     )
@@ -312,6 +321,7 @@ class MCPMiddleware:
         tenant_context = None
         auth_tenant_id: str | None = None
         auth_api_key_id: str | None = None
+        mcp_pre_authenticated = False
         if MCP_AUTH_TOKEN:
             # Legacy authentication mode - validate against static token
             if not auth_token:
@@ -320,8 +330,9 @@ class MCPMiddleware:
             if auth_token != MCP_AUTH_TOKEN:
                 await self._send_error(send, 401, "Invalid authentication token")
                 return
-            # Legacy mode doesn't use tenant schemas
+            # Legacy mode: mark as pre-authenticated so tenant extension won't re-validate
             tenant_context = None
+            mcp_pre_authenticated = True
         else:
             # Use TenantExtension.authenticate_mcp() for auth
             try:
@@ -368,19 +379,30 @@ class MCPMiddleware:
         # - Header/env bank_id → multi-bank app (bank_id param, all tools)
         target_app = self.single_bank_app if bank_id_from_path else self.multi_bank_app
 
-        # Set bank_id, api_key, tenant_id, and api_key_id context
+        # Set bank_id, api_key, tenant_id, api_key_id, and mcp_authenticated context
         bank_id_token = _current_bank_id.set(bank_id)
         # Store the auth token for tenant extension to validate
         api_key_token = _current_api_key.set(auth_token) if auth_token else None
         # Store tenant_id and api_key_id from authentication for usage metering
         tenant_id_token = _current_tenant_id.set(auth_tenant_id) if auth_tenant_id else None
         api_key_id_token = _current_api_key_id.set(auth_api_key_id) if auth_api_key_id else None
+        # Store MCP pre-authentication flag to skip tenant re-validation
+        mcp_auth_token = _current_mcp_authenticated.set(mcp_pre_authenticated)
         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"] = ""
 
+            # Ensure Accept header includes required MIME types for MCP SDK.
+            # Some clients (e.g., Claude Code) don't send Accept, causing
+            # the SDK to reject with 406 Not Acceptable.
+            accept_header = self._get_header(new_scope, "accept")
+            if not accept_header or "text/event-stream" not in accept_header:
+                headers = [(k, v) for k, v in new_scope.get("headers", []) if k.lower() != b"accept"]
+                headers.append((b"accept", b"application/json, text/event-stream"))
+                new_scope["headers"] = headers
+
             # Wrap send to rewrite the SSE endpoint URL to include bank_id if using path-based routing.
             # Only rewrite SSE (text/event-stream) responses to avoid corrupting tool results
             # that might contain the literal string "data: /messages".
@@ -410,6 +432,7 @@ class MCPMiddleware:
                 _current_tenant_id.reset(tenant_id_token)
             if api_key_id_token is not None:
                 _current_api_key_id.reset(api_key_id_token)
+            _current_mcp_authenticated.reset(mcp_auth_token)
             if schema_token is not None:
                 _current_schema.reset(schema_token)
 

hindsight-api-slim/hindsight_api/config.py

@@ -193,6 +193,7 @@ ENV_EMBEDDINGS_LITELLM_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_MODEL"
 ENV_RERANKER_LITELLM_API_BASE = "HINDSIGHT_API_RERANKER_LITELLM_API_BASE"
 ENV_RERANKER_LITELLM_API_KEY = "HINDSIGHT_API_RERANKER_LITELLM_API_KEY"
 ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"
+ENV_RERANKER_LITELLM_MAX_TOKENS_PER_DOC = "HINDSIGHT_API_RERANKER_LITELLM_MAX_TOKENS_PER_DOC"
 
 # LiteLLM SDK configuration (direct API access, no proxy needed)
 ENV_EMBEDDINGS_LITELLM_SDK_API_KEY = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_API_KEY"
@@ -211,6 +212,9 @@ ENV_RERANKER_LOCAL_MODEL = "HINDSIGHT_API_RERANKER_LOCAL_MODEL"
 ENV_RERANKER_LOCAL_FORCE_CPU = "HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU"
 ENV_RERANKER_LOCAL_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT"
 ENV_RERANKER_LOCAL_TRUST_REMOTE_CODE = "HINDSIGHT_API_RERANKER_LOCAL_TRUST_REMOTE_CODE"
+ENV_RERANKER_LOCAL_FP16 = "HINDSIGHT_API_RERANKER_LOCAL_FP16"
+ENV_RERANKER_LOCAL_BUCKET_BATCHING = "HINDSIGHT_API_RERANKER_LOCAL_BUCKET_BATCHING"
+ENV_RERANKER_LOCAL_BATCH_SIZE = "HINDSIGHT_API_RERANKER_LOCAL_BATCH_SIZE"
 ENV_RERANKER_TEI_URL = "HINDSIGHT_API_RERANKER_TEI_URL"
 ENV_RERANKER_TEI_BATCH_SIZE = "HINDSIGHT_API_RERANKER_TEI_BATCH_SIZE"
 ENV_RERANKER_TEI_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_TEI_MAX_CONCURRENT"
@@ -238,6 +242,7 @@ ENV_GRAPH_RETRIEVER = "HINDSIGHT_API_GRAPH_RETRIEVER"
 ENV_MPFP_TOP_K_NEIGHBORS = "HINDSIGHT_API_MPFP_TOP_K_NEIGHBORS"
 ENV_RECALL_MAX_CONCURRENT = "HINDSIGHT_API_RECALL_MAX_CONCURRENT"
 ENV_RECALL_CONNECTION_BUDGET = "HINDSIGHT_API_RECALL_CONNECTION_BUDGET"
+ENV_RECALL_MAX_QUERY_TOKENS = "HINDSIGHT_API_RECALL_MAX_QUERY_TOKENS"
 ENV_MENTAL_MODEL_REFRESH_CONCURRENCY = "HINDSIGHT_API_MENTAL_MODEL_REFRESH_CONCURRENCY"
 
 # OpenTelemetry tracing configuration
@@ -262,6 +267,7 @@ ENV_RETAIN_EXTRACT_CAUSAL_LINKS = "HINDSIGHT_API_RETAIN_EXTRACT_CAUSAL_LINKS"
 ENV_RETAIN_EXTRACTION_MODE = "HINDSIGHT_API_RETAIN_EXTRACTION_MODE"
 ENV_RETAIN_MISSION = "HINDSIGHT_API_RETAIN_MISSION"
 ENV_RETAIN_CUSTOM_INSTRUCTIONS = "HINDSIGHT_API_RETAIN_CUSTOM_INSTRUCTIONS"
+ENV_RETAIN_DEFAULT_STRATEGY = "HINDSIGHT_API_RETAIN_DEFAULT_STRATEGY"
 ENV_RETAIN_BATCH_TOKENS = "HINDSIGHT_API_RETAIN_BATCH_TOKENS"
 ENV_RETAIN_ENTITY_LOOKUP = "HINDSIGHT_API_RETAIN_ENTITY_LOOKUP"
 ENV_RETAIN_BATCH_ENABLED = "HINDSIGHT_API_RETAIN_BATCH_ENABLED"
@@ -280,6 +286,7 @@ ENV_FILE_STORAGE_AZURE_CONTAINER = "HINDSIGHT_API_FILE_STORAGE_AZURE_CONTAINER"
 ENV_FILE_STORAGE_AZURE_ACCOUNT_NAME = "HINDSIGHT_API_FILE_STORAGE_AZURE_ACCOUNT_NAME"
 ENV_FILE_STORAGE_AZURE_ACCOUNT_KEY = "HINDSIGHT_API_FILE_STORAGE_AZURE_ACCOUNT_KEY"
 ENV_FILE_PARSER = "HINDSIGHT_API_FILE_PARSER"
+ENV_FILE_PARSER_ALLOWLIST = "HINDSIGHT_API_FILE_PARSER_ALLOWLIST"
 ENV_FILE_PARSER_IRIS_TOKEN = "HINDSIGHT_API_FILE_PARSER_IRIS_TOKEN"
 ENV_FILE_PARSER_IRIS_ORG_ID = "HINDSIGHT_API_FILE_PARSER_IRIS_ORG_ID"
 ENV_FILE_CONVERSION_MAX_BATCH_SIZE_MB = "HINDSIGHT_API_FILE_CONVERSION_MAX_BATCH_SIZE_MB"
@@ -292,7 +299,13 @@ ENV_ENABLE_OBSERVATIONS = "HINDSIGHT_API_ENABLE_OBSERVATIONS"
 ENV_CONSOLIDATION_BATCH_SIZE = "HINDSIGHT_API_CONSOLIDATION_BATCH_SIZE"
 ENV_CONSOLIDATION_LLM_BATCH_SIZE = "HINDSIGHT_API_CONSOLIDATION_LLM_BATCH_SIZE"
 ENV_CONSOLIDATION_MAX_TOKENS = "HINDSIGHT_API_CONSOLIDATION_MAX_TOKENS"
+ENV_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS = "HINDSIGHT_API_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS"
+ENV_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION = (
+    "HINDSIGHT_API_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION"
+)
 ENV_OBSERVATIONS_MISSION = "HINDSIGHT_API_OBSERVATIONS_MISSION"
+ENV_ENABLE_OBSERVATION_HISTORY = "HINDSIGHT_API_ENABLE_OBSERVATION_HISTORY"
+ENV_ENABLE_MENTAL_MODEL_HISTORY = "HINDSIGHT_API_ENABLE_MENTAL_MODEL_HISTORY"
 
 # Webhook configuration (global, static - server-level only)
 ENV_WEBHOOK_URL = "HINDSIGHT_API_WEBHOOK_URL"
@@ -325,7 +338,9 @@ ENV_WORKER_CONSOLIDATION_MAX_SLOTS = "HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLO
 # Reflect agent settings
 ENV_REFLECT_MAX_ITERATIONS = "HINDSIGHT_API_REFLECT_MAX_ITERATIONS"
 ENV_REFLECT_MAX_CONTEXT_TOKENS = "HINDSIGHT_API_REFLECT_MAX_CONTEXT_TOKENS"
+ENV_REFLECT_WALL_TIMEOUT = "HINDSIGHT_API_REFLECT_WALL_TIMEOUT"
 ENV_REFLECT_MISSION = "HINDSIGHT_API_REFLECT_MISSION"
+ENV_REFLECT_SOURCE_FACTS_MAX_TOKENS = "HINDSIGHT_API_REFLECT_SOURCE_FACTS_MAX_TOKENS"
 
 # Disposition settings
 ENV_DISPOSITION_SKEPTICISM = "HINDSIGHT_API_DISPOSITION_SKEPTICISM"
@@ -340,15 +355,19 @@ DEFAULT_LLM_PROVIDER = "openai"
 # Provider-specific default models
 PROVIDER_DEFAULT_MODELS = {
     "openai": "gpt-4o-mini",
-    "anthropic": "claude-haiku-4-5-20251001",
+    "anthropic": "claude-haiku-4-5",
     "gemini": "gemini-2.5-flash",
     "groq": "openai/gpt-oss-120b",
+    "minimax": "MiniMax-M2.7",
     "ollama": "gemma3:12b",
     "lmstudio": "local-model",
     "vertexai": "google/gemini-2.5-flash-lite",
     "openai-codex": "gpt-5.2-codex",
     "claude-code": "claude-sonnet-4-5-20250929",
     "mock": "mock-model",
+    "none": "none",
+    "litellm": "gpt-4o-mini",
+    "bedrock": "us.amazon.nova-2-lite-v1:0",
 }
 DEFAULT_LLM_MODEL = "gpt-4o-mini"  # Fallback if provider not in table
 DEFAULT_LLM_MAX_CONCURRENT = 32
@@ -379,6 +398,9 @@ DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4  # Limit concurrent CPU-bound rerankin
 DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE = (
     False  # Security: disabled by default, required for some models like jina-reranker-v2
 )
+DEFAULT_RERANKER_LOCAL_FP16 = False  # FP16 inference: opt-in, faster on MPS/CUDA (not CPU)
+DEFAULT_RERANKER_LOCAL_BUCKET_BATCHING = False  # Length-sorted bucket batching: opt-in, 36-54% speedup
+DEFAULT_RERANKER_LOCAL_BATCH_SIZE = 32  # Batch size for local reranker predict() calls
 DEFAULT_RERANKER_TEI_BATCH_SIZE = 128
 DEFAULT_RERANKER_TEI_MAX_CONCURRENT = 8
 DEFAULT_RERANKER_MAX_CANDIDATES = 300
@@ -400,6 +422,7 @@ DEFAULT_TEXT_SEARCH_EXTENSION = "native"  # Options: "native", "vchord", "pg_tex
 DEFAULT_LITELLM_API_BASE = "http://localhost:4000"
 DEFAULT_EMBEDDINGS_LITELLM_MODEL = "text-embedding-3-small"
 DEFAULT_RERANKER_LITELLM_MODEL = "cohere/rerank-english-v3.0"
+DEFAULT_RERANKER_LITELLM_MAX_TOKENS_PER_DOC: int | None = None
 
 # LiteLLM SDK defaults
 DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL = "cohere/embed-english-v3.0"
@@ -418,6 +441,7 @@ DEFAULT_GRAPH_RETRIEVER = "link_expansion"  # Options: "link_expansion", "mpfp",
 DEFAULT_MPFP_TOP_K_NEIGHBORS = 20  # Fan-out limit per node in MPFP graph traversal
 DEFAULT_RECALL_MAX_CONCURRENT = 32  # Max concurrent recall operations per worker
 DEFAULT_RECALL_CONNECTION_BUDGET = 4  # Max concurrent DB connections per recall operation
+DEFAULT_RECALL_MAX_QUERY_TOKENS = 500  # Maximum tokens allowed in recall query
 DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY = 8  # Max concurrent mental model refreshes
 
 # Retain settings
@@ -425,9 +449,11 @@ DEFAULT_RETAIN_MAX_COMPLETION_TOKENS = 64000  # Max tokens for fact extraction L
 DEFAULT_RETAIN_CHUNK_SIZE = 3000  # Max chars per chunk for fact extraction
 DEFAULT_RETAIN_EXTRACT_CAUSAL_LINKS = True  # Extract causal links between facts
 DEFAULT_RETAIN_EXTRACTION_MODE = "concise"  # Extraction mode: "concise", "verbose", or "custom"
-RETAIN_EXTRACTION_MODES = ("concise", "verbose", "custom")  # Allowed extraction modes
+RETAIN_EXTRACTION_MODES = ("concise", "verbose", "custom", "verbatim", "chunks")  # Allowed extraction modes
 DEFAULT_RETAIN_MISSION = None  # Declarative spec of what to retain (injected into any extraction mode)
 DEFAULT_RETAIN_CUSTOM_INSTRUCTIONS = None  # Custom extraction guidelines (only used when mode="custom")
+DEFAULT_RETAIN_DEFAULT_STRATEGY = None  # Default strategy name (None = no strategy override)
+DEFAULT_RETAIN_STRATEGIES: dict | None = None  # Named retain strategies (dict of name → config overrides)
 DEFAULT_RETAIN_BATCH_TOKENS = 10_000  # ~40KB of text  # Max chars per sub-batch for async retain auto-splitting
 DEFAULT_RETAIN_ENTITY_LOOKUP = "trigram"  # "full" or "trigram"
 DEFAULT_RETAIN_BATCH_ENABLED = False  # Use LLM Batch API for fact extraction (only when async=True)
@@ -435,7 +461,8 @@ DEFAULT_RETAIN_BATCH_POLL_INTERVAL_SECONDS = 60  # Batch API polling interval in
 
 # File storage defaults
 DEFAULT_FILE_STORAGE_TYPE = "native"  # PostgreSQL BYTEA storage
-DEFAULT_FILE_PARSER = "markitdown"  # File parser to use (markitdown is the only supported parser)
+DEFAULT_FILE_PARSER = "markitdown"  # Default parser fallback chain (comma-separated, e.g. "iris,markitdown")
+DEFAULT_FILE_PARSER_ALLOWLIST = None  # Allowlist of parsers clients may request (None = all registered parsers)
 DEFAULT_FILE_CONVERSION_MAX_BATCH_SIZE_MB = 100  # Max total batch size in MB (all files combined)
 DEFAULT_FILE_CONVERSION_MAX_BATCH_SIZE = 10  # Max files per batch upload
 DEFAULT_ENABLE_FILE_UPLOAD_API = True  # Enable file upload endpoint
@@ -443,9 +470,17 @@ DEFAULT_FILE_DELETE_AFTER_RETAIN = True  # Delete file bytes after retain (saves
 
 # Observations defaults (consolidated knowledge from facts)
 DEFAULT_ENABLE_OBSERVATIONS = True  # Observations enabled by default
+DEFAULT_ENABLE_OBSERVATION_HISTORY = True  # Observation history tracking enabled by default
+DEFAULT_ENABLE_MENTAL_MODEL_HISTORY = True  # Mental model history tracking enabled by default
 DEFAULT_CONSOLIDATION_BATCH_SIZE = 50  # Memories to load per batch (internal memory optimization)
 DEFAULT_CONSOLIDATION_LLM_BATCH_SIZE = 8  # Facts per LLM call (1 = no batching; >1 = batch mode)
 DEFAULT_CONSOLIDATION_MAX_TOKENS = 512  # Max tokens for recall when finding related observations
+DEFAULT_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS = (
+    -1
+)  # Total token budget for source facts in consolidation recall (-1 = unlimited)
+DEFAULT_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION = (
+    256  # Max tokens of source facts per observation in consolidation prompt (-1 = unlimited)
+)
 DEFAULT_OBSERVATIONS_MISSION = None  # Declarative spec of what observations are for this bank
 
 # Database migrations
@@ -469,6 +504,8 @@ DEFAULT_WORKER_CONSOLIDATION_MAX_SLOTS = 2  # Max concurrent consolidation tasks
 # Reflect agent settings
 DEFAULT_REFLECT_MAX_ITERATIONS = 10  # Max tool call iterations before forcing response
 DEFAULT_REFLECT_MAX_CONTEXT_TOKENS = 100_000  # Max accumulated context tokens before forcing final prompt
+DEFAULT_REFLECT_WALL_TIMEOUT = 300  # Wall-clock timeout in seconds for the entire reflect operation (5 minutes)
+DEFAULT_REFLECT_SOURCE_FACTS_MAX_TOKENS = -1  # Token budget for source facts in search_observations (-1 = disabled)
 
 # Disposition defaults (None = not set, fall back to bank DB value or 3)
 DEFAULT_DISPOSITION_SKEPTICISM = None
@@ -540,6 +577,11 @@ class JsonFormatter(logging.Formatter):
         return json.dumps(log_entry)
 
 
+def _parse_str_list(value: str) -> list[str]:
+    """Parse a comma-separated string into a non-empty list of stripped tokens."""
+    return [v.strip() for v in value.split(",") if v.strip()]
+
+
 def _validate_extraction_mode(mode: str) -> str:
     """Validate and normalize extraction mode."""
     mode_lower = mode.lower()
@@ -642,6 +684,9 @@ class HindsightConfig:
     reranker_local_force_cpu: bool
     reranker_local_max_concurrent: int
     reranker_local_trust_remote_code: bool
+    reranker_local_fp16: bool
+    reranker_local_bucket_batching: bool
+    reranker_local_batch_size: int
     reranker_tei_url: str | None
     reranker_tei_batch_size: int
     reranker_tei_max_concurrent: int
@@ -652,6 +697,7 @@ class HindsightConfig:
     reranker_litellm_api_base: str
     reranker_litellm_api_key: str | None
     reranker_litellm_model: str
+    reranker_litellm_max_tokens_per_doc: int | None
     reranker_litellm_sdk_api_key: str | None
     reranker_litellm_sdk_model: str
     reranker_litellm_sdk_api_base: str | None
@@ -673,6 +719,7 @@ class HindsightConfig:
     mpfp_top_k_neighbors: int
     recall_max_concurrent: int
     recall_connection_budget: int
+    recall_max_query_tokens: int
     mental_model_refresh_concurrency: int
 
     # Retain settings
@@ -682,6 +729,8 @@ class HindsightConfig:
     retain_extraction_mode: str
     retain_mission: str | None
     retain_custom_instructions: str | None
+    retain_default_strategy: str | None
+    retain_strategies: dict | None
     retain_batch_tokens: int
     retain_batch_enabled: bool
     retain_batch_poll_interval_seconds: int
@@ -699,7 +748,8 @@ class HindsightConfig:
     file_storage_azure_container: str | None  # Azure container name (required for azure storage)
     file_storage_azure_account_name: str | None  # Azure storage account name
     file_storage_azure_account_key: str | None  # Azure storage account key
-    file_parser: str  # File parser to use (e.g., "markitdown", "iris")
+    file_parser: list[str]  # Ordered fallback chain of parsers (e.g. ["iris", "markitdown"])
+    file_parser_allowlist: list[str] | None  # Parsers clients may request (None = all registered)
     file_parser_iris_token: str | None  # Vectorize API token for iris parser (VECTORIZE_TOKEN)
     file_parser_iris_org_id: str | None  # Vectorize org ID for iris parser (VECTORIZE_ORG_ID)
     file_conversion_max_batch_size_mb: int  # Max total batch size in MB (all files combined)
@@ -709,9 +759,13 @@ class HindsightConfig:
 
     # Observations settings (consolidated knowledge from facts)
     enable_observations: bool
+    enable_observation_history: bool
+    enable_mental_model_history: bool
     consolidation_batch_size: int
     consolidation_llm_batch_size: int
     consolidation_max_tokens: int
+    consolidation_source_facts_max_tokens: int
+    consolidation_source_facts_max_tokens_per_observation: int
     observations_mission: str | None
 
     # Entity labels (controlled vocabulary of key:value classification labels extracted at retain time)
@@ -723,6 +777,7 @@ class HindsightConfig:
 
     # Reflect agent settings
     reflect_mission: str | None
+    reflect_source_facts_max_tokens: int
 
     # Disposition settings (hierarchical - can be overridden per bank; None = fall back to DB)
     disposition_skepticism: int | None
@@ -754,6 +809,7 @@ class HindsightConfig:
     # Reflect agent settings
     reflect_max_iterations: int
     reflect_max_context_tokens: int
+    reflect_wall_timeout: int
 
     # OpenTelemetry tracing configuration
     otel_traces_enabled: bool
@@ -807,14 +863,20 @@ class HindsightConfig:
         "retain_extraction_mode",
         "retain_mission",
         "retain_custom_instructions",
+        "retain_default_strategy",
+        "retain_strategies",
         # Entity labels (controlled vocabulary for entity classification)
         "entity_labels",
         "entities_allow_free_form",
         # Consolidation settings
         "enable_observations",
+        "consolidation_llm_batch_size",
+        "consolidation_source_facts_max_tokens",
+        "consolidation_source_facts_max_tokens_per_observation",
         "observations_mission",
         # Reflect settings
         "reflect_mission",
+        "reflect_source_facts_max_tokens",
         # Disposition settings
         "disposition_skepticism",
         "disposition_literalism",
@@ -897,9 +959,19 @@ class HindsightConfig:
                 f"Invalid text_search_extension: {self.text_search_extension}. Must be one of: {', '.join(valid_text_search)}"
             )
 
+        # When LLM provider is "none", force chunks-only mode and disable LLM-dependent features
+        if self.llm_provider == "none":
+            self.retain_extraction_mode = "chunks"
+            self.enable_observations = False
+            logger.info(
+                "LLM provider set to 'none': forcing retain_extraction_mode='chunks', "
+                "disabling observations/consolidation. Reflect will return HTTP 400."
+            )
+
         # RETAIN_MAX_COMPLETION_TOKENS must be greater than RETAIN_CHUNK_SIZE
         # to ensure the LLM has enough output capacity to extract facts from chunks
-        if self.retain_max_completion_tokens <= self.retain_chunk_size:
+        # (not applicable when provider is "none" since no LLM calls are made)
+        if self.llm_provider != "none" and self.retain_max_completion_tokens <= self.retain_chunk_size:
             raise ValueError(
                 f"Invalid configuration: HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS "
                 f"({self.retain_max_completion_tokens}) must be greater than "
@@ -1056,6 +1128,15 @@ class HindsightConfig:
                 ENV_RERANKER_LOCAL_TRUST_REMOTE_CODE, str(DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE)
             ).lower()
             in ("true", "1"),
+            reranker_local_fp16=os.getenv(ENV_RERANKER_LOCAL_FP16, str(DEFAULT_RERANKER_LOCAL_FP16)).lower()
+            in ("true", "1"),
+            reranker_local_bucket_batching=os.getenv(
+                ENV_RERANKER_LOCAL_BUCKET_BATCHING, str(DEFAULT_RERANKER_LOCAL_BUCKET_BATCHING)
+            ).lower()
+            in ("true", "1"),
+            reranker_local_batch_size=int(
+                os.getenv(ENV_RERANKER_LOCAL_BATCH_SIZE, str(DEFAULT_RERANKER_LOCAL_BATCH_SIZE))
+            ),
             reranker_tei_url=os.getenv(ENV_RERANKER_TEI_URL),
             reranker_tei_batch_size=int(os.getenv(ENV_RERANKER_TEI_BATCH_SIZE, str(DEFAULT_RERANKER_TEI_BATCH_SIZE))),
             reranker_tei_max_concurrent=int(
@@ -1071,6 +1152,9 @@ class HindsightConfig:
             or os.getenv(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE),
             reranker_litellm_api_key=os.getenv(ENV_RERANKER_LITELLM_API_KEY) or os.getenv(ENV_LITELLM_API_KEY),
             reranker_litellm_model=os.getenv(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL),
+            reranker_litellm_max_tokens_per_doc=int(v)
+            if (v := os.getenv(ENV_RERANKER_LITELLM_MAX_TOKENS_PER_DOC))
+            else DEFAULT_RERANKER_LITELLM_MAX_TOKENS_PER_DOC,
             # LiteLLM SDK reranker (direct API access)
             reranker_litellm_sdk_api_key=os.getenv(ENV_RERANKER_LITELLM_SDK_API_KEY),
             reranker_litellm_sdk_model=os.getenv(ENV_RERANKER_LITELLM_SDK_MODEL, DEFAULT_RERANKER_LITELLM_SDK_MODEL),
@@ -1097,6 +1181,7 @@ class HindsightConfig:
             recall_connection_budget=int(
                 os.getenv(ENV_RECALL_CONNECTION_BUDGET, str(DEFAULT_RECALL_CONNECTION_BUDGET))
             ),
+            recall_max_query_tokens=int(os.getenv(ENV_RECALL_MAX_QUERY_TOKENS, str(DEFAULT_RECALL_MAX_QUERY_TOKENS))),
             mental_model_refresh_concurrency=int(
                 os.getenv(ENV_MENTAL_MODEL_REFRESH_CONCURRENCY, str(DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY))
             ),
@@ -1117,6 +1202,8 @@ class HindsightConfig:
             ),
             retain_mission=os.getenv(ENV_RETAIN_MISSION) or DEFAULT_RETAIN_MISSION,
             retain_custom_instructions=os.getenv(ENV_RETAIN_CUSTOM_INSTRUCTIONS) or DEFAULT_RETAIN_CUSTOM_INSTRUCTIONS,
+            retain_default_strategy=os.getenv(ENV_RETAIN_DEFAULT_STRATEGY) or DEFAULT_RETAIN_DEFAULT_STRATEGY,
+            retain_strategies=DEFAULT_RETAIN_STRATEGIES,
             retain_batch_tokens=int(os.getenv(ENV_RETAIN_BATCH_TOKENS, str(DEFAULT_RETAIN_BATCH_TOKENS))),
             retain_entity_lookup=os.getenv(ENV_RETAIN_ENTITY_LOOKUP, DEFAULT_RETAIN_ENTITY_LOOKUP),
             retain_batch_enabled=os.getenv(ENV_RETAIN_BATCH_ENABLED, str(DEFAULT_RETAIN_BATCH_ENABLED)).lower()
@@ -1136,7 +1223,10 @@ class HindsightConfig:
             file_storage_azure_container=os.getenv(ENV_FILE_STORAGE_AZURE_CONTAINER) or None,
             file_storage_azure_account_name=os.getenv(ENV_FILE_STORAGE_AZURE_ACCOUNT_NAME) or None,
             file_storage_azure_account_key=os.getenv(ENV_FILE_STORAGE_AZURE_ACCOUNT_KEY) or None,
-            file_parser=os.getenv(ENV_FILE_PARSER, DEFAULT_FILE_PARSER),
+            file_parser=_parse_str_list(os.getenv(ENV_FILE_PARSER, DEFAULT_FILE_PARSER)),
+            file_parser_allowlist=_parse_str_list(os.getenv(ENV_FILE_PARSER_ALLOWLIST))
+            if os.getenv(ENV_FILE_PARSER_ALLOWLIST)
+            else None,
             file_parser_iris_token=os.getenv(ENV_FILE_PARSER_IRIS_TOKEN) or None,
             file_parser_iris_org_id=os.getenv(ENV_FILE_PARSER_IRIS_ORG_ID) or None,
             file_conversion_max_batch_size_mb=int(
@@ -1153,6 +1243,14 @@ class HindsightConfig:
             == "true",
             # Observations settings (consolidated knowledge from facts)
             enable_observations=os.getenv(ENV_ENABLE_OBSERVATIONS, str(DEFAULT_ENABLE_OBSERVATIONS)).lower() == "true",
+            enable_observation_history=os.getenv(
+                ENV_ENABLE_OBSERVATION_HISTORY, str(DEFAULT_ENABLE_OBSERVATION_HISTORY)
+            ).lower()
+            == "true",
+            enable_mental_model_history=os.getenv(
+                ENV_ENABLE_MENTAL_MODEL_HISTORY, str(DEFAULT_ENABLE_MENTAL_MODEL_HISTORY)
+            ).lower()
+            == "true",
             consolidation_batch_size=int(
                 os.getenv(ENV_CONSOLIDATION_BATCH_SIZE, str(DEFAULT_CONSOLIDATION_BATCH_SIZE))
             ),
@@ -1162,6 +1260,15 @@ class HindsightConfig:
             consolidation_max_tokens=int(
                 os.getenv(ENV_CONSOLIDATION_MAX_TOKENS, str(DEFAULT_CONSOLIDATION_MAX_TOKENS))
             ),
+            consolidation_source_facts_max_tokens=int(
+                os.getenv(ENV_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS, str(DEFAULT_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS))
+            ),
+            consolidation_source_facts_max_tokens_per_observation=int(
+                os.getenv(
+                    ENV_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION,
+                    str(DEFAULT_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION),
+                )
+            ),
             observations_mission=os.getenv(ENV_OBSERVATIONS_MISSION) or DEFAULT_OBSERVATIONS_MISSION,
             entity_labels=None,
             entities_allow_free_form=True,
@@ -1187,7 +1294,11 @@ class HindsightConfig:
             reflect_max_context_tokens=int(
                 os.getenv(ENV_REFLECT_MAX_CONTEXT_TOKENS, str(DEFAULT_REFLECT_MAX_CONTEXT_TOKENS))
             ),
+            reflect_wall_timeout=int(os.getenv(ENV_REFLECT_WALL_TIMEOUT, str(DEFAULT_REFLECT_WALL_TIMEOUT))),
             reflect_mission=os.getenv(ENV_REFLECT_MISSION) or None,
+            reflect_source_facts_max_tokens=int(
+                os.getenv(ENV_REFLECT_SOURCE_FACTS_MAX_TOKENS, str(DEFAULT_REFLECT_SOURCE_FACTS_MAX_TOKENS))
+            ),
             # Disposition settings (None = fall back to DB value)
             disposition_skepticism=int(os.getenv(ENV_DISPOSITION_SKEPTICISM))
             if os.getenv(ENV_DISPOSITION_SKEPTICISM)

hindsight-api-slim/hindsight_api/config_resolver.py

@@ -10,7 +10,7 @@ multiple API servers.
 
 import json
 import logging
-from dataclasses import asdict
+from dataclasses import asdict, replace
 from typing import Any
 
 import asyncpg
@@ -239,6 +239,14 @@ class ConfigResolver:
                 logger.warning(f"Failed to check permissions for bank {bank_id}: {e}")
                 # Continue without permission check (fail open for backward compatibility)
 
+        # Validate retain_strategies: reject empty string keys
+        if "retain_strategies" in normalized_updates and normalized_updates["retain_strategies"]:
+            empty_keys = [k for k in normalized_updates["retain_strategies"] if not str(k).strip()]
+            if empty_keys:
+                raise ValueError(
+                    "Strategy names must not be empty strings. Remove entries with empty names before saving."
+                )
+
         # Merge with existing config (JSONB || operator)
         async with self.pool.acquire() as conn:
             await conn.execute(
@@ -273,3 +281,35 @@ class ConfigResolver:
             )
 
         logger.info(f"Reset bank config for {bank_id} to defaults")
+
+
+def apply_strategy(config: HindsightConfig, strategy_name: str) -> HindsightConfig:
+    """
+    Apply a named retain strategy's overrides on top of a resolved config.
+
+    A strategy is a named set of hierarchical field overrides stored in
+    config.retain_strategies. Any field in _HIERARCHICAL_FIELDS can be
+    overridden, including retain_extraction_mode, retain_chunk_size,
+    entity_labels, entities_allow_free_form, etc.
+
+    Unknown strategy names log a warning and return config unchanged.
+    Unknown or non-hierarchical fields in the strategy are silently ignored.
+    """
+    strategies = config.retain_strategies or {}
+    if strategy_name not in strategies:
+        logger.warning(f"Unknown retain strategy '{strategy_name}', using resolved config as-is")
+        return config
+
+    overrides = strategies[strategy_name]
+    if not isinstance(overrides, dict):
+        logger.warning(f"Retain strategy '{strategy_name}' is not a dict, skipping")
+        return config
+
+    configurable = HindsightConfig.get_configurable_fields()
+    filtered = {k: v for k, v in overrides.items() if k in configurable}
+
+    if not filtered:
+        return config
+
+    logger.debug(f"Applying retain strategy '{strategy_name}': {list(filtered.keys())}")
+    return replace(config, **filtered)

hindsight-api-slim/hindsight_api/engine/consolidation/consolidator.py

@@ -24,9 +24,10 @@ from datetime import datetime, timezone
 from itertools import combinations
 from typing import TYPE_CHECKING, Any
 
-from pydantic import BaseModel
+from pydantic import BaseModel, field_validator
 
 from ...config import get_config
+from ..llm_wrapper import sanitize_llm_output
 from ..memory_engine import fq_table
 from ..retain import embedding_utils
 from .prompts import build_batch_consolidation_prompt
@@ -45,12 +46,22 @@ class _CreateAction(BaseModel):
     text: str
     source_fact_ids: list[str]  # memory UUIDs from the NEW FACTS list
 
+    @field_validator("text", mode="before")
+    @classmethod
+    def sanitize_text(cls, v: str) -> str:
+        return sanitize_llm_output(v) or ""
+
 
 class _UpdateAction(BaseModel):
     text: str
     observation_id: str  # UUID of the existing observation to update
     source_fact_ids: list[str]  # memory UUIDs from the NEW FACTS list
 
+    @field_validator("text", mode="before")
+    @classmethod
+    def sanitize_text(cls, v: str) -> str:
+        return sanitize_llm_output(v) or ""
+
 
 class _DeleteAction(BaseModel):
     observation_id: str  # UUID of the observation to remove
@@ -69,6 +80,7 @@ class _BatchLLMResult:
     deletes: list[_DeleteAction] = field(default_factory=list)
     obs_count: int = 0
     prompt_chars: int = 0
+    failed: bool = False
 
 
 @dataclass
@@ -150,6 +162,7 @@ async def run_consolidation_job(
     memory_engine: "MemoryEngine",
     bank_id: str,
     request_context: "RequestContext",
+    operation_id: str | None = None,
 ) -> dict[str, Any]:
     """
     Run consolidation job for a bank.
@@ -207,6 +220,7 @@ async def run_consolidation_job(
             FROM {fq_table("memory_units")}
             WHERE bank_id = $1
               AND consolidated_at IS NULL
+              AND consolidation_failed_at IS NULL
               AND fact_type IN ('experience', 'world')
             """,
             bank_id,
@@ -228,6 +242,7 @@ async def run_consolidation_job(
         "observations_deleted": 0,
         "actions_executed": 0,
         "skipped": 0,
+        "memories_failed": 0,
     }
 
     # Track all unique tags from consolidated memories for mental model refresh filtering
@@ -245,6 +260,7 @@ async def run_consolidation_job(
                 FROM {fq_table("memory_units")}
                 WHERE bank_id = $1
                   AND consolidated_at IS NULL
+                  AND consolidation_failed_at IS NULL
                   AND fact_type IN ('experience', 'world')
                 ORDER BY created_at ASC
                 LIMIT $2
@@ -286,94 +302,148 @@ async def run_consolidation_job(
                 if memory_tags:
                     consolidated_tags.update(memory_tags)
 
-            async with pool.acquire() as conn:
-                # Determine observation_scopes for this batch. All memories in a batch share
-                # the same tags (enforced by tag_groups), so we only check the first memory.
-                # asyncpg returns JSONB columns as raw JSON strings, so parse if needed.
-                _obs_raw = llm_batch[0].get("observation_scopes") if llm_batch else None
-                _obs_parsed = json.loads(_obs_raw) if isinstance(_obs_raw, str) else _obs_raw
+            # Process llm_batch with adaptive splitting: on LLM failure, halve the sub-batch
+            # and retry, down to batch_size=1. Only if a single-memory batch still fails is
+            # the memory marked with consolidation_failed_at and excluded from future runs
+            # until explicitly retried via the API.
+            all_results: list[dict[str, Any]] = []
+            all_deleted = 0
+            succeeded_ids: list[Any] = []
+            failed_ids: list[Any] = []
 
-                # Resolve the scope spec into a concrete list[list[str]] (or None for combined).
-                if _obs_parsed == "per_tag":
-                    _memory_tags = llm_batch[0].get("tags") or []
-                    obs_tags_list = [[tag] for tag in _memory_tags] if _memory_tags else None
-                elif _obs_parsed == "all_combinations":
-                    _memory_tags = llm_batch[0].get("tags") or []
-                    obs_tags_list = (
-                        [
-                            list(combo)
-                            for r in range(1, len(_memory_tags) + 1)
-                            for combo in combinations(_memory_tags, r)
-                        ]
-                        if _memory_tags
-                        else None
-                    )
-                elif _obs_parsed == "combined" or _obs_parsed is None:
-                    obs_tags_list = None  # single combined pass (default behaviour)
-                else:
-                    # explicit list[list[str]]
-                    obs_tags_list = _obs_parsed
+            pending: list[list[dict[str, Any]]] = [llm_batch]
+            while pending:
+                sub_batch = pending.pop(0)
 
-                batch_deleted: int = 0
-                if obs_tags_list:
-                    # Multi-pass: run one observation consolidation pass per tag set
-                    results = []
-                    for obs_tags in obs_tags_list:
-                        pass_results, pass_deleted = await _process_memory_batch(
+                async with pool.acquire() as conn:
+                    # Determine observation_scopes for this sub-batch. All memories share
+                    # the same tags (enforced by tag_groups), so we only check the first memory.
+                    # asyncpg returns JSONB columns as raw JSON strings, so parse if needed.
+                    _obs_raw = sub_batch[0].get("observation_scopes") if sub_batch else None
+                    _obs_parsed = json.loads(_obs_raw) if isinstance(_obs_raw, str) else _obs_raw
+
+                    # Resolve the scope spec into a concrete list[list[str]] (or None for combined).
+                    if _obs_parsed == "per_tag":
+                        _memory_tags = sub_batch[0].get("tags") or []
+                        obs_tags_list = [[tag] for tag in _memory_tags] if _memory_tags else None
+                    elif _obs_parsed == "all_combinations":
+                        _memory_tags = sub_batch[0].get("tags") or []
+                        obs_tags_list = (
+                            [
+                                list(combo)
+                                for r in range(1, len(_memory_tags) + 1)
+                                for combo in combinations(_memory_tags, r)
+                            ]
+                            if _memory_tags
+                            else None
+                        )
+                    elif _obs_parsed == "combined" or _obs_parsed is None:
+                        obs_tags_list = None  # single combined pass (default behaviour)
+                    else:
+                        # explicit list[list[str]]
+                        obs_tags_list = _obs_parsed
+
+                    sub_deleted: int = 0
+                    sub_llm_failed = False
+                    if obs_tags_list:
+                        # Multi-pass: run one observation consolidation pass per tag set
+                        sub_results: list[dict[str, Any]] = []
+                        for obs_tags in obs_tags_list:
+                            pass_results, pass_deleted, pass_failed = await _process_memory_batch(
+                                conn=conn,
+                                memory_engine=memory_engine,
+                                llm_config=llm_config,
+                                bank_id=bank_id,
+                                memories=sub_batch,
+                                request_context=request_context,
+                                perf=perf,
+                                config=config,
+                                obs_tags_override=obs_tags,
+                            )
+                            sub_deleted += pass_deleted
+                            sub_llm_failed = sub_llm_failed or pass_failed
+                            # Merge results: prefer non-skipped actions
+                            if not sub_results:
+                                sub_results = pass_results
+                            else:
+                                for i, (existing, new) in enumerate(zip(sub_results, pass_results)):
+                                    if existing.get("action") == "skipped" and new.get("action") != "skipped":
+                                        sub_results[i] = new
+                                    elif existing.get("action") != "skipped" and new.get("action") != "skipped":
+                                        # Both did something — combine into "multiple"
+                                        existing_created = existing.get(
+                                            "created", 1 if existing.get("action") == "created" else 0
+                                        )
+                                        existing_updated = existing.get(
+                                            "updated", 1 if existing.get("action") == "updated" else 0
+                                        )
+                                        new_created = new.get("created", 1 if new.get("action") == "created" else 0)
+                                        new_updated = new.get("updated", 1 if new.get("action") == "updated" else 0)
+                                        total = existing_created + existing_updated + new_created + new_updated
+                                        sub_results[i] = {
+                                            "action": "multiple",
+                                            "created": existing_created + new_created,
+                                            "updated": existing_updated + new_updated,
+                                            "merged": 0,
+                                            "total_actions": total,
+                                        }
+                    else:
+                        # Normal single pass using the memory's own tags
+                        sub_results, sub_deleted, sub_llm_failed = await _process_memory_batch(
                             conn=conn,
                             memory_engine=memory_engine,
                             llm_config=llm_config,
                             bank_id=bank_id,
-                            memories=llm_batch,
+                            memories=sub_batch,
                             request_context=request_context,
                             perf=perf,
                             config=config,
-                            obs_tags_override=obs_tags,
                         )
-                        batch_deleted += pass_deleted
-                        # Merge results: prefer non-skipped actions
-                        if not results:
-                            results = pass_results
-                        else:
-                            for i, (existing, new) in enumerate(zip(results, pass_results)):
-                                if existing.get("action") == "skipped" and new.get("action") != "skipped":
-                                    results[i] = new
-                                elif existing.get("action") != "skipped" and new.get("action") != "skipped":
-                                    # Both did something — combine into "multiple"
-                                    existing_created = existing.get(
-                                        "created", 1 if existing.get("action") == "created" else 0
-                                    )
-                                    existing_updated = existing.get(
-                                        "updated", 1 if existing.get("action") == "updated" else 0
-                                    )
-                                    new_created = new.get("created", 1 if new.get("action") == "created" else 0)
-                                    new_updated = new.get("updated", 1 if new.get("action") == "updated" else 0)
-                                    total = existing_created + existing_updated + new_created + new_updated
-                                    results[i] = {
-                                        "action": "multiple",
-                                        "created": existing_created + new_created,
-                                        "updated": existing_updated + new_updated,
-                                        "merged": 0,
-                                        "total_actions": total,
-                                    }
-                else:
-                    # Normal single pass using the memory's own tags
-                    results, batch_deleted = await _process_memory_batch(
-                        conn=conn,
-                        memory_engine=memory_engine,
-                        llm_config=llm_config,
-                        bank_id=bank_id,
-                        memories=llm_batch,
-                        request_context=request_context,
-                        perf=perf,
-                        config=config,
-                    )
-                stats["observations_deleted"] += batch_deleted
 
-                await conn.executemany(
-                    f"UPDATE {fq_table('memory_units')} SET consolidated_at = NOW() WHERE id = $1",
-                    [(m["id"],) for m in llm_batch],
+                all_deleted += sub_deleted
+
+                if sub_llm_failed and len(sub_batch) > 1:
+                    # Split and retry with smaller batches
+                    mid = len(sub_batch) // 2
+                    logger.warning(
+                        f"[CONSOLIDATION] bank={bank_id} LLM failed for sub-batch of {len(sub_batch)},"
+                        f" splitting into {mid}/{len(sub_batch) - mid}"
+                    )
+                    pending[0:0] = [sub_batch[:mid], sub_batch[mid:]]
+                elif sub_llm_failed:
+                    # batch_size=1 and still failing — mark as permanently failed for now
+                    failed_ids.append(sub_batch[0]["id"])
+                    all_results.append({"action": "failed"})
+                    logger.warning(
+                        f"[CONSOLIDATION] bank={bank_id} LLM failed for single memory"
+                        f" {sub_batch[0]['id']}, marking consolidation_failed_at"
+                    )
+                else:
+                    succeeded_ids.extend(m["id"] for m in sub_batch)
+                    all_results.extend(sub_results)
+
+            # Commit consolidated_at / consolidation_failed_at in a single DB round-trip
+            async with pool.acquire() as conn:
+                if succeeded_ids:
+                    await conn.executemany(
+                        f"UPDATE {fq_table('memory_units')} SET consolidated_at = NOW() WHERE id = $1",
+                        [(mem_id,) for mem_id in succeeded_ids],
+                    )
+                if failed_ids:
+                    await conn.executemany(
+                        f"UPDATE {fq_table('memory_units')} SET consolidation_failed_at = NOW() WHERE id = $1",
+                        [(mem_id,) for mem_id in failed_ids],
+                    )
+
+            stats["observations_deleted"] += all_deleted
+            results = all_results
+
+            # Checkpoint: abort if the operation (and thus the bank) was deleted mid-run.
+            if operation_id and not await memory_engine._check_op_alive(operation_id):
+                logger.info(
+                    f"[CONSOLIDATION] bank={bank_id} operation {operation_id} cancelled (bank deleted), stopping early"
                 )
+                return {"status": "cancelled", "bank_id": bank_id, **stats}
 
             for result in results:
                 stats["memories_processed"] += 1
@@ -394,6 +464,8 @@ async def run_consolidation_job(
                     stats["actions_executed"] += result.get("total_actions", 0)
                 elif action == "skipped":
                     stats["skipped"] += 1
+                elif action == "failed":
+                    stats["memories_failed"] += 1
 
             # Per-LLM-batch log
             llm_batch_time = time.time() - llm_batch_start
@@ -406,6 +478,7 @@ async def run_consolidation_job(
             batch_created = stats["observations_created"] - snap_stats["observations_created"]
             batch_updated = stats["observations_updated"] - snap_stats["observations_updated"]
             batch_skipped = stats["skipped"] - snap_stats["skipped"]
+            batch_failed = stats["memories_failed"] - snap_stats["memories_failed"]
             llm_calls_made = perf.llm_calls - snap_llm_calls
             logger.info(
                 f"[CONSOLIDATION] bank={bank_id} llm_batch #{llm_batch_num}"
@@ -413,7 +486,8 @@ async def run_consolidation_job(
                 f" | {stats['memories_processed']}/{total_count} processed"
                 f" | {', '.join(timing_parts)}"
                 f" | created={batch_created} updated={batch_updated} skipped={batch_skipped}"
-                f" | input_tokens=~{input_tokens}"
+                + (f" failed={batch_failed}" if batch_failed else "")
+                + f" | input_tokens=~{input_tokens}"
                 f" | avg={llm_batch_time / len(llm_batch):.3f}s/memory"
             )
 
@@ -565,7 +639,7 @@ async def _process_memory_batch(
     perf: ConsolidationPerfLog | None = None,
     config: Any = None,
     obs_tags_override: list[str] | None = None,
-) -> tuple[list[dict[str, Any]], int]:
+) -> tuple[list[dict[str, Any]], int, bool]:
     """
     Process a batch of memories in a single LLM call.
 
@@ -728,7 +802,7 @@ async def _process_memory_batch(
         else:
             results.append({"action": "skipped", "reason": "no_durable_knowledge"})
 
-    return results, deleted_count
+    return results, deleted_count, llm_result.failed
 
 
 def _min_date(dates: "Any") -> "datetime | None":
@@ -766,13 +840,17 @@ async def _execute_update_action(
         logger.debug(f"Update skipped: observation {observation_id} not found in recall results")
         return
 
-    history = [
-        {
-            "previous_text": model.text,
-            "changed_at": datetime.now(timezone.utc).isoformat(),
-            "source_memory_ids": [str(mid) for mid in source_memory_ids],
-        }
-    ]
+    from ...config import get_config
+
+    history_entry = {
+        "previous_text": model.text,
+        "previous_tags": list(model.tags or []),
+        "previous_occurred_start": model.occurred_start,
+        "previous_occurred_end": model.occurred_end,
+        "previous_mentioned_at": model.mentioned_at,
+        "changed_at": datetime.now(timezone.utc).isoformat(),
+        "new_source_memory_ids": [str(mid) for mid in source_memory_ids],
+    }
 
     source_ids = list(model.source_fact_ids or []) + source_memory_ids
 
@@ -787,13 +865,18 @@ async def _execute_update_action(
     if perf:
         perf.record_timing("embedding", time.time() - t0)
 
+    config = get_config()
+    history_clause = (
+        "history = COALESCE(history, '[]'::jsonb) || $3::jsonb," if config.enable_observation_history else ""
+    )
+
     t0 = time.time()
     await conn.execute(
         f"""
         UPDATE {fq_table("memory_units")}
         SET text = $1,
             embedding = $2::vector,
-            history = $3,
+            {history_clause}
             source_memory_ids = $4,
             proof_count = $5,
             tags = $10,
@@ -805,7 +888,7 @@ async def _execute_update_action(
         """,
         new_text,
         embedding_str,
-        json.dumps(history),
+        json.dumps([history_entry]),
         source_ids,
         len(source_ids),
         uuid.UUID(observation_id),
@@ -917,10 +1000,9 @@ async def _find_related_observations(
     """
     # Use recall to find related observations with token budget
     # max_tokens naturally limits how many observations are returned
-    from ...config import get_config
     from ...tracing import get_tracer, is_tracing_enabled
 
-    config = get_config()
+    config = await memory_engine._config_resolver.resolve_full_config(bank_id, request_context)
 
     # SECURITY: Use all_strict matching if tags provided to prevent cross-scope consolidation
     tags_match = "all_strict" if tags else "any"
@@ -945,7 +1027,8 @@ async def _find_related_observations(
             tags=tags,  # Filter by source memory's tags
             tags_match=tags_match,  # Use strict matching for security
             include_source_facts=True,  # Embed source facts so we avoid a separate DB fetch
-            max_source_facts_tokens=-1,  # No token limit — we need all source facts for consolidation
+            max_source_facts_tokens=config.consolidation_source_facts_max_tokens,
+            max_source_facts_tokens_per_observation=config.consolidation_source_facts_max_tokens_per_observation,
             _quiet=True,  # Suppress logging
         )
     finally:
@@ -1009,14 +1092,17 @@ async def _consolidate_batch_with_llm(
         observations_text = "[]"
 
     def _fact_line(m: dict[str, Any]) -> str:
-        parts = [f"[{m['id']}] {m['text']}"]
+        text = f"[{m['id']}] {m['text']}"
+        temporal_parts = []
         if m.get("occurred_start"):
-            parts.append(f"occurred_start={m['occurred_start']}")
+            temporal_parts.append(f"occurred_start={m['occurred_start']}")
         if m.get("occurred_end"):
-            parts.append(f"occurred_end={m['occurred_end']}")
+            temporal_parts.append(f"occurred_end={m['occurred_end']}")
         if m.get("mentioned_at"):
-            parts.append(f"mentioned_at={m['mentioned_at']}")
-        return " | ".join(parts)
+            temporal_parts.append(f"mentioned_at={m['mentioned_at']}")
+        if temporal_parts:
+            text += f" ({', '.join(temporal_parts)})"
+        return text
 
     facts_lines = "\n".join(_fact_line(m) for m in memories)
 
@@ -1050,7 +1136,7 @@ async def _consolidate_batch_with_llm(
     logger.error(
         f"[CONSOLIDATION] LLM batch call failed after {max_attempts} attempts, skipping batch. Last error: {last_exc}"
     )
-    return _BatchLLMResult(obs_count=len(union_observations), prompt_chars=len(prompt))
+    return _BatchLLMResult(obs_count=len(union_observations), prompt_chars=len(prompt), failed=True)
 
 
 async def _create_observation_directly(

hindsight-api-slim/hindsight_api/engine/consolidation/prompts.py

@@ -29,14 +29,31 @@ Compare the facts against existing observations:
 - Same topic as an existing observation → UPDATE it (observation_id + source_fact_ids)
 - New topic with durable knowledge → CREATE a new observation (source_fact_ids)
 - Cross-reference facts within the batch: a later fact may resolve a vague reference in an earlier one
-- Purely ephemeral facts → omit them (no create/update needed)"""
+- Purely ephemeral facts → omit them unless the MISSION above explicitly targets such data (e.g. timestamped events, session state, screen content)"""
 
 # Output format — JSON braces escaped as {{ }} so .format() leaves them literal
 _BATCH_OUTPUT_FORMAT = """
 Output a JSON object with three arrays.
 
-Example (showing the required UUID format for all IDs):
-{{"creates": [{{"text": "Alice lives in Berlin", "source_fact_ids": ["a1b2c3d4-e5f6-7890-abcd-ef1234567890", "b2c3d4e5-f6a7-8901-bcde-f12345678901"]}}],
+## EXAMPLE
+
+Input facts:
+[a1b2c3d4-e5f6-7890-abcd-ef1234567890] Alice mentioned she works long hours, often past midnight | Involving: Alice (occurred_start=2024-01-15, mentioned_at=2024-01-15)
+[b2c3d4e5-f6a7-8901-bcde-f12345678901] Alice said she's exhausted from the project deadlines | Involving: Alice (occurred_start=2024-01-20, mentioned_at=2024-01-20)
+
+Good observation text — clean prose, no metadata, each fact tracked distinctly:
+  "Alice works long hours, often past midnight."
+  "Alice feels exhausted from project deadlines."
+
+Bad observation text — NEVER do this (verbatim copy of fact text with metadata):
+  "Alice mentioned she works long hours, often past midnight | Involving: Alice (occurred_start=2024-01-15, mentioned_at=2024-01-15)"
+
+Observation text rules:
+- Write clean prose — NEVER copy raw fact lines or their metadata (temporal fields, "Involving:", "When:" labels, UUIDs).
+- Parenthesized metadata like (occurred_start=...) and pipe-separated labels like "| Involving: ..." are fact formatting — strip them entirely from observation text.
+- How many observations to create and how much to aggregate is driven by the MISSION above.
+
+{{"creates": [{{"text": "Alice works long hours, often past midnight.", "source_fact_ids": ["a1b2c3d4-e5f6-7890-abcd-ef1234567890"]}}, {{"text": "Alice feels exhausted from project deadlines.", "source_fact_ids": ["b2c3d4e5-f6a7-8901-bcde-f12345678901"]}}],
   "updates": [{{"text": "Alice works at Acme Corp as a senior engineer", "observation_id": "c3d4e5f6-a7b8-9012-cdef-123456789012", "source_fact_ids": ["d4e5f6a7-b8c9-0123-defa-234567890123"]}}],
   "deletes": [{{"observation_id": "e5f6a7b8-c9d0-1234-efab-345678901234"}}]}}
 

hindsight-api-slim/hindsight_api/engine/cross_encoder.py

@@ -20,8 +20,10 @@ from ..config import (
     DEFAULT_RERANKER_COHERE_MODEL,
     DEFAULT_RERANKER_FLASHRANK_CACHE_DIR,
     DEFAULT_RERANKER_FLASHRANK_MODEL,
+    DEFAULT_RERANKER_LITELLM_MAX_TOKENS_PER_DOC,
     DEFAULT_RERANKER_LITELLM_MODEL,
     DEFAULT_RERANKER_LITELLM_SDK_MODEL,
+    DEFAULT_RERANKER_LOCAL_BATCH_SIZE,
     DEFAULT_RERANKER_LOCAL_FORCE_CPU,
     DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT,
     DEFAULT_RERANKER_LOCAL_MODEL,
@@ -110,6 +112,9 @@ class LocalSTCrossEncoder(CrossEncoderModel):
         max_concurrent: int = 4,
         force_cpu: bool = False,
         trust_remote_code: bool = False,
+        fp16: bool = False,
+        bucket_batching: bool = False,
+        batch_size: int = DEFAULT_RERANKER_LOCAL_BATCH_SIZE,
     ):
         """
         Initialize local SentenceTransformers cross-encoder.
@@ -124,10 +129,20 @@ class LocalSTCrossEncoder(CrossEncoderModel):
             trust_remote_code: Allow loading models with custom code (security risk).
                               Required for some models like jina-reranker-v2-base-multilingual.
                               Default: False (disabled for security)
+            fp16: Use FP16 (half precision) inference. Faster on MPS and CUDA,
+                  may be slower on CPU. Default: False (opt-in via env var).
+            bucket_batching: Sort pairs by token length before batching to reduce
+                            padding waste. 36-54% speedup, quality-identical.
+                            Default: False (opt-in via env var).
+            batch_size: Batch size for predict() calls. Optimal values vary by
+                       hardware and model (MPS: 32, CUDA: 128+). Default: 32.
         """
         self.model_name = model_name or DEFAULT_RERANKER_LOCAL_MODEL
         self.force_cpu = force_cpu
         self.trust_remote_code = trust_remote_code
+        self.fp16 = fp16
+        self.bucket_batching = bucket_batching
+        self.batch_size = batch_size
         self._model = None
         LocalSTCrossEncoder._max_concurrent = max_concurrent
 
@@ -175,6 +190,24 @@ class LocalSTCrossEncoder(CrossEncoderModel):
             except Exception as e:
                 logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")
 
+        # Patch transformers 5.x compatibility for models using XLM-RoBERTa
+        # (e.g., jina-reranker-v2-base-multilingual). transformers 5.x removed
+        # create_position_ids_from_input_ids as a module-level function; the custom
+        # code in these models still references it. This monkey-patch restores it.
+        try:
+            import transformers.models.xlm_roberta.modeling_xlm_roberta as xlm_module
+            from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaEmbeddings
+
+            if not hasattr(xlm_module, "create_position_ids_from_input_ids"):
+                setattr(
+                    xlm_module,
+                    "create_position_ids_from_input_ids",
+                    XLMRobertaEmbeddings.create_position_ids_from_input_ids,
+                )
+                logger.info("Reranker: applied transformers 5.x compatibility patch for XLM-RoBERTa")
+        except Exception:
+            pass
+
         # Suppress verbose transformers warnings during model loading
         # This suppresses the "UNEXPECTED" warnings from CrossEncoder which are harmless
         # but look alarming to users (e.g., "embeddings.position_ids | UNEXPECTED")
@@ -199,6 +232,12 @@ class LocalSTCrossEncoder(CrossEncoderModel):
                 # Restore original logging level
                 transformers_logger.setLevel(original_level)
 
+        # FP16 inference: convert model weights to half precision.
+        # Empirically validated: 27-36% faster on MPS, quality-identical (20/20 overlap).
+        if self.fp16 and device != "cpu":
+            self._model.model.half()
+            logger.info("Reranker: FP16 inference enabled")
+
         # Initialize shared executor (limited workers naturally limits concurrency)
         if LocalSTCrossEncoder._executor is None:
             LocalSTCrossEncoder._executor = ThreadPoolExecutor(
@@ -210,8 +249,32 @@ class LocalSTCrossEncoder(CrossEncoderModel):
             logger.info("Reranker: local provider initialized (using existing executor)")
 
     def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """Synchronous prediction wrapper for thread pool execution."""
-        scores = self._model.predict(pairs, show_progress_bar=False)
+        """Synchronous prediction wrapper for thread pool execution.
+
+        Supports two optimizations (controlled via .env):
+        - bucket_batching: sort pairs by token length to reduce padding waste (36-54% speedup)
+        - batch_size: explicit batch size for predict() calls (MPS optimal: 32)
+        """
+        import numpy as np
+
+        if self.bucket_batching and len(pairs) > 1:
+            # Sort pairs by approximate token length to create homogeneous batches.
+            # This eliminates padding waste — short pairs aren't padded to the length
+            # of the longest pair in the batch. Quality-identical by construction.
+            lengths = [len(pairs[i][0]) + len(pairs[i][1]) for i in range(len(pairs))]
+            sorted_indices = sorted(range(len(pairs)), key=lambda i: lengths[i])
+            sorted_pairs = [pairs[i] for i in sorted_indices]
+
+            sorted_scores = self._model.predict(sorted_pairs, batch_size=self.batch_size, show_progress_bar=False)
+            sorted_scores = sorted_scores.tolist() if hasattr(sorted_scores, "tolist") else list(sorted_scores)
+
+            # Restore original order
+            scores = [0.0] * len(pairs)
+            for new_pos, orig_idx in enumerate(sorted_indices):
+                scores[orig_idx] = sorted_scores[new_pos]
+            return scores
+
+        scores = self._model.predict(pairs, batch_size=self.batch_size, show_progress_bar=False)
         return scores.tolist() if hasattr(scores, "tolist") else list(scores)
 
     async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
@@ -820,6 +883,17 @@ class FlashRankCrossEncoder(CrossEncoderModel):
         return await loop.run_in_executor(FlashRankCrossEncoder._executor, self._predict_sync, pairs)
 
 
+def _truncate_to_tokens(text: str, max_tokens: int) -> str:
+    """Truncate text to at most max_tokens using the shared tiktoken encoder."""
+    from .memory_engine import _get_tiktoken_encoding
+
+    enc = _get_tiktoken_encoding()
+    tokens = enc.encode(text)
+    if len(tokens) <= max_tokens:
+        return text
+    return enc.decode(tokens[:max_tokens])
+
+
 class LiteLLMCrossEncoder(CrossEncoderModel):
     """
     LiteLLM cross-encoder implementation using LiteLLM proxy's /rerank endpoint.
@@ -843,6 +917,7 @@ class LiteLLMCrossEncoder(CrossEncoderModel):
         api_key: str | None = None,
         model: str = DEFAULT_RERANKER_LITELLM_MODEL,
         timeout: float = 60.0,
+        max_tokens_per_doc: int | None = DEFAULT_RERANKER_LITELLM_MAX_TOKENS_PER_DOC,
     ):
         """
         Initialize LiteLLM cross-encoder client.
@@ -853,11 +928,15 @@ class LiteLLMCrossEncoder(CrossEncoderModel):
             model: Reranking model name (default: cohere/rerank-english-v3.0)
                    Use provider prefix (e.g., cohere/, together_ai/, voyage/)
             timeout: Request timeout in seconds (default: 60.0)
+            max_tokens_per_doc: If set, truncate each document to this many tokens before
+                                sending to the reranker (uses tiktoken cl100k_base encoding).
+                                Useful for models with small context windows (e.g. 1024 tokens).
         """
         self.api_base = api_base.rstrip("/")
         self.api_key = api_key
         self.model = model
         self.timeout = timeout
+        self.max_tokens_per_doc = max_tokens_per_doc
         self._async_client: httpx.AsyncClient | None = None
 
     @property
@@ -905,6 +984,8 @@ class LiteLLMCrossEncoder(CrossEncoderModel):
 
         for query, indexed_texts in query_groups.items():
             texts = [text for _, text in indexed_texts]
+            if self.max_tokens_per_doc is not None:
+                texts = [_truncate_to_tokens(t, self.max_tokens_per_doc) for t in texts]
             indices = [idx for idx, _ in indexed_texts]
 
             # LiteLLM /rerank follows Cohere API format
@@ -950,6 +1031,7 @@ class LiteLLMSDKCrossEncoder(CrossEncoderModel):
         model: str = DEFAULT_RERANKER_LITELLM_SDK_MODEL,
         api_base: str | None = None,
         timeout: float = 60.0,
+        max_tokens_per_doc: int | None = DEFAULT_RERANKER_LITELLM_MAX_TOKENS_PER_DOC,
     ):
         """
         Initialize LiteLLM SDK cross-encoder client.
@@ -959,11 +1041,15 @@ class LiteLLMSDKCrossEncoder(CrossEncoderModel):
             model: Model name with provider prefix (e.g., "deepinfra/Qwen3-reranker-8B")
             api_base: Custom base URL for API (optional)
             timeout: Request timeout in seconds (default: 60.0)
+            max_tokens_per_doc: If set, truncate each document to this many tokens before
+                                sending to the reranker (uses tiktoken cl100k_base encoding).
+                                Useful for models with small context windows (e.g. 1024 tokens).
         """
         self.api_key = api_key
         self.model = model
         self.api_base = api_base
         self.timeout = timeout
+        self.max_tokens_per_doc = max_tokens_per_doc
         self._initialized = False
         self._litellm = None  # Will be set during initialization
 
@@ -1017,6 +1103,8 @@ class LiteLLMSDKCrossEncoder(CrossEncoderModel):
 
         for query, indexed_texts in query_groups.items():
             texts = [text for _, text in indexed_texts]
+            if self.max_tokens_per_doc is not None:
+                texts = [_truncate_to_tokens(t, self.max_tokens_per_doc) for t in texts]
             indices = [idx for idx, _ in indexed_texts]
 
             # Build kwargs for rerank call
@@ -1050,6 +1138,97 @@ class LiteLLMSDKCrossEncoder(CrossEncoderModel):
         return all_scores
 
 
+class JinaMLXCrossEncoder(CrossEncoderModel):
+    """
+    Jina Reranker v3 MLX implementation for Apple Silicon.
+
+    Uses jinaai/jina-reranker-v3-mlx — a 0.6B parameter multilingual listwise reranker
+    optimized for Apple Silicon via the MLX framework. No transformers/PyTorch dependency.
+
+    The model is downloaded automatically from HuggingFace Hub on first use.
+    Requires: mlx>=0.31.0, mlx-lm>=0.31.1, safetensors>=0.6.2
+    """
+
+    HF_REPO_ID = "jinaai/jina-reranker-v3-mlx"
+
+    def __init__(self, model_path: str | None = None):
+        """
+        Args:
+            model_path: Local path to the downloaded model directory.
+                        If None, the model is downloaded from HuggingFace Hub.
+        """
+        self.model_path = model_path
+        self._reranker = None
+
+    @property
+    def provider_name(self) -> str:
+        return "jina-mlx"
+
+    async def initialize(self) -> None:
+        if self._reranker is not None:
+            return
+
+        try:
+            import mlx.core  # noqa: F401
+            import mlx_lm  # noqa: F401
+        except ImportError:
+            raise ImportError(
+                "mlx and mlx-lm are required for JinaMLXCrossEncoder. "
+                "Install with: pip install mlx>=0.31.0 mlx-lm>=0.31.1 safetensors>=0.6.2"
+            )
+
+        loop = asyncio.get_event_loop()
+        await loop.run_in_executor(None, self._load_model)
+
+    def _load_model(self) -> None:
+        """Download (if needed) and load the MLX reranker. Runs in a thread."""
+        import os
+
+        from huggingface_hub import snapshot_download
+
+        from .jina_mlx_reranker import MLXReranker
+
+        model_path = self.model_path
+        if model_path is None:
+            logger.info(f"Reranker: downloading {self.HF_REPO_ID} from HuggingFace Hub...")
+            model_path = snapshot_download(repo_id=self.HF_REPO_ID)
+
+        logger.info(f"Reranker: loading jina-reranker-v3-mlx from {model_path}")
+        self._reranker = MLXReranker(
+            model_path=model_path,
+            projector_path=os.path.join(model_path, "projector.safetensors"),
+        )
+        logger.info("Reranker: jina-mlx provider initialized")
+
+    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """Score pairs grouped by query. Runs in a thread."""
+        if not pairs:
+            return []
+
+        query_groups: dict[str, list[tuple[int, str]]] = {}
+        for idx, (query, doc) in enumerate(pairs):
+            query_groups.setdefault(query, []).append((idx, doc))
+
+        all_scores = [0.0] * len(pairs)
+
+        for query, indexed_docs in query_groups.items():
+            docs = [doc for _, doc in indexed_docs]
+            indices = [idx for idx, _ in indexed_docs]
+            results = self._reranker.rerank(query, docs)
+            for result in results:
+                original_idx = result["index"]
+                all_scores[indices[original_idx]] = result["relevance_score"]
+
+        return all_scores
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        if self._reranker is None:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(None, self._predict_sync, pairs)
+
+
 def create_cross_encoder_from_env() -> CrossEncoderModel:
     """
     Create a CrossEncoderModel instance based on configuration.
@@ -1079,6 +1258,9 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
             max_concurrent=config.reranker_local_max_concurrent,
             force_cpu=config.reranker_local_force_cpu,
             trust_remote_code=config.reranker_local_trust_remote_code,
+            fp16=config.reranker_local_fp16,
+            bucket_batching=config.reranker_local_bucket_batching,
+            batch_size=config.reranker_local_batch_size,
         )
     elif provider == "cohere":
         api_key = config.reranker_cohere_api_key
@@ -1098,6 +1280,7 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
             api_base=config.reranker_litellm_api_base,
             api_key=config.reranker_litellm_api_key,
             model=config.reranker_litellm_model,
+            max_tokens_per_doc=config.reranker_litellm_max_tokens_per_doc,
         )
     elif provider == "litellm-sdk":
         api_key = config.reranker_litellm_sdk_api_key
@@ -1109,6 +1292,7 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
             api_key=api_key,
             model=config.reranker_litellm_sdk_model,
             api_base=config.reranker_litellm_sdk_api_base,
+            max_tokens_per_doc=config.reranker_litellm_max_tokens_per_doc,
         )
     elif provider == "zeroentropy":
         api_key = config.reranker_zeroentropy_api_key
@@ -1122,7 +1306,9 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
         )
     elif provider == "rrf":
         return RRFPassthroughCrossEncoder()
+    elif provider == "jina-mlx":
+        return JinaMLXCrossEncoder()
     else:
         raise ValueError(
-            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'zeroentropy', 'flashrank', 'litellm', 'litellm-sdk', 'rrf'"
+            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'zeroentropy', 'flashrank', 'litellm', 'litellm-sdk', 'rrf', 'jina-mlx'"
         )

hindsight-api-slim/hindsight_api/engine/db_utils.py

@@ -58,10 +58,16 @@ async def retry_with_backoff(
             last_exception = e
             if attempt < max_retries:
                 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..."
-                )
+                if isinstance(e, asyncpg.exceptions.DeadlockDetectedError):
+                    logger.warning(
+                        f"Deadlock detected during parallel document processing — this is expected and will resolve automatically "
+                        f"(attempt {attempt + 1}/{max_retries + 1}, retrying in {delay:.1f}s)"
+                    )
+                else:
+                    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}")

hindsight-api-slim/hindsight_api/engine/entity_resolver.py

@@ -75,6 +75,7 @@ class EntityResolver:
         """
         self.pool = pool
         self.entity_lookup = entity_lookup
+        self._pg_trgm_checked = False
         # Keyed by asyncio task id so concurrent retain batches never mix their
         # pending updates.  flush_pending_stats() pops only the calling task's items.
         self._pending_stats: dict[int, list[_EntityStat]] = {}
@@ -85,6 +86,19 @@ class EntityResolver:
         task = asyncio.current_task()
         return id(task) if task is not None else 0
 
+    def discard_pending_stats(self) -> None:
+        """
+        Discard accumulated entity stats and co-occurrence counts for the current task.
+
+        Call this on any exception path between resolve_entities_batch /
+        link_units_to_entities_batch and flush_pending_stats() to prevent the
+        per-task dicts from growing unbounded when tasks fail before flushing.
+        Safe to call even if no entries exist for the current task.
+        """
+        key = self._task_key()
+        self._pending_stats.pop(key, None)
+        self._pending_cooccurrences.pop(key, None)
+
     async def flush_pending_stats(self) -> None:
         """
         Flush accumulated entity stats and co-occurrence counts for the current task.
@@ -202,6 +216,20 @@ class EntityResolver:
         taxonomy_lookup: set[str] | None = None,
     ) -> list[str]:
         if self.entity_lookup == "trigram":
+            # Auto-detect pg_trgm availability on first call and fall back to
+            # "full" strategy if the extension is not installed.  See #626.
+            if not self._pg_trgm_checked:
+                self._pg_trgm_checked = True
+                has_trgm = await conn.fetchval("SELECT EXISTS(SELECT 1 FROM pg_extension WHERE extname = 'pg_trgm')")
+                if not has_trgm:
+                    logger.warning(
+                        "pg_trgm extension is not available — falling back to 'full' "
+                        "entity lookup strategy. Install pg_trgm for faster entity "
+                        "resolution on large banks. See: "
+                        "https://github.com/vectorize-io/hindsight/issues/626"
+                    )
+                    self.entity_lookup = "full"
+                    return await self._resolve_entities_batch_full(conn, bank_id, entities_data, unit_event_date)
             return await self._resolve_entities_batch_trigram(conn, bank_id, entities_data, unit_event_date)
         return await self._resolve_entities_batch_full(conn, bank_id, entities_data, unit_event_date)
 
@@ -459,10 +487,12 @@ class EntityResolver:
             entity_dates = [g.event_date for _, g in sorted_groups]
 
             # INSERT ... ON CONFLICT DO NOTHING — no row lock on already-existing entities.
+            # mention_count starts at 0 here; flush_pending_stats() is the sole source of
+            # truth for mention counting (one stat per original mention in the batch).
             inserted_rows = await conn.fetch(
                 f"""
                 INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
-                SELECT $1, name, COALESCE(event_date, now()), COALESCE(event_date, now()), 1
+                SELECT $1, name, COALESCE(event_date, now()), COALESCE(event_date, now()), 0
                 FROM unnest($2::text[], $3::timestamptz[]) AS t(name, event_date)
                 ON CONFLICT (bank_id, LOWER(canonical_name))
                 DO NOTHING
@@ -475,27 +505,52 @@ class EntityResolver:
             id_by_name: dict[str, str] = {row["name_lower"]: row["id"] for row in inserted_rows}
 
             # Fallback SELECT for names that conflicted (another worker won the race).
-            missing = [n for n, _ in sorted_groups if n not in id_by_name]
-            if missing:
+            #
+            # IMPORTANT: we must let PostgreSQL do the lowercasing on BOTH sides of the
+            # comparison.  Python's str.lower() and PostgreSQL's LOWER() differ for some
+            # Unicode characters — most notably Turkish İ (U+0130):
+            #   Python:     'İstanbul'.lower()  == 'i\u0307stanbul'  (i + combining dot, 2 chars)
+            #   PostgreSQL: LOWER('İstanbul')   == 'istanbul'         (plain i, 1 char)
+            # Passing a Python-lowercased name to "LOWER(canonical_name) = ANY($2::text[])"
+            # would fail to match the stored entity, leaving entity_id as None and causing
+            # a NOT NULL constraint violation on unit_entities.entity_id.
+            #
+            # Fix: pass the original (mixed-case) input names and use
+            # "LOWER(canonical_name) = ANY(SELECT LOWER(n) FROM unnest($2) AS n)" so
+            # PostgreSQL lowercases both sides identically.  The query also returns the
+            # original input_name so we can index id_by_name by Python's lower() of that
+            # name, which is what the assignment loop below uses as its lookup key.
+            missing_original = [g.name for name_lower, g in sorted_groups if name_lower not in id_by_name]
+            if missing_original:
                 existing_rows = await conn.fetch(
                     f"""
-                    SELECT id, LOWER(canonical_name) AS name_lower
-                    FROM {fq_table("entities")}
-                    WHERE bank_id = $1 AND LOWER(canonical_name) = ANY($2::text[])
+                    SELECT e.id, LOWER(e.canonical_name) AS name_lower, inputs.input_name
+                    FROM {fq_table("entities")} e
+                    JOIN (
+                        SELECT LOWER(n) AS input_name_lower, n AS input_name
+                        FROM unnest($2::text[]) AS n
+                    ) AS inputs ON LOWER(e.canonical_name) = inputs.input_name_lower
+                    WHERE e.bank_id = $1
                     """,
                     bank_id,
-                    missing,
+                    missing_original,
                 )
                 for row in existing_rows:
                     id_by_name[row["name_lower"]] = row["id"]
+                    # Also index by Python's lower() of the original input name so the
+                    # assignment loop (which uses Python-lowercased keys) finds it even
+                    # when Python and PostgreSQL produce different lowercase strings.
+                    id_by_name[row["input_name"].lower()] = row["id"]
 
-            # Assign entity IDs back and queue for post-txn stats flush.
+            # Assign entity IDs back and queue one stat per original mention so that
+            # flush_pending_stats() increments mention_count by the true mention count,
+            # not just 1 per unique name.
             for name_lower, g in sorted_groups:
                 entity_id = id_by_name.get(name_lower)
                 if entity_id:
                     for original_idx in g.indices:
                         entity_ids[original_idx] = entity_id
-                    pending.append(_EntityStat(entity_id=entity_id, event_date=g.event_date))
+                        pending.append(_EntityStat(entity_id=entity_id, event_date=g.event_date))
 
         # Accumulate into the resolver's pending list; the orchestrator flushes
         # these with await entity_resolver.flush_pending_stats() after the txn.

hindsight-api-slim/hindsight_api/engine/jina_mlx_reranker.py

@@ -0,0 +1,144 @@
+"""
+MLX implementation of jina-reranker-v3 for Apple Silicon.
+
+This file is adapted from the official model repository:
+  https://huggingface.co/jinaai/jina-reranker-v3-mlx/blob/main/rerank.py
+
+License: CC BY-NC 4.0 (contact Jina AI for commercial usage)
+
+Changes from upstream:
+- Removed the __main__ example block
+- Type annotations added to public methods
+- top_n parameter added to rerank() (upstream only exposed it implicitly)
+"""
+
+import numpy as np
+
+
+class _MLPProjector:
+    def __init__(self):
+        import mlx.nn as nn
+
+        self.linear1 = nn.Linear(1024, 512, bias=False)
+        self.linear2 = nn.Linear(512, 512, bias=False)
+
+    def __call__(self, x):
+        import mlx.nn as nn
+
+        x = self.linear1(x)
+        x = nn.relu(x)
+        x = self.linear2(x)
+        return x
+
+
+def _load_projector(projector_path: str) -> _MLPProjector:
+    import mlx.core as mx
+    from safetensors import safe_open
+
+    projector = _MLPProjector()
+    with safe_open(projector_path, framework="numpy") as f:
+        projector.linear1.weight = mx.array(f.get_tensor("linear1.weight"))
+        projector.linear2.weight = mx.array(f.get_tensor("linear2.weight"))
+    return projector
+
+
+def _sanitize(text: str, special_tokens: dict[str, str]) -> str:
+    for token in special_tokens.values():
+        text = text.replace(token, "")
+    return text
+
+
+def _format_prompt(query: str, docs: list[str], special_tokens: dict[str, str]) -> str:
+    query = _sanitize(query, special_tokens)
+    docs = [_sanitize(d, special_tokens) for d in docs]
+
+    doc_token = special_tokens["doc_embed_token"]
+    query_token = special_tokens["query_embed_token"]
+
+    prefix = (
+        "<|im_start|>system\n"
+        "You are a search relevance expert who can determine a ranking of the passages based on how relevant they are to the query. "
+        "If the query is a question, how relevant a passage is depends on how well it answers the question. "
+        "If not, try to analyze the intent of the query and assess how well each passage satisfies the intent. "
+        "If an instruction is provided, you should follow the instruction when determining the ranking."
+        "<|im_end|>\n<|im_start|>user\n"
+    )
+    suffix = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
+
+    body = (
+        f"I will provide you with {len(docs)} passages, each indicated by a numerical identifier. "
+        f"Rank the passages based on their relevance to query: {query}\n"
+    )
+    body += "\n".join(f'<passage id="{i}">\n{doc}{doc_token}\n</passage>' for i, doc in enumerate(docs))
+    body += f"\n<query>\n{query}{query_token}\n</query>"
+    return prefix + body + suffix
+
+
+class MLXReranker:
+    """
+    MLX-accelerated jina-reranker-v3 for Apple Silicon.
+
+    Loads the model from a local directory (use huggingface_hub.snapshot_download
+    to fetch jinaai/jina-reranker-v3-mlx if you don't have it already).
+    """
+
+    _SPECIAL_TOKENS = {
+        "query_embed_token": "<|rerank_token|>",
+        "doc_embed_token": "<|embed_token|>",
+    }
+    _DOC_TOKEN_ID = 151670
+    _QUERY_TOKEN_ID = 151671
+
+    def __init__(self, model_path: str, projector_path: str):
+        from mlx_lm import load
+
+        self.model, self.tokenizer = load(model_path)
+        self.model.eval()
+        self.projector = _load_projector(projector_path)
+
+    def rerank(self, query: str, documents: list[str], top_n: int | None = None) -> list[dict]:
+        """
+        Rank documents by relevance to a query.
+
+        Returns a list of dicts with keys: document, relevance_score, index.
+        Sorted by descending relevance_score.
+        """
+        import mlx.core as mx
+
+        prompt = _format_prompt(query, documents, self._SPECIAL_TOKENS)
+        input_ids = self.tokenizer.encode(prompt)
+        hidden_states = self.model.model([input_ids])[0]  # [seq_len, hidden_size]
+
+        input_ids_np = np.array(input_ids)
+        query_positions = np.where(input_ids_np == self._QUERY_TOKEN_ID)[0]
+        doc_positions = np.where(input_ids_np == self._DOC_TOKEN_ID)[0]
+
+        if len(query_positions) == 0:
+            raise ValueError("Query embed token not found in prompt")
+        if len(doc_positions) == 0:
+            raise ValueError("Document embed tokens not found in prompt")
+
+        query_hidden = mx.expand_dims(hidden_states[int(query_positions[0])], axis=0)
+        doc_hidden = mx.stack([hidden_states[int(p)] for p in doc_positions])
+
+        query_emb = self.projector(query_hidden)  # [1, 512]
+        doc_emb = self.projector(doc_hidden)  # [num_docs, 512]
+
+        query_exp = mx.broadcast_to(mx.expand_dims(query_emb, 0), (1, len(documents), 512))
+        doc_exp = mx.expand_dims(doc_emb, 0)
+
+        scores = mx.sum(doc_exp * query_exp, axis=-1) / (
+            mx.sqrt(mx.sum(doc_exp * doc_exp, axis=-1)) * mx.sqrt(mx.sum(query_exp * query_exp, axis=-1))
+        )  # [1, num_docs]
+        scores_np = np.array(scores[0])
+
+        order = np.argsort(scores_np)[::-1]
+        n = min(top_n, len(documents)) if top_n is not None else len(documents)
+        return [
+            {
+                "document": documents[order[i]],
+                "relevance_score": float(scores_np[order[i]]),
+                "index": int(order[i]),
+            }
+            for i in range(n)
+        ]