fix

feat: add iris as file parser
2026-02-18 13:44:37 +01:00 · 2026-02-18 12:06:04 +01:00
703 changed files with 23015 additions and 67910 deletions
@@ -5,7 +5,7 @@
 # Supported providers: openai, groq, ollama, gemini, anthropic, lmstudio, vertexai
 HINDSIGHT_API_LLM_PROVIDER=openai
 HINDSIGHT_API_LLM_API_KEY=your-api-key-here
-HINDSIGHT_API_LLM_MODEL=gpt-4o-mini
+HINDSIGHT_API_LLM_MODEL=o3-mini
 HINDSIGHT_API_LLM_BASE_URL=https://api.openai.com/v1

 # Example: Anthropic Claude configuration
@@ -31,9 +31,6 @@ jobs:
      - 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
        with:
          path: hindsight-docs/build
@@ -46,14 +46,6 @@ jobs:
      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)
    - name: Publish hindsight-client to PyPI
      uses: pypa/gh-action-pypi-publish@release/v1
@@ -85,18 +77,6 @@ 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
@@ -108,8 +88,6 @@ jobs:
          hindsight/dist/*
          hindsight-integrations/litellm/dist/*
          hindsight-embed/dist/*
-          hindsight-integrations/crewai/dist/*
-          hindsight-integrations/pydantic-ai/dist/*
        retention-days: 1

  release-typescript-client:
@@ -259,55 +237,6 @@ jobs:
        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
@@ -339,14 +268,11 @@ jobs:
    - name: Build
      run: npm run build --workspace=hindsight-control-plane

-    - name: Verify standalone build
-      run: test -f hindsight-control-plane/standalone/server.js || (echo 'standalone/server.js missing - build failed' && exit 1)
-
    - name: Publish to npm
      working-directory: ./hindsight-control-plane
      run: |
        set +e
-        OUTPUT=$(npm publish --access public --ignore-scripts 2>&1)
+        OUTPUT=$(npm publish --access public 2>&1)
        EXIT_CODE=$?
        echo "$OUTPUT"
        if [ $EXIT_CODE -ne 0 ]; then
@@ -387,10 +313,6 @@ jobs:
            target: aarch64-apple-darwin
            artifact_name: hindsight
            asset_name: hindsight-darwin-arm64
-          - os: ubuntu-24.04-arm
-            target: aarch64-unknown-linux-gnu
-            artifact_name: hindsight
-            asset_name: hindsight-linux-arm64

    steps:
    - uses: actions/checkout@v4
@@ -565,7 +487,7 @@ 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-openclaw-integration, release-ai-sdk-integration, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
    permissions:
      contents: write

@@ -600,12 +522,6 @@ jobs:
        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
      with:
@@ -644,7 +560,6 @@ jobs:
        cp artifacts/python-packages/hindsight-api/dist/* release-assets/ || true
        cp artifacts/python-packages/hindsight/dist/* release-assets/ || true
        cp artifacts/python-packages/hindsight-integrations/litellm/dist/* release-assets/ || true
-        cp artifacts/python-packages/hindsight-integrations/pydantic-ai/dist/* release-assets/ || true
        cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
        # TypeScript client
        cp artifacts/typescript-client/*.tgz release-assets/ || true
@@ -652,8 +567,6 @@ jobs:
        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
@@ -97,29 +97,6 @@ jobs:
      working-directory: ./hindsight-integrations/ai-sdk
      run: npm run build

-  build-chat-integration:
-    runs-on: ubuntu-latest
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Set up Node.js
-      uses: actions/setup-node@v4
-      with:
-        node-version: '22'
-
-    - name: Install dependencies
-      working-directory: ./hindsight-integrations/chat
-      run: npm ci
-
-    - name: Run tests
-      working-directory: ./hindsight-integrations/chat
-      run: npm test
-
-    - name: Build
-      working-directory: ./hindsight-integrations/chat
-      run: npm run build
-
  build-control-plane:
    runs-on: ubuntu-latest

@@ -194,9 +171,9 @@ jobs:
  test-rust-cli:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      HINDSIGHT_API_URL: http://localhost:8888
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
@@ -204,12 +181,6 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install Rust
      uses: dtolnay/rust-toolchain@stable

@@ -256,46 +227,25 @@ jobs:
      working-directory: ./hindsight-api
      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match

-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
    - name: Create .env file
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
        EOF

    - name: Start API server
      run: |
        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
+        for i in {1..60}; do
          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
            echo "API server is ready after ${i}s"
            break
          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
            cat /tmp/api-server.log
            exit 1
          fi
@@ -390,21 +340,12 @@ jobs:

    # Only test slim variants to save disk space (they're much smaller)
    # Slim variants require external embedding providers
-    - name: Setup GCP credentials for smoke test
-      if: matrix.variant == 'slim'
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Smoke test - verify container starts
      if: matrix.variant == 'slim'
      env:
-        HINDSIGHT_API_LLM_PROVIDER: vertexai
-        HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID: ${{ env.HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID }}
-        HINDSIGHT_API_EMBEDDINGS_PROVIDER: cohere
+        GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
+        HINDSIGHT_API_EMBEDDINGS_PROVIDER: openai
+        HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        HINDSIGHT_API_RERANKER_PROVIDER: cohere
        HINDSIGHT_API_COHERE_API_KEY: ${{ secrets.COHERE_API_KEY }}
      run: ./docker/test-image.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"
@@ -412,13 +353,14 @@ jobs:
  test-api:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
      GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
      OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
      COHERE_API_KEY: ${{ secrets.COHERE_API_KEY }}
      HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
@@ -426,12 +368,6 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -478,9 +414,9 @@ jobs:
  test-python-client:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      HINDSIGHT_API_URL: http://localhost:8888
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
@@ -489,12 +425,6 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -522,46 +452,25 @@ jobs:
      working-directory: ./hindsight-api
      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match

-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
    - name: Create .env file
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
        EOF

    - name: Start API server
      run: |
        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
+        for i in {1..60}; do
          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
            echo "API server is ready after ${i}s"
            break
          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
            cat /tmp/api-server.log
            exit 1
          fi
@@ -581,9 +490,9 @@ jobs:
  test-typescript-client:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      HINDSIGHT_API_URL: http://localhost:8888
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
@@ -592,12 +501,6 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -630,46 +533,25 @@ jobs:
      working-directory: ./hindsight-clients/typescript
      run: npm run build

-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
    - name: Create .env file
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
        EOF

    - name: Start API server
      run: |
        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
+        for i in {1..60}; do
          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
            echo "API server is ready after ${i}s"
            break
          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
            cat /tmp/api-server.log
            exit 1
          fi
@@ -686,36 +568,12 @@ jobs:
        echo "=== API Server Logs ==="
        cat /tmp/api-server.log || echo "No API server log found"

-  build-rust-cli-arm64:
-    runs-on: ubuntu-24.04-arm
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Install Rust
-      uses: dtolnay/rust-toolchain@stable
-      with:
-        targets: aarch64-unknown-linux-gnu
-
-    - name: Cache cargo
-      uses: actions/cache@v4
-      with:
-        path: |
-          ~/.cargo/registry
-          ~/.cargo/git
-          hindsight-cli/target
-        key: linux-arm64-cargo-${{ hashFiles('**/Cargo.lock') }}
-
-    - name: Build CLI
-      working-directory: hindsight-cli
-      run: cargo build --release --target aarch64-unknown-linux-gnu
-
  test-rust-client:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      HINDSIGHT_API_URL: http://localhost:8888
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
@@ -724,12 +582,6 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -761,46 +613,25 @@ jobs:
      working-directory: ./hindsight-api
      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match

-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
    - name: Create .env file
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
        EOF

    - name: Start API server
      run: |
        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
+        for i in {1..60}; do
          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
            echo "API server is ready after ${i}s"
            break
          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
            cat /tmp/api-server.log
            exit 1
          fi
@@ -820,9 +651,9 @@ jobs:
  test-go-client:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      HINDSIGHT_API_URL: http://localhost:8888
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
@@ -831,12 +662,6 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -862,56 +687,31 @@ jobs:
      working-directory: ./hindsight-api
      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match

-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
    - name: Create .env file
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
        EOF

    - name: Start API server
      run: |
        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
+        for i in {1..60}; do
          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
            echo "API server is ready after ${i}s"
            break
          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
            cat /tmp/api-server.log
            exit 1
          fi
          sleep 1
        done

-    - name: Build Go client
-      working-directory: ./hindsight-clients/go
-      run: go build ./...
-
    - name: Run Go client tests
      working-directory: ./hindsight-clients/go
      run: go test -v -tags=integration
@@ -922,120 +722,12 @@ jobs:
        echo "=== API Server Logs ==="
        cat /tmp/api-server.log || echo "No API server log found"

-  test-openclaw-integration:
-    runs-on: ubuntu-latest
-    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
-      HINDSIGHT_API_URL: http://localhost:8888
-      HINDSIGHT_EMBED_PACKAGE_PATH: ${{ github.workspace }}/hindsight-embed
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-        prune-cache: false
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Set up Node.js
-      uses: actions/setup-node@v4
-      with:
-        node-version: '22'
-
-    - name: Build API
-      working-directory: ./hindsight-api
-      run: uv build
-
-    - name: Install API dependencies
-      working-directory: ./hindsight-api
-      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
-
-    - name: Install embed dependencies
-      working-directory: ./hindsight-embed
-      run: uv sync --frozen --index-strategy unsafe-best-match
-
-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
-    - name: Install openclaw integration dependencies
-      working-directory: ./hindsight-integrations/openclaw
-      run: npm ci
-
-    - name: Create .env file
-      run: |
-        cat > .env << EOF
-        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
-        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
-        EOF
-
-    - name: Start API server
-      run: |
-        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
-        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
-          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
-            echo "API server is ready after ${i}s"
-            break
-          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
-            cat /tmp/api-server.log
-            exit 1
-          fi
-          sleep 1
-        done
-
-    - name: Run openclaw integration tests
-      working-directory: ./hindsight-integrations/openclaw
-      run: npm run test:integration
-
-    - name: Show API server logs
-      if: always()
-      run: |
-        echo "=== API Server Logs ==="
-        cat /tmp/api-server.log || echo "No API server log found"
-
  test-integration:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      HINDSIGHT_API_URL: http://localhost:8888
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
@@ -1043,12 +735,6 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -1096,22 +782,21 @@ jobs:
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
        EOF

    - name: Start API server
      run: |
        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
+        for i in {1..60}; do
          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
            echo "API server is ready after ${i}s"
            break
          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
            cat /tmp/api-server.log
            exit 1
          fi
@@ -1128,35 +813,6 @@ jobs:
        echo "=== API Server Logs ==="
        cat /tmp/api-server.log || echo "No API server log found"

-  test-crewai-integration:
-    runs-on: ubuntu-latest
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-        prune-cache: false
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Build crewai integration
-      working-directory: ./hindsight-integrations/crewai
-      run: uv build
-
-    - name: Install dependencies
-      working-directory: ./hindsight-integrations/crewai
-      run: uv sync --frozen
-
-    - name: Run tests
-      working-directory: ./hindsight-integrations/crewai
-      run: uv run pytest tests -v
-
  test-litellm-integration:
    runs-on: ubuntu-latest

@@ -1186,53 +842,18 @@ jobs:
      working-directory: ./hindsight-integrations/litellm
      run: uv run pytest tests -v

-  test-pydantic-ai-integration:
-    runs-on: ubuntu-latest
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-        prune-cache: false
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Build pydantic-ai integration
-      working-directory: ./hindsight-integrations/pydantic-ai
-      run: uv build
-
-    - name: Install dependencies
-      working-directory: ./hindsight-integrations/pydantic-ai
-      run: uv sync --frozen
-
-    - name: Run tests
-      working-directory: ./hindsight-integrations/pydantic-ai
-      run: uv run pytest tests -v
-
  test-embed:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      # Prefer CPU-only PyTorch in CI
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu

    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -1268,25 +889,19 @@ jobs:
  test-hindsight-all:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      # For test_server_integration.py compatibility
-      HINDSIGHT_LLM_PROVIDER: vertexai
-      HINDSIGHT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_LLM_PROVIDER: groq
+      HINDSIGHT_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_LLM_MODEL: openai/gpt-oss-20b
      # Prefer CPU-only PyTorch in CI
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu

    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -1321,15 +936,11 @@ jobs:

  test-doc-examples:
    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        language: [python, node, cli, go]
-    name: test-doc-examples (${{ matrix.language }})
+    needs: test-rust-cli
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      HINDSIGHT_API_URL: http://localhost:8888
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
@@ -1337,32 +948,14 @@ jobs:
    steps:
    - uses: actions/checkout@v4

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
-    - name: Install Rust
-      if: matrix.language == 'cli'
-      uses: dtolnay/rust-toolchain@stable
-
-    - name: Cache cargo
-      if: matrix.language == 'cli'
-      uses: actions/cache@v4
+    - name: Download CLI artifact
+      uses: actions/download-artifact@v4
      with:
-        path: |
-          ~/.cargo/registry
-          ~/.cargo/git
-          hindsight-cli/target
-        key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}
+        name: hindsight-cli
+        path: /usr/local/bin

-    - name: Build CLI
-      if: matrix.language == 'cli'
-      working-directory: hindsight-cli
-      run: |
-        cargo build --release
-        cp target/release/hindsight /usr/local/bin/hindsight
+    - name: Make CLI executable
+      run: chmod +x /usr/local/bin/hindsight

    - name: Install uv
      uses: astral-sh/setup-uv@v5
@@ -1376,7 +969,6 @@ jobs:
        python-version-file: ".python-version"

    - name: Set up Node.js
-      if: matrix.language == 'node'
      uses: actions/setup-node@v4
      with:
        node-version: '20'
@@ -1390,56 +982,33 @@ jobs:
        uv sync --frozen --no-install-project --index-strategy unsafe-best-match

    - name: Install Python client dependencies
-      if: matrix.language == 'python'
      working-directory: ./hindsight-clients/python
      run: uv sync --frozen --extra test --index-strategy unsafe-best-match

    - name: Install TypeScript client
-      if: matrix.language == 'node'
      run: |
        npm ci --workspace=hindsight-clients/typescript
        npm run build --workspace=hindsight-clients/typescript

-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading reranker model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
    - name: Create .env file
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json
-        HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID
        EOF

    - name: Start API server
      run: |
        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
        echo "Waiting for API server to be ready..."
-        for i in {1..120}; do
+        for i in {1..60}; do
          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
            echo "API server is ready after ${i}s"
            break
          fi
-          if [ $i -eq 120 ]; then
-            echo "API server failed to start after 120s"
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
            cat /tmp/api-server.log
            exit 1
          fi
@@ -1447,11 +1016,10 @@ jobs:
        done

    - name: Configure CLI
-      if: matrix.language == 'cli'
      run: hindsight configure --api-url http://localhost:8888

-    - name: Run doc examples (${{ matrix.language }})
-      run: ./scripts/test-doc-examples.sh --lang ${{ matrix.language }}
+    - name: Run all doc examples
+      run: ./scripts/test-doc-examples.sh

    - name: Show API server logs
      if: always()
@@ -1462,9 +1030,9 @@ jobs:
  test-upgrade:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_API_LLM_PROVIDER: vertexai
-      HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY: /tmp/gcp-credentials.json
-      HINDSIGHT_API_LLM_MODEL: google/gemini-2.5-flash-lite
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu

@@ -1473,12 +1041,6 @@ jobs:
      with:
        fetch-depth: 0  # Full history needed for git clone of tags

-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ secrets.GCP_VERTEXAI_CREDENTIALS }}' > /tmp/gcp-credentials.json
-        PROJECT_ID=$(jq -r '.project_id' /tmp/gcp-credentials.json)
-        echo "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=$PROJECT_ID" >> $GITHUB_ENV
-
    - name: Fetch tags
      run: git fetch --tags

@@ -154,7 +154,7 @@ Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`

 3. **Run migrations locally**:
   ```bash
-   # Set database URL and run migrations for the base schema plus all tenants
+   # Set database URL and run migrations
   uv run hindsight-admin run-db-migration

   # Run on a specific tenant schema
@@ -317,10 +317,10 @@ npm install
 Required env vars:
 - `HINDSIGHT_API_LLM_PROVIDER`: openai, anthropic, gemini, groq, ollama, lmstudio
 - `HINDSIGHT_API_LLM_API_KEY`: Your API key
- `HINDSIGHT_API_LLM_MODEL`: Model name (e.g., gpt-4o-mini, claude-sonnet-4-20250514)
+- `HINDSIGHT_API_LLM_MODEL`: Model name (e.g., o3-mini, claude-sonnet-4-20250514)

 Optional (uses local models by default):
 - `HINDSIGHT_API_EMBEDDINGS_PROVIDER`: local (default) or tei
 - `HINDSIGHT_API_RERANKER_PROVIDER`: local (default) or tei
 - `HINDSIGHT_API_DATABASE_URL`: External PostgreSQL (uses embedded pg0 by default)
- `HINDSIGHT_API_ENABLE_BANK_CONFIG_API`: Enable per-bank config API (default: true)
+- `HINDSIGHT_API_ENABLE_BANK_CONFIG_API`: Enable per-bank config API (default: false, disabled for security)
@@ -36,7 +36,7 @@ Hindsight is being used in production at Fortune 500 enterprises and by a growin

 ## Adding Hindsight to Your AI Agents

-The easiest way to use Hindsight with an existing agent is with the LLM Wrapper. You can add memory to your agent with 2 lines of code. That will swap your current LLM client out with the Hindsight wrapper. After that, memories will be stored and retrieved automatically as you make LLM calls.
+The easiest way use Hindsight with an existing agent is with the LLM Wrapper. You can add memory to your agent with 2 lines of code. That will swap your current LLM client out with the Hindsight wrapper. After that, memories will be stored and retrieved automatically as you make LLM calls.

 If you need more control over how and when your agent stores and recalls memories, there's also a simple API you can integrate with using the SDKs or directly via HTTP.

@@ -181,7 +181,7 @@ Satisfying these requirements in Hindsight is straightforward. When new user inp

 ![Overview](./hindsight-docs/static/img/hindsight-overview.webp)

-Most agent memory implementations rely on basic vector search or sometimes use a knowledge graph. Hindsight uses biomimetic data structures to organize agent memories in a way that is more like how human memory works:
+Most agent memory implementation rely on basic vector search or sometimes use a knowledge graph. Hindsight uses biomimetic data structures to organize agent memories in a way that is more like how human memory works:

 - **World:** Facts about the world ("The stove gets hot")
 - **Experiences:** Agent's own experiences ("I touched the stove and it really hurt")
@@ -307,5 +307,3 @@ MIT — see [LICENSE](./LICENSE)
 ---

 Built by [Vectorize.io](https://vectorize.io)
-
-<img src="https://umami-pixel.chris-latimer.workers.dev/?id=a8b043e6-6964-454d-80df-69b69d3f0d50&host=github.com&url=/vectorize-io/hindsight" width="1" height="1" alt="" />
@@ -170,11 +170,6 @@ RUN chown -R hindsight:hindsight /app

 USER hindsight

-# Create pg0 data directory as hindsight user so that Docker seeds new named
-# volumes with correct ownership (UID 1000) on first use, avoiding the
-# "Permission denied" error when mounting a fresh root-owned volume.
-RUN mkdir -p /home/hindsight/.pg0
-
 ENV PATH="/app/api/.venv/bin:${PATH}"

 # Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
@@ -326,11 +321,6 @@ RUN chown -R hindsight:hindsight /app

 USER hindsight

-# Create pg0 data directory as hindsight user so that Docker seeds new named
-# volumes with correct ownership (UID 1000) on first use, avoiding the
-# "Permission denied" error when mounting a fresh root-owned volume.
-RUN mkdir -p /home/hindsight/.pg0
-
 ENV PATH="/app/api/.venv/bin:${PATH}"

 # Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
@@ -77,32 +77,18 @@ 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
-    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
+    for i in {1..60}; do
+        if curl -sf http://localhost:8888/health &>/dev/null; then
            break
        fi
        sleep 1
    done
-
-    if [ "$api_ready" != "true" ]; then
-        echo "❌ API did not become healthy within ${API_STARTUP_WAIT_SECONDS}s"
-        exit 1
-    fi
 else
    echo "API disabled (HINDSIGHT_ENABLE_API=false)"
 fi
@@ -111,7 +97,7 @@ fi
 if [ "$ENABLE_CP" = "true" ]; then
    echo "🎛️  Starting Control Plane..."
    cd /app/control-plane
-    PORT="${HINDSIGHT_CP_PORT:-9999}" node server.js &
+    PORT=9999 node server.js &
    CP_PID=$!
    PIDS+=($CP_PID)
 else
@@ -124,7 +110,7 @@ echo "✅ Hindsight is running!"
 echo ""
 echo "📍 Access:"
 if [ "$ENABLE_CP" = "true" ]; then
-    echo "   Control Plane: http://localhost:${HINDSIGHT_CP_PORT:-9999}"
+    echo "   Control Plane: http://localhost:9999"
 fi
 if [ "$ENABLE_API" = "true" ]; then
    echo "   API:           http://localhost:8888"
@@ -13,9 +13,9 @@
 #   target  - Optional: 'cp-only' for control plane, otherwise assumes API image (default: api)
 #
 # Environment variables:
-#   HINDSIGHT_API_LLM_API_KEY                   - Required for API/standalone images (LLM verification)
-#   HINDSIGHT_API_LLM_PROVIDER                  - LLM provider (default: openai)
-#   HINDSIGHT_API_LLM_MODEL                     - LLM model (default: gpt-4o-mini)
+#   GROQ_API_KEY                                - Required for API/standalone images (LLM verification)
+#   HINDSIGHT_API_LLM_PROVIDER                  - LLM provider (default: groq)
+#   HINDSIGHT_API_LLM_MODEL                     - LLM model (default: llama-3.3-70b-versatile)
 #   HINDSIGHT_API_EMBEDDINGS_PROVIDER           - Embeddings provider (optional, for slim images: openai, cohere, tei)
 #   HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY     - OpenAI API key for embeddings (optional)
 #   HINDSIGHT_API_RERANKER_PROVIDER             - Reranker provider (optional, for slim images: cohere, tei)
@@ -34,7 +34,7 @@
 #   ./docker/test-image.sh hindsight-control-plane:test cp-only
 #
 #   # Test slim image with external providers
-#   export HINDSIGHT_API_LLM_API_KEY=sk_xxx
+#   export GROQ_API_KEY=gsk_xxx
 #   export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
 #   export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=sk-xxx
 #   export HINDSIGHT_API_RERANKER_PROVIDER=cohere
@@ -60,8 +60,8 @@ IMAGE="${1:-}"
 TARGET="${2:-api}"
 TIMEOUT="${SMOKE_TEST_TIMEOUT:-120}"
 CONTAINER_NAME="${SMOKE_TEST_CONTAINER_NAME:-hindsight-smoke-test}"
-LLM_PROVIDER="${HINDSIGHT_API_LLM_PROVIDER:-openai}"
-LLM_MODEL="${HINDSIGHT_API_LLM_MODEL:-gpt-4o-mini}"
+LLM_PROVIDER="${HINDSIGHT_API_LLM_PROVIDER:-groq}"
+LLM_MODEL="${HINDSIGHT_API_LLM_MODEL:-llama-3.3-70b-versatile}"

 # Validate arguments
 if [ -z "$IMAGE" ]; then
@@ -88,9 +88,9 @@ else
 fi

 # Check for required environment variables
-if [ "$NEEDS_LLM" = true ] && [ "$LLM_PROVIDER" != "vertexai" ] && [ -z "${HINDSIGHT_API_LLM_API_KEY:-}" ]; then
-    echo -e "${RED}Error: HINDSIGHT_API_LLM_API_KEY environment variable is required for API/standalone images${NC}"
-    echo "Set it with: export HINDSIGHT_API_LLM_API_KEY=your-api-key"
+if [ "$NEEDS_LLM" = true ] && [ -z "${GROQ_API_KEY:-}" ]; then
+    echo -e "${RED}Error: GROQ_API_KEY environment variable is required for API/standalone images${NC}"
+    echo "Set it with: export GROQ_API_KEY=your-api-key"
    exit 2
 fi

@@ -123,25 +123,9 @@ else
    # Build docker run command with required and optional env vars
    DOCKER_CMD="docker run -d --name $CONTAINER_NAME"
    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_PROVIDER=$LLM_PROVIDER"
-    if [ -n "${HINDSIGHT_API_LLM_API_KEY:-}" ]; then
-        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_API_KEY=${HINDSIGHT_API_LLM_API_KEY}"
-    fi
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_API_KEY=${GROQ_API_KEY}"
    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_MODEL=$LLM_MODEL"

-    # Add Vertex AI config if provider is vertexai
-    if [ "$LLM_PROVIDER" = "vertexai" ]; then
-        if [ -n "${HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY:-}" ]; then
-            DOCKER_CMD="$DOCKER_CMD -v ${HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY}:/tmp/gcp-credentials.json:ro"
-            DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json"
-        fi
-        if [ -n "${HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID:-}" ]; then
-            DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=${HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID}"
-        fi
-        if [ -n "${HINDSIGHT_API_LLM_VERTEXAI_REGION:-}" ]; then
-            DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_VERTEXAI_REGION=${HINDSIGHT_API_LLM_VERTEXAI_REGION}"
-        fi
-    fi
-
    # Add optional embeddings provider config
    if [ -n "${HINDSIGHT_API_EMBEDDINGS_PROVIDER:-}" ]; then
        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_EMBEDDINGS_PROVIDER=${HINDSIGHT_API_EMBEDDINGS_PROVIDER}"
@@ -6,17 +6,24 @@
 # It expects API keys to be set in environment variables.
 #
 # Usage:
+#   export GROQ_API_KEY=gsk_xxx
 #   export OPENAI_API_KEY=sk-xxx
 #   export COHERE_API_KEY=xxx
 #   ./docker/test-slim-local.sh
 #
 # Or inline:
-#   OPENAI_API_KEY=sk_xxx COHERE_API_KEY=xxx ./docker/test-slim-local.sh
+#   GROQ_API_KEY=gsk_xxx OPENAI_API_KEY=sk_xxx COHERE_API_KEY=xxx ./docker/test-slim-local.sh
 #

 set -euo pipefail

 # Check for required API keys
+if [ -z "${GROQ_API_KEY:-}" ]; then
+    echo "❌ Error: GROQ_API_KEY environment variable is required"
+    echo "Set it with: export GROQ_API_KEY=gsk_xxx"
+    exit 1
+fi
+
 if [ -z "${OPENAI_API_KEY:-}" ]; then
    echo "❌ Error: OPENAI_API_KEY environment variable is required"
    echo "Set it with: export OPENAI_API_KEY=sk-xxx"
@@ -34,10 +41,7 @@ IMAGE="${1:-hindsight-slim:test}"
 echo "Testing image: $IMAGE"
 echo ""

-# Set up LLM and external providers
-export HINDSIGHT_API_LLM_PROVIDER=openai
-export HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY
-export HINDSIGHT_API_LLM_MODEL=gpt-4o-mini
+# Set up external providers
 export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
 export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=$OPENAI_API_KEY
 export HINDSIGHT_API_RERANKER_PROVIDER=cohere
@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.4.17
-appVersion: "0.4.17"
+version: 0.4.11
+appVersion: "0.4.11"
 keywords:
  - ai
  - memory
@@ -46,4 +46,4 @@ __all__ = [
    "RemoteTEICrossEncoder",
    "LLMConfig",
 ]
-__version__ = "0.4.17"
+__version__ = "0.4.11"
@@ -14,8 +14,7 @@ from typing import Any
 import asyncpg
 import typer

-from ..config import DEFAULT_DATABASE_SCHEMA, HindsightConfig
-from ..extensions import TenantExtension, load_extension
+from ..config import HindsightConfig
 from ..pg0 import parse_pg0_url, resolve_database_url


@@ -215,81 +214,20 @@ def restore(
    typer.echo("Restore complete")


-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,
-    )
+async def _run_migration(db_url: str, schema: str = "public") -> None:
+    """Resolve database URL and run migrations."""
+    from ..migrations import 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)
-
-    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
+    run_migrations(resolved_url, schema=schema)


@app.command(name="run-db-migration")
 def run_db_migration(
-    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.",
-    ),
+    schema: str = typer.Option("public", "--schema", "-s", help="Database schema to run migrations on"),
 ):
    """Run database migrations to the latest version."""
    config = HindsightConfig.from_env()
@@ -299,21 +237,11 @@ def run_db_migration(
        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
        raise typer.Exit(1)

-    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...")
+    typer.echo(f"Running database migrations (schema: {schema})...")

-    schemas = asyncio.run(
-        _run_migration(
-            config.database_url,
-            schema=schema,
-            base_schema=config.database_schema,
-            embedding_dimension=embedding_dimension,
-        )
-    )
+    asyncio.run(_run_migration(config.database_url, schema))

-    typer.echo(f"Database migrations completed successfully for {len(schemas)} schema(s)")
+    typer.echo("Database migrations completed successfully")


 async def _decommission_worker(db_url: str, worker_id: str, schema: str = "public") -> int:
@@ -1,88 +0,0 @@
-"""Add text_signals column to memory_units for enriched BM25 indexing.
-
-text_signals stores a denormalized space-separated string of entity names
-(and future signals) to improve full-text search recall without polluting
-the stored fact text.
-
- vchord: text_signals included in tokenize() at insert time
- native: search_vector GENERATED column regenerated to include text_signals
- pg_textsearch: no change (index only supports a single base column)
-
-Revision ID: a2b3c4d5e6f7
-Revises: z1u2v3w4x5y6
-Create Date: 2026-02-28
-"""
-
-import os
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "a2b3c4d5e6f7"
-down_revision: str | Sequence[str] | None = "aa2b3c4d5e6f"
-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 _detect_text_search_extension() -> str:
-    return os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
-
-
-def upgrade() -> None:
-    schema = _get_schema_prefix()
-    table = f"{schema}memory_units"
-    text_search_ext = _detect_text_search_extension()
-
-    # Add text_signals column (nullable TEXT, populated at retain time)
-    op.execute(f"ALTER TABLE {table} ADD COLUMN IF NOT EXISTS text_signals TEXT")
-
-    if text_search_ext == "native":
-        # Native PostgreSQL: drop and recreate the GENERATED tsvector column to include text_signals
-        op.execute(f"ALTER TABLE {table} DROP COLUMN IF EXISTS search_vector")
-        op.execute(f"""
-            ALTER TABLE {table}
-            ADD COLUMN search_vector tsvector
-            GENERATED ALWAYS AS (
-                to_tsvector('english',
-                    COALESCE(text, '') || ' ' ||
-                    COALESCE(context, '') || ' ' ||
-                    COALESCE(text_signals, '')
-                )
-            ) STORED
-        """)
-        # Recreate GIN index (was dropped with the column)
-        op.execute(f"""
-            CREATE INDEX IF NOT EXISTS idx_memory_units_text_search
-            ON {table} USING gin(search_vector)
-        """)
-
-    # vchord: tokenize() call in fact_storage.py is updated to include text_signals at insert time
-    # pg_textsearch: no change — index operates on the base `text` column only
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-    table = f"{schema}memory_units"
-    text_search_ext = _detect_text_search_extension()
-
-    if text_search_ext == "native":
-        op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_text_search")
-        op.execute(f"ALTER TABLE {table} DROP COLUMN IF EXISTS search_vector")
-        op.execute(f"""
-            ALTER TABLE {table}
-            ADD COLUMN search_vector tsvector
-            GENERATED ALWAYS AS (
-                to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))
-            ) STORED
-        """)
-        op.execute(f"""
-            CREATE INDEX idx_memory_units_text_search
-            ON {table} USING gin(search_vector)
-        """)
-
-    op.execute(f"ALTER TABLE {table} DROP COLUMN IF EXISTS text_signals")
@@ -1,54 +0,0 @@
-"""Add GIN index on source_memory_ids for observation lookup performance
-
-Without this index, queries using the array overlap operator (&&) or array
-containment (@>) on source_memory_ids require a full sequential scan over all
-observation memory_units. At ~77k observations this was measured at 45ms per
-query, becoming a bottleneck during consolidation recall (57-64s timeouts) and
-user recall (18-27s average).
-
-The GIN index reduces these queries to index scans: 45ms → 0.049ms (927x
-speedup). Recall dropped from 18-27s to ~6s, and consolidation recall
-stabilised from timeout to ~15s.
-
-Created with CONCURRENTLY so the migration does not block reads or writes.
-CONCURRENTLY requires running outside a transaction block, so the migration
-emits an explicit COMMIT before the statement and uses IF NOT EXISTS for
-idempotency.
-
-Revision ID: a2b3c4d5e6f8
-Revises: f7g8h9i0j1k2
-Create Date: 2026-03-04
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "a2b3c4d5e6f8"
-down_revision: str | Sequence[str] | None = "f7g8h9i0j1k2"
-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()
-
-    # CREATE INDEX CONCURRENTLY cannot run inside a transaction block.
-    # Commit the current Alembic transaction first.
-    op.execute("COMMIT")
-    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"
-    )
-
-
-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")
@@ -1,36 +0,0 @@
-"""Make event_date nullable in memory_units to support timestamp-free content
-
-Revision ID: aa2b3c4d5e6f
-Revises: z1u2v3w4x5y6
-Create Date: 2026-03-02
-
-When callers retain content without a timestamp (e.g. fictional documents, static text),
-the event_date column should be allowed to be NULL rather than defaulting to utcnow().
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "aa2b3c4d5e6f"
-down_revision: str | Sequence[str] | None = "z1u2v3w4x5y6"
-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 ALTER COLUMN event_date DROP NOT NULL")
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-    # Backfill NULLs with now() before restoring the NOT NULL constraint
-    op.execute(f"UPDATE {schema}memory_units SET event_date = now() WHERE event_date IS NULL")
-    op.execute(f"ALTER TABLE {schema}memory_units ALTER COLUMN event_date SET NOT NULL")
@@ -1,68 +0,0 @@
-"""Add partial indexes on memory_units temporal date fields for fast temporal retrieval
-
-Revision ID: b3c4d5e6f7g8
-Revises: c1a2b3d4e5f6
-Create Date: 2026-03-02
-
-The temporal retrieval entry-point query filters memory_units by occurred_start,
-occurred_end, and mentioned_at using OR conditions. Without dedicated indexes the
-planner falls back to a sequential scan of all bank rows after applying the
-(bank_id, fact_type) index, then re-checks each date field.
-
-These three partial indexes give the planner bitmap-index scan options for the
-three most common date predicates, dramatically reducing the row set before any
-embedding computation is required.
-
-All indexes are created CONCURRENTLY so the migration does not block writes on
-memory_units during production deployments. CONCURRENTLY requires running outside
-a transaction block; see migrations.py for how this is handled safely.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "b3c4d5e6f7g8"
-down_revision: str | Sequence[str] | None = "c1a2b3d4e5f6"
-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()
-    # Partial index on occurred_start (covers "occurred_start BETWEEN $4 AND $5")
-    op.execute("COMMIT")
-    op.execute(
-        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_bank_occurred_start "
-        f"ON {schema}memory_units(bank_id, fact_type, occurred_start) "
-        f"WHERE occurred_start IS NOT NULL"
-    )
-    # Partial index on occurred_end (covers "occurred_end BETWEEN $4 AND $5")
-    op.execute("COMMIT")
-    op.execute(
-        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_bank_occurred_end "
-        f"ON {schema}memory_units(bank_id, fact_type, occurred_end) "
-        f"WHERE occurred_end IS NOT NULL"
-    )
-    # Partial index on mentioned_at (covers "mentioned_at BETWEEN $4 AND $5")
-    op.execute("COMMIT")
-    op.execute(
-        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_bank_mentioned_at "
-        f"ON {schema}memory_units(bank_id, fact_type, mentioned_at) "
-        f"WHERE mentioned_at 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_bank_mentioned_at")
-    op.execute("COMMIT")
-    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_bank_occurred_end")
-    op.execute("COMMIT")
-    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_bank_occurred_start")
@@ -1,34 +0,0 @@
-"""Backfill observation_scopes column if missing.
-
-This migration ensures observation_scopes exists even on databases that had
-revision z1u2v3w4x5y6 applied when it referred to the old text_signals migration
-(before it was renamed to a2b3c4d5e6f7). The ADD COLUMN IF NOT EXISTS makes this
-a no-op on databases that already have the column.
-
-Revision ID: b4c5d6e7f8a9
-Revises: a2b3c4d5e6f7
-Create Date: 2026-03-02
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "b4c5d6e7f8a9"
-down_revision: str | Sequence[str] | None = "a2b3c4d5e6f7"
-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}memory_units ADD COLUMN IF NOT EXISTS observation_scopes JSONB")
-
-
-def downgrade() -> None:
-    pass  # intentionally no-op — safe to leave the column in place
@@ -1,46 +0,0 @@
-"""Enable pg_trgm extension and add GIN trigram index on entities.canonical_name
-
-Revision ID: c1a2b3d4e5f6
-Revises: b4c5d6e7f8a9
-Create Date: 2026-03-02
-
-Index is created CONCURRENTLY so the migration does not block writes on entities
-during production deployments. CONCURRENTLY requires running outside a transaction
-block; see migrations.py for how this is handled safely.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "c1a2b3d4e5f6"
-down_revision: str | Sequence[str] | None = "b4c5d6e7f8a9"
-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:
-    # pg_trgm ships with every standard PostgreSQL installation 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")
-
-    schema = _get_schema_prefix()
-    # GIN index on canonical_name enables sub-millisecond trigram similarity queries
-    # (% operator, similarity()) instead of full-table scans across all bank entities.
-    op.execute("COMMIT")
-    op.execute(
-        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS entities_canonical_name_trgm_idx "
-        f"ON {schema}entities USING GIN (canonical_name gin_trgm_ops)"
-    )
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-    op.execute("COMMIT")
-    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}entities_canonical_name_trgm_idx")
-    # Note: not dropping pg_trgm extension as other indexes may depend on it
@@ -1,30 +0,0 @@
-"""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")
@@ -1,83 +0,0 @@
-"""Add covering and composite indexes to speed up link expansion graph retrieval.
-
-Two indexes target the two bottlenecks identified by EXPLAIN ANALYZE on a 17M-row
-memory_links table:
-
-1. idx_memory_links_to_type_weight  (to_unit_id, link_type, weight DESC)
-   The semantic incoming direction — finding facts that consider seeds as their
-   nearest neighbour — currently hits an expensive BitmapAnd of two separate
-   bitmap scans (to_unit_id bitmap ∩ link_type bitmap).  A composite index
-   on (to_unit_id, link_type) turns this into a single index scan and reduces
-   latency from ~36 ms to < 5 ms per query.
-
-2. idx_memory_links_entity_covering  (from_unit_id) INCLUDE (to_unit_id, entity_id)
-   WHERE link_type = 'entity'
-   The entity co-occurrence expansion uses COUNT(DISTINCT ml.entity_id) and
-   joins on ml.to_unit_id.  Without a covering index the planner must read
-   ~2 500 heap pages to fetch entity_id and to_unit_id after the bitmap index
-   scan, adding ~230 ms of random I/O.  INCLUDE adds those two columns to the
-   index leaf pages so the entire query can be served from the index (index-only
-   scan), eliminating the heap reads entirely.
-   Partial index (WHERE link_type = 'entity') keeps index size ~40 % smaller.
-
-Both indexes are created with CONCURRENTLY so the migration does not block
-concurrent reads or writes on memory_links.  CONCURRENTLY requires running
-outside a transaction block, so the migration emits an explicit COMMIT before
-each statement and uses IF NOT EXISTS for idempotency.
-
-Revision ID: d2e3f4a5b6c7
-Revises: b3c4d5e6f7g8
-Create Date: 2026-03-02
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "d2e3f4a5b6c7"
-down_revision: str | Sequence[str] | None = "b3c4d5e6f7g8"
-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()
-
-    # CREATE INDEX CONCURRENTLY cannot run inside a transaction block.
-    # Commit the current Alembic transaction, then issue each CONCURRENTLY
-    # statement in its own implicit autocommit transaction.
-    # IF NOT EXISTS makes each statement idempotent if the migration is retried.
-
-    # Index for the semantic *incoming* direction in link_expansion_retrieval.py.
-    # Replaces the BitmapAnd of idx_memory_links_to_unit ∩ idx_memory_links_link_type
-    # with a single composite index scan.
-    op.execute("COMMIT")
-    op.execute(
-        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_links_to_type_weight "
-        f"ON {schema}memory_links(to_unit_id, link_type, weight DESC)"
-    )
-
-    # Covering index for entity co-occurrence expansion.
-    # Enables an index-only scan: entity_id and to_unit_id are read from the
-    # index leaf pages instead of the heap, eliminating ~2 500 random heap-page
-    # reads per expansion query.
-    op.execute("COMMIT")
-    op.execute(
-        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_links_entity_covering "
-        f"ON {schema}memory_links(from_unit_id) "
-        f"INCLUDE (to_unit_id, entity_id) "
-        f"WHERE link_type = 'entity'"
-    )
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-    op.execute("COMMIT")
-    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_links_entity_covering")
-    op.execute("COMMIT")
-    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_links_to_type_weight")
@@ -1,62 +0,0 @@
-"""Add webhooks table and next_retry_at to async_operations.
-
-Webhook deliveries are handled as async_operations tasks (operation_type='webhook_delivery')
-rather than a dedicated webhook_deliveries table.
-
-Revision ID: e4f5a6b7c8d9
-Revises: d2e3f4a5b6c7
-Create Date: 2026-03-04
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "e4f5a6b7c8d9"
-down_revision: str | Sequence[str] | None = "d2e3f4a5b6c7"
-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"""
-        CREATE TABLE IF NOT EXISTS {schema}webhooks (
-            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            bank_id TEXT,
-            url TEXT NOT NULL,
-            secret TEXT,
-            event_types TEXT[] NOT NULL DEFAULT '{{}}',
-            enabled BOOLEAN NOT NULL DEFAULT TRUE,
-            created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
-            updated_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
-        )
-        """
-    )
-
-    # Index for bank-scoped webhook lookup
-    op.execute(f"CREATE INDEX IF NOT EXISTS idx_webhooks_bank_id ON {schema}webhooks(bank_id)")
-
-    # Add next_retry_at to async_operations for task-owned retry scheduling
-    op.execute(f"ALTER TABLE {schema}async_operations ADD COLUMN IF NOT EXISTS next_retry_at TIMESTAMPTZ NULL")
-
-    # Index for polling: status + next_retry_at
-    op.execute(
-        f"CREATE INDEX IF NOT EXISTS idx_async_operations_status_retry "
-        f"ON {schema}async_operations(status, next_retry_at)"
-    )
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_status_retry")
-    op.execute(f"ALTER TABLE {schema}async_operations DROP COLUMN IF EXISTS next_retry_at")
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_webhooks_bank_id")
-    op.execute(f"DROP TABLE IF EXISTS {schema}webhooks")
@@ -1,33 +0,0 @@
-"""Add http_config JSONB column to webhooks table.
-
-Stores HTTP delivery configuration (method, timeout, headers, params) as a
-single JSONB column rather than separate columns.
-
-Revision ID: f7g8h9i0j1k2
-Revises: e4f5a6b7c8d9
-Create Date: 2026-03-04
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "f7g8h9i0j1k2"
-down_revision: str | Sequence[str] | None = "e4f5a6b7c8d9"
-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}webhooks ADD COLUMN IF NOT EXISTS http_config JSONB NOT NULL DEFAULT '{{}}'")
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-    op.execute(f"ALTER TABLE {schema}webhooks DROP COLUMN IF EXISTS http_config")
@@ -1,35 +0,0 @@
-"""Add observation_scopes column to memory_units table
-
-Revision ID: z1u2v3w4x5y6
-Revises: a1b2c3d4e5f6
-Create Date: 2026-02-25
-
-Adds observation_scopes JSONB column to memory_units to control how observations
-are scoped during consolidation. Accepts "per_tag", "combined", or an explicit
-list of tag-set lists for custom multi-pass consolidation.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "z1u2v3w4x5y6"
-down_revision: str | Sequence[str] | None = "a1b2c3d4e5f6"
-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 observation_scopes JSONB")
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS observation_scopes")
@@ -8,48 +8,12 @@ from contextvars import ContextVar
 from fastmcp import FastMCP

 from hindsight_api import MemoryEngine
-from hindsight_api.config import _get_raw_config
 from hindsight_api.engine.memory_engine import _current_schema
 from hindsight_api.extensions import MCPExtension, load_extension
 from hindsight_api.extensions.tenant import AuthenticationError
 from hindsight_api.mcp_tools import MCPToolsConfig, register_mcp_tools
 from hindsight_api.models import RequestContext

-# All tools available in the system (explicit list — no wildcards)
-_ALL_TOOLS: frozenset[str] = frozenset(
-    {
-        "retain",
-        "recall",
-        "reflect",
-        "list_banks",
-        "create_bank",
-        "list_mental_models",
-        "get_mental_model",
-        "create_mental_model",
-        "update_mental_model",
-        "delete_mental_model",
-        "refresh_mental_model",
-        "list_directives",
-        "create_directive",
-        "delete_directive",
-        "list_memories",
-        "get_memory",
-        "delete_memory",
-        "list_documents",
-        "get_document",
-        "delete_document",
-        "list_operations",
-        "get_operation",
-        "cancel_operation",
-        "list_tags",
-        "get_bank",
-        "get_bank_stats",
-        "update_bank",
-        "delete_bank",
-        "clear_memories",
-    }
-)
-
 # Configure logging from HINDSIGHT_API_LOG_LEVEL environment variable
 _log_level_str = os.environ.get("HINDSIGHT_API_LOG_LEVEL", "info").lower()
 _log_level_map = {
@@ -114,15 +78,21 @@ def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
                   If False, only expose bank-scoped tools without bank_id parameters.

    Returns:
-        Configured FastMCP server instance
+        Configured FastMCP server instance with stateless_http enabled
    """
-    mcp = FastMCP("hindsight-mcp-server")
+    # Use stateless_http=True for Claude Code compatibility
+    mcp = FastMCP("hindsight-mcp-server", stateless_http=True)

-    global_config = _get_raw_config()
-
-    # Tools available for this mode (multi-bank exposes all tools; single-bank excludes bank-management tools)
-    _SINGLE_BANK_TOOLS: frozenset[str] = frozenset(
-        {
+    # Configure and register tools using shared module
+    config = MCPToolsConfig(
+        bank_id_resolver=get_current_bank_id,
+        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
+        include_bank_id_param=multi_bank,
+        tools=None
+        if multi_bank
+        else {
            "retain",
            "recall",
            "reflect",
@@ -132,40 +102,8 @@ def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
            "update_mental_model",
            "delete_mental_model",
            "refresh_mental_model",
-            "list_directives",
-            "create_directive",
-            "delete_directive",
-            "list_memories",
-            "get_memory",
-            "delete_memory",
-            "list_documents",
-            "get_document",
-            "delete_document",
-            "list_operations",
-            "get_operation",
-            "cancel_operation",
-            "list_tags",
-            "get_bank",
-            "update_bank",
-            "delete_bank",
-            "clear_memories",
-        }
-    )
-    base_tools: frozenset[str] | None = None if multi_bank else _SINGLE_BANK_TOOLS
-
-    # Apply global mcp_enabled_tools filter (env-level allowlist)
-    if global_config.mcp_enabled_tools is not None:
-        allowed = frozenset(global_config.mcp_enabled_tools)
-        base_tools = (base_tools if base_tools is not None else _ALL_TOOLS) & allowed
-
-    # Configure and register tools using shared module
-    config = MCPToolsConfig(
-        bank_id_resolver=get_current_bank_id,
-        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
-        include_bank_id_param=multi_bank,
-        tools=base_tools,
+        },  # Scoped tools for single-bank mode (excludes bank management: list_banks, create_bank)
+        retain_fire_and_forget=False,  # HTTP MCP supports sync/async modes
    )

    register_mcp_tools(mcp, memory, config)
@@ -273,9 +211,9 @@ class MCPMiddleware:
        else:
            # Create servers internally (for direct construction / tests)
            self.multi_bank_server = create_mcp_server(memory, multi_bank=True)
-            self.multi_bank_app = self.multi_bank_server.http_app(path="/", stateless_http=True)
+            self.multi_bank_app = self.multi_bank_server.http_app(path="/")
            self.single_bank_server = create_mcp_server(memory, multi_bank=False)
-            self.single_bank_app = self.single_bank_server.http_app(path="/", stateless_http=True)
+            self.single_bank_app = self.single_bank_server.http_app(path="/")

    def _get_header(self, scope: dict, name: str) -> str | None:
        """Extract a header value from ASGI scope."""
@@ -331,7 +269,7 @@ class MCPMiddleware:
                auth_tenant_id = auth_context.tenant_id
                auth_api_key_id = auth_context.api_key_id
            except AuthenticationError as e:
-                await self._send_error(send, 401, str(e), extra_headers=e.headers)
+                await self._send_error(send, 401, str(e))
                return

        # Set schema from tenant context so downstream DB queries use the correct schema
@@ -413,17 +351,14 @@ class MCPMiddleware:
            if schema_token is not None:
                _current_schema.reset(schema_token)

-    async def _send_error(self, send, status: int, message: str, extra_headers: dict[str, str] | None = None):
+    async def _send_error(self, send, status: int, message: str):
        """Send an error response."""
        body = json.dumps({"error": message}).encode()
-        headers = [(b"content-type", b"application/json")]
-        for key, value in (extra_headers or {}).items():
-            headers.append((key.encode(), value.encode()))
        await send(
            {
                "type": "http.response.start",
                "status": status,
-                "headers": headers,
+                "headers": [(b"content-type", b"application/json")],
            }
        )
        await send(
@@ -444,9 +379,9 @@ def create_mcp_servers(memory: MemoryEngine):
        Tuple of (multi_bank_server, single_bank_server, multi_bank_app, single_bank_app)
    """
    multi_bank_server = create_mcp_server(memory, multi_bank=True)
-    multi_bank_app = multi_bank_server.http_app(path="/", stateless_http=True)
+    multi_bank_app = multi_bank_server.http_app(path="/")

    single_bank_server = create_mcp_server(memory, multi_bank=False)
-    single_bank_app = single_bank_server.http_app(path="/", stateless_http=True)
+    single_bank_app = single_bank_server.http_app(path="/")

    return multi_bank_server, single_bank_server, multi_bank_app, single_bank_app
@@ -218,10 +218,6 @@ ENV_RERANKER_MAX_CANDIDATES = "HINDSIGHT_API_RERANKER_MAX_CANDIDATES"
 ENV_RERANKER_FLASHRANK_MODEL = "HINDSIGHT_API_RERANKER_FLASHRANK_MODEL"
 ENV_RERANKER_FLASHRANK_CACHE_DIR = "HINDSIGHT_API_RERANKER_FLASHRANK_CACHE_DIR"

-# ZeroEntropy configuration (reranker only)
-ENV_RERANKER_ZEROENTROPY_API_KEY = "HINDSIGHT_API_RERANKER_ZEROENTROPY_API_KEY"
-ENV_RERANKER_ZEROENTROPY_MODEL = "HINDSIGHT_API_RERANKER_ZEROENTROPY_MODEL"
-
 ENV_VECTOR_EXTENSION = "HINDSIGHT_API_VECTOR_EXTENSION"
 ENV_TEXT_SEARCH_EXTENSION = "HINDSIGHT_API_TEXT_SEARCH_EXTENSION"

@@ -232,12 +228,13 @@ ENV_LOG_LEVEL = "HINDSIGHT_API_LOG_LEVEL"
 ENV_LOG_FORMAT = "HINDSIGHT_API_LOG_FORMAT"
 ENV_WORKERS = "HINDSIGHT_API_WORKERS"
 ENV_MCP_ENABLED = "HINDSIGHT_API_MCP_ENABLED"
-ENV_MCP_ENABLED_TOOLS = "HINDSIGHT_API_MCP_ENABLED_TOOLS"
 ENV_ENABLE_BANK_CONFIG_API = "HINDSIGHT_API_ENABLE_BANK_CONFIG_API"
 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_MCP_LOCAL_BANK_ID = "HINDSIGHT_API_MCP_LOCAL_BANK_ID"
+ENV_MCP_INSTRUCTIONS = "HINDSIGHT_API_MCP_INSTRUCTIONS"
 ENV_MENTAL_MODEL_REFRESH_CONCURRENCY = "HINDSIGHT_API_MENTAL_MODEL_REFRESH_CONCURRENCY"

 # OpenTelemetry tracing configuration
@@ -252,18 +249,13 @@ ENV_LLM_VERTEXAI_PROJECT_ID = "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID"
 ENV_LLM_VERTEXAI_REGION = "HINDSIGHT_API_LLM_VERTEXAI_REGION"
 ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY = "HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY"

-# Gemini safety settings
-ENV_LLM_GEMINI_SAFETY_SETTINGS = "HINDSIGHT_API_LLM_GEMINI_SAFETY_SETTINGS"
-
 # Retain settings
 ENV_RETAIN_MAX_COMPLETION_TOKENS = "HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS"
 ENV_RETAIN_CHUNK_SIZE = "HINDSIGHT_API_RETAIN_CHUNK_SIZE"
 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_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"
 ENV_RETAIN_BATCH_POLL_INTERVAL_SECONDS = "HINDSIGHT_API_RETAIN_BATCH_POLL_INTERVAL_SECONDS"

@@ -280,7 +272,6 @@ 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"
@@ -291,21 +282,7 @@ ENV_FILE_DELETE_AFTER_RETAIN = "HINDSIGHT_API_FILE_DELETE_AFTER_RETAIN"
 # Observations settings (consolidated knowledge from facts)
 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"
-ENV_WEBHOOK_SECRET = "HINDSIGHT_API_WEBHOOK_SECRET"
-ENV_WEBHOOK_EVENT_TYPES = "HINDSIGHT_API_WEBHOOK_EVENT_TYPES"
-ENV_WEBHOOK_DELIVERY_POLL_INTERVAL_SECONDS = "HINDSIGHT_API_WEBHOOK_DELIVERY_POLL_INTERVAL_SECONDS"

 # Optimization flags
 ENV_SKIP_LLM_VERIFICATION = "HINDSIGHT_API_SKIP_LLM_VERIFICATION"
@@ -331,13 +308,6 @@ 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_MISSION = "HINDSIGHT_API_REFLECT_MISSION"
-
-# Disposition settings
-ENV_DISPOSITION_SKEPTICISM = "HINDSIGHT_API_DISPOSITION_SKEPTICISM"
-ENV_DISPOSITION_LITERALISM = "HINDSIGHT_API_DISPOSITION_LITERALISM"
-ENV_DISPOSITION_EMPATHY = "HINDSIGHT_API_DISPOSITION_EMPATHY"

 # Default values
 DEFAULT_DATABASE_URL = "pg0"
@@ -346,18 +316,18 @@ DEFAULT_LLM_PROVIDER = "openai"

 # Provider-specific default models
 PROVIDER_DEFAULT_MODELS = {
-    "openai": "gpt-4o-mini",
+    "openai": "o3-mini",
    "anthropic": "claude-haiku-4-5-20251001",
    "gemini": "gemini-2.5-flash",
    "groq": "openai/gpt-oss-120b",
    "ollama": "gemma3:12b",
    "lmstudio": "local-model",
-    "vertexai": "google/gemini-2.5-flash-lite",
+    "vertexai": "gemini-2.0-flash-001",
    "openai-codex": "gpt-5.2-codex",
    "claude-code": "claude-sonnet-4-5-20250929",
    "mock": "mock-model",
 }
-DEFAULT_LLM_MODEL = "gpt-4o-mini"  # Fallback if provider not in table
+DEFAULT_LLM_MODEL = "o3-mini"  # Fallback if provider not in table
 DEFAULT_LLM_MAX_CONCURRENT = 32
 DEFAULT_LLM_MAX_RETRIES = 10  # Max retry attempts for LLM API calls
 DEFAULT_LLM_INITIAL_BACKOFF = 1.0  # Initial backoff in seconds for retry exponential backoff
@@ -369,9 +339,6 @@ DEFAULT_LLM_VERTEXAI_PROJECT_ID = None  # Required for Vertex AI
 DEFAULT_LLM_VERTEXAI_REGION = "us-central1"
 DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY = None  # Optional, uses ADC if not set

-# Gemini safety settings defaults
-DEFAULT_LLM_GEMINI_SAFETY_SETTINGS = None  # None = use Gemini default safety settings
-
 DEFAULT_EMBEDDINGS_PROVIDER = "local"
 DEFAULT_EMBEDDINGS_LOCAL_MODEL = "BAAI/bge-small-en-v1.5"
 DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU = False  # Force CPU mode for local embeddings (avoids MPS/XPC issues on macOS)
@@ -395,8 +362,6 @@ DEFAULT_RERANKER_FLASHRANK_CACHE_DIR = None  # Use default cache directory
 DEFAULT_EMBEDDINGS_COHERE_MODEL = "embed-english-v3.0"
 DEFAULT_RERANKER_COHERE_MODEL = "rerank-english-v3.0"

-DEFAULT_RERANKER_ZEROENTROPY_MODEL = "zerank-2"
-
 # Vector extension (pgvector, vchord, or pgvectorscale)
 DEFAULT_VECTOR_EXTENSION = "pgvector"  # Options: "pgvector", "vchord", "pgvectorscale"

@@ -419,12 +384,12 @@ DEFAULT_LOG_LEVEL = "info"
 DEFAULT_LOG_FORMAT = "text"  # Options: "text", "json"
 DEFAULT_WORKERS = 1
 DEFAULT_MCP_ENABLED = True
-DEFAULT_MCP_ENABLED_TOOLS: list[str] | None = None  # None = all tools enabled
-DEFAULT_ENABLE_BANK_CONFIG_API = True
+DEFAULT_ENABLE_BANK_CONFIG_API = False  # Disabled by default for security
 DEFAULT_GRAPH_RETRIEVER = "link_expansion"  # Options: "link_expansion", "mpfp", "bfs"
 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_MCP_LOCAL_BANK_ID = "mcp"
 DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY = 8  # Max concurrent mental model refreshes

 # Retain settings
@@ -433,17 +398,14 @@ 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
-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_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)
 DEFAULT_RETAIN_BATCH_POLL_INTERVAL_SECONDS = 60  # Batch API polling interval in seconds

 # File storage defaults
 DEFAULT_FILE_STORAGE_TYPE = "native"  # PostgreSQL BYTEA storage
-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_PARSER = "markitdown"  # File parser to use (markitdown is the only supported parser)
 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
@@ -451,18 +413,8 @@ 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
+DEFAULT_CONSOLIDATION_MAX_TOKENS = 1024  # Max tokens for recall when finding related observations

 # Database migrations
 DEFAULT_RUN_MIGRATIONS_ON_STARTUP = True
@@ -484,12 +436,6 @@ 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
-
-# Disposition defaults (None = not set, fall back to bank DB value or 3)
-DEFAULT_DISPOSITION_SKEPTICISM = None
-DEFAULT_DISPOSITION_LITERALISM = None
-DEFAULT_DISPOSITION_EMPATHY = None

 # OpenTelemetry tracing configuration
 DEFAULT_OTEL_TRACES_ENABLED = False  # Disabled by default for backward compatibility
@@ -519,12 +465,6 @@ Use this tool PROACTIVELY to:
 # Default embedding dimension (used by initial migration, adjusted at runtime)
 EMBEDDING_DIMENSION = DEFAULT_EMBEDDING_DIMENSION

-# Webhook configuration defaults
-DEFAULT_WEBHOOK_URL = None  # None = no global webhook configured
-DEFAULT_WEBHOOK_SECRET = None  # None = no signing
-DEFAULT_WEBHOOK_EVENT_TYPES = "consolidation.completed"  # Comma-separated; default = all supported events
-DEFAULT_WEBHOOK_DELIVERY_POLL_INTERVAL_SECONDS = 30  # How often to poll for pending deliveries
-

 class JsonFormatter(logging.Formatter):
    """JSON formatter for structured logging.
@@ -556,11 +496,6 @@ 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()
@@ -606,9 +541,6 @@ class HindsightConfig:
    llm_vertexai_region: str
    llm_vertexai_service_account_key: str | None

-    # Gemini safety settings (None = use Gemini defaults; list of dicts with category/threshold)
-    llm_gemini_safety_settings: list | None
-
    # Per-operation LLM configuration (None = use default LLM config)
    retain_llm_provider: str | None
    retain_llm_api_key: str | None
@@ -676,8 +608,6 @@ class HindsightConfig:
    reranker_litellm_sdk_api_key: str | None
    reranker_litellm_sdk_model: str
    reranker_litellm_sdk_api_base: str | None
-    reranker_zeroentropy_api_key: str | None
-    reranker_zeroentropy_model: str

    # Server
    host: str
@@ -686,7 +616,6 @@ class HindsightConfig:
    log_level: str
    log_format: str
    mcp_enabled: bool
-    mcp_enabled_tools: list[str] | None  # None = all tools; explicit list = allowlist
    enable_bank_config_api: bool

    # Recall
@@ -701,12 +630,10 @@ class HindsightConfig:
    retain_chunk_size: int
    retain_extract_causal_links: bool
    retain_extraction_mode: str
-    retain_mission: str | None
    retain_custom_instructions: str | None
    retain_batch_tokens: int
    retain_batch_enabled: bool
    retain_batch_poll_interval_seconds: int
-    retain_entity_lookup: str  # "full" or "trigram"

    # File storage (static - server-level only)
    file_storage_type: str  # "native" (PostgreSQL) or "s3" (S3-compatible)
@@ -720,8 +647,7 @@ 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: 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: str  # File parser to use (e.g., "markitdown", "iris")
    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)
@@ -731,29 +657,8 @@ 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)
-    # List of label group dicts: [{key, description, type, optional, values: [{value, description}]}]
-    entity_labels: list | None
-    # Whether to extract regular named entities alongside entity labels (default: True)
-    # When False: only label entities are extracted (or no entities at all if no labels configured)
-    entities_allow_free_form: bool
-
-    # Reflect agent settings
-    reflect_mission: str | None
-
-    # Disposition settings (hierarchical - can be overridden per bank; None = fall back to DB)
-    disposition_skepticism: int | None
-    disposition_literalism: int | None
-    disposition_empathy: int | None

    # Optimization flags
    skip_llm_verification: bool
@@ -779,7 +684,6 @@ class HindsightConfig:

    # Reflect agent settings
    reflect_max_iterations: int
-    reflect_max_context_tokens: int

    # OpenTelemetry tracing configuration
    otel_traces_enabled: bool
@@ -788,12 +692,6 @@ class HindsightConfig:
    otel_service_name: str
    otel_deployment_environment: str

-    # Webhook configuration (static - server-level only, not per-bank)
-    webhook_url: str | None  # Global webhook URL (None = disabled)
-    webhook_secret: str | None  # HMAC signing secret (None = unsigned)
-    webhook_event_types: list[str]  # Event types to deliver globally
-    webhook_delivery_poll_interval_seconds: int  # How often the delivery worker polls
-
    # Class-level sets for configuration categorization

    # CREDENTIAL_FIELDS: Never exposed via API, never configurable per-tenant/bank
@@ -826,30 +724,12 @@ class HindsightConfig:
    # These fields are manually tagged as safe to expose and modify.
    # Excludes credentials, infrastructure config, provider/model selection, and performance tuning.
    _CONFIGURABLE_FIELDS = {
-        # MCP tool access control
-        "mcp_enabled_tools",
        # Retention settings (behavioral)
        "retain_chunk_size",
        "retain_extraction_mode",
-        "retain_mission",
        "retain_custom_instructions",
-        # 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",
-        # Disposition settings
-        "disposition_skepticism",
-        "disposition_literalism",
-        "disposition_empathy",
-        # Gemini safety settings (controls content filtering for Gemini/VertexAI providers)
-        "llm_gemini_safety_settings",
    }

    @property
@@ -970,8 +850,6 @@ class HindsightConfig:
            llm_vertexai_region=os.getenv(ENV_LLM_VERTEXAI_REGION, DEFAULT_LLM_VERTEXAI_REGION),
            llm_vertexai_service_account_key=os.getenv(ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY)
            or DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY,
-            # Gemini safety settings (JSON-encoded list of {category, threshold} dicts)
-            llm_gemini_safety_settings=json.loads(os.getenv(ENV_LLM_GEMINI_SAFETY_SETTINGS, "null")),
            # Per-operation LLM config (None = use default)
            retain_llm_provider=os.getenv(ENV_RETAIN_LLM_PROVIDER) or None,
            retain_llm_api_key=os.getenv(ENV_RETAIN_LLM_API_KEY) or None,
@@ -1104,9 +982,6 @@ class HindsightConfig:
            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),
            reranker_litellm_sdk_api_base=os.getenv(ENV_RERANKER_LITELLM_SDK_API_BASE) or None,
-            # ZeroEntropy reranker
-            reranker_zeroentropy_api_key=os.getenv(ENV_RERANKER_ZEROENTROPY_API_KEY),
-            reranker_zeroentropy_model=os.getenv(ENV_RERANKER_ZEROENTROPY_MODEL, DEFAULT_RERANKER_ZEROENTROPY_MODEL),
            # Server
            host=os.getenv(ENV_HOST, DEFAULT_HOST),
            port=int(os.getenv(ENV_PORT, DEFAULT_PORT)),
@@ -1114,9 +989,6 @@ class HindsightConfig:
            log_level=os.getenv(ENV_LOG_LEVEL, DEFAULT_LOG_LEVEL),
            log_format=os.getenv(ENV_LOG_FORMAT, DEFAULT_LOG_FORMAT).lower(),
            mcp_enabled=os.getenv(ENV_MCP_ENABLED, str(DEFAULT_MCP_ENABLED)).lower() == "true",
-            mcp_enabled_tools=[t.strip() for t in os.getenv(ENV_MCP_ENABLED_TOOLS).split(",") if t.strip()]
-            if os.getenv(ENV_MCP_ENABLED_TOOLS)
-            else DEFAULT_MCP_ENABLED_TOOLS,
            enable_bank_config_api=os.getenv(ENV_ENABLE_BANK_CONFIG_API, str(DEFAULT_ENABLE_BANK_CONFIG_API)).lower()
            == "true",
            # Recall
@@ -1144,10 +1016,8 @@ class HindsightConfig:
            retain_extraction_mode=_validate_extraction_mode(
                os.getenv(ENV_RETAIN_EXTRACTION_MODE, DEFAULT_RETAIN_EXTRACTION_MODE)
            ),
-            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_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()
            == "true",
            retain_batch_poll_interval_seconds=int(
@@ -1165,10 +1035,7 @@ 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=_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=os.getenv(ENV_FILE_PARSER, DEFAULT_FILE_PARSER),
            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(
@@ -1185,35 +1052,12 @@ 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))
            ),
-            consolidation_llm_batch_size=int(
-                os.getenv(ENV_CONSOLIDATION_LLM_BATCH_SIZE, str(DEFAULT_CONSOLIDATION_LLM_BATCH_SIZE))
-            ),
            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,
            # Database migrations
            run_migrations_on_startup=os.getenv(ENV_RUN_MIGRATIONS_ON_STARTUP, "true").lower() == "true",
            # Database connection pool
@@ -1233,20 +1077,6 @@ class HindsightConfig:
            ),
            # Reflect agent settings
            reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
-            reflect_max_context_tokens=int(
-                os.getenv(ENV_REFLECT_MAX_CONTEXT_TOKENS, str(DEFAULT_REFLECT_MAX_CONTEXT_TOKENS))
-            ),
-            reflect_mission=os.getenv(ENV_REFLECT_MISSION) or None,
-            # Disposition settings (None = fall back to DB value)
-            disposition_skepticism=int(os.getenv(ENV_DISPOSITION_SKEPTICISM))
-            if os.getenv(ENV_DISPOSITION_SKEPTICISM)
-            else DEFAULT_DISPOSITION_SKEPTICISM,
-            disposition_literalism=int(os.getenv(ENV_DISPOSITION_LITERALISM))
-            if os.getenv(ENV_DISPOSITION_LITERALISM)
-            else DEFAULT_DISPOSITION_LITERALISM,
-            disposition_empathy=int(os.getenv(ENV_DISPOSITION_EMPATHY))
-            if os.getenv(ENV_DISPOSITION_EMPATHY)
-            else DEFAULT_DISPOSITION_EMPATHY,
            # OpenTelemetry tracing configuration
            otel_traces_enabled=os.getenv(ENV_OTEL_TRACES_ENABLED, str(DEFAULT_OTEL_TRACES_ENABLED)).lower()
            in ("true", "1", "yes"),
@@ -1254,20 +1084,6 @@ class HindsightConfig:
            otel_exporter_otlp_headers=os.getenv(ENV_OTEL_EXPORTER_OTLP_HEADERS) or None,
            otel_service_name=os.getenv(ENV_OTEL_SERVICE_NAME, DEFAULT_OTEL_SERVICE_NAME),
            otel_deployment_environment=os.getenv(ENV_OTEL_DEPLOYMENT_ENVIRONMENT, DEFAULT_OTEL_DEPLOYMENT_ENVIRONMENT),
-            # Webhook configuration (static, server-level only)
-            webhook_url=os.getenv(ENV_WEBHOOK_URL) or DEFAULT_WEBHOOK_URL,
-            webhook_secret=os.getenv(ENV_WEBHOOK_SECRET) or DEFAULT_WEBHOOK_SECRET,
-            webhook_event_types=[
-                t.strip()
-                for t in os.getenv(ENV_WEBHOOK_EVENT_TYPES, DEFAULT_WEBHOOK_EVENT_TYPES).split(",")
-                if t.strip()
-            ],
-            webhook_delivery_poll_interval_seconds=int(
-                os.getenv(
-                    ENV_WEBHOOK_DELIVERY_POLL_INTERVAL_SECONDS,
-                    str(DEFAULT_WEBHOOK_DELIVERY_POLL_INTERVAL_SECONDS),
-                )
-            ),
        )
        config.validate()
        return config
@@ -16,7 +16,6 @@ from typing import Any
 import asyncpg

 from hindsight_api.config import HindsightConfig, _get_raw_config, normalize_config_dict
-from hindsight_api.engine.memory_engine import fq_table
 from hindsight_api.extensions.tenant import TenantExtension
 from hindsight_api.models import RequestContext

@@ -150,8 +149,8 @@ class ConfigResolver:
        try:
            async with self.pool.acquire() as conn:
                row = await conn.fetchrow(
-                    f"""
-                    SELECT config FROM {fq_table("banks")} WHERE bank_id = $1
+                    """
+                    SELECT config FROM banks WHERE bank_id = $1
                    """,
                    bank_id,
                )
@@ -242,8 +241,8 @@ class ConfigResolver:
        # Merge with existing config (JSONB || operator)
        async with self.pool.acquire() as conn:
            await conn.execute(
-                f"""
-                UPDATE {fq_table("banks")}
+                """
+                UPDATE banks
                SET config = config || $1::jsonb,
                    updated_at = now()
                WHERE bank_id = $2
@@ -263,9 +262,9 @@ class ConfigResolver:
        """
        async with self.pool.acquire() as conn:
            await conn.execute(
-                f"""
-                UPDATE {fq_table("banks")}
-                SET config = '{{}}'::jsonb,
+                """
+                UPDATE banks
+                SET config = '{}'::jsonb,
                    updated_at = now()
                WHERE bank_id = $1
                """,
@@ -1,83 +1,85 @@
 """Prompts for the consolidation engine."""

-# Default mission when no bank-specific mission is set
-_DEFAULT_MISSION = "Track every detail: names, numbers, dates, places, and relationships. Prefer specifics over abstractions, never generalise."
+CONSOLIDATION_SYSTEM_PROMPT = """You are a memory consolidation system. Your job is to convert facts into durable knowledge (observations) and merge with existing knowledge when appropriate.

-# Processing rules — always present regardless of mission
-_PROCESSING_RULES = """Processing rules (always apply):
- REDUNDANT: same info worded differently → UPDATE the existing observation.
- CONTRADICTION/UPDATE: capture both states with temporal markers ("used to X, now Y").
- RESOLVE REFERENCES: when a new fact provides a concrete value resolving a vague placeholder in an existing observation (e.g. "home country", "hometown", "birthplace", "native language", "her ex", "that city"), UPDATE the observation to embed the resolved value explicitly. Example: new fact says "grandma in Sweden" + existing observation says "moved from her home country" → update to "home country is Sweden".
- NEVER merge observations about different people or unrelated topics."""
+You must output ONLY valid JSON with no markdown code blocks or additional text. However, the "text" field within each observation should use markdown formatting (headers, lists, bold, etc.) for clarity and readability.

-# Data section — format placeholders {facts_text} and {observations_text} are substituted at call time
-_BATCH_DATA_SECTION = """
-NEW FACTS:
-{facts_text}
+## EXTRACT DURABLE KNOWLEDGE, NOT EPHEMERAL STATE
+Facts often describe events or actions. Extract the DURABLE KNOWLEDGE implied by the fact, not the transient state.

-EXISTING OBSERVATIONS (JSON array, pooled from recalls across all facts above):
+Examples of extracting durable knowledge:
+- "User moved to Room 203" -> "Room 203 exists" (location exists, not where user is now)
+- "User visited Acme Corp at Room 105" -> "Acme Corp is located in Room 105"
+- "User took the elevator to floor 3" -> "Floor 3 is accessible by elevator"
+- "User met Sarah at the lobby" -> "Sarah can be found at the lobby"
+
+DO NOT track current user position/state as knowledge - that changes constantly.
+DO track permanent facts learned from the user's actions.
+
+## PRESERVE SPECIFIC DETAILS
+Keep names, locations, numbers, and other specifics. Do NOT:
+- Abstract into general principles
+- Generate business insights
+- Make knowledge generic
+
+GOOD examples:
+- Fact: "John likes pizza" -> "John likes pizza"
+- Fact: "Alice works at Google" -> "Alice works at Google"
+
+BAD examples:
+- "John likes pizza" -> "Understanding dietary preferences helps..." (TOO ABSTRACT)
+- "User is at Room 203" -> "User is currently at Room 203" (EPHEMERAL STATE)
+
+## MERGE RULES (when comparing to existing observations):
+1. REDUNDANT: Same information worded differently → update existing
+2. CONTRADICTION: Opposite information about same topic → update with temporal markers showing change
+   Example: "Alex used to love pizza but now hates it" OR "Alex's pizza preference changed from love to hate"
+3. UPDATE: New state replacing old state → update showing the transition with "used to", "now", "changed from X to Y"
+
+## CRITICAL RULES:
+- NEVER merge facts about DIFFERENT people
+- NEVER merge unrelated topics (food preferences vs work vs hobbies)
+- When merging contradictions, the "text" field MUST capture BOTH states with temporal markers:
+  * Use "used to X, now Y" OR "changed from X to Y" OR "X but now Y"
+  * DO NOT just state the new fact - you MUST show the change
+- Keep observations focused on ONE specific topic per person
+- The "text" field MUST contain durable knowledge, not ephemeral state
+- Do NOT include "tags" in output - tags are handled automatically"""
+
+CONSOLIDATION_USER_PROMPT = """Analyze this new fact and consolidate into knowledge.
+{mission_section}
+NEW FACT: {fact_text}
+
+EXISTING OBSERVATIONS (JSON array with source memories and dates):
 {observations_text}

 Each observation includes:
 - id: unique identifier for updating
 - text: the observation content
 - proof_count: number of supporting memories
+- tags: visibility scope (handled automatically)
+- created_at/updated_at: when observation was created/modified
 - occurred_start/occurred_end: temporal range of source facts
 - source_memories: array of supporting facts with their text and dates

-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 unless the MISSION above explicitly targets such data (e.g. timestamped events, session state, screen content)"""
+Instructions:
+1. Extract DURABLE KNOWLEDGE from the new fact (not ephemeral state)
+2. Review source_memories in existing observations to understand evidence
+3. Check dates to detect contradictions or updates
+4. Compare with observations:
+   - Same topic → UPDATE with learning_id
+   - New topic → CREATE new observation
+   - Purely ephemeral → return []

-# Output format — JSON braces escaped as {{ }} so .format() leaves them literal
-_BATCH_OUTPUT_FORMAT = """
-Output a JSON object with three arrays.
+Output JSON array of actions (the "text" field should use markdown formatting for structure):
+[
+  {{"action": "update", "learning_id": "uuid-from-observations", "text": "## Updated Knowledge\n\n**Key point**: details here\n\n- Supporting detail 1\n- Supporting detail 2", "reason": "..."}},
+  {{"action": "create", "text": "## New Durable Knowledge\n\nDescription with **emphasis** and proper structure", "reason": "..."}}
+]

-## EXAMPLE
+Return [] if fact contains no durable knowledge.

-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"}}]}}
-
-Rules:
- "source_fact_ids": copy the EXACT UUID strings shown in brackets [uuid] from NEW FACTS — never use integers or positions.
- "observation_id": copy the EXACT "id" UUID string from EXISTING OBSERVATIONS.
- One create/update may reference multiple facts when they jointly support the observation.
- "deletes": only when an observation is directly superseded or contradicted by new facts.
- Do NOT include "tags" — handled automatically.
- Return {{"creates": [], "updates": [], "deletes": []}} if nothing durable is found."""
-
-
-def build_batch_consolidation_prompt(observations_mission: str | None = None) -> str:
-    """
-    Build the consolidation prompt for batch mode (multiple facts per LLM call).
-
-    The mission defines *what* to track (customisable per bank).
-    Processing rules and output format are always present regardless of mission.
-    """
-    mission = observations_mission or _DEFAULT_MISSION
-
-    return (
-        "You are a memory consolidation system. Synthesize facts into observations "
-        "and merge with existing observations when appropriate.\n\n"
-        f"## MISSION\n{mission}\n\n"
-        f"{_PROCESSING_RULES}" + _BATCH_DATA_SECTION + _BATCH_OUTPUT_FORMAT
-    )
+IMPORTANT: Format the "text" field with markdown for better readability:
+- Use headers, lists, bold/italic, tables where appropriate
+- CRITICAL: Add blank lines before and after block elements (tables, code blocks, lists)
+- Ensure proper spacing for markdown to render correctly"""
@@ -29,7 +29,6 @@ from ..config import (
    DEFAULT_RERANKER_PROVIDER,
    DEFAULT_RERANKER_TEI_BATCH_SIZE,
    DEFAULT_RERANKER_TEI_MAX_CONCURRENT,
-    DEFAULT_RERANKER_ZEROENTROPY_MODEL,
    ENV_RERANKER_COHERE_API_KEY,
    ENV_RERANKER_COHERE_MODEL,
    ENV_RERANKER_FLASHRANK_CACHE_DIR,
@@ -43,7 +42,6 @@ from ..config import (
    ENV_RERANKER_TEI_BATCH_SIZE,
    ENV_RERANKER_TEI_MAX_CONCURRENT,
    ENV_RERANKER_TEI_URL,
-    ENV_RERANKER_ZEROENTROPY_API_KEY,
 )

 logger = logging.getLogger(__name__)
@@ -558,104 +556,6 @@ class CohereCrossEncoder(CrossEncoderModel):
        return all_scores


-class ZeroEntropyCrossEncoder(CrossEncoderModel):
-    """
-    ZeroEntropy cross-encoder implementation using the ZeroEntropy Rerank API.
-
-    Supports zerank-2 (flagship) and zerank-2-small models.
-    See: https://docs.zeroentropy.dev/models
-    """
-
-    RERANK_URL = "https://api.zeroentropy.dev/v1/models/rerank"
-
-    def __init__(
-        self,
-        api_key: str,
-        model: str = DEFAULT_RERANKER_ZEROENTROPY_MODEL,
-        timeout: float = 60.0,
-    ):
-        """
-        Initialize ZeroEntropy cross-encoder client.
-
-        Args:
-            api_key: ZeroEntropy API key
-            model: ZeroEntropy rerank model name (default: zerank-2)
-            timeout: Request timeout in seconds (default: 60.0)
-        """
-        self.api_key = api_key
-        self.model = model
-        self.timeout = timeout
-        self._async_client: httpx.AsyncClient | None = None
-
-    @property
-    def provider_name(self) -> str:
-        return "zeroentropy"
-
-    async def initialize(self) -> None:
-        """Initialize the async HTTP client."""
-        if self._async_client is not None:
-            return
-
-        logger.info(f"Reranker: initializing ZeroEntropy provider with model {self.model}")
-        self._async_client = httpx.AsyncClient(
-            timeout=self.timeout,
-            headers={
-                "Authorization": f"Bearer {self.api_key}",
-                "Content-Type": "application/json",
-            },
-        )
-        logger.info("Reranker: ZeroEntropy provider initialized")
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """
-        Score query-document pairs using the ZeroEntropy Rerank API.
-
-        Args:
-            pairs: List of (query, document) tuples to score
-
-        Returns:
-            List of relevance scores
-        """
-        if self._async_client is None:
-            raise RuntimeError("Reranker not initialized. Call initialize() first.")
-
-        if not pairs:
-            return []
-
-        # Group pairs by query for efficient batching
-        query_groups: dict[str, list[tuple[int, str]]] = {}
-        for idx, (query, text) in enumerate(pairs):
-            if query not in query_groups:
-                query_groups[query] = []
-            query_groups[query].append((idx, text))
-
-        all_scores = [0.0] * len(pairs)
-
-        for query, indexed_texts in query_groups.items():
-            texts = [text for _, text in indexed_texts]
-            indices = [idx for idx, _ in indexed_texts]
-
-            response = await self._async_client.post(
-                self.RERANK_URL,
-                json={
-                    "model": self.model,
-                    "query": query,
-                    "documents": texts,
-                    "top_n": len(texts),
-                },
-            )
-            response.raise_for_status()
-            result = response.json()
-
-            # Map scores back to original positions
-            for item in result.get("results", []):
-                original_idx = item["index"]
-                score = item["relevance_score"]
-                all_scores[indices[original_idx]] = score
-
-        return all_scores
-
-
 class RRFPassthroughCrossEncoder(CrossEncoderModel):
    """
    Passthrough cross-encoder that preserves RRF scores without neural reranking.
@@ -1110,19 +1010,9 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
            model=config.reranker_litellm_sdk_model,
            api_base=config.reranker_litellm_sdk_api_base,
        )
-    elif provider == "zeroentropy":
-        api_key = config.reranker_zeroentropy_api_key
-        if not api_key:
-            raise ValueError(
-                f"{ENV_RERANKER_ZEROENTROPY_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'zeroentropy'"
-            )
-        return ZeroEntropyCrossEncoder(
-            api_key=api_key,
-            model=config.reranker_zeroentropy_model,
-        )
    elif provider == "rrf":
        return RRFPassthroughCrossEncoder()
    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', 'flashrank', 'litellm', 'litellm-sdk', 'rrf'"
        )
@@ -20,7 +20,6 @@ RETRYABLE_EXCEPTIONS = (
    asyncpg.exceptions.InterfaceError,
    asyncpg.exceptions.ConnectionDoesNotExistError,
    asyncpg.exceptions.TooManyConnectionsError,
-    asyncpg.exceptions.DeadlockDetectedError,
    OSError,
    ConnectionError,
    asyncio.TimeoutError,
@@ -58,16 +57,10 @@ async def retry_with_backoff(
            last_exception = e
            if attempt < max_retries:
                delay = min(base_delay * (2**attempt), max_delay)
-                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..."
-                    )
+                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}")
@@ -794,7 +794,6 @@ class LiteLLMSDKEmbeddings(Embeddings):
                "model": self.model,
                "input": ["test"],
                "api_key": self.api_key,
-                "encoding_format": "float",
            }
            if self.api_base:
                embed_kwargs["api_base"] = self.api_base
@@ -841,7 +840,6 @@ class LiteLLMSDKEmbeddings(Embeddings):
                    "model": self.model,
                    "input": batch,
                    "api_key": self.api_key,
-                    "encoding_format": "float",
                }
                if self.api_base:
                    embed_kwargs["api_base"] = self.api_base
@@ -5,10 +5,6 @@ Uses spaCy for entity extraction and implements resolution logic
 to disambiguate entities across memory units.
 """

-import asyncio
-import logging
-from collections import defaultdict
-from dataclasses import dataclass, field
 from datetime import UTC, datetime
 from difflib import SequenceMatcher

@@ -16,43 +12,6 @@ import asyncpg

 from .db_utils import acquire_with_retry
 from .memory_engine import fq_table
-from .retain.entity_labels import build_labels_lookup as _build_labels_lookup_from_config
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class _EntityToCreate:
-    """An entity that needs to be inserted (no matching candidate found)."""
-
-    idx: int
-    name: str
-    event_date: datetime | None
-
-
-@dataclass
-class _EntityStat:
-    """Stat accumulation entry for a resolved entity (post-transaction update)."""
-
-    entity_id: str
-    event_date: datetime | None
-
-
-@dataclass
-class _EntityStatAgg:
-    """Aggregated stats used when flushing pending updates."""
-
-    count: int = 0
-    max_date: datetime | None = None
-
-
-@dataclass
-class _CooccurrencePair:
-    """A (entity_id_1, entity_id_2) pair observed in a retain batch (for post-txn flush)."""
-
-    entity_id_1: str
-    entity_id_2: str
-

 # Load spaCy model (singleton)
 _nlp = None
@@ -63,95 +22,14 @@ class EntityResolver:
    Resolves entities to canonical IDs with disambiguation.
    """

-    def __init__(self, pool: asyncpg.Pool, entity_lookup: str = "full"):
+    def __init__(self, pool: asyncpg.Pool):
        """
        Initialize entity resolver.

        Args:
            pool: asyncpg connection pool
-            entity_lookup: Lookup strategy — "full" loads all bank entities then
-                matches in Python; "trigram" uses pg_trgm GIN index to fetch only
-                similar candidates per entity name (much faster for large banks).
        """
        self.pool = pool
-        self.entity_lookup = entity_lookup
-        # 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]] = {}
-        self._pending_cooccurrences: dict[int, list[_CooccurrencePair]] = {}
-
-    def _task_key(self) -> int:
-        """Return a unique key for the current asyncio task (or 0 for non-task context)."""
-        task = asyncio.current_task()
-        return id(task) if task is not None else 0
-
-    async def flush_pending_stats(self) -> None:
-        """
-        Flush accumulated entity stats and co-occurrence counts for the current task.
-
-        Must be called AFTER the retain transaction commits.  Pops only the items
-        accumulated by the calling asyncio task so concurrent retain batches never
-        flush each other's uncommitted entity IDs.
-        """
-        if self.pool is None:
-            return
-
-        key = self._task_key()
-        stats = self._pending_stats.pop(key, [])
-        cooccurrences = self._pending_cooccurrences.pop(key, [])
-
-        if not stats and not cooccurrences:
-            return
-
-        async with acquire_with_retry(self.pool) as conn:
-            if stats:
-                # Aggregate: sum counts and find max date per entity_id.
-                agg: dict[str, _EntityStatAgg] = defaultdict(_EntityStatAgg)
-                for s in stats:
-                    entry = agg[s.entity_id]
-                    entry.count += 1
-                    if s.event_date is not None:
-                        entry.max_date = s.event_date if entry.max_date is None else max(entry.max_date, s.event_date)
-
-                # Sort by entity_id so all concurrent workers acquire row locks in
-                # the same order — prevents circular lock dependencies (deadlocks).
-                rows = sorted((eid, a.count, a.max_date) for eid, a in agg.items())
-                await conn.executemany(
-                    f"""
-                    UPDATE {fq_table("entities")} SET
-                        mention_count = mention_count + $2,
-                        last_seen     = GREATEST(last_seen, $3)
-                    WHERE id = $1::uuid
-                    """,
-                    rows,
-                )
-
-            if cooccurrences:
-                # Aggregate: count occurrences per (entity_id_1, entity_id_2) pair.
-                coo_agg: dict[tuple[str, str], int] = {}
-                for c in cooccurrences:
-                    pair = (c.entity_id_1, c.entity_id_2)
-                    coo_agg[pair] = coo_agg.get(pair, 0) + 1
-
-                now = datetime.now(UTC)
-                # Sort by (entity_id_1, entity_id_2) for consistent lock ordering.
-                await conn.executemany(
-                    f"""
-                    INSERT INTO {fq_table("entity_cooccurrences")}
-                        (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
-                    VALUES ($1, $2, $3, $4)
-                    ON CONFLICT (entity_id_1, entity_id_2)
-                    DO UPDATE SET
-                        cooccurrence_count = {fq_table("entity_cooccurrences")}.cooccurrence_count + EXCLUDED.cooccurrence_count,
-                        last_cooccurred    = GREATEST({fq_table("entity_cooccurrences")}.last_cooccurred, EXCLUDED.last_cooccurred)
-                    """,
-                    sorted((e1, e2, count, now) for (e1, e2), count in coo_agg.items()),
-                )
-
-    @staticmethod
-    def _build_labels_lookup(entity_labels: list | None) -> set[str]:
-        """Build a set of valid 'key:value' entity label strings for fast lookup."""
-        return _build_labels_lookup_from_config(entity_labels)

    async def resolve_entities_batch(
        self,
@@ -160,7 +38,6 @@ class EntityResolver:
        context: str,
        unit_event_date,
        conn=None,
-        entity_labels: list | None = None,
    ) -> list[str]:
        """
        Resolve multiple entities in batch (MUCH faster than sequential).
@@ -181,34 +58,15 @@ class EntityResolver:
        if not entities_data:
            return []

-        taxonomy_lookup = self._build_labels_lookup(entity_labels)
        if conn is None:
            async with acquire_with_retry(self.pool) as conn:
-                return await self._resolve_entities_batch_impl(
-                    conn, bank_id, entities_data, context, unit_event_date, taxonomy_lookup
-                )
+                return await self._resolve_entities_batch_impl(conn, bank_id, entities_data, context, unit_event_date)
        else:
-            return await self._resolve_entities_batch_impl(
-                conn, bank_id, entities_data, context, unit_event_date, taxonomy_lookup
-            )
+            return await self._resolve_entities_batch_impl(conn, bank_id, entities_data, context, unit_event_date)

    async def _resolve_entities_batch_impl(
-        self,
-        conn,
-        bank_id: str,
-        entities_data: list[dict],
-        context: str,
-        unit_event_date,
-        taxonomy_lookup: set[str] | None = None,
+        self, conn, bank_id: str, entities_data: list[dict], context: str, unit_event_date
    ) -> list[str]:
-        if self.entity_lookup == "trigram":
-            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)
-
-    async def _resolve_entities_batch_full(
-        self, conn, bank_id: str, entities_data: list[dict], unit_event_date
-    ) -> list[str]:
-        """Original strategy: load all bank entities then match in Python."""
        # Query ALL candidates for this bank
        all_entities = await conn.fetch(
            f"""
@@ -272,103 +130,10 @@ class EntityResolver:
                    matching.append((ent_id, canonical_name, metadata, last_seen, mention_count))
            all_candidates[entity_text] = matching

-        return await self._resolve_from_candidates(
-            conn, bank_id, entities_data, unit_event_date, all_candidates, cooccurrence_map
-        )
-
-    async def _resolve_entities_batch_trigram(
-        self, conn, bank_id: str, entities_data: list[dict], unit_event_date
-    ) -> list[str]:
-        """
-        Trigram strategy: fetch only similar candidates per entity name using pg_trgm.
-
-        Instead of loading all bank entities (O(N)), uses a GIN trigram index to fetch
-        only the small set of candidates that are textually similar to each input name.
-        Reduces DB data transfer from 165K rows to ~5-20 rows per entity.
-        """
-        entity_texts = list(set(e["text"] for e in entities_data))
-
-        # Fetch candidates for all unique entity texts in a single batched query.
-        # The trigram % operator uses the GIN index; the substring conditions cover
-        # exact prefix/suffix matches that trigrams might miss at low similarity.
-        rows = await conn.fetch(
-            f"""
-            SELECT DISTINCT ON (e.id)
-                e.id, e.canonical_name, e.metadata, e.last_seen, e.mention_count,
-                q.query_text
-            FROM unnest($2::text[]) AS q(query_text)
-            JOIN {fq_table("entities")} e ON (
-                e.bank_id = $1
-                AND (
-                    e.canonical_name % q.query_text
-                    OR LOWER(e.canonical_name) LIKE '%' || LOWER(q.query_text) || '%'
-                    OR LOWER(q.query_text) LIKE '%' || LOWER(e.canonical_name) || '%'
-                )
-            )
-            """,
-            bank_id,
-            entity_texts,
-        )
-
-        # Group candidates by query_text
-        all_candidates: dict[str, list] = {t: [] for t in entity_texts}
-        candidate_ids: set = set()
-        for row in rows:
-            query_text = row["query_text"]
-            all_candidates[query_text].append(
-                (row["id"], row["canonical_name"], row["metadata"], row["last_seen"], row["mention_count"])
-            )
-            candidate_ids.add(row["id"])
-
-        # Fetch co-occurrences only for the candidate entities (not all bank entities)
-        cooccurrence_map: dict[str, set[str]] = {}
-        if candidate_ids:
-            candidate_id_list = list(candidate_ids)
-            cooc_rows = await conn.fetch(
-                f"""
-                SELECT ec.entity_id_1, ec.entity_id_2
-                FROM {fq_table("entity_cooccurrences")} ec
-                WHERE ec.entity_id_1 = ANY($1::uuid[])
-                   OR ec.entity_id_2 = ANY($1::uuid[])
-                """,
-                candidate_id_list,
-            )
-            # Build name lookup for co-occurrence mapping
-            id_to_name = {
-                row["id"]: row["canonical_name"].lower()
-                for cands in all_candidates.values()
-                for row in [{"id": c[0], "canonical_name": c[1]} for c in cands]
-            }
-            for row in cooc_rows:
-                eid1, eid2 = row["entity_id_1"], row["entity_id_2"]
-                if eid1 not in cooccurrence_map:
-                    cooccurrence_map[eid1] = set()
-                if eid2 not in cooccurrence_map:
-                    cooccurrence_map[eid2] = set()
-                if eid2 in id_to_name:
-                    cooccurrence_map[eid1].add(id_to_name[eid2])
-                if eid1 in id_to_name:
-                    cooccurrence_map[eid2].add(id_to_name[eid1])
-
-        return await self._resolve_from_candidates(
-            conn, bank_id, entities_data, unit_event_date, all_candidates, cooccurrence_map
-        )
-
-    async def _resolve_from_candidates(
-        self,
-        conn,
-        bank_id: str,
-        entities_data: list[dict],
-        unit_event_date,
-        all_candidates: dict[str, list],
-        cooccurrence_map: dict[str, set[str]],
-    ) -> list[str]:
-        """Shared scoring + upsert logic used by both lookup strategies."""
-
        # Resolve each entity using pre-fetched candidates
        entity_ids = [None] * len(entities_data)
-        entities_to_update: list[_EntityStat] = []
-        entities_to_create: list[_EntityToCreate] = []
+        entities_to_update = []  # (entity_id, event_date)
+        entities_to_create = []  # (idx, entity_data, event_date)

        for idx, entity_data in enumerate(entities_data):
            entity_text = entity_data["text"]
@@ -380,7 +145,7 @@ class EntityResolver:

            if not candidates:
                # Will create new entity
-                entities_to_create.append(_EntityToCreate(idx=idx, name=entity_text, event_date=entity_event_date))
+                entities_to_create.append((idx, entity_data, entity_event_date))
                continue

            # Score candidates
@@ -424,87 +189,73 @@ class EntityResolver:

            if best_score > threshold:
                entity_ids[idx] = best_candidate
-                entities_to_update.append(_EntityStat(entity_id=best_candidate, event_date=entity_event_date))
+                entities_to_update.append((best_candidate, entity_event_date))
            else:
-                entities_to_create.append(
-                    _EntityToCreate(idx=idx, name=entity_data["text"], event_date=entity_event_date)
-                )
+                entities_to_create.append((idx, entity_data, entity_event_date))

-        # Existing entities: IDs already known from the candidate SELECT above.
-        # No in-transaction UPDATE — mention_count/last_seen are stats deferred to
-        # flush_pending_stats() which the orchestrator calls after the transaction.
-        pending: list[_EntityStat] = list(entities_to_update)
+        # Batch update existing entities
+        if entities_to_update:
+            await conn.executemany(
+                f"""
+                UPDATE {fq_table("entities")} SET
+                    mention_count = mention_count + 1,
+                    last_seen = $2
+                WHERE id = $1::uuid
+                """,
+                entities_to_update,
+            )

-        # New entities: INSERT with DO NOTHING to avoid row locks on concurrent races.
-        # ON CONFLICT DO NOTHING returns nothing for rows that conflicted; we handle
-        # that rare case with a fallback SELECT.
+        # Batch create new entities using COPY + INSERT for maximum speed
+        # This handles duplicates via ON CONFLICT and returns all IDs
        if entities_to_create:
-            # Group by lowercase name — deduplicate within the batch.
-            @dataclass
-            class _NameGroup:
-                name: str
-                event_date: datetime | None
-                indices: list[int] = field(default_factory=list)
+            # Group entities by canonical name (lowercase) to handle duplicates within batch
+            # For duplicates, we only insert once and reuse the ID, but track the count
+            unique_entities = {}  # lowercase_name -> (entity_data, event_date, [indices])
+            for idx, entity_data, event_date in entities_to_create:
+                name_lower = entity_data["text"].lower()
+                if name_lower not in unique_entities:
+                    unique_entities[name_lower] = (entity_data, event_date, [idx])
+                else:
+                    # Same entity appears multiple times - add index to list
+                    unique_entities[name_lower][2].append(idx)

-            groups: dict[str, _NameGroup] = {}
-            for e in entities_to_create:
-                name_lower = e.name.lower()
-                if name_lower not in groups:
-                    groups[name_lower] = _NameGroup(name=e.name, event_date=e.event_date)
-                groups[name_lower].indices.append(e.idx)
+            # Batch insert unique entities and get their IDs
+            # Use a single query with unnest for speed
+            entity_names = []
+            entity_dates = []
+            entity_counts = []  # Track how many times each entity appears in this batch
+            indices_map = []  # Maps result index -> list of original indices

-            # Sort by lowercase name for deterministic ordering.
-            sorted_groups = sorted(groups.items())
-            entity_names = [g.name for _, g in sorted_groups]
-            entity_dates = [g.event_date for _, g in sorted_groups]
+            for name_lower, (entity_data, event_date, indices) in unique_entities.items():
+                entity_names.append(entity_data["text"])
+                entity_dates.append(event_date)
+                entity_counts.append(len(indices))  # Count of occurrences in this batch
+                indices_map.append(indices)

-            # 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(
+            # Batch INSERT ... ON CONFLICT with RETURNING
+            # Uses the batch count for mention_count instead of always 1
+            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()), 0
-                FROM unnest($2::text[], $3::timestamptz[]) AS t(name, event_date)
+                SELECT $1, name, event_date, event_date, cnt
+                FROM unnest($2::text[], $3::timestamptz[], $4::int[]) AS t(name, event_date, cnt)
                ON CONFLICT (bank_id, LOWER(canonical_name))
-                DO NOTHING
-                RETURNING id, LOWER(canonical_name) AS name_lower
+                DO UPDATE SET
+                    mention_count = {fq_table("entities")}.mention_count + EXCLUDED.mention_count,
+                    last_seen = EXCLUDED.last_seen
+                RETURNING id
                """,
                bank_id,
                entity_names,
                entity_dates,
+                entity_counts,
            )
-            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:
-                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[])
-                    """,
-                    bank_id,
-                    missing,
-                )
-                for row in existing_rows:
-                    id_by_name[row["name_lower"]] = row["id"]
-
-            # 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))
-
-        # Accumulate into the resolver's pending list; the orchestrator flushes
-        # these with await entity_resolver.flush_pending_stats() after the txn.
-        key = self._task_key()
-        self._pending_stats.setdefault(key, []).extend(pending)
+            # Map returned IDs back to original indices
+            for result_idx, row in enumerate(rows):
+                entity_id = row["id"]
+                for original_idx in indices_map[result_idx]:
+                    entity_ids[original_idx] = entity_id

        return entity_ids

@@ -657,7 +408,7 @@ class EntityResolver:
        entity_id = await conn.fetchval(
            f"""
            INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
-            VALUES ($1, $2, COALESCE($3, now()), COALESCE($4, now()), 1)
+            VALUES ($1, $2, $3, $4, 1)
            ON CONFLICT (bank_id, LOWER(canonical_name))
            DO UPDATE SET
                mention_count = {fq_table("entities")}.mention_count + 1,
@@ -790,14 +541,19 @@ class EntityResolver:
                        entity_id_1, entity_id_2 = entity_id_2, entity_id_1
                    cooccurrence_pairs.add((entity_id_1, entity_id_2))

-        # Accumulate co-occurrence pairs for post-transaction flush.
-        # The actual INSERT/UPDATE is deferred to flush_pending_stats() to avoid
-        # row-level lock contention (ON CONFLICT DO UPDATE inside a long transaction
-        # serialises concurrent writers on popular entity pairs).
+        # Batch update co-occurrences
        if cooccurrence_pairs:
-            key = self._task_key()
-            self._pending_cooccurrences.setdefault(key, []).extend(
-                _CooccurrencePair(entity_id_1=e1, entity_id_2=e2) for e1, e2 in cooccurrence_pairs
+            now = datetime.now(UTC)
+            await conn.executemany(
+                f"""
+                INSERT INTO {fq_table("entity_cooccurrences")} (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
+                VALUES ($1, $2, $3, $4)
+                ON CONFLICT (entity_id_1, entity_id_2)
+                DO UPDATE SET
+                    cooccurrence_count = {fq_table("entity_cooccurrences")}.cooccurrence_count + 1,
+                    last_cooccurred = EXCLUDED.last_cooccurred
+                """,
+                [(e1, e2, 1, now) for e1, e2 in cooccurrence_pairs],
            )

    async def get_units_by_entity(self, entity_id: str, limit: int = 100) -> list[str]:
@@ -12,7 +12,6 @@ from typing import TYPE_CHECKING, Any
 if TYPE_CHECKING:
    from hindsight_api.engine.memory_engine import Budget
    from hindsight_api.engine.response_models import RecallResult, ReflectResult
-    from hindsight_api.engine.search.tags import TagsMatch
    from hindsight_api.models import RequestContext


@@ -338,8 +337,6 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        search_query: str | None = None,
-        tags: list[str] | None = None,
-        tags_match: "TagsMatch" = "any_strict",
        limit: int = 100,
        offset: int = 0,
        request_context: "RequestContext",
@@ -349,9 +346,7 @@ class MemoryEngineInterface(ABC):

        Args:
            bank_id: The memory bank ID.
-            search_query: Case-insensitive substring filter on document ID.
-            tags: Filter by tags.
-            tags_match: How to match tags (any, all, any_strict, all_strict).
+            search_query: Search query.
            limit: Maximum results.
            offset: Pagination offset.
            request_context: Request context for authentication.
@@ -48,28 +48,6 @@ _llm_max_concurrent = int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_
 _global_llm_semaphore = asyncio.Semaphore(_llm_max_concurrent)


-def sanitize_llm_output(text: str | None) -> str | None:
-    """
-    Sanitize text by removing characters that break downstream systems.
-
-    Removes:
-    - ASCII control characters (0x00-0x08, 0x0B-0x0C, 0x0E-0x1F, 0x7F): break
-      json.loads and PostgreSQL UTF-8 encoding; tab (0x09), newline (0x0A), and
-      carriage return (0x0D) are preserved as they are valid in text and JSON.
-    - Unicode surrogates (U+D800-U+DFFF): Invalid in UTF-8, break LLM APIs
-
-    Surrogate characters are used in UTF-16 encoding but cannot be encoded
-    in UTF-8. They can appear in Python strings from improperly decoded data
-    (e.g., from JavaScript or broken files). Control characters commonly appear
-    in LLM output embedded inside JSON string values.
-    """
-    if text is None:
-        return None
-    if not text:
-        return text
-    return re.sub(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f\ud800-\udfff]", "", text)
-
-
 class OutputTooLongError(Exception):
    """
    Bridge exception raised when LLM output exceeds token limits.
@@ -82,59 +60,6 @@ class OutputTooLongError(Exception):
    pass


-def parse_llm_json(raw: str) -> Any:
-    """
-    Robustly parse JSON returned by an LLM.
-
-    Handles common LLM output quirks:
-    1. Markdown code fences (```json ... ```) — strip them before parsing.
-    2. Embedded control characters (\\x00-\\x1f, \\x7f) — replace with space
-       and retry if the initial parse fails.
-
-    Args:
-        raw: Raw text returned by the LLM.
-
-    Returns:
-        Parsed Python object (dict, list, etc.).
-
-    Raises:
-        json.JSONDecodeError: If the text cannot be parsed even after cleanup.
-    """
-    text = raw.strip()
-
-    # Strip markdown code fences (some models wrap JSON in ```json ... ```)
-    if text.startswith("```"):
-        text = text.split("\n", 1)[1] if "\n" in text else text[3:]
-        if text.endswith("```"):
-            text = text[:-3]
-        text = text.strip()
-
-    try:
-        return json.loads(text)
-    except json.JSONDecodeError:
-        # Some models (e.g. Gemini) embed raw control characters inside JSON
-        # string values. Replacing them with a space usually produces valid JSON.
-        cleaned = re.sub(r"[\x00-\x1f\x7f]", " ", text)
-        return json.loads(cleaned)
-
-
-_PROVIDERS_WITHOUT_API_KEY = frozenset(
-    {
-        "ollama",
-        "lmstudio",
-        "openai-codex",
-        "claude-code",
-        "mock",
-        "vertexai",
-    }
-)
-
-
-def requires_api_key(provider: str) -> bool:
-    """Return True if the given provider requires an API key to operate."""
-    return provider.lower() not in _PROVIDERS_WITHOUT_API_KEY
-
-
 def create_llm_provider(
    provider: str,
    api_key: str,
@@ -146,7 +71,6 @@ def create_llm_provider(
    vertexai_project_id: str | None = None,
    vertexai_region: str | None = None,
    vertexai_credentials: Any = None,
-    gemini_safety_settings: list | None = None,
 ) -> Any:  # Returns LLMInterface
    """
    Factory function to create the appropriate LLM provider implementation.
@@ -215,7 +139,6 @@ def create_llm_provider(
            vertexai_project_id=vertexai_project_id,
            vertexai_region=vertexai_region,
            vertexai_credentials=vertexai_credentials,
-            gemini_safety_settings=gemini_safety_settings,
        )

    elif provider_lower == "anthropic":
@@ -258,7 +181,6 @@ class LLMProvider:
        reasoning_effort: str = "low",
        groq_service_tier: str | None = None,
        openai_service_tier: str | None = None,
-        gemini_safety_settings: list | None = None,
    ):
        """
        Initialize LLM provider.
@@ -271,7 +193,6 @@ class LLMProvider:
            reasoning_effort: Reasoning effort level for supported providers.
            groq_service_tier: Groq service tier ("on_demand", "flex", "auto") - from config.
            openai_service_tier: OpenAI service tier (None or "flex") - from config.
-            gemini_safety_settings: Safety settings for Gemini/VertexAI providers.
        """
        self.provider = provider.lower()
        self.api_key = api_key
@@ -281,8 +202,6 @@ class LLMProvider:
        # Service tiers from hierarchical config (not env vars)
        self.groq_service_tier = groq_service_tier
        self.openai_service_tier = openai_service_tier
-        # Gemini safety settings (instance default; can be overridden per-request via context var)
-        self.gemini_safety_settings = gemini_safety_settings

        # Validate provider
        valid_providers = [
@@ -351,18 +270,6 @@ class LLMProvider:
                f"model={self.model}, auth={'service_account' if service_account_key else 'ADC'}"
            )

-        # For Gemini/VertexAI providers: read safety settings from global config if not explicitly provided
-        # Use _get_raw_config() to bypass StaticConfigProxy (which blocks configurable fields),
-        # since LLMProvider initialization legitimately needs the server-level default.
-        if self.provider in ("gemini", "vertexai") and self.gemini_safety_settings is None:
-            from ..config import _get_raw_config
-
-            try:
-                raw_config = _get_raw_config()
-                self.gemini_safety_settings = raw_config.llm_gemini_safety_settings
-            except Exception:
-                pass  # Config may not be initialized in test environments
-
        # Create provider implementation using factory
        self._provider_impl = create_llm_provider(
            provider=self.provider,
@@ -375,7 +282,6 @@ class LLMProvider:
            vertexai_project_id=vertexai_project_id,
            vertexai_region=vertexai_region,
            vertexai_credentials=vertexai_credentials,
-            gemini_safety_settings=self.gemini_safety_settings,
        )

        # Backward compatibility: Keep mock provider properties
@@ -544,14 +450,6 @@ class LLMProvider:

            return result

-    def set_response_callback(self, fn: Any) -> None:
-        """Set a callback invoked on each call() instead of the fixed mock response."""
-        if self.provider == "mock":
-            from .providers.mock_llm import MockLLM
-
-            if isinstance(self._provider_impl, MockLLM):
-                self._provider_impl.set_response_callback(fn)
-
    def set_mock_response(self, response: Any) -> None:
        """Set the response to return from mock calls."""
        # Backward compatibility: Store in both wrapper and provider implementation
@@ -644,23 +542,6 @@ class LLMProvider:
        # SDK will automatically check for authentication when first used
        # No need to verify here - let it fail gracefully on first call with helpful error

-    def with_config(self, config: Any) -> "ConfiguredLLMProvider":
-        """
-        Return a configured wrapper for a specific bank operation.
-
-        The wrapper applies per-bank overrides (e.g. Gemini safety settings)
-        to every ``call()`` / ``call_with_tools()`` invocation without
-        changing the underlying provider or its long-lived client connection.
-
-        Args:
-            config: Resolved ``HindsightConfig`` for the current bank/request.
-
-        Returns:
-            A ``ConfiguredLLMProvider`` that delegates to this provider with
-            the supplied config applied.
-        """
-        return ConfiguredLLMProvider(self, config.llm_gemini_safety_settings)
-
    async def cleanup(self) -> None:
        """Clean up resources."""
        pass
@@ -671,9 +552,8 @@ class LLMProvider:
        provider = os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq")
        api_key = os.getenv("HINDSIGHT_API_LLM_API_KEY", "")

-        # API key not needed for openai-codex (uses OAuth), claude-code (uses Keychain OAuth),
-        # ollama (local), or vertexai (uses GCP service account credentials)
-        if not api_key and provider not in ("openai-codex", "claude-code", "ollama", "vertexai"):
+        # API key not needed for openai-codex (uses OAuth) or claude-code (uses Keychain OAuth)
+        if not api_key and provider not in ("openai-codex", "claude-code"):
            raise ValueError(
                "HINDSIGHT_API_LLM_API_KEY environment variable is required (unless using openai-codex or claude-code)"
            )
@@ -689,9 +569,8 @@ class LLMProvider:
        provider = os.getenv("HINDSIGHT_API_ANSWER_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
        api_key = os.getenv("HINDSIGHT_API_ANSWER_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY", ""))

-        # API key not needed for openai-codex (uses OAuth), claude-code (uses Keychain OAuth),
-        # ollama (local), or vertexai (uses GCP service account credentials)
-        if not api_key and provider not in ("openai-codex", "claude-code", "ollama", "vertexai"):
+        # API key not needed for openai-codex (uses OAuth) or claude-code (uses Keychain OAuth)
+        if not api_key and provider not in ("openai-codex", "claude-code"):
            raise ValueError(
                "HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_ANSWER_LLM_API_KEY environment variable is required "
                "(unless using openai-codex or claude-code)"
@@ -708,9 +587,8 @@ class LLMProvider:
        provider = os.getenv("HINDSIGHT_API_JUDGE_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
        api_key = os.getenv("HINDSIGHT_API_JUDGE_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY", ""))

-        # API key not needed for openai-codex (uses OAuth), claude-code (uses Keychain OAuth),
-        # ollama (local), or vertexai (uses GCP service account credentials)
-        if not api_key and provider not in ("openai-codex", "claude-code", "ollama", "vertexai"):
+        # API key not needed for openai-codex (uses OAuth) or claude-code (uses Keychain OAuth)
+        if not api_key and provider not in ("openai-codex", "claude-code"):
            raise ValueError(
                "HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_JUDGE_LLM_API_KEY environment variable is required "
                "(unless using openai-codex or claude-code)"
@@ -722,58 +600,5 @@ class LLMProvider:
        return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="high")


-class ConfiguredLLMProvider:
-    """
-    Thin wrapper around LLMProvider that applies bank-specific config to every call.
-
-    Obtained via ``LLMProvider.with_config(resolved_config)``.  The wrapper
-    sets any provider-specific overrides (currently Gemini safety settings)
-    immediately before each call using a ContextVar token, then resets it
-    afterwards — so nesting is safe and the configuration cannot leak across
-    operations.
-
-    All attribute access falls through to the underlying provider so callers
-    that read ``llm.provider``, ``llm.model``, etc. continue to work without
-    any changes.
-    """
-
-    def __init__(self, provider: "LLMProvider", gemini_safety_settings: list | None) -> None:
-        # Use object.__setattr__ to avoid triggering __getattr__
-        object.__setattr__(self, "_provider", provider)
-        object.__setattr__(self, "_gemini_safety_settings", gemini_safety_settings)
-
-    # ── attribute passthrough ──────────────────────────────────────────────────
-
-    def __getattr__(self, name: str) -> Any:
-        return getattr(object.__getattribute__(self, "_provider"), name)
-
-    # ── overridden call methods ────────────────────────────────────────────────
-
-    async def call(self, messages: list[dict[str, Any]], **kwargs: Any) -> Any:
-        from .providers.gemini_llm import _safety_settings_ctx
-
-        token = _safety_settings_ctx.set(object.__getattribute__(self, "_gemini_safety_settings"))
-        try:
-            return await object.__getattribute__(self, "_provider").call(messages=messages, **kwargs)
-        finally:
-            _safety_settings_ctx.reset(token)
-
-    async def call_with_tools(
-        self,
-        messages: list[dict[str, Any]],
-        tools: list[dict[str, Any]],
-        **kwargs: Any,
-    ) -> "LLMToolCallResult":
-        from .providers.gemini_llm import _safety_settings_ctx
-
-        token = _safety_settings_ctx.set(object.__getattribute__(self, "_gemini_safety_settings"))
-        try:
-            return await object.__getattribute__(self, "_provider").call_with_tools(
-                messages=messages, tools=tools, **kwargs
-            )
-        finally:
-            _safety_settings_ctx.reset(token)
-
-
 # Backwards compatibility alias
 LLMConfig = LLMProvider
@@ -0,0 +1,14 @@
+"""
+Mental models module for Hindsight.
+
+Mental models contain directives - hard rules that are injected into reflect prompts.
+Directives are user-defined and their observations are user-provided (not LLM-generated).
+
+Other types of consolidated knowledge are handled by:
+- Learnings: Automatic bottom-up consolidation from facts
+- Pinned Reflections: User-curated living documents
+"""
+
+from .models import MentalModel, MentalModelSubtype
+
+__all__ = ["MentalModel", "MentalModelSubtype"]
@@ -0,0 +1,53 @@
+"""
+Pydantic models for mental models.
+"""
+
+from datetime import datetime, timezone
+from enum import Enum
+
+from pydantic import BaseModel, Field
+
+
+class MentalModelSubtype(str, Enum):
+    """Subtype of mental model.
+
+    Currently only DIRECTIVE is supported. Other types of consolidated knowledge
+    are handled by:
+    - Learnings: Automatic bottom-up consolidation from facts
+    - Pinned Reflections: User-curated living documents
+    """
+
+    DIRECTIVE = "directive"  # User-defined hard rules, observations user-provided
+
+
+class MentalModel(BaseModel):
+    """
+    A mental model representing synthesized understanding.
+
+    Mental models are the agent's consolidated knowledge. Unlike raw facts,
+    mental models provide:
+    - A one-liner description for quick scanning/retrieval
+    - A full summary for deep understanding
+    - Links to related mental models
+    """
+
+    id: str = Field(description="Unique identifier within the bank")
+    bank_id: str = Field(description="Bank this mental model belongs to")
+    subtype: MentalModelSubtype = Field(description="How this model was created")
+    name: str = Field(description="Human-readable name")
+    description: str = Field(description="One-liner for quick scanning and retrieval matching")
+    summary: str | None = Field(default=None, description="Full synthesized understanding")
+
+    # References
+    entity_id: str | None = Field(default=None, description="Reference to entities table when type=entity")
+    source_facts: list[str] = Field(default_factory=list, description="Fact IDs used to generate summary")
+    links: list[str] = Field(default_factory=list, description="Related mental model IDs")
+
+    # Tags for scoped visibility (similar to document tags)
+    tags: list[str] = Field(default_factory=list, description="Tags for scoped visibility filtering")
+
+    # Timestamps
+    last_updated: datetime | None = Field(default=None, description="When summary was last regenerated")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this model was created"
+    )
@@ -1,31 +1,10 @@
 """File parser implementations."""

-import logging
-from dataclasses import dataclass
-
 from .base import FileParser, UnsupportedFileTypeError
 from .iris import IrisParser
 from .markitdown import MarkitdownParser

-__all__ = [
-    "FileParser",
-    "UnsupportedFileTypeError",
-    "IrisParser",
-    "MarkitdownParser",
-    "FileParserRegistry",
-    "ConvertResult",
-]
-
-
-@dataclass
-class ConvertResult:
-    """Result of a successful file conversion."""
-
-    content: str
-    parser_name: str
-
-
-logger = logging.getLogger(__name__)
+__all__ = ["FileParser", "UnsupportedFileTypeError", "IrisParser", "MarkitdownParser", "FileParserRegistry"]


 class FileParserRegistry:
@@ -78,51 +57,6 @@ class FileParserRegistry:

        raise ValueError(f"No parser found for {filename}. Available parsers: {list(self._parsers.keys())}")

-    async def convert_with_fallback(
-        self,
-        parsers: list[str],
-        file_data: bytes,
-        filename: str,
-        content_type: str | None = None,
-    ) -> ConvertResult:
-        """
-        Try each parser in order, falling back on failure or empty content.
-
-        Moves to the next parser if the current one raises UnsupportedFileTypeError
-        or returns empty content. Any other exception (RuntimeError, network error,
-        etc.) also triggers a fallback so the chain is exhausted before failing.
-
-        Args:
-            parsers: Ordered list of parser names to try
-            file_data: Raw file bytes
-            filename: Original filename
-            content_type: MIME type (optional)
-
-        Returns:
-            ConvertResult with the parsed content and the name of the parser that succeeded
-
-        Raises:
-            ValueError: If a parser name is not registered
-            RuntimeError: If all parsers fail or return empty content
-        """
-        last_error: Exception | None = None
-        for name in parsers:
-            parser = self.get_parser(name, filename, content_type)
-            try:
-                content = await parser.convert(file_data, filename)
-                if content and content.strip():
-                    return ConvertResult(content=content, parser_name=name)
-                logger.warning(f"Parser '{name}' returned empty content for '{filename}', trying next")
-                last_error = RuntimeError(f"Parser '{name}' returned no content for '{filename}'")
-            except UnsupportedFileTypeError as e:
-                logger.warning(f"Parser '{name}' does not support '{filename}', trying next: {e}")
-                last_error = e
-            except Exception as e:
-                logger.warning(f"Parser '{name}' failed for '{filename}', trying next: {e}")
-                last_error = e
-
-        raise last_error or RuntimeError(f"No parsers available for '{filename}'")
-
    def list_parsers(self) -> list[str]:
        """Get list of registered parser names."""
        return list(self._parsers.keys())
@@ -62,7 +62,7 @@ class IrisParser(FileParser):
        """
        content_type = mimetypes.guess_type(filename)[0] or "application/octet-stream"

-        async with httpx.AsyncClient(timeout=httpx.Timeout(30.0, read=120.0)) as client:
+        async with httpx.AsyncClient() as client:
            # Step 1: Request a presigned upload URL
            init_resp = await client.post(
                f"{_IRIS_BASE_URL}/org/{self._org_id}/files",
@@ -75,10 +75,9 @@ class IrisParser(FileParser):
            upload_url: str = init_data["uploadUrl"]

            # Step 2: Upload the file bytes to the presigned URL (no auth header)
-            # Ensure file_data is plain bytes (GCS storage may return obstore.Bytes)
            upload_resp = await client.put(
                upload_url,
-                content=bytes(file_data),
+                content=file_data,
                headers={"Content-Type": content_type},
            )
            _raise_for_status(upload_resp, filename, "file upload")
@@ -238,24 +238,21 @@ class ClaudeCodeLLM(LLMInterface):
                )

                # Record trace span
-                try:
-                    from hindsight_api.tracing import get_span_recorder
+                from hindsight_api.tracing import get_span_recorder

-                    span_recorder = get_span_recorder()
-                    span_recorder.record_llm_call(
-                        provider=self.provider,
-                        model=self.model,
-                        scope=scope,
-                        messages=messages,
-                        response_content=result if isinstance(result, str) else result.model_dump_json(),
-                        input_tokens=estimated_input,
-                        output_tokens=estimated_output,
-                        duration=duration,
-                        finish_reason=None,
-                        error=None,
-                    )
-                except Exception:
-                    pass  # logging failure must never affect the operation
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=result if isinstance(result, str) else json.dumps(result),
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    duration=duration,
+                    finish_reason=None,
+                    error=None,
+                )

                # Log slow calls
                if duration > 10.0:
@@ -11,7 +11,6 @@ import json
 import logging
 import os
 import time
-from contextvars import ContextVar
 from typing import Any

 from google import genai
@@ -19,18 +18,11 @@ from google.genai import errors as genai_errors
 from google.genai import types as genai_types

 from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
-from hindsight_api.engine.llm_wrapper import parse_llm_json
 from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
 from hindsight_api.metrics import get_metrics_collector

 logger = logging.getLogger(__name__)

-# Per-request Gemini safety settings override.
-# Set exclusively by ConfiguredLLMProvider.call() / call_with_tools() via token-based
-# set/reset, so it is properly scoped to each individual LLM call and never leaks.
-_safety_settings_ctx: ContextVar[list | None] = ContextVar("gemini_safety_settings", default=None)
-
-
 # Vertex AI imports (optional)
 try:
    import google.auth
@@ -65,9 +57,6 @@ class GeminiLLM(LLMInterface):
        self._client = None
        self._is_vertexai = self.provider == "vertexai"

-        # Safety settings: None means use Gemini's defaults
-        self._safety_settings: list | None = kwargs.get("gemini_safety_settings")
-
        if self._is_vertexai:
            self._init_vertexai(**kwargs)
        else:
@@ -226,29 +215,16 @@ class GeminiLLM(LLMInterface):
        if temperature is not None:
            config_kwargs["temperature"] = temperature

-        # Apply safety settings: context var (per-request bank override) takes precedence over instance default
-        effective_safety_settings = _safety_settings_ctx.get()
-        if effective_safety_settings is None:
-            effective_safety_settings = self._safety_settings
-        if effective_safety_settings is not None:
-            config_kwargs["safety_settings"] = [
-                genai_types.SafetySetting(category=s["category"], threshold=s["threshold"])
-                for s in effective_safety_settings
-            ]
-
        generation_config = genai_types.GenerateContentConfig(**config_kwargs) if config_kwargs else None

        last_exception = None

        for attempt in range(max_retries + 1):
            try:
-                response = await asyncio.wait_for(
-                    self._client.aio.models.generate_content(
-                        model=self.model,
-                        contents=gemini_contents,
-                        config=generation_config,
-                    ),
-                    timeout=90.0,  # Safety net for network hangs; valid slow responses are <90s
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=generation_config,
                )

                content = response.text
@@ -271,7 +247,7 @@ class GeminiLLM(LLMInterface):

                # Parse structured output if requested
                if response_format is not None:
-                    json_data = parse_llm_json(content)
+                    json_data = json.loads(content)
                    if skip_validation:
                        result = json_data
                    else:
@@ -429,57 +405,31 @@ class GeminiLLM(LLMInterface):
        # Convert messages
        system_instruction = None
        gemini_contents = []
-        msg_list = list(messages)
-        i = 0
-        while i < len(msg_list):
-            msg = msg_list[i]
+        for msg in messages:
            role = msg.get("role", "user")
            content = msg.get("content", "")

            if role == "system":
                system_instruction = (system_instruction + "\n\n" + content) if system_instruction else content
-                i += 1
            elif role == "tool":
-                # Gemini requires ALL tool responses for a given model turn to be grouped
-                # into a single Content with multiple FunctionResponse parts.
-                # Consecutive role="tool" messages correspond to one model turn's tool calls.
-                parts = []
-                while i < len(msg_list) and msg_list[i].get("role") == "tool":
-                    tool_msg = msg_list[i]
-                    tool_content = tool_msg.get("content", "")
-                    parts.append(
-                        genai_types.Part(
-                            function_response=genai_types.FunctionResponse(
-                                name=tool_msg.get("name", ""),
-                                response={"result": tool_content},
+                # Gemini uses function_response
+                gemini_contents.append(
+                    genai_types.Content(
+                        role="user",
+                        parts=[
+                            genai_types.Part(
+                                function_response=genai_types.FunctionResponse(
+                                    name=msg.get("name", ""),
+                                    response={"result": content},
+                                )
                            )
-                        )
+                        ],
                    )
-                    i += 1
-                gemini_contents.append(genai_types.Content(role="user", parts=parts))
+                )
            elif role == "assistant":
-                tool_calls_in_msg = msg.get("tool_calls", [])
-                if tool_calls_in_msg:
-                    # Convert OpenAI-style tool_calls to Gemini function_call parts
-                    # This is required for proper multi-turn conversation history
-                    parts = []
-                    if content:
-                        parts.append(genai_types.Part(text=content))
-                    for tc in tool_calls_in_msg:
-                        fn = tc.get("function", {})
-                        fn_name = fn.get("name", "")
-                        fn_args_str = fn.get("arguments", "{}")
-                        fn_args = parse_llm_json(fn_args_str)
-                        parts.append(
-                            genai_types.Part(function_call=genai_types.FunctionCall(name=fn_name, args=fn_args))
-                        )
-                    gemini_contents.append(genai_types.Content(role="model", parts=parts))
-                else:
-                    gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
-                i += 1
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
            else:
                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
-                i += 1

        config_kwargs: dict[str, Any] = {"tools": gemini_tools}
        if system_instruction:
@@ -487,50 +437,15 @@ class GeminiLLM(LLMInterface):
        if temperature is not None:
            config_kwargs["temperature"] = temperature

-        # Map OpenAI-style tool_choice to Gemini FunctionCallingConfig
-        if tool_choice == "required":
-            config_kwargs["tool_config"] = genai_types.ToolConfig(
-                function_calling_config=genai_types.FunctionCallingConfig(
-                    mode="ANY",
-                )
-            )
-        elif isinstance(tool_choice, dict) and tool_choice.get("type") == "function":
-            fn_name = tool_choice.get("function", {}).get("name")
-            if fn_name:
-                config_kwargs["tool_config"] = genai_types.ToolConfig(
-                    function_calling_config=genai_types.FunctionCallingConfig(
-                        mode="ANY",
-                        allowed_function_names=[fn_name],
-                    )
-                )
-        elif tool_choice == "none":
-            config_kwargs["tool_config"] = genai_types.ToolConfig(
-                function_calling_config=genai_types.FunctionCallingConfig(mode="NONE")
-            )
-        # "auto" is the default (no tool_config needed)
-
-        # Apply safety settings: context var (per-request bank override) takes precedence over instance default
-        effective_safety_settings = _safety_settings_ctx.get()
-        if effective_safety_settings is None:
-            effective_safety_settings = self._safety_settings
-        if effective_safety_settings is not None:
-            config_kwargs["safety_settings"] = [
-                genai_types.SafetySetting(category=s["category"], threshold=s["threshold"])
-                for s in effective_safety_settings
-            ]
-
        config = genai_types.GenerateContentConfig(**config_kwargs)

        last_exception = None
        for attempt in range(max_retries + 1):
            try:
-                response = await asyncio.wait_for(
-                    self._client.aio.models.generate_content(
-                        model=self.model,
-                        contents=gemini_contents,
-                        config=config,
-                    ),
-                    timeout=90.0,  # Safety net for network hangs; valid slow responses are <90s
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=config,
                )

                # Extract content and tool calls
@@ -6,7 +6,6 @@ without making actual API calls to external LLM services.
 """

 import logging
-from collections.abc import Callable
 from typing import Any

 from ..llm_interface import LLMInterface
@@ -67,7 +66,6 @@ class MockLLM(LLMInterface):
        self._mock_calls: list[dict] = []
        self._mock_response: Any = None
        self._mock_exception: Exception | None = None
-        self._response_callback: Callable[[list[dict], str], Any] | None = None

    async def verify_connection(self) -> None:
        """
@@ -149,9 +147,7 @@ class MockLLM(LLMInterface):
        )

        # Return mock response
-        if self._response_callback is not None:
-            result = self._response_callback(messages, scope)
-        elif self._mock_response is not None:
+        if self._mock_response is not None:
            result = self._mock_response
        elif response_format is not None:
            # Try to create a minimal valid instance of the response format
@@ -218,15 +214,7 @@ class MockLLM(LLMInterface):

        span_recorder = get_span_recorder()

-        if self._response_callback is not None:
-            cb_result = self._response_callback(messages, scope)
-            if isinstance(cb_result, LLMToolCallResult):
-                result = cb_result
-            else:
-                result = LLMToolCallResult(
-                    content=str(cb_result) if cb_result is not None else "mock response", finish_reason="stop"
-                )
-        elif self._mock_response is not None:
+        if self._mock_response is not None:
            if isinstance(self._mock_response, LLMToolCallResult):
                result = self._mock_response
            elif isinstance(self._mock_response, list):
@@ -270,16 +258,6 @@ class MockLLM(LLMInterface):
        """Clean up resources (no-op for mock provider)."""
        pass

-    def set_response_callback(self, fn: Callable[[list[dict], str], Any]) -> None:
-        """
-        Set a callback invoked on each call() instead of _mock_response.
-
-        The callback receives (messages, scope) and returns the response.
-        Useful for returning different responses per call (e.g., cycling
-        through a corpus in a benchmark).
-        """
-        self._response_callback = fn
-
    def set_mock_response(self, response: Any) -> None:
        """
        Set the response to return from mock calls.
@@ -522,19 +522,6 @@ class OpenAICompatibleLLM(LLMInterface):
        """
        start_time = time.time()

-        # Normalize named tool_choice dicts to "required" + filter tools.
-        # Some providers (e.g. LM Studio, Ollama) reject the OpenAI named format
-        # {"type": "function", "function": {"name": "..."}}.  The semantics are
-        # identical to tool_choice="required" with the tools list restricted to
-        # just the requested tool, so we apply that transformation universally.
-        if isinstance(tool_choice, dict) and tool_choice.get("type") == "function":
-            forced_name = tool_choice.get("function", {}).get("name")
-            if forced_name:
-                filtered = [t for t in tools if t.get("function", {}).get("name") == forced_name]
-                if filtered:
-                    tools = filtered
-                tool_choice = "required"
-
        # Build call parameters
        call_params: dict[str, Any] = {
            "model": self.model,
@@ -92,17 +92,11 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
        self._search_dates = None

    def load(self) -> None:
-        """Load dateparser and warm up internal data structures.
-
-        Triggers the real initialization cost (regex tables, timezone data) at
-        load time so the first actual recall doesn't pay the cold-start penalty.
-        """
+        """Load dateparser (lazy import)."""
        if self._search_dates is None:
            from dateparser.search import search_dates

            self._search_dates = search_dates
-            # Warm up: fire a dummy call to trigger lazy-loaded internal tables.
-            self._search_dates("today")

    def analyze(self, query: str, reference_date: datetime | None = None) -> QueryAnalysis:
        """
@@ -14,26 +14,32 @@ import re
 import time
 from typing import TYPE_CHECKING, Any, Awaitable, Callable

-import tiktoken
-
 from .models import DirectiveInfo, LLMCall, ReflectAgentResult, TokenUsageSummary, ToolCall
 from .prompts import FINAL_SYSTEM_PROMPT, _extract_directive_rules, build_final_prompt, build_system_prompt_for_tools
 from .tools_schema import get_reflect_tools


 def _build_directives_applied(directives: list[dict[str, Any]] | None) -> list[DirectiveInfo]:
-    """Build list of DirectiveInfo from directives."""
+    """Build list of DirectiveInfo from directive mental models.
+
+    Handles multiple directive formats:
+    1. New format: directives have direct 'content' field
+    2. Fallback: directives have 'description' field
+    """
    if not directives:
        return []

-    return [
-        DirectiveInfo(
-            id=directive.get("id", ""),
-            name=directive.get("name", ""),
-            content=directive.get("content", ""),
-        )
-        for directive in directives
-    ]
+    result = []
+    for directive in directives:
+        directive_id = directive.get("id", "")
+        directive_name = directive.get("name", "")
+
+        # Get content from 'content' field or fallback to 'description'
+        content = directive.get("content", "") or directive.get("description", "")
+
+        result.append(DirectiveInfo(id=directive_id, name=directive_name, content=content))
+
+    return result


 if TYPE_CHECKING:
@@ -261,46 +267,6 @@ OUTPUT:"""
        return None, 0, 0


-_TIKTOKEN_ENCODING = tiktoken.get_encoding("cl100k_base")
-
-
-def _count_messages_tokens(messages: list[dict[str, Any]]) -> int:
-    """Estimate the token count of the messages list using cl100k_base encoding."""
-    total = 0
-    for msg in messages:
-        content = msg.get("content") or ""
-        if isinstance(content, str):
-            total += len(_TIKTOKEN_ENCODING.encode(content))
-        elif isinstance(content, list):
-            for part in content:
-                if isinstance(part, dict) and isinstance(part.get("text"), str):
-                    total += len(_TIKTOKEN_ENCODING.encode(part["text"]))
-        # Tool call arguments and results also count
-        for tc in msg.get("tool_calls") or []:
-            if isinstance(tc, dict):
-                func = tc.get("function", {})
-                total += len(_TIKTOKEN_ENCODING.encode(func.get("arguments", "")))
-    return total
-
-
-def _is_context_overflow_error(exc: Exception) -> bool:
-    """Return True if the exception signals the LLM context window was exceeded."""
-    msg = str(exc).lower()
-    return any(
-        phrase in msg
-        for phrase in (
-            "context_length_exceeded",
-            "context length exceeded",
-            "maximum context length",
-            "prompt_too_long",
-            "prompt is too long",
-            "resource_exhausted",
-            "input is too long",
-            "too many tokens",
-        )
-    )
-
-
 async def run_reflect_agent(
    llm_config: "LLMProvider",
    bank_id: str,
@@ -308,7 +274,7 @@ async def run_reflect_agent(
    bank_profile: dict[str, Any],
    search_mental_models_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
    search_observations_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
-    recall_fn: Callable[[str, int, int], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
    context: str | None = None,
    max_iterations: int = DEFAULT_MAX_ITERATIONS,
@@ -317,7 +283,6 @@ async def run_reflect_agent(
    directives: list[dict[str, Any]] | None = None,
    has_mental_models: bool = False,
    budget: str | None = None,
-    max_context_tokens: int = 100_000,
 ) -> ReflectAgentResult:
    """
    Execute the reflect agent loop using native tool calling.
@@ -425,15 +390,12 @@ async def run_reflect_agent(
            f"total={elapsed_ms}ms"
        )

-    consecutive_errors = 0
    for iteration in range(max_iterations):
        is_last = iteration == max_iterations - 1

        if is_last:
            # Force text response on last iteration - no tools
-            prompt = build_final_prompt(
-                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
-            )
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
            llm_start = time.time()
            response, usage = await llm_config.call(
                messages=[
@@ -478,91 +440,17 @@ async def run_reflect_agent(
                directives_applied=directives_applied,
            )

-        # Proactive context-window guard: if accumulated messages would exceed the
-        # configured token budget, bail out early and synthesize from what we have.
-        estimated_tokens = _count_messages_tokens(messages)
-        if estimated_tokens >= max_context_tokens and (
-            bool(available_memory_ids) or bool(available_mental_model_ids) or bool(available_observation_ids)
-        ):
-            logger.warning(
-                f"[REFLECT {reflect_id}] Context budget exceeded on iteration {iteration + 1}: "
-                f"~{estimated_tokens} tokens >= {max_context_tokens} limit. Forcing final synthesis."
-            )
-            prompt = build_final_prompt(
-                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
-            )
-            llm_start = time.time()
-            response, usage = await llm_config.call(
-                messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
-                    {"role": "user", "content": prompt},
-                ],
-                scope="reflect",
-                max_completion_tokens=max_tokens,
-                return_usage=True,
-            )
-            llm_duration = int((time.time() - llm_start) * 1000)
-            total_input_tokens += usage.input_tokens
-            total_output_tokens += usage.output_tokens
-            llm_trace.append(
-                {
-                    "scope": "final",
-                    "duration_ms": llm_duration,
-                    "input_tokens": usage.input_tokens,
-                    "output_tokens": usage.output_tokens,
-                }
-            )
-            answer = _clean_answer_text(response.strip())
-
-            structured_output = None
-            if response_schema and answer:
-                structured_output, struct_in, struct_out = await _generate_structured_output(
-                    answer, response_schema, llm_config, reflect_id
-                )
-                total_input_tokens += struct_in
-                total_output_tokens += struct_out
-
-            _log_completion(answer, iteration + 1, forced=True)
-            return ReflectAgentResult(
-                text=answer,
-                structured_output=structured_output,
-                iterations=iteration + 1,
-                tools_called=total_tools_called,
-                tool_trace=tool_trace,
-                llm_trace=_get_llm_trace(),
-                usage=_get_usage(),
-                directives_applied=directives_applied,
-            )
-
        # Call LLM with tools
        llm_start = time.time()

-        # Determine tool_choice for this iteration.
-        # Force the full hierarchical retrieval path before allowing auto:
-        # With mental models:
-        #   0 → search_mental_models, 1 → search_observations, 2 → recall, 3+ → auto
-        # Without mental models:
-        #   0 → search_observations, 1 → recall, 2+ → auto
-        if iteration == 0 and has_mental_models:
-            iter_tool_choice: str | dict = {"type": "function", "function": {"name": "search_mental_models"}}
-        elif iteration == 0:
-            iter_tool_choice = {"type": "function", "function": {"name": "search_observations"}}
-        elif iteration == 1 and has_mental_models:
-            iter_tool_choice = {"type": "function", "function": {"name": "search_observations"}}
-        elif iteration == 1 or (iteration == 2 and has_mental_models):
-            iter_tool_choice = {"type": "function", "function": {"name": "recall"}}
-        else:
-            iter_tool_choice = "auto"
-
        try:
            result = await llm_config.call_with_tools(
                messages=messages,
                tools=tools,
                scope="reflect_tool_call",
-                tool_choice=iter_tool_choice,
+                tool_choice="required" if iteration == 0 else "auto",  # Force tool use on first iteration
            )
            llm_duration = int((time.time() - llm_start) * 1000)
-            consecutive_errors = 0
            total_input_tokens += result.input_tokens
            total_output_tokens += result.output_tokens
            llm_trace.append(
@@ -576,25 +464,15 @@ async def run_reflect_agent(

        except Exception as e:
            err_duration = int((time.time() - llm_start) * 1000)
-            consecutive_errors += 1
            logger.warning(f"[REFLECT {reflect_id}] LLM error on iteration {iteration + 1}: {e} ({err_duration}ms)")
            llm_trace.append({"scope": f"agent_{iteration + 1}_err", "duration_ms": err_duration})
+            # Guardrail: If no evidence gathered yet, retry
            has_gathered_evidence = (
                bool(available_memory_ids) or bool(available_mental_model_ids) or bool(available_observation_ids)
            )
-            # Context overflow errors must never be retried — retrying would only make them worse.
-            # Skip straight to final synthesis with whatever evidence we have.
-            if _is_context_overflow_error(e):
-                logger.warning(
-                    f"[REFLECT {reflect_id}] Context window exceeded on iteration {iteration + 1}, "
-                    "forcing final synthesis from gathered evidence."
-                )
-            # For other errors: retry if no evidence yet (but cap consecutive errors to avoid long hangs)
-            elif not has_gathered_evidence and iteration < max_iterations - 1 and consecutive_errors < 2:
+            if not has_gathered_evidence and iteration < max_iterations - 1:
                continue
-            prompt = build_final_prompt(
-                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
-            )
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
            llm_start = time.time()
            response, usage = await llm_config.call(
                messages=[
@@ -665,9 +543,7 @@ async def run_reflect_agent(
                    directives_applied=directives_applied,
                )
            # Empty response, force final
-            prompt = build_final_prompt(
-                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
-            )
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
            llm_start = time.time()
            response, usage = await llm_config.call(
                messages=[
@@ -931,9 +807,9 @@ async def _process_done_tool(
        answer = "No answer provided."

    # Validate IDs (only include IDs that were actually retrieved)
-    used_memory_ids = [mid for mid in (args.get("memory_ids") or []) if mid in available_memory_ids]
-    used_mental_model_ids = [mid for mid in (args.get("mental_model_ids") or []) if mid in available_mental_model_ids]
-    used_observation_ids = [oid for oid in (args.get("observation_ids") or []) if oid in available_observation_ids]
+    used_memory_ids = [mid for mid in args.get("memory_ids", []) if mid in available_memory_ids]
+    used_mental_model_ids = [mid for mid in args.get("mental_model_ids", []) if mid in available_mental_model_ids]
+    used_observation_ids = [oid for oid in args.get("observation_ids", []) if oid in available_observation_ids]

    # Generate structured output if schema provided
    structured_output = None
@@ -969,7 +845,7 @@ async def _execute_tool_with_timing(
    tc: "LLMToolCall",
    search_mental_models_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
    search_observations_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
-    recall_fn: Callable[[str, int, int], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
 ) -> tuple[dict[str, Any], int]:
    """Execute a tool call and return result with timing."""
@@ -1041,7 +917,7 @@ async def _execute_tool(
    args: dict[str, Any],
    search_mental_models_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
    search_observations_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
-    recall_fn: Callable[[str, int, int], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
 ) -> dict[str, Any]:
    """Execute a single tool by name."""
@@ -1067,8 +943,7 @@ async def _execute_tool(
        if not query:
            return {"error": "recall requires a query parameter"}
        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)  # Default 2048, min 1000
-        max_chunk_tokens = max(int(args.get("max_chunk_tokens") or 1000), 1000)  # Always enabled, min 1000
-        return await recall_fn(query, max_tokens, max_chunk_tokens)
+        return await recall_fn(query, max_tokens)

    elif tool_name == "expand":
        memory_ids = args.get("memory_ids", [])
@@ -1096,9 +971,9 @@ def _summarize_input(tool_name: str, args: dict[str, Any]) -> str:
    elif tool_name == "recall":
        query = args.get("query", "")
        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
+        # Show actual value used (default 2048, min 1000)
        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)
-        max_chunk_tokens = max(int(args.get("max_chunk_tokens") or 1000), 1000)
-        return f"(query={query_preview}, max_tokens={max_tokens}, max_chunk_tokens={max_chunk_tokens})"
+        return f"(query={query_preview}, max_tokens={max_tokens})"
    elif tool_name == "expand":
        memory_ids = args.get("memory_ids", [])
        depth = args.get("depth", "chunk")
@@ -10,30 +10,59 @@ The reflect agent uses hierarchical retrieval:
 import json
 from typing import Any

-import tiktoken
-
-_TIKTOKEN_ENCODING = tiktoken.get_encoding("cl100k_base")
-
-# Fraction of max_context_tokens reserved for tool results in the final synthesis prompt.
-# The remainder covers the system prompt, question, bank context, and output tokens.
-_FINAL_PROMPT_CONTEXT_FRACTION = 0.8
-

 def _extract_directive_rules(directives: list[dict[str, Any]]) -> list[str]:
-    """Extract directive rules as a list of strings."""
+    """
+    Extract directive rules as a list of strings.
+
+    Args:
+        directives: List of directives with name and content
+
+    Returns:
+        List of directive rule strings
+    """
    rules = []
    for directive in directives:
-        name = directive.get("name", "")
+        directive_name = directive.get("name", "")
+        # New format: directives have direct content field
        content = directive.get("content", "")
        if content:
-            rules.append(f"**{name}**: {content}" if name else content)
+            if directive_name:
+                rules.append(f"**{directive_name}**: {content}")
+            else:
+                rules.append(content)
+        else:
+            # Legacy format: check for observations
+            observations = directive.get("observations", [])
+            if observations:
+                for obs in observations:
+                    # Support both Pydantic Observation objects and dicts
+                    if hasattr(obs, "title"):
+                        title = obs.title
+                        obs_content = obs.content
+                    else:
+                        title = obs.get("title", "")
+                        obs_content = obs.get("content", "")
+                    if title and obs_content:
+                        rules.append(f"**{title}**: {obs_content}")
+                    elif obs_content:
+                        rules.append(obs_content)
+            elif directive_name:
+                # Fallback to description
+                desc = directive.get("description", "")
+                if desc:
+                    rules.append(f"**{directive_name}**: {desc}")
    return rules


 def build_directives_section(directives: list[dict[str, Any]]) -> str:
-    """Build the directives section for the system prompt.
+    """
+    Build the directives section for the system prompt.

    Directives are hard rules that MUST be followed in all responses.
+
+    Args:
+        directives: List of directive mental models with observations
    """
    if not directives:
        return ""
@@ -140,12 +169,6 @@ def build_system_prompt_for_tools(

    parts.extend(
        [
-            "## LANGUAGE RULE (default - directives take precedence)",
-            "- By default, detect the language of the user's question and respond in that SAME language.",
-            "- If the question is in Chinese, respond in Chinese. If in Japanese, respond in Japanese.",
-            "- IMPORTANT: The DIRECTIVES section above has HIGHER PRIORITY than this rule.",
-            "  If a directive specifies a language (e.g. 'Always respond in French'), follow the directive.",
-            "",
            "## CRITICAL RULES",
            "- ONLY use information from tool results - no external knowledge or guessing",
            "- You SHOULD synthesize, infer, and reason from the retrieved memories",
@@ -182,7 +205,6 @@ def build_system_prompt_for_tools(
                "### 3. RAW FACTS (recall) - Ground Truth",
                "- Individual memories (world facts and experiences)",
                "- Use when: no mental models/observations exist, they're stale, or you need specific details",
-                "- MANDATORY: If search_mental_models and search_observations both return 0 results, you MUST call recall() before giving up",
                "- This is the source of truth that other levels are built from",
                "",
            ]
@@ -200,7 +222,6 @@ def build_system_prompt_for_tools(
                "### 2. RAW FACTS (recall) - Ground Truth",
                "- Individual memories (world facts and experiences)",
                "- Use when: no observations exist, they're stale, or you need specific details",
-                "- MANDATORY: If search_observations returns 0 results or count=0, you MUST call recall() before giving up",
                "- This is the source of truth that observations are built from",
                "",
            ]
@@ -278,7 +299,7 @@ def build_system_prompt_for_tools(
        parts.extend(
            [
                "1. First, try search_observations() - check for consolidated knowledge",
-                "2. If search_observations returns 0 results OR observations are stale, you MUST call recall() for raw facts",
+                "2. If observations are stale OR you need specific details, use recall() for raw facts",
                "3. Use expand() if you need more context on specific memories",
                "4. When ready, call done() with your answer and supporting IDs",
            ]
@@ -294,7 +315,6 @@ def build_system_prompt_for_tools(
            "- Format for clarity and readability with proper spacing and hierarchy",
            "- NEVER include memory IDs, UUIDs, or 'Memory references' in the answer text",
            "- Put IDs ONLY in the memory_ids/mental_model_ids/observation_ids arrays, not in the answer",
-            "- CRITICAL: This is a NON-CONVERSATIONAL system. NEVER ask follow-up questions, offer further assistance, or suggest next steps. Your answer must be complete and self-contained. The user cannot reply.",
        ]
    )

@@ -402,7 +422,6 @@ def build_final_prompt(
    context_history: list[dict],
    bank_profile: dict,
    additional_context: str | None = None,
-    max_context_tokens: int = 100_000,
 ) -> str:
    """Build the final prompt when forcing a text response (no tools)."""
    parts = []
@@ -432,32 +451,18 @@ def build_final_prompt(
    if additional_context:
        parts.append(f"\n## Additional Context\n{additional_context}")

-    # Tool call history — include as many entries as fit within the token budget,
-    # preferring the most recent calls (they tend to be the most targeted).
+    # Tool call history
    if context_history:
        parts.append("\n## Retrieved Data (synthesize and reason from this data)")
-        token_budget = int(max_context_tokens * _FINAL_PROMPT_CONTEXT_FRACTION)
-        # Render entries newest-first, then reverse so the prompt reads chronologically.
-        rendered: list[str] = []
-        truncated = False
-        for entry in reversed(context_history):
+        for entry in context_history:
            tool = entry["tool"]
            output = entry["output"]
+            # Format as proper JSON for LLM readability
            try:
                output_str = json.dumps(output, indent=2, default=str)
            except (TypeError, ValueError):
                output_str = str(output)
-            block = f"\n### From {tool}:\n```json\n{output_str}\n```"
-            block_tokens = len(_TIKTOKEN_ENCODING.encode(block))
-            if block_tokens > token_budget:
-                truncated = True
-                break
-            rendered.append(block)
-            token_budget -= block_tokens
-        for block in reversed(rendered):
-            parts.append(block)
-        if truncated:
-            parts.append("\n*Note: Some earlier tool results were omitted to stay within the context window.*")
+            parts.append(f"\n### From {tool}:\n```json\n{output_str}\n```")
    else:
        parts.append("\n## Retrieved Data\nNo data was retrieved.")

@@ -505,6 +510,4 @@ CRITICAL: Output ONLY the final synthesized answer. Do NOT include:
 - Meta-commentary about what you're doing ("I'll search...", "Let me analyze...")
 - Explanations of your reasoning process
 - Descriptions of your approach
-Just provide the direct answer with proper markdown formatting.
-
-CRITICAL: This is a NON-CONVERSATIONAL system. NEVER ask follow-up questions, offer to search again, suggest alternatives, or end with anything like "Would you like me to..." or "Let me know if...". The user cannot reply. Your answer must be complete and self-contained."""
+Just provide the direct answer with proper markdown formatting."""
@@ -9,7 +9,7 @@ Implements hierarchical retrieval:

 import logging
 import uuid
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
 from typing import TYPE_CHECKING, Any

 if TYPE_CHECKING:
@@ -20,6 +20,9 @@ if TYPE_CHECKING:

 logger = logging.getLogger(__name__)

+# Observation is considered stale if not updated in this many days
+STALE_THRESHOLD_DAYS = 7
+

 async def tool_search_mental_models(
    conn: "Connection",
@@ -30,7 +33,6 @@ async def tool_search_mental_models(
    tags: list[str] | None = None,
    tags_match: str = "any",
    exclude_ids: list[str] | None = None,
-    pending_consolidation: int = 0,
 ) -> dict[str, Any]:
    """
    Search user-curated mental models by semantic similarity.
@@ -85,6 +87,7 @@ async def tool_search_mental_models(
        *params,
    )

+    now = datetime.now(timezone.utc)
    mental_models = []

    for row in rows:
@@ -92,10 +95,11 @@ async def tool_search_mental_models(
        if last_refreshed_at and last_refreshed_at.tzinfo is None:
            last_refreshed_at = last_refreshed_at.replace(tzinfo=timezone.utc)

-        # A mental model is stale when there are memories that haven't been consolidated yet —
-        # the same signal used for observations staleness.
-        is_stale = pending_consolidation > 0
-        staleness_reason = f"{pending_consolidation} memories pending consolidation" if is_stale else None
+        # Calculate freshness
+        is_stale = False
+        if last_refreshed_at:
+            age = now - last_refreshed_at
+            is_stale = age > timedelta(days=STALE_THRESHOLD_DAYS)

        mental_models.append(
            {
@@ -106,7 +110,6 @@ async def tool_search_mental_models(
                "relevance": round(row["relevance"], 4),
                "updated_at": last_refreshed_at.isoformat() if last_refreshed_at else None,
                "is_stale": is_stale,
-                "staleness_reason": staleness_reason,
            }
        )

@@ -129,7 +132,7 @@ async def tool_search_observations(
    pending_consolidation: int = 0,
 ) -> dict[str, Any]:
    """
-    Search consolidated observations using recall with include_source_facts.
+    Search consolidated observations using recall with include_observations.

    Observations are auto-generated from memories. Returns freshness info
    so the agent knows if it should also verify with recall().
@@ -146,24 +149,72 @@ async def tool_search_observations(
        pending_consolidation: Number of memories waiting to be consolidated

    Returns:
-        Dict with matching observations including freshness info and source memories
+        Dict with matching observations including freshness info
    """
+    from ..memory_engine import fq_table
+
+    # Use recall to search observations (they come back in results field when fact_type=["observation"])
    result = await memory_engine.recall_async(
        bank_id=bank_id,
        query=query,
-        fact_type=["observation"],
-        max_tokens=max_tokens,
+        fact_type=["observation"],  # Only retrieve observations
+        max_tokens=max_tokens,  # Token budget controls how many observations are returned
        enable_trace=False,
        request_context=request_context,
        tags=tags,
        tags_match=tags_match,
-        include_source_facts=True,
-        max_source_facts_tokens=-1,  # No token limit — include all source facts
        _connection_budget=1,
        _quiet=True,
    )

-    is_stale = pending_consolidation > 0
+    observations = []
+
+    # When fact_type=["observation"], results come back in `results` field as MemoryFact objects
+    # We need to fetch additional fields (proof_count, source_memory_ids) from the database
+    if result.results:
+        obs_ids = [m.id for m in result.results]
+
+        # Fetch proof_count and source_memory_ids for these observations
+        pool = await memory_engine._get_pool()
+        async with pool.acquire() as conn:
+            obs_rows = await conn.fetch(
+                f"""
+                SELECT id, proof_count, source_memory_ids
+                FROM {fq_table("memory_units")}
+                WHERE id = ANY($1::uuid[])
+                """,
+                obs_ids,
+            )
+            obs_data = {str(row["id"]): row for row in obs_rows}
+
+        for m in result.results:
+            # Get additional data from DB lookup
+            extra = obs_data.get(m.id, {})
+            proof_count = extra.get("proof_count", 1) if extra else 1
+            source_ids = extra.get("source_memory_ids", []) if extra else []
+            # Convert UUIDs to strings
+            source_memory_ids = [str(sid) for sid in (source_ids or [])]
+
+            # Determine staleness
+            is_stale = False
+            staleness_reason = None
+            if pending_consolidation > 0:
+                is_stale = True
+                staleness_reason = f"{pending_consolidation} memories pending consolidation"
+
+            observations.append(
+                {
+                    "id": str(m.id),
+                    "text": m.text,
+                    "proof_count": proof_count,
+                    "source_memory_ids": source_memory_ids,
+                    "tags": m.tags or [],
+                    "is_stale": is_stale,
+                    "staleness_reason": staleness_reason,
+                }
+            )
+
+    # Return freshness info (more understandable than raw pending_consolidation count)
    if pending_consolidation == 0:
        freshness = "up_to_date"
    elif pending_consolidation < 10:
@@ -173,10 +224,8 @@ async def tool_search_observations(

    return {
        "query": query,
-        "count": len(result.results),
-        "observations": [m.model_dump() for m in result.results],
-        "source_facts": {k: v.model_dump() for k, v in (result.source_facts or {}).items()},
-        "is_stale": is_stale,
+        "count": len(observations),
+        "observations": observations,
        "freshness": freshness,
    }

@@ -187,10 +236,10 @@ async def tool_recall(
    query: str,
    request_context: "RequestContext",
    max_tokens: int = 2048,
+    max_results: int = 50,
    tags: list[str] | None = None,
    tags_match: str = "any",
    connection_budget: int = 1,
-    max_chunk_tokens: int = 1000,
 ) -> dict[str, Any]:
    """
    Search memories using TEMPR retrieval.
@@ -204,19 +253,18 @@ async def tool_recall(
        query: Search query
        request_context: Request context for authentication
        max_tokens: Maximum tokens for results (default 2048)
+        max_results: Maximum number of results
        tags: Filter by tags (includes untagged memories)
        tags_match: How to match tags - "any" (OR), "all" (AND), or "exact"
        connection_budget: Max DB connections for this recall (default 1 for internal ops)
-        max_chunk_tokens: Maximum tokens for raw source chunk text (default 1000, always included)

    Returns:
-        Dict with list of matching memories including raw chunk text
+        Dict with list of matching memories
    """
-    include_chunks = True
    result = await memory_engine.recall_async(
        bank_id=bank_id,
        query=query,
-        fact_type=["experience", "world"],
+        fact_type=["experience", "world"],  # Exclude opinions and observations
        max_tokens=max_tokens,
        enable_trace=False,
        request_context=request_context,
@@ -224,14 +272,24 @@ async def tool_recall(
        tags_match=tags_match,
        _connection_budget=connection_budget,
        _quiet=True,  # Suppress logging for internal operations
-        include_chunks=include_chunks,
-        max_chunk_tokens=max_chunk_tokens,
    )

+    memories = []
+    for m in result.results[:max_results]:
+        memories.append(
+            {
+                "id": str(m.id),
+                "text": m.text,
+                "type": m.fact_type,
+                "entities": m.entities or [],
+                "occurred": m.occurred_start,  # Already ISO format string
+            }
+        )
+
    return {
        "query": query,
-        "memories": [m.model_dump() for m in result.results],
-        "chunks": {k: v.model_dump() for k, v in (result.chunks or {}).items()},
+        "count": len(memories),
+        "memories": memories,
    }


@@ -47,8 +47,7 @@ TOOL_SEARCH_OBSERVATIONS = {
        "description": (
            "Search consolidated observations (auto-generated knowledge). These are automatically "
            "synthesized from memories. Returns observations with freshness info (updated_at, is_stale). "
-            "If an observation is STALE, you should ALSO use recall() to verify with current facts. "
-            "IMPORTANT: If search_mental_models is available, you MUST call it FIRST before using this tool."
+            "If an observation is STALE, you should ALSO use recall() to verify with current facts."
        ),
        "parameters": {
            "type": "object",
@@ -96,10 +95,6 @@ TOOL_RECALL = {
                    "type": "integer",
                    "description": "Optional limit on result size (default 2048). Use higher values for broader searches.",
                },
-                "max_chunk_tokens": {
-                    "type": "integer",
-                    "description": "Maximum tokens for raw source chunk text included alongside each memory fact (default 1000, min 1000). Chunks provide the surrounding context the fact was extracted from. Increase for broader context.",
-                },
            },
            "required": ["reason", "query"],
        },
@@ -144,7 +139,7 @@ TOOL_DONE_ANSWER = {
            "properties": {
                "answer": {
                    "type": "string",
-                    "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array. LANGUAGE: By default, write in the SAME language as the user's question. However, if a language directive in the system prompt specifies a different language, follow that directive instead.",
+                    "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
                },
                "memory_ids": {
                    "type": "array",
@@ -195,11 +190,7 @@ def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
                "properties": {
                    "answer": {
                        "type": "string",
-                        "description": (
-                            "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. "
-                            "NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array. "
-                            f"MANDATORY: Your answer MUST comply with ALL directives:\n{rules_list}"
-                        ),
+                        "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
                    },
                    "memory_ids": {
                        "type": "array",
@@ -159,10 +159,6 @@ class MemoryFact(BaseModel):
        None, description="ID of the chunk this fact was extracted from (format: bank_id_document_id_chunk_index)"
    )
    tags: list[str] | None = Field(None, description="Visibility scope tags associated with this fact")
-    source_fact_ids: list[str] | None = Field(
-        None,
-        description="IDs of source facts this observation was derived from (observation type only, when source_facts is enabled)",
-    )


 class ChunkInfo(BaseModel):
@@ -230,9 +226,6 @@ class RecallResult(BaseModel):
    chunks: dict[str, ChunkInfo] | None = Field(
        None, description="Chunks for facts, keyed by '{document_id}_{chunk_index}'"
    )
-    source_facts: dict[str, MemoryFact] | None = Field(
-        None, description="Source facts for observation-type results, keyed by fact ID"
-    )


 class ReflectResult(BaseModel):
@@ -5,6 +5,7 @@ This package contains modular components for the retain operation:
 - types: Type definitions for retain pipeline
 - fact_extraction: Extract facts from content
 - embedding_processing: Augment texts and generate embeddings
+- deduplication: Check for duplicate facts
 - entity_processing: Process and resolve entities
 - link_creation: Create temporal, semantic, entity, and causal links
 - chunk_storage: Handle chunk storage
@@ -13,6 +14,7 @@ This package contains modular components for the retain operation:

 from . import (
    chunk_storage,
+    deduplication,
    embedding_processing,
    entity_processing,
    fact_extraction,
@@ -33,6 +35,7 @@ __all__ = [
    # Modules
    "fact_extraction",
    "embedding_processing",
+    "deduplication",
    "entity_processing",
    "link_creation",
    "chunk_storage",
@@ -0,0 +1,85 @@
+"""
+Deduplication logic for retain pipeline.
+
+Checks for duplicate facts using semantic similarity and temporal proximity.
+"""
+
+import logging
+from collections import defaultdict
+from datetime import UTC
+
+from .types import ProcessedFact
+
+logger = logging.getLogger(__name__)
+
+
+async def check_duplicates_batch(conn, bank_id: str, facts: list[ProcessedFact], duplicate_checker_fn) -> list[bool]:
+    """
+    Check which facts are duplicates using batched time-window queries.
+
+    Groups facts by 12-hour time buckets to efficiently check for duplicates
+    within a 24-hour window.
+
+    Args:
+        conn: Database connection
+        bank_id: Bank identifier
+        facts: List of ProcessedFact objects to check
+        duplicate_checker_fn: Async function(conn, bank_id, texts, embeddings, date, time_window_hours)
+                              that returns List[bool] indicating duplicates
+
+    Returns:
+        List of boolean flags (same length as facts) indicating if each fact is a duplicate
+    """
+    if not facts:
+        return []
+
+    # Group facts by event_date (rounded to 12-hour buckets) for efficient batching
+    time_buckets = defaultdict(list)
+    for idx, fact in enumerate(facts):
+        # Use occurred_start if available, otherwise use mentioned_at
+        # For deduplication purposes, we need a time reference
+        fact_date = fact.occurred_start if fact.occurred_start is not None else fact.mentioned_at
+
+        # Defensive: if both are None (shouldn't happen), use now()
+        if fact_date is None:
+            from datetime import datetime
+
+            fact_date = datetime.now(UTC)
+
+        # Round to 12-hour bucket to group similar times
+        bucket_key = fact_date.replace(hour=(fact_date.hour // 12) * 12, minute=0, second=0, microsecond=0)
+        time_buckets[bucket_key].append((idx, fact))
+
+    # Process each bucket in batch
+    all_is_duplicate = [False] * len(facts)
+
+    for bucket_date, bucket_items in time_buckets.items():
+        indices = [item[0] for item in bucket_items]
+        texts = [item[1].fact_text for item in bucket_items]
+        embeddings = [item[1].embedding for item in bucket_items]
+
+        # Check duplicates for this time bucket
+        dup_flags = await duplicate_checker_fn(conn, bank_id, texts, embeddings, bucket_date, time_window_hours=24)
+
+        # Map results back to original indices
+        for idx, is_dup in zip(indices, dup_flags):
+            all_is_duplicate[idx] = is_dup
+
+    return all_is_duplicate
+
+
+def filter_duplicates(facts: list[ProcessedFact], is_duplicate_flags: list[bool]) -> list[ProcessedFact]:
+    """
+    Filter out duplicate facts based on duplicate flags.
+
+    Args:
+        facts: List of ProcessedFact objects
+        is_duplicate_flags: Boolean flags indicating which facts are duplicates
+
+    Returns:
+        List of non-duplicate facts
+    """
+    if len(facts) != len(is_duplicate_flags):
+        raise ValueError(f"Mismatch between facts ({len(facts)}) and flags ({len(is_duplicate_flags)})")
+
+    return [fact for fact, is_dup in zip(facts, is_duplicate_flags) if not is_dup]
@@ -27,21 +27,11 @@ def augment_texts_with_dates(facts: list[ExtractedFact], format_date_fn) -> list
    """
    augmented_texts = []
    for fact in facts:
-        # Use occurred_start as the representative date, fall back to mentioned_at
+        # Use occurred_start as the representative date
        fact_date = fact.occurred_start or fact.mentioned_at
-        # Augment text with date and entity names for embedding (but store original text in DB)
-        # Entity names (including key:value labels) improve retrieval without polluting stored content
-        if fact_date is not None:
-            readable_date = format_date_fn(fact_date)
-            if fact.occurred_end and fact.occurred_end != fact.occurred_start:
-                readable_end = format_date_fn(fact.occurred_end)
-                augmented_text = f"{fact.fact_text} (happened from {readable_date} to {readable_end})"
-            else:
-                augmented_text = f"{fact.fact_text} (happened in {readable_date})"
-        else:
-            augmented_text = fact.fact_text
-        if fact.entities:
-            augmented_text = f"{augmented_text} [{', '.join(fact.entities)}]"
+        readable_date = format_date_fn(fact_date)
+        # Augment text with date for embedding (but store original text in DB)
+        augmented_text = f"{fact.fact_text} (happened in {readable_date})"
        augmented_texts.append(augmented_text)
    return augmented_texts

@@ -41,9 +41,10 @@ async def generate_embeddings_batch(embeddings_backend, texts: list[str]) -> lis
        List of embeddings in same order as input texts
    """
    try:
+        # Run embeddings in thread pool to avoid blocking event loop
        loop = asyncio.get_event_loop()
        embeddings = await loop.run_in_executor(
-            None,
+            None,  # Use default thread pool
            embeddings_backend.encode,
            texts,
        )
@@ -1,194 +0,0 @@
-"""
-Entity labels models and helpers for retain pipeline.
-
-Defines a controlled vocabulary of key:value classification labels
-(e.g., 'pedagogy:scaffolding', 'interest:active') that are extracted
-at retain time and stored as entities.
-"""
-
-from typing import Literal
-
-from pydantic import BaseModel, Field, create_model
-
-
-class LabelValue(BaseModel):
-    """A single allowed value for a label group."""
-
-    value: str
-    description: str = ""
-
-
-class LabelGroup(BaseModel):
-    """A label group (dimension) with its type and allowed values."""
-
-    key: str
-    description: str = ""
-    type: Literal["value", "multi-values", "text"] = "value"
-    optional: bool = True
-    tag: bool = False
-    values: list[LabelValue] = []
-
-
-class EntityLabelsConfig(BaseModel):
-    """Entity labels configuration for a bank (controlled vocabulary)."""
-
-    attributes: list[LabelGroup] = []
-
-
-def parse_entity_labels(raw: dict | list | None) -> EntityLabelsConfig | None:
-    """
-    Parse raw entity labels config into EntityLabelsConfig.
-
-    Accepts:
-    - None → returns None
-    - list → list of attribute dicts (each may use legacy free_values/multi_value or new type field)
-    - dict → {attributes: [...]}
-
-    Legacy migration (backward-compat):
-    - free_values=True          → type="text"
-    - multi_value=True          → type="multi-values"
-    - neither / free_values=False → type="value"
-
-    Args:
-        raw: Raw entity labels config from bank config
-
-    Returns:
-        EntityLabelsConfig or None if raw is None/empty
-    """
-    if raw is None:
-        return None
-
-    if isinstance(raw, list):
-        if not raw:
-            return None
-        attributes = [LabelGroup.model_validate(_migrate_label_group(a)) for a in raw]
-        return EntityLabelsConfig(attributes=attributes)
-
-    if isinstance(raw, dict):
-        attrs_raw = raw.get("attributes", [])
-        if not attrs_raw:
-            return None
-        attributes = [LabelGroup.model_validate(_migrate_label_group(a)) for a in attrs_raw]
-        return EntityLabelsConfig(attributes=attributes)
-
-    return None
-
-
-def _migrate_label_group(raw: dict) -> dict:
-    """Migrate legacy free_values/multi_value fields to the new type field."""
-    if not isinstance(raw, dict) or "type" in raw:
-        return raw
-    patched = dict(raw)
-    if patched.get("free_values"):
-        patched["type"] = "text"
-    elif patched.get("multi_value"):
-        patched["type"] = "multi-values"
-    else:
-        patched["type"] = "value"
-    # Remove legacy keys so Pydantic doesn't error on unknown fields
-    patched.pop("free_values", None)
-    patched.pop("multi_value", None)
-    return patched
-
-
-def build_labels_model(labels_cfg: EntityLabelsConfig) -> type[BaseModel] | None:
-    """
-    Build a dynamic Pydantic model for structured label extraction.
-
-    Each LabelGroup becomes a typed field based on its type:
-    - type="text"                        → str | None  (always optional)
-    - type="value",   optional=True      → Literal["v1","v2"] | None
-    - type="value",   optional=False     → Literal["v1","v2"]  (required)
-    - type="multi-values"                → list[Literal["v1","v2"]]
-
-    Args:
-        labels_cfg: Parsed EntityLabelsConfig
-
-    Returns:
-        Dynamic Pydantic model class, or None if no groups defined
-    """
-    fields: dict = {}
-    for group in labels_cfg.attributes:
-        if not group.key:
-            continue
-        description = group.description or group.key
-
-        if group.type == "text":
-            # Free-form: any string value accepted, always optional
-            fields[group.key] = (str | None, Field(default=None, description=description))
-        else:
-            # Enum-constrained: must have defined values
-            if not group.values:
-                continue
-            values = tuple(v.value for v in group.values if v.value)
-            if not values:
-                continue
-            # Literal[("v1", "v2")] is equivalent to Literal["v1", "v2"] in Python 3.11+
-            literal_type = Literal[values]  # type: ignore[valid-type]
-            if group.type == "multi-values":
-                fields[group.key] = (
-                    list[literal_type],  # type: ignore[valid-type]
-                    Field(default_factory=list, description=description),
-                )
-            elif group.optional:
-                fields[group.key] = (
-                    literal_type | None,  # type: ignore[valid-type]
-                    Field(default=None, description=description),
-                )
-            else:
-                fields[group.key] = (
-                    literal_type,  # type: ignore[valid-type]
-                    Field(description=description),
-                )
-
-    if not fields:
-        return None
-
-    return create_model("Labels", **fields)
-
-
-def is_label_entity(text: str, labels_cfg: EntityLabelsConfig, labels_lookup: set[str]) -> bool:
-    """
-    Return True if entity text belongs to any configured label group.
-
-    For enum groups: checks the pre-built lookup set.
-    For text groups: checks that the text starts with a known key prefix.
-    """
-    if text.lower() in labels_lookup:
-        return True
-    for group in labels_cfg.attributes:
-        if group.type == "text" and group.key and text.lower().startswith(f"{group.key.lower()}:"):
-            return True
-    return False
-
-
-def build_labels_lookup(labels_cfg: EntityLabelsConfig | list | None) -> set[str]:
-    """
-    Build a set of valid 'key:value' label strings (lowercase) for fast lookup.
-
-    Accepts either EntityLabelsConfig or raw list/None for backwards compatibility.
-
-    Args:
-        labels_cfg: EntityLabelsConfig, raw list of attribute dicts, or None
-
-    Returns:
-        Set of lowercase 'key:value' strings
-    """
-    if labels_cfg is None:
-        return set()
-
-    # Accept raw list/dict for backwards compatibility
-    if not isinstance(labels_cfg, EntityLabelsConfig):
-        parsed = parse_entity_labels(labels_cfg)
-        if parsed is None:
-            return set()
-        labels_cfg = parsed
-
-    valid = set()
-    for group in labels_cfg.attributes:
-        if group.type == "text":
-            continue  # No fixed vocabulary — all values accepted in post-processing
-        for v in group.values:
-            if group.key and v.value:
-                valid.add(f"{group.key}:{v.value}".lower())
-    return valid
@@ -20,7 +20,6 @@ async def process_entities_batch(
    facts: list[ProcessedFact],
    log_buffer: list[str] = None,
    user_entities_per_content: dict[int, list[dict]] = None,
-    entity_labels: list | None = None,
 ) -> list[EntityLink]:
    """
    Process entities for all facts and create entity links.
@@ -91,7 +90,6 @@ async def process_entities_batch(
        fact_dates,
        entities_per_fact,
        log_buffer,  # Pass log_buffer for detailed logging
-        entity_labels=entity_labels,
    )

    return entity_links
@@ -10,31 +10,23 @@ import json
 import logging
 import re
 from datetime import datetime, timedelta
-from typing import Literal, cast
+from typing import Literal

-from pydantic import BaseModel, ConfigDict, Field, create_model, field_validator
+from pydantic import BaseModel, ConfigDict, Field, field_validator

 from ...config import get_config
-from ..llm_wrapper import LLMConfig, OutputTooLongError, sanitize_llm_output
+from ..llm_wrapper import LLMConfig, OutputTooLongError
 from ..response_models import TokenUsage
-from .entity_labels import (
-    EntityLabelsConfig,
-    build_labels_lookup,
-    build_labels_model,
-    is_label_entity,
-    parse_entity_labels,
-)


-def _infer_temporal_date(fact_text: str, event_date: datetime | None) -> str | None:
+def _infer_temporal_date(fact_text: str, event_date: datetime) -> str | None:
    """
    Infer a temporal date from fact text when LLM didn't provide occurred_start.

    This is a fallback for when the LLM fails to extract temporal information
    from relative time expressions like "last night", "yesterday", etc.
    """
-    if event_date is None:
-        return None
+    import re

    fact_lower = fact_text.lower()

@@ -66,7 +58,25 @@ def _infer_temporal_date(fact_text: str, event_date: datetime | None) -> str | N


 def _sanitize_text(text: str | None) -> str | None:
-    return sanitize_llm_output(text)
+    """
+    Sanitize text by removing characters that break downstream systems.
+
+    Removes:
+    - Null bytes (\\x00): Invalid in PostgreSQL UTF-8 encoding
+    - Unicode surrogates (U+D800-U+DFFF): Invalid in UTF-8, break LLM APIs
+
+    Surrogate characters are used in UTF-16 encoding but cannot be encoded
+    in UTF-8. They can appear in Python strings from improperly decoded data
+    (e.g., from JavaScript or broken files). Null bytes commonly appear in
+    OCR output, PDF extraction, or copy-paste from binary sources.
+    """
+    if text is None:
+        return None
+    if not text:
+        return text
+    # Remove null bytes and surrogate characters
+    text = text.replace("\x00", "")
+    return re.sub(r"[\ud800-\udfff]", "", text)


 class Entity(BaseModel):
@@ -92,6 +102,7 @@ class Fact(BaseModel):
    # Optional temporal fields
    occurred_start: str | None = None
    occurred_end: str | None = None
+    mentioned_at: str | None = None

    # Optional location field
    where: str | None = Field(
@@ -429,9 +440,11 @@ def _chunk_conversation(turns: list[dict], max_chars: int) -> list[str]:
 # Uses {extraction_guidelines} placeholder for mode-specific instructions
 _BASE_FACT_EXTRACTION_PROMPT = """Extract SIGNIFICANT facts from text. Be SELECTIVE - only extract facts worth remembering long-term.

-LANGUAGE: MANDATORY — Detect the language of the input text and produce ALL output in that EXACT same language. You are STRICTLY FORBIDDEN from translating or switching to any other language. Every single word of your output must be in the same language as the input. Do NOT output in a different language under any circumstance.
+LANGUAGE REQUIREMENT: Detect the language of the input text. All extracted facts, entity names, descriptions, and other output MUST be in the SAME language as the input. Do not translate to another language.

-{retain_mission_section}{extraction_guidelines}
+{fact_types_instruction}
+
+{extraction_guidelines}

 ══════════════════════════════════════════════════════════════════════════
 FACT FORMAT - BE CONCISE
@@ -470,9 +483,7 @@ TEMPORAL HANDLING
 ══════════════════════════════════════════════════════════════════════════

 Use "Event Date" from input as reference for relative dates.
- CRITICAL: Convert ALL relative temporal expressions to absolute dates in the fact text itself.
-  "yesterday" → write the resolved date (e.g. "on November 12, 2024"), NOT the word "yesterday"
-  "last night", "this morning", "today", "tonight" → convert to the resolved absolute date
+- "yesterday" relative to Event Date, not today
 - For events: set occurred_start AND occurred_end (same for point events)
 - For conversation facts: NO occurred dates

@@ -510,7 +521,7 @@ CONSOLIDATE related statements into ONE fact when possible."""
 _CONCISE_EXAMPLES = """

 ══════════════════════════════════════════════════════════════════════════
-EXAMPLES (shown in English for illustration; for non-English input, ALL output values MUST be in the input language)
+EXAMPLES
 ══════════════════════════════════════════════════════════════════════════

 Example 1 - Selective extraction (Event Date: June 10, 2024):
@@ -538,16 +549,16 @@ about experiences ARE important to remember, even if they seem small (e.g., how
 tasted, how someone looked, how loud music was). Extract these if they characterize
 an experience or person."""

-# Assembled concise prompt
+# Assembled concise prompt (backward compatible - exact same output as before)
 CONCISE_FACT_EXTRACTION_PROMPT = _BASE_FACT_EXTRACTION_PROMPT.format(
-    retain_mission_section="{retain_mission_section}",
+    fact_types_instruction="{fact_types_instruction}",
    extraction_guidelines=_CONCISE_GUIDELINES,
    examples=_CONCISE_EXAMPLES,
 )

 # Custom prompt uses same base but without examples
 CUSTOM_FACT_EXTRACTION_PROMPT = _BASE_FACT_EXTRACTION_PROMPT.format(
-    retain_mission_section="{retain_mission_section}",
+    fact_types_instruction="{fact_types_instruction}",
    extraction_guidelines="{custom_instructions}",
    examples="",  # No examples for custom mode
 )
@@ -556,7 +567,10 @@ CUSTOM_FACT_EXTRACTION_PROMPT = _BASE_FACT_EXTRACTION_PROMPT.format(
 # Verbose extraction prompt - detailed, comprehensive facts (legacy mode)
 VERBOSE_FACT_EXTRACTION_PROMPT = """Extract facts from text into structured format with FIVE required dimensions - BE EXTREMELY DETAILED.

-LANGUAGE: MANDATORY — Detect the language of the input text and produce ALL output in that EXACT same language. You are STRICTLY FORBIDDEN from translating or switching to any other language. Every single word of your output must be in the same language as the input. Do NOT output in a different language under any circumstance.
+LANGUAGE REQUIREMENT: Detect the language of the input text. All extracted facts, entity names, descriptions,
+and other output MUST be in the SAME language as the input. Do not translate to English if the input is in another language.
+
+{fact_types_instruction}

 ══════════════════════════════════════════════════════════════════════════
 FACT FORMAT - ALL FIVE DIMENSIONS REQUIRED - MAXIMUM VERBOSITY
@@ -681,185 +695,59 @@ Example: "Lost job → couldn't pay rent → moved apartment"
 - Fact 2: Moved apartment, causal_relations: [{target_index: 1, relation_type: "caused_by"}]"""


-def _build_labels_prompt_section(labels_cfg: EntityLabelsConfig | list | None, free_form_entities: bool = True) -> str:
-    """Build the entity labels classification section for the extraction prompt."""
-    if labels_cfg is None:
-        return ""
-
-    # Accept raw list for backwards compatibility
-    if isinstance(labels_cfg, list):
-        if not labels_cfg:
-            return ""
-        labels_cfg = parse_entity_labels(labels_cfg)
-        if labels_cfg is None:
-            return ""
-
-    if not labels_cfg.attributes:
-        return ""
-
-    if free_form_entities:
-        entities_instruction = "Classify each fact using the structured 'labels' field below. Continue extracting regular named entities in the 'entities' field."
-    else:
-        entities_instruction = "Classify each fact using the structured 'labels' field below. Do NOT add regular named entities — labels-only mode."
-
-    lines = [
-        "\n\n══════════════════════════════════════════════════════════════════════════",
-        "ENTITY LABELS - CLASSIFICATION ATTRIBUTES",
-        "══════════════════════════════════════════════════════════════════════════",
-        "",
-        entities_instruction,
-        "",
-        "For each fact, fill the 'labels' object. Each field is a label group:",
-        "",
-    ]
-
-    for attr in labels_cfg.attributes:
-        if attr.type == "text":
-            # Free-text: no predefined values — LLM writes any relevant string or null
-            lines.append(f"- {attr.key} (free text or null): {attr.description}")
-        else:
-            mode = "multi-value (list)" if attr.type == "multi-values" else "single value or null"
-            lines.append(f"- {attr.key} ({mode}): {attr.description}")
-            for v in attr.values:
-                desc = f" — {v.description}" if v.description else ""
-                lines.append(f'    • "{v.value}"{desc}')
-        lines.append("")
-
-    lines.append("Only assign labels when clearly applicable. Leave null/empty if the fact does not match.")
-    return "\n".join(lines)
-
-
 def _build_extraction_prompt_and_schema(config) -> tuple[str, type]:
    """
    Build extraction prompt and response schema based on config.

-    When a taxonomy is configured, dynamically builds a Pydantic model with a
-    typed `taxonomy_entities` field using an Enum built from valid taxonomy values.
-    This enables JSON schema enforcement for structured outputs.
-
    Returns:
        Tuple of (prompt, response_schema)
    """
+    fact_types_instruction = "Extract ONLY 'world' and 'assistant' type facts."
    extraction_mode = config.retain_extraction_mode
    extract_causal_links = config.retain_extract_causal_links

-    # Build retain_mission section if set - injected before the mode-specific guidelines
-    retain_mission = getattr(config, "retain_mission", None)
-    if retain_mission:
-        retain_mission_section = (
-            f"══════════════════════════════════════════════════════════════════════════\n"
-            f"FOCUS — What to retain for this bank\n"
-            f"══════════════════════════════════════════════════════════════════════════\n\n"
-            f"{retain_mission}\n\n"
-        )
-    else:
-        retain_mission_section = ""
-
    # Select base prompt based on extraction mode
    if extraction_mode == "custom":
        if not config.retain_custom_instructions:
            base_prompt = CONCISE_FACT_EXTRACTION_PROMPT
-            prompt = base_prompt.format(
-                retain_mission_section=retain_mission_section,
-            )
+            prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
        else:
            base_prompt = CUSTOM_FACT_EXTRACTION_PROMPT
            prompt = base_prompt.format(
-                retain_mission_section=retain_mission_section,
+                fact_types_instruction=fact_types_instruction,
                custom_instructions=config.retain_custom_instructions,
            )
    elif extraction_mode == "verbose":
-        prompt = VERBOSE_FACT_EXTRACTION_PROMPT
+        base_prompt = VERBOSE_FACT_EXTRACTION_PROMPT
+        prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
    else:
        base_prompt = CONCISE_FACT_EXTRACTION_PROMPT
-        prompt = base_prompt.format(
-            retain_mission_section=retain_mission_section,
-        )
+        prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)

    # Add causal relationships section if enabled
    if extract_causal_links:
        prompt = prompt + CAUSAL_RELATIONSHIPS_SECTION
-        base_fact_class = ExtractedFactVerbose if extraction_mode == "verbose" else ExtractedFact
-        base_response_class = FactExtractionResponseVerbose if extraction_mode == "verbose" else FactExtractionResponse
+        response_schema = FactExtractionResponseVerbose if extraction_mode == "verbose" else FactExtractionResponse
    else:
-        base_fact_class = ExtractedFactNoCausal
-        base_response_class = FactExtractionResponseNoCausal
-
-    # Add entity labels section if configured and build dynamic schema
-    entity_labels_raw = getattr(config, "entity_labels", None)
-    labels_cfg = parse_entity_labels(entity_labels_raw)
-    free_form_entities = getattr(config, "entities_allow_free_form", True)
-    labels_section = _build_labels_prompt_section(labels_cfg, free_form_entities)
-    if labels_section:
-        prompt = prompt + labels_section
-
-    response_schema = base_response_class
-
-    if labels_cfg and labels_cfg.attributes:
-        LabelsModel = build_labels_model(labels_cfg)
-        if LabelsModel is not None:
-            dynamic_fields: dict = {
-                "labels": (
-                    LabelsModel,
-                    Field(
-                        description="Classification labels for this fact. Fill each applicable field; leave others null/empty."
-                    ),
-                )
-            }
-            if not free_form_entities:
-                dynamic_fields["entities"] = (
-                    list[Entity] | None,
-                    Field(default=None, description="Leave empty — labels-only mode"),
-                )
-            # Inherit parent's required fields and add 'labels' so it appears in the JSON schema
-            # required array (the base class json_schema_extra overrides required entirely)
-            base_extra = base_fact_class.model_config.get("json_schema_extra")
-            base_required = cast(dict, base_extra).get("required", []) if isinstance(base_extra, dict) else []
-            DynamicFact = create_model(
-                "LabelsFact",
-                __base__=base_fact_class,
-                __config__=ConfigDict(
-                    json_schema_mode="validation",
-                    json_schema_extra={"required": [*base_required, "labels"]},
-                ),
-                **dynamic_fields,
-            )
-            DynamicResponse = create_model("LabelsResponse", facts=(list[DynamicFact], ...))  # type: ignore[valid-type]
-            response_schema = DynamicResponse
+        response_schema = FactExtractionResponseNoCausal

    return prompt, response_schema


-def _build_user_message(
-    chunk: str,
-    chunk_index: int,
-    total_chunks: int,
-    event_date: datetime | None,
-    context: str,
-    metadata: dict[str, str] | None = None,
-) -> str:
+def _build_user_message(chunk: str, chunk_index: int, total_chunks: int, event_date: datetime, context: str) -> str:
    """Build user message for fact extraction."""
    from .orchestrator import parse_datetime_flexible

    sanitized_chunk = _sanitize_text(chunk)
    sanitized_context = _sanitize_text(context) if context else "none"
-
-    if event_date is not None:
-        event_date = parse_datetime_flexible(event_date)
-        event_date_str = f"{event_date.strftime('%A, %B %d, %Y')} ({event_date.isoformat()})"
-    else:
-        event_date_str = "Unknown"
-
-    metadata_section = ""
-    if metadata:
-        metadata_lines = "\n".join(f"  {k}: {v}" for k, v in metadata.items())
-        metadata_section = f"\nMetadata:\n{metadata_lines}"
+    event_date = parse_datetime_flexible(event_date)
+    event_date_formatted = event_date.strftime("%A, %B %d, %Y")

    return f"""Extract facts from the following text chunk.

 Chunk: {chunk_index + 1}/{total_chunks}
-Event Date: {event_date_str}
-Context: {sanitized_context}{metadata_section}
+Event Date: {event_date_formatted} ({event_date.isoformat()})
+Context: {sanitized_context}

 Text:
 {sanitized_chunk}"""
@@ -896,12 +784,11 @@ async def _extract_facts_from_chunk(
    chunk: str,
    chunk_index: int,
    total_chunks: int,
-    event_date: datetime | None,
+    event_date: datetime,
    context: str,
    llm_config: "LLMConfig",
    config,
    agent_name: str = None,
-    metadata: dict[str, str] | None = None,
 ) -> tuple[list[dict[str, str]], TokenUsage]:
    """
    Extract facts from a single chunk (internal helper for parallel processing).
@@ -923,18 +810,19 @@ async def _extract_facts_from_chunk(
    extract_causal_links = config.retain_extract_causal_links

    # Build user message using helper function
-    user_message = _build_user_message(chunk, chunk_index, total_chunks, event_date, context, metadata)
+    user_message = _build_user_message(chunk, chunk_index, total_chunks, event_date, context)

    # Retry logic for JSON validation errors
-    # Use retain-specific overrides if set, otherwise fall back to global LLM config
-    llm_max_retries = (
-        config.retain_llm_max_retries if config.retain_llm_max_retries is not None else config.llm_max_retries
-    )
-    last_error: Exception | None = None
+    max_retries = 2
+    last_error = None

    usage = TokenUsage()  # Track cumulative usage across retries
-    for attempt in range(llm_max_retries):
+    for attempt in range(max_retries):
        try:
+            # Use retain-specific overrides if set, otherwise fall back to global LLM config
+            max_retries = (
+                config.retain_llm_max_retries if config.retain_llm_max_retries is not None else config.llm_max_retries
+            )
            initial_backoff = (
                config.retain_llm_initial_backoff
                if config.retain_llm_initial_backoff is not None
@@ -950,7 +838,7 @@ async def _extract_facts_from_chunk(
                scope="retain_extract_facts",
                temperature=0.1,
                max_completion_tokens=config.retain_max_completion_tokens,
-                max_retries=llm_max_retries,
+                max_retries=max_retries,
                initial_backoff=initial_backoff,
                max_backoff=max_backoff,
                skip_validation=True,  # Get raw JSON, we'll validate leniently
@@ -964,14 +852,14 @@ async def _extract_facts_from_chunk(

            # Handle malformed LLM responses
            if not isinstance(extraction_response_json, dict):
-                if attempt < llm_max_retries - 1:
+                if attempt < max_retries - 1:
                    logger.warning(
-                        f"LLM returned non-dict JSON on attempt {attempt + 1}/{llm_max_retries}: {type(extraction_response_json).__name__}. Retrying..."
+                        f"LLM returned non-dict JSON on attempt {attempt + 1}/{max_retries}: {type(extraction_response_json).__name__}. Retrying..."
                    )
                    continue
                else:
                    logger.warning(
-                        f"LLM returned non-dict JSON after {llm_max_retries} attempts: {type(extraction_response_json).__name__}. "
+                        f"LLM returned non-dict JSON after {max_retries} attempts: {type(extraction_response_json).__name__}. "
                        f"Raw: {str(extraction_response_json)[:500]}"
                    )
                    return [], usage
@@ -1081,9 +969,9 @@ async def _extract_facts_from_chunk(
                # Add entities if present (validate as Entity objects)
                # LLM sometimes returns strings instead of {"text": "..."} format
                entities = get_value("entities")
-                validated_entities = []
                if entities:
                    # Validate and normalize each entity
+                    validated_entities = []
                    for ent in entities:
                        if isinstance(ent, str):
                            # Normalize string to Entity object
@@ -1093,48 +981,8 @@ async def _extract_facts_from_chunk(
                                validated_entities.append(Entity.model_validate(ent))
                            except Exception as e:
                                logger.warning(f"Invalid entity {ent}: {e}")
-
-                # Post-process label entities from structured labels object
-                entity_labels_raw = getattr(config, "entity_labels", None)
-                labels_cfg = parse_entity_labels(entity_labels_raw)
-                free_form_entities = getattr(config, "entities_allow_free_form", True)
-                if labels_cfg and labels_cfg.attributes:
-                    labels_lookup = build_labels_lookup(labels_cfg)
-                    labels_data = llm_fact.get("labels") or {}
-                    if isinstance(labels_data, dict):
-                        existing_texts_lower = {e.text.lower() for e in validated_entities}
-                        for group in labels_cfg.attributes:
-                            value = labels_data.get(group.key)
-                            if not value:
-                                continue
-                            values_list = value if isinstance(value, list) else [value]
-                            for v in values_list:
-                                if not isinstance(v, str) or not v.strip() or v.lower() in ("none", "null", "n/a"):
-                                    continue
-                                label_str = f"{group.key}:{v.strip()}"
-                                if group.type == "text":
-                                    if label_str.lower() not in existing_texts_lower:
-                                        validated_entities.append(Entity(text=label_str))
-                                        existing_texts_lower.add(label_str.lower())
-                                elif (
-                                    label_str.lower() in labels_lookup and label_str.lower() not in existing_texts_lower
-                                ):
-                                    validated_entities.append(Entity(text=label_str))
-                                    existing_texts_lower.add(label_str.lower())
-                                else:
-                                    logger.warning(f"Label '{label_str}' not in valid label values, skipping")
-
-                    # In labels-only mode, keep only label entities
-                    if not free_form_entities:
-                        validated_entities = [
-                            e for e in validated_entities if is_label_entity(e.text, labels_cfg, labels_lookup)
-                        ]
-                elif not free_form_entities:
-                    # No labels but free_form disabled: clear all entities
-                    validated_entities = []
-
-                if validated_entities:
-                    fact_data["entities"] = validated_entities
+                    if validated_entities:
+                        fact_data["entities"] = validated_entities

                # Add per-fact causal relations (only if enabled in config)
                if extract_causal_links:
@@ -1173,9 +1021,8 @@ async def _extract_facts_from_chunk(
                    if validated_relations:
                        fact_data["causal_relations"] = validated_relations

-                # Set mentioned_at to the event_date (when the conversation/document occurred),
-                # or None when the caller opted into no timestamp.
-                fact_data["mentioned_at"] = event_date.isoformat() if event_date is not None else None
+                # Always set mentioned_at to the event_date (when the conversation/document occurred)
+                fact_data["mentioned_at"] = event_date.isoformat()

                # Build Fact model instance
                try:
@@ -1187,9 +1034,9 @@ async def _extract_facts_from_chunk(
                    continue

            # If we got malformed facts and haven't exhausted retries, try again
-            if has_malformed_facts and len(chunk_facts) < len(raw_facts) * 0.8 and attempt < llm_max_retries - 1:
+            if has_malformed_facts and len(chunk_facts) < len(raw_facts) * 0.8 and attempt < max_retries - 1:
                logger.warning(
-                    f"Got {len(raw_facts) - len(chunk_facts)} malformed facts out of {len(raw_facts)} on attempt {attempt + 1}/{llm_max_retries}. Retrying..."
+                    f"Got {len(raw_facts) - len(chunk_facts)} malformed facts out of {len(raw_facts)} on attempt {attempt + 1}/{max_retries}. Retrying..."
                )
                continue

@@ -1222,30 +1069,27 @@ async def _extract_facts_from_chunk(

            if "json_validate_failed" in str(e):
                logger.warning(
-                    f"          [1.3.{chunk_index + 1}] Attempt {attempt + 1}/{llm_max_retries} failed with JSON validation error: {e}"
+                    f"          [1.3.{chunk_index + 1}] Attempt {attempt + 1}/{max_retries} failed with JSON validation error: {e}"
                )
-                if attempt < llm_max_retries - 1:
+                if attempt < max_retries - 1:
                    logger.info(f"          [1.3.{chunk_index + 1}] Retrying...")
                    continue
            # If it's not a JSON validation error or we're out of retries, re-raise
            raise

-    # If we exhausted all retries, raise the last error or a descriptive fallback
-    if last_error is not None:
-        raise last_error
-    raise RuntimeError(f"Fact extraction failed after {llm_max_retries} attempts: LLM did not return valid JSON")
+    # If we exhausted all retries, raise the last error
+    raise last_error


 async def _extract_facts_with_auto_split(
    chunk: str,
    chunk_index: int,
    total_chunks: int,
-    event_date: datetime | None,
+    event_date: datetime,
    context: str,
    llm_config: LLMConfig,
    config,
    agent_name: str = None,
-    metadata: dict[str, str] | None = None,
 ) -> tuple[list[dict[str, str]], TokenUsage]:
    """
    Extract facts from a chunk with automatic splitting if output exceeds token limits.
@@ -1262,7 +1106,6 @@ async def _extract_facts_with_auto_split(
        llm_config: LLM configuration to use
        config: Resolved HindsightConfig for this bank
        agent_name: Optional agent name (memory owner)
-        metadata: Optional document metadata key-value pairs

    Returns:
        Tuple of (facts list, token usage) extracted from the chunk (possibly from sub-chunks)
@@ -1282,7 +1125,6 @@ async def _extract_facts_with_auto_split(
            llm_config=llm_config,
            config=config,
            agent_name=agent_name,
-            metadata=metadata,
        )
    except OutputTooLongError:
        # Output exceeded token limits - split the chunk in half and retry
@@ -1328,7 +1170,6 @@ async def _extract_facts_with_auto_split(
                llm_config=llm_config,
                config=config,
                agent_name=agent_name,
-                metadata=metadata,
            ),
            _extract_facts_with_auto_split(
                chunk=second_half,
@@ -1339,7 +1180,6 @@ async def _extract_facts_with_auto_split(
                llm_config=llm_config,
                config=config,
                agent_name=agent_name,
-                metadata=metadata,
            ),
        ]

@@ -1359,12 +1199,11 @@ async def _extract_facts_with_auto_split(

 async def extract_facts_from_text(
    text: str,
-    event_date: datetime | None,
+    event_date: datetime,
    llm_config: LLMConfig,
    agent_name: str,
    config,
    context: str = "",
-    metadata: dict[str, str] | None = None,
 ) -> tuple[list[Fact], list[tuple[str, int]], TokenUsage]:
    """
    Extract semantic facts from conversational or narrative text using LLM.
@@ -1382,7 +1221,6 @@ async def extract_facts_from_text(
        agent_name: Agent name (memory owner)
        config: Resolved HindsightConfig for this bank
        context: Context about the conversation/document
-        metadata: Optional document metadata key-value pairs

    Returns:
        Tuple of (facts, chunks, usage) where:
@@ -1410,7 +1248,6 @@ async def extract_facts_from_text(
            llm_config=llm_config,
            config=config,
            agent_name=agent_name,
-            metadata=metadata,
        )
        for i, chunk in enumerate(chunks)
    ]
@@ -1437,8 +1274,8 @@ from .types import ExtractedFact as ExtractedFactType

 logger = logging.getLogger(__name__)

-# Each fact gets 10ms offset to preserve ordering within a document
-SECONDS_PER_FACT = 0.01
+# Each fact gets 10 seconds offset to preserve ordering within a document
+SECONDS_PER_FACT = 10


 async def extract_facts_from_contents_batch_api(
@@ -1520,7 +1357,7 @@ async def extract_facts_from_contents_batch_api(

            # Build user message using helper function
            user_message = _build_user_message(
-                chunk, chunk_index_in_content, len(chunks), item.event_date, item.context, item.metadata or None
+                chunk, chunk_index_in_content, len(chunks), item.event_date, item.context
            )

            # Build request body using helper function
@@ -1732,8 +1569,8 @@ async def extract_facts_from_contents_batch_api(

            # Entities
            entities = get_value("entities")
-            validated_entities = []
            if entities:
+                validated_entities = []
                for ent in entities:
                    if isinstance(ent, str):
                        validated_entities.append(Entity(text=ent))
@@ -1742,45 +1579,8 @@ async def extract_facts_from_contents_batch_api(
                            validated_entities.append(Entity.model_validate(ent))
                        except Exception:
                            pass
-
-            # Post-process label entities from structured labels object
-            entity_labels_raw = getattr(config, "entity_labels", None)
-            labels_cfg_batch = parse_entity_labels(entity_labels_raw)
-            free_form_entities_batch = getattr(config, "entities_allow_free_form", True)
-            if labels_cfg_batch and labels_cfg_batch.attributes:
-                labels_lookup_batch = build_labels_lookup(labels_cfg_batch)
-                labels_data = llm_fact.get("labels") or {}
-                if isinstance(labels_data, dict):
-                    existing_texts_lower = {e.text.lower() for e in validated_entities}
-                    for group in labels_cfg_batch.attributes:
-                        value = labels_data.get(group.key)
-                        if not value:
-                            continue
-                        values_list = value if isinstance(value, list) else [value]
-                        for v in values_list:
-                            if not isinstance(v, str) or not v.strip() or v.lower() in ("none", "null", "n/a"):
-                                continue
-                            label_str = f"{group.key}:{v.strip()}"
-                            if group.type == "text":
-                                if label_str.lower() not in existing_texts_lower:
-                                    validated_entities.append(Entity(text=label_str))
-                                    existing_texts_lower.add(label_str.lower())
-                            elif (
-                                label_str.lower() in labels_lookup_batch
-                                and label_str.lower() not in existing_texts_lower
-                            ):
-                                validated_entities.append(Entity(text=label_str))
-                                existing_texts_lower.add(label_str.lower())
-
-                if not free_form_entities_batch:
-                    validated_entities = [
-                        e for e in validated_entities if is_label_entity(e.text, labels_cfg_batch, labels_lookup_batch)
-                    ]
-            elif not free_form_entities_batch:
-                validated_entities = []
-
-            if validated_entities:
-                fact_data["entities"] = validated_entities
+                if validated_entities:
+                    fact_data["entities"] = validated_entities

            # Causal relations
            if extract_causal_links:
@@ -1811,9 +1611,8 @@ async def extract_facts_from_contents_batch_api(
                if validated_relations:
                    fact_data["causal_relations"] = validated_relations

-            # Set mentioned_at to the event_date (when the conversation/document occurred),
-            # or None when the caller opted into no timestamp.
-            fact_data["mentioned_at"] = event_date.isoformat() if event_date is not None else None
+            # Always set mentioned_at
+            fact_data["mentioned_at"] = event_date.isoformat()

            try:
                fact = Fact(fact=combined_text, fact_type=fact_type, **fact_data)
@@ -1872,7 +1671,6 @@ async def extract_facts_from_contents_batch_api(
                mentioned_at=content.event_date,
                metadata=content.metadata,
                tags=content.tags,
-                observation_scopes=content.observation_scopes,
            )

            extracted_facts.append(extracted_fact)
@@ -1881,9 +1679,6 @@ async def extract_facts_from_contents_batch_api(
    # Step 7: Add temporal offsets
    _add_temporal_offsets(extracted_facts, contents)

-    # Step 8: Auto-tag facts from label groups with tag=True
-    _inject_label_tags(extracted_facts, config)
-
    logger.info(f"Batch API extracted {len(extracted_facts)} facts from {len(all_chunks_info)} chunks")

    return extracted_facts, chunks_metadata, total_usage
@@ -1942,7 +1737,6 @@ async def extract_facts_from_contents(
            llm_config=llm_config,
            agent_name=agent_name,
            config=config,
-            metadata=item.metadata or None,
        )
        fact_extraction_tasks.append(task)

@@ -2006,7 +1800,6 @@ async def extract_facts_from_contents(
                        mentioned_at=content.event_date,
                        metadata=content.metadata,
                        tags=content.tags,
-                        observation_scopes=content.observation_scopes,
                    )

                    extracted_facts.append(extracted_fact)
@@ -2016,9 +1809,6 @@ async def extract_facts_from_contents(
    # Step 4: Add time offsets to preserve ordering within each content
    _add_temporal_offsets(extracted_facts, contents)

-    # Step 5: Auto-tag facts from label groups with tag=True
-    _inject_label_tags(extracted_facts, config)
-
    return extracted_facts, chunks_metadata, total_usage


@@ -2074,24 +1864,3 @@ def _add_temporal_offsets(facts: list[ExtractedFactType], contents: list[RetainC
            fact.occurred_end = parse_datetime_flexible(fact.occurred_end) + offset
        if fact.mentioned_at:
            fact.mentioned_at = parse_datetime_flexible(fact.mentioned_at) + offset
-
-
-def _inject_label_tags(facts: list[ExtractedFactType], config) -> None:
-    """
-    For label groups with tag=True, add extracted key:value label entities
-    to each fact's tags list. Modifies facts in place.
-
-    This lets entity labels double as tags, enabling filtering via the
-    existing tags API without any extra query infrastructure.
-    """
-    labels_cfg = parse_entity_labels(getattr(config, "entity_labels", None))
-    if not labels_cfg:
-        return
-    tag_group_keys = {g.key.lower() for g in labels_cfg.attributes if g.tag}
-    if not tag_group_keys:
-        return
-    for fact in facts:
-        label_tags = [e for e in fact.entities if ":" in e and e.split(":", 1)[0].lower() in tag_group_keys]
-        if label_tags:
-            existing = set(fact.tags)
-            fact.tags = fact.tags + [t for t in label_tags if t not in existing]
@@ -47,8 +47,6 @@ async def insert_facts_batch(
    chunk_ids = []
    document_ids = []
    tags_list = []
-    observation_scopes_list = []
-    text_signals_list = []

    for fact in facts:
        fact_texts.append(_sanitize_text(fact.fact_text))
@@ -70,19 +68,6 @@ async def insert_facts_batch(
        document_ids.append(fact.document_id if fact.document_id else document_id)
        # Convert tags to JSON string for proper batch insertion (PostgreSQL unnest doesn't handle 2D arrays well)
        tags_list.append(json.dumps(fact.tags if fact.tags else []))
-        # observation_scopes: stored as JSONB (string or 2D array), None if not provided
-        observation_scopes_list.append(
-            json.dumps(fact.observation_scopes) if fact.observation_scopes is not None else None
-        )
-        # Build text_signals: entity names + date tokens for enriched BM25 indexing
-        signal_parts = []
-        if fact.entities:
-            signal_parts.extend(e.name for e in fact.entities)
-        if fact.occurred_start:
-            signal_parts.append(fact.occurred_start.strftime("%B %-d %Y"))
-        if fact.occurred_end and fact.occurred_end != fact.occurred_start:
-            signal_parts.append(fact.occurred_end.strftime("%B %-d %Y"))
-        text_signals_list.append(" ".join(signal_parts) if signal_parts else None)

    # Batch insert all facts
    # Note: tags are passed as JSON strings and converted back to varchar[] via jsonb_array_elements_text + array_agg
@@ -90,19 +75,16 @@ async def insert_facts_batch(
    config = get_config()
    if config.text_search_extension == "vchord":
        # VectorChord: manually tokenize and insert search_vector
-        # text_signals (entity names etc.) are included in the tokenize input for enriched BM25
        query = f"""
            WITH input_data AS (
                SELECT * FROM unnest(
                    $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
-                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[], $15::jsonb[], $16::text[]
+                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[]
                ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json,
-                       observation_scopes_json, text_signals)
+                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json)
            )
            INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags,
-                                     observation_scopes, text_signals, search_vector)
+                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags, search_vector)
            SELECT
                $1,
                text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
@@ -111,30 +93,23 @@ async def insert_facts_batch(
                    (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
                    '{{}}'::varchar[]
                ),
-                observation_scopes_json,
-                text_signals,
-                tokenize(
-                    COALESCE(text, '') || ' ' || COALESCE(context, '') || ' ' || COALESCE(text_signals, ''),
-                    'llmlingua2'
-                )::bm25_catalog.bm25vector
+                tokenize(COALESCE(text, '') || ' ' || COALESCE(context, ''), 'llmlingua2')::bm25_catalog.bm25vector
            FROM input_data
            RETURNING id
        """
    else:  # native or pg_textsearch
-        # Native PostgreSQL: search_vector is GENERATED ALWAYS (expression includes text_signals), don't include it
+        # Native PostgreSQL: search_vector is GENERATED ALWAYS, don't include it
        # pg_textsearch: indexes operate on base columns directly, don't populate search_vector
        query = f"""
            WITH input_data AS (
                SELECT * FROM unnest(
                    $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
-                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[], $15::jsonb[], $16::text[]
+                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[]
                ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json,
-                       observation_scopes_json, text_signals)
+                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json)
            )
            INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags,
-                                     observation_scopes, text_signals)
+                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags)
            SELECT
                $1,
                text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
@@ -142,9 +117,7 @@ async def insert_facts_batch(
                COALESCE(
                    (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
                    '{{}}'::varchar[]
-                ),
-                observation_scopes_json,
-                text_signals
+                )
            FROM input_data
            RETURNING id
        """
@@ -165,8 +138,6 @@ async def insert_facts_batch(
        chunk_ids,
        document_ids,
        tags_list,
-        observation_scopes_list,
-        text_signals_list,
    )

    unit_ids = [str(row["id"]) for row in results]
@@ -47,9 +47,6 @@ def compute_temporal_links(

    links = []
    for unit_id, unit_event_date in new_units.items():
-        # Units without event_date can't form temporal links
-        if unit_event_date is None:
-            continue
        # Normalize unit_event_date for consistent comparison
        unit_event_date_norm = _normalize_datetime(unit_event_date)

@@ -99,11 +96,7 @@ def compute_temporal_query_bounds(
        return None, None

    # Normalize all dates to be timezone-aware to avoid comparison issues
-    # Filter out None values — units without event_date can't form temporal links
-    all_dates = [_normalize_datetime(d) for d in new_units.values() if d is not None]
-
-    if not all_dates:
-        return None, None
+    all_dates = [_normalize_datetime(d) for d in new_units.values()]

    try:
        min_date = min(all_dates) - timedelta(hours=time_window_hours)
@@ -150,7 +143,6 @@ async def extract_entities_batch_optimized(
    fact_dates: list,
    llm_entities: list[list[dict]],
    log_buffer: list[str] = None,
-    entity_labels: list | None = None,
 ) -> list[tuple]:
    """
    Process LLM-extracted entities for ALL facts in batch.
@@ -240,7 +232,6 @@ async def extract_entities_batch_optimized(
                context=context,
                unit_event_date=None,  # Not used when per-entity dates provided
                conn=conn,  # Use main transaction connection
-                entity_labels=entity_labels,
            )

            _log(
@@ -441,23 +432,20 @@ async def create_temporal_links_batch_per_fact(
        min_date, max_date = compute_temporal_query_bounds(new_units, time_window_hours)

        fetch_neighbors_start = time_mod.time()
-        if min_date is not None and max_date is not None:
-            all_candidates = await conn.fetch(
-                f"""
-                SELECT id, event_date
-                FROM {fq_table("memory_units")}
-                WHERE bank_id = $1
-                  AND event_date BETWEEN $2 AND $3
-                  AND id::text != ALL($4)
-                ORDER BY event_date DESC
-                """,
-                bank_id,
-                min_date,
-                max_date,
-                unit_ids,
-            )
-        else:
-            all_candidates = []
+        all_candidates = await conn.fetch(
+            f"""
+            SELECT id, event_date
+            FROM {fq_table("memory_units")}
+            WHERE bank_id = $1
+              AND event_date BETWEEN $2 AND $3
+              AND id::text != ALL($4)
+            ORDER BY event_date DESC
+            """,
+            bank_id,
+            min_date,
+            max_date,
+            unit_ids,
+        )
        _log(
            log_buffer,
            f"      [7.2] Fetch {len(all_candidates)} candidate neighbors (1 query): {time_mod.time() - fetch_neighbors_start:.3f}s",
@@ -472,15 +460,11 @@ async def create_temporal_links_batch_per_fact(
            # Convert new_units dict to candidate format for within-batch linking
            new_unit_items = list(new_units.items())
            for i, (unit_id, event_date) in enumerate(new_unit_items):
-                if event_date is None:
-                    continue  # Skip units without event_date for temporal linking
                unit_event_date_norm = _normalize_datetime(event_date)

                # Compare with other new units (only those after this one to avoid duplicates)
                for j in range(i + 1, len(new_unit_items)):
                    other_id, other_event_date = new_unit_items[j]
-                    if other_event_date is None:
-                        continue  # Skip units without event_date
                    other_event_date_norm = _normalize_datetime(other_event_date)

                    # Check if within time window
@@ -498,13 +482,14 @@ async def create_temporal_links_batch_per_fact(
            # Batch inserts to avoid timeout on large batches
            BATCH_SIZE = 1000
            for batch_start in range(0, len(links), BATCH_SIZE):
+                batch = links[batch_start : batch_start + BATCH_SIZE]
                await conn.executemany(
                    f"""
                    INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
                    VALUES ($1, $2, $3, $4, $5)
                    ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
                    """,
-                    links[batch_start : batch_start + BATCH_SIZE],
+                    batch,
                )
            _log(log_buffer, f"      [7.4] Insert {len(links)} temporal links: {time_mod.time() - insert_start:.3f}s")

@@ -552,46 +537,82 @@ async def create_semantic_links_batch(

        import numpy as np

-        # Use pgvector ANN search (HNSW index) for each new unit instead of fetching
-        # all existing embeddings into Python. At large scale (100K+ units) the old
-        # approach would transfer 100K × 384 floats (~150 MB) per retain call; the
-        # ANN query completes in <5 ms and transfers only top_k rows.
-        ann_start = time_mod.time()
-        all_links = []
-
-        # Build UUID exclude list once for all ANN queries
-        import uuid as uuid_mod
-
-        exclude_uuids = [uuid_mod.UUID(uid) if isinstance(uid, str) else uid for uid in unit_ids]
-
-        for unit_id, new_embedding in zip(unit_ids, embeddings):
-            emb_str = str(list(new_embedding) if not isinstance(new_embedding, list) else new_embedding)
-            rows = await conn.fetch(
-                f"""
-                SELECT id::text,
-                       1 - (embedding <=> $1::vector) AS similarity
-                FROM {fq_table("memory_units")}
-                WHERE bank_id = $2
-                  AND embedding IS NOT NULL
-                  AND id != ALL($3::uuid[])
-                ORDER BY embedding <=> $1::vector
-                LIMIT $4
-                """,
-                emb_str,
-                bank_id,
-                exclude_uuids,
-                top_k,
-            )
-            for row in rows:
-                sim = float(min(1.0, max(0.0, row["similarity"])))
-                if sim >= threshold:
-                    all_links.append((unit_id, str(row["id"]), "semantic", sim, None))
-
+        # Fetch ALL existing units with embeddings in ONE query
+        fetch_start = time_mod.time()
+        all_existing = await conn.fetch(
+            f"""
+            SELECT id, embedding
+            FROM {fq_table("memory_units")}
+            WHERE bank_id = $1
+              AND embedding IS NOT NULL
+              AND id::text != ALL($2)
+            """,
+            bank_id,
+            unit_ids,
+        )
        _log(
            log_buffer,
-            f"      [8.1] ANN search for {len(unit_ids)} new units → {len(all_links)} candidate links: {time_mod.time() - ann_start:.3f}s",
+            f"      [8.1] Fetch {len(all_existing)} existing embeddings (1 query): {time_mod.time() - fetch_start:.3f}s",
        )

+        # Convert to numpy for vectorized similarity computation
+        compute_start = time_mod.time()
+        all_links = []
+
+        if all_existing:
+            # Convert existing embeddings to numpy array
+            existing_ids = [str(row["id"]) for row in all_existing]
+            # Stack embeddings as 2D array: (num_embeddings, embedding_dim)
+            embedding_arrays = []
+            for row in all_existing:
+                raw_emb = row["embedding"]
+                # Handle different pgvector formats
+                if isinstance(raw_emb, str):
+                    # Parse string format: "[1.0, 2.0, ...]"
+                    import json
+
+                    emb = np.array(json.loads(raw_emb), dtype=np.float32)
+                elif isinstance(raw_emb, (list, tuple)):
+                    emb = np.array(raw_emb, dtype=np.float32)
+                else:
+                    # Try direct conversion (works for numpy arrays, pgvector objects, etc.)
+                    emb = np.array(raw_emb, dtype=np.float32)
+
+                # Ensure it's 1D
+                if emb.ndim != 1:
+                    raise ValueError(f"Expected 1D embedding, got shape {emb.shape}")
+                embedding_arrays.append(emb)
+
+            if not embedding_arrays:
+                existing_embeddings = np.array([])
+            elif len(embedding_arrays) == 1:
+                # Single embedding: reshape to (1, dim)
+                existing_embeddings = embedding_arrays[0].reshape(1, -1)
+            else:
+                # Multiple embeddings: vstack
+                existing_embeddings = np.vstack(embedding_arrays)
+
+            # For each new unit, compute similarities with ALL existing units
+            for unit_id, new_embedding in zip(unit_ids, embeddings):
+                new_emb_array = np.array(new_embedding)
+
+                # Compute cosine similarities (dot product for normalized vectors)
+                similarities = np.dot(existing_embeddings, new_emb_array)
+
+                # Find top-k above threshold
+                # Get indices of similarities above threshold
+                above_threshold = np.where(similarities >= threshold)[0]
+
+                if len(above_threshold) > 0:
+                    # Sort by similarity (descending) and take top-k
+                    sorted_indices = above_threshold[np.argsort(-similarities[above_threshold])][:top_k]
+
+                    for idx in sorted_indices:
+                        similar_id = existing_ids[idx]
+                        # Clamp to [0, 1] to handle floating point precision issues
+                        similarity = float(min(1.0, max(0.0, similarities[idx])))
+                        all_links.append((unit_id, similar_id, "semantic", similarity, None))
+
        # Also compute similarities WITHIN the new batch (new units to each other)
        # Apply the same top_k limit per unit as we do for existing units
        if len(unit_ids) > 1:
@@ -622,7 +643,7 @@ async def create_semantic_links_batch(

        _log(
            log_buffer,
-            f"      [8.2] Within-batch similarities added {len(all_links)} total semantic links",
+            f"      [8.2] Compute similarities & generate {len(all_links)} semantic links: {time_mod.time() - compute_start:.3f}s",
        )

        if all_links:
@@ -630,13 +651,14 @@ async def create_semantic_links_batch(
            # Batch inserts to avoid timeout on large batches
            BATCH_SIZE = 1000
            for batch_start in range(0, len(all_links), BATCH_SIZE):
+                batch = all_links[batch_start : batch_start + BATCH_SIZE]
                await conn.executemany(
                    f"""
                    INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
                    VALUES ($1, $2, $3, $4, $5)
                    ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
                    """,
-                    all_links[batch_start : batch_start + BATCH_SIZE],
+                    batch,
                )
            _log(
                log_buffer, f"      [8.3] Insert {len(all_links)} semantic links: {time_mod.time() - insert_start:.3f}s"
@@ -652,18 +674,18 @@ async def create_semantic_links_batch(
        raise


-async def insert_entity_links_batch(conn, links: list[EntityLink], chunk_size: int = 5000):
+async def insert_entity_links_batch(conn, links: list[EntityLink], chunk_size: int = 50000):
    """
-    Insert all entity links using COPY to temp table + chunked INSERT for reliability.
+    Insert all entity links using COPY to temp table + INSERT for maximum speed.

-    Uses PostgreSQL COPY (via copy_records_to_table) for bulk loading into a
-    temp table, then INSERT ... ON CONFLICT in chunks of chunk_size. Chunking
-    prevents single-query timeouts on very large tables (100M+ rows).
+    Uses PostgreSQL COPY (via copy_records_to_table) for bulk loading,
+    then INSERT ... ON CONFLICT from temp table. This is the fastest
+    method for bulk inserts with conflict handling.

    Args:
        conn: Database connection
        links: List of EntityLink objects
-        chunk_size: Number of rows per INSERT chunk (default 5000)
+        chunk_size: Number of rows per batch (default 50000)
    """
    if not links:
        return
@@ -672,11 +694,10 @@ async def insert_entity_links_batch(conn, links: list[EntityLink], chunk_size: i

    total_start = time_mod.time()

-    # Create temp table with serial for stable chunked access
+    # Create temp table for bulk loading
    create_start = time_mod.time()
    await conn.execute("""
        CREATE TEMP TABLE IF NOT EXISTS _temp_entity_links (
-            _row_num SERIAL,
            from_unit_id uuid,
            to_unit_id uuid,
            link_type text,
@@ -693,7 +714,9 @@ async def insert_entity_links_batch(conn, links: list[EntityLink], chunk_size: i

    # Convert EntityLink objects to tuples for COPY
    convert_start = time_mod.time()
-    records = [(link.from_unit_id, link.to_unit_id, link.link_type, link.weight, link.entity_id) for link in links]
+    records = []
+    for link in links:
+        records.append((link.from_unit_id, link.to_unit_id, link.link_type, link.weight, link.entity_id))
    logger.debug(f"      [9.3] Convert {len(records)} records: {time_mod.time() - convert_start:.3f}s")

    # Bulk load using COPY (fastest method)
@@ -705,25 +728,15 @@ async def insert_entity_links_batch(conn, links: list[EntityLink], chunk_size: i
    )
    logger.debug(f"      [9.4] COPY {len(records)} records to temp table: {time_mod.time() - copy_start:.3f}s")

-    # Insert from temp table in chunks to avoid single-query timeouts on large tables
+    # Insert from temp table with ON CONFLICT (single query for all rows)
    insert_start = time_mod.time()
-    total_rows = len(records)
-    chunks = 0
-    for chunk_start in range(0, total_rows, chunk_size):
-        chunk_end = chunk_start + chunk_size
-        await conn.execute(
-            f"""
-            INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
-            SELECT from_unit_id, to_unit_id, link_type, weight, entity_id
-            FROM _temp_entity_links
-            WHERE _row_num > $1 AND _row_num <= $2
-            ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
-            """,
-            chunk_start,
-            chunk_end,
-        )
-        chunks += 1
-    logger.debug(f"      [9.5] INSERT {total_rows} rows in {chunks} chunks: {time_mod.time() - insert_start:.3f}s")
+    await conn.execute(f"""
+        INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
+        SELECT from_unit_id, to_unit_id, link_type, weight, entity_id
+        FROM _temp_entity_links
+        ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
+    """)
+    logger.debug(f"      [9.5] INSERT from temp table: {time_mod.time() - insert_start:.3f}s")
    logger.debug(f"      [9.TOTAL] Entity links batch insert: {time_mod.time() - total_start:.3f}s")


@@ -7,11 +7,10 @@ Coordinates all retain pipeline modules to store memories efficiently.
 import logging
 import time
 import uuid
-from collections.abc import Awaitable, Callable
 from datetime import UTC, datetime
 from typing import Any

-from ..db_utils import acquire_with_retry, retry_with_backoff
+from ..db_utils import acquire_with_retry
 from . import bank_utils


@@ -53,11 +52,10 @@ def parse_datetime_flexible(value: Any) -> datetime:
        raise TypeError(f"Expected datetime or string, got {type(value).__name__}")


-import asyncpg
-
 from ..response_models import TokenUsage
 from . import (
    chunk_storage,
+    deduplication,
    embedding_processing,
    entity_processing,
    fact_extraction,
@@ -75,6 +73,7 @@ async def retain_batch(
    llm_config,
    entity_resolver,
    format_date_fn,
+    duplicate_checker_fn,
    bank_id: str,
    contents_dicts: list[RetainContentDict],
    config,
@@ -85,7 +84,6 @@ async def retain_batch(
    document_tags: list[str] | None = None,
    operation_id: str | None = None,
    schema: str | None = None,
-    outbox_callback: Callable[["asyncpg.Connection"], Awaitable[None]] | None = None,
 ) -> tuple[list[list[str]], TokenUsage]:
    """
    Process a batch of content through the retain pipeline.
@@ -96,6 +94,7 @@ async def retain_batch(
        llm_config: LLM configuration for fact extraction
        entity_resolver: Entity resolver for entity processing
        format_date_fn: Function to format datetime to readable string
+        duplicate_checker_fn: Function to check for duplicate facts
        bank_id: Bank identifier
        contents_dicts: List of content dictionaries
        config: Resolved HindsightConfig for this bank
@@ -129,14 +128,12 @@ async def retain_batch(
        item_tags = item.get("tags", []) or []
        merged_tags = list(set(item_tags + (document_tags or [])))

-        # Handle event_date: distinguish "not provided" (default to now) from
-        # "explicitly None" (caller opted into no timestamp).
-        if "event_date" in item and item["event_date"] is None:
-            event_date_value = None  # Caller explicitly signalled "unknown date"
-        elif item.get("event_date"):
-            event_date_value = parse_datetime_flexible(item["event_date"])
+        # Handle event_date: parse flexibly (handles both datetime objects and ISO strings)
+        event_date_value = item.get("event_date")
+        if event_date_value:
+            event_date_value = parse_datetime_flexible(event_date_value)
        else:
-            event_date_value = utcnow()  # Backward-compatible default
+            event_date_value = utcnow()

        content = RetainContent(
            content=item["content"],
@@ -145,7 +142,6 @@ async def retain_batch(
            metadata=item.get("metadata", {}),
            entities=item.get("entities", []),
            tags=merged_tags,
-            observation_scopes=item.get("observation_scopes"),
        )
        contents.append(content)

@@ -160,20 +156,13 @@ async def retain_batch(
    )

    if not extracted_facts:
-        # Still need to create document if document_id was provided or chunks exist
-        from collections import defaultdict
-
-        docs_tracked = 0
+        # Still need to create document if document_id was provided
        async with acquire_with_retry(pool) as conn:
            async with conn.transaction():
-                # Group contents by document_id (consistent with normal path)
-                contents_by_doc_early = defaultdict(list)
-                for idx, content_dict in enumerate(contents_dicts):
-                    doc_id = content_dict.get("document_id")
-                    contents_by_doc_early[doc_id].append((idx, content_dict))
+                await fact_storage.ensure_bank_exists(conn, bank_id)

+                # Handle document tracking even with no facts
                if document_id:
-                    # Legacy: single document_id parameter
                    combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
                    # Collect tags from all content items and merge with document_tags
                    all_tags = set(document_tags or [])
@@ -198,57 +187,45 @@ async def retain_batch(
                    await fact_storage.handle_document_tracking(
                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params, merged_tags
                    )
-                    docs_tracked += 1
                else:
-                    # Handle per-item document_ids and/or chunks (mirrors normal path logic)
-                    has_any_doc_ids = any(item.get("document_id") for item in contents_dicts)
+                    # Check for per-item document_ids
+                    from collections import defaultdict

-                    if has_any_doc_ids or chunks:
-                        for original_doc_id, doc_contents in contents_by_doc_early.items():
-                            should_create_doc = (original_doc_id is not None) or chunks
-                            if not should_create_doc:
-                                continue
+                    contents_by_doc = defaultdict(list)
+                    for idx, content_dict in enumerate(contents_dicts):
+                        doc_id = content_dict.get("document_id")
+                        if doc_id:
+                            contents_by_doc[doc_id].append((idx, content_dict))

-                            actual_doc_id = original_doc_id
-                            if actual_doc_id is None:
-                                # No document_id but have chunks - generate one
-                                actual_doc_id = str(uuid.uuid4())
+                    for doc_id, doc_contents in contents_by_doc.items():
+                        combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
+                        # Collect tags from all content items for this document and merge with document_tags
+                        all_tags = set(document_tags or [])
+                        for _, item in doc_contents:
+                            item_tags = item.get("tags", []) or []
+                            all_tags.update(item_tags)
+                        merged_tags = list(all_tags)

-                            combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
-                            all_tags = set(document_tags or [])
-                            for _, item in doc_contents:
-                                item_tags = item.get("tags", []) or []
-                                all_tags.update(item_tags)
-                            merged_tags = list(all_tags)
-
-                            retain_params = {}
-                            if doc_contents:
-                                first_item = doc_contents[0][1]
-                                if first_item.get("context"):
-                                    retain_params["context"] = first_item["context"]
-                                if first_item.get("event_date"):
-                                    retain_params["event_date"] = (
-                                        first_item["event_date"].isoformat()
-                                        if hasattr(first_item["event_date"], "isoformat")
-                                        else str(first_item["event_date"])
-                                    )
-                                if first_item.get("metadata"):
-                                    retain_params["metadata"] = first_item["metadata"]
-                            await fact_storage.handle_document_tracking(
-                                conn,
-                                bank_id,
-                                actual_doc_id,
-                                combined_content,
-                                is_first_batch,
-                                retain_params,
-                                merged_tags,
-                            )
-                            docs_tracked += 1
+                        retain_params = {}
+                        if doc_contents:
+                            first_item = doc_contents[0][1]
+                            if first_item.get("context"):
+                                retain_params["context"] = first_item["context"]
+                            if first_item.get("event_date"):
+                                retain_params["event_date"] = (
+                                    first_item["event_date"].isoformat()
+                                    if hasattr(first_item["event_date"], "isoformat")
+                                    else str(first_item["event_date"])
+                                )
+                            if first_item.get("metadata"):
+                                retain_params["metadata"] = first_item["metadata"]
+                        await fact_storage.handle_document_tracking(
+                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params, merged_tags
+                        )

        total_time = time.time() - start_time
-        doc_status = f"{docs_tracked} document(s) tracked" if docs_tracked > 0 else "no document tracked"
        logger.info(
-            f"RETAIN_BATCH COMPLETE: 0 facts extracted from {len(contents)} contents in {total_time:.3f}s ({doc_status}, no facts)"
+            f"RETAIN_BATCH COMPLETE: 0 facts extracted from {len(contents)} contents in {total_time:.3f}s (document tracked, no facts)"
        )
        return [[] for _ in contents], usage

@@ -269,6 +246,9 @@ async def retain_batch(
        for extracted_fact, embedding in zip(extracted_facts, embeddings)
    ]

+    # Track document IDs for logging
+    document_ids_added = []
+
    # Group contents by document_id for document tracking and chunk storage
    from collections import defaultdict

@@ -277,268 +257,267 @@ async def retain_batch(
        doc_id = content_dict.get("document_id")
        contents_by_doc[doc_id].append((idx, content_dict))

-    # Step 4: Database transaction (retried on deadlock)
-    result_unit_ids: list[list[str]] = []
+    # Step 4: Database transaction
+    async with acquire_with_retry(pool) as conn:
+        async with conn.transaction():
+            # Ensure bank exists
+            await fact_storage.ensure_bank_exists(conn, bank_id)

-    log_buffer_pre_db = len(log_buffer)
+            # Handle document tracking for all documents
+            step_start = time.time()
+            # Map None document_id to generated UUIDs
+            doc_id_mapping = {}  # Maps original doc_id (including None) to actual doc_id used

-    async def _run_db_work() -> None:
-        nonlocal result_unit_ids
+            if document_id:
+                # Legacy: single document_id parameter
+                combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
+                retain_params = {}
+                # Collect tags from all content items and merge with document_tags
+                all_tags = set(document_tags or [])
+                for item in contents_dicts:
+                    item_tags = item.get("tags", []) or []
+                    all_tags.update(item_tags)
+                merged_tags = list(all_tags)

-        # Reset per-fact mutations and log buffer so each retry attempt starts clean
-        del log_buffer[log_buffer_pre_db:]
-        document_ids_added: list[str] = []
-        for pf in processed_facts:
-            pf.document_id = None
-            pf.chunk_id = None
+                if contents_dicts:
+                    first_item = contents_dicts[0]
+                    if first_item.get("context"):
+                        retain_params["context"] = first_item["context"]
+                    if first_item.get("event_date"):
+                        retain_params["event_date"] = (
+                            first_item["event_date"].isoformat()
+                            if hasattr(first_item["event_date"], "isoformat")
+                            else str(first_item["event_date"])
+                        )
+                    if first_item.get("metadata"):
+                        retain_params["metadata"] = first_item["metadata"]

-        async with acquire_with_retry(pool) as conn:
-            async with conn.transaction():
-                # Handle document tracking for all documents
-                step_start = time.time()
-                # Map None document_id to generated UUIDs
-                doc_id_mapping = {}  # Maps original doc_id (including None) to actual doc_id used
+                await fact_storage.handle_document_tracking(
+                    conn, bank_id, document_id, combined_content, is_first_batch, retain_params, merged_tags
+                )
+                document_ids_added.append(document_id)
+                doc_id_mapping[None] = document_id  # For backwards compatibility
+            else:
+                # Handle per-item document_ids (create documents if any item has document_id or if chunks exist)
+                has_any_doc_ids = any(item.get("document_id") for item in contents_dicts)

-                if document_id:
-                    # Legacy: single document_id parameter
-                    combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
-                    retain_params = {}
-                    # Collect tags from all content items and merge with document_tags
-                    all_tags = set(document_tags or [])
-                    for item in contents_dicts:
-                        item_tags = item.get("tags", []) or []
-                        all_tags.update(item_tags)
-                    merged_tags = list(all_tags)
+                if has_any_doc_ids or chunks:
+                    for original_doc_id, doc_contents in contents_by_doc.items():
+                        actual_doc_id = original_doc_id

-                    if contents_dicts:
-                        first_item = contents_dicts[0]
-                        if first_item.get("context"):
-                            retain_params["context"] = first_item["context"]
-                        if first_item.get("event_date"):
-                            retain_params["event_date"] = (
-                                first_item["event_date"].isoformat()
-                                if hasattr(first_item["event_date"], "isoformat")
-                                else str(first_item["event_date"])
+                        # Only create document record if:
+                        # 1. Item has explicit document_id, OR
+                        # 2. There are chunks (need document for chunk storage)
+                        should_create_doc = (original_doc_id is not None) or chunks
+
+                        if should_create_doc:
+                            if actual_doc_id is None:
+                                # No document_id but have chunks - generate one
+                                actual_doc_id = str(uuid.uuid4())
+
+                            # Store mapping for later use
+                            doc_id_mapping[original_doc_id] = actual_doc_id
+
+                            # Combine content for this document
+                            combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
+
+                            # Collect tags from all content items for this document and merge with document_tags
+                            all_tags = set(document_tags or [])
+                            for _, item in doc_contents:
+                                item_tags = item.get("tags", []) or []
+                                all_tags.update(item_tags)
+                            merged_tags = list(all_tags)
+
+                            # Extract retain params from first content item
+                            retain_params = {}
+                            if doc_contents:
+                                first_item = doc_contents[0][1]
+                                if first_item.get("context"):
+                                    retain_params["context"] = first_item["context"]
+                                if first_item.get("event_date"):
+                                    retain_params["event_date"] = (
+                                        first_item["event_date"].isoformat()
+                                        if hasattr(first_item["event_date"], "isoformat")
+                                        else str(first_item["event_date"])
+                                    )
+                                if first_item.get("metadata"):
+                                    retain_params["metadata"] = first_item["metadata"]
+
+                            await fact_storage.handle_document_tracking(
+                                conn,
+                                bank_id,
+                                actual_doc_id,
+                                combined_content,
+                                is_first_batch,
+                                retain_params,
+                                merged_tags,
                            )
-                        if first_item.get("metadata"):
-                            retain_params["metadata"] = first_item["metadata"]
+                            document_ids_added.append(actual_doc_id)

-                    await fact_storage.handle_document_tracking(
-                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params, merged_tags
-                    )
-                    document_ids_added.append(document_id)
-                    doc_id_mapping[None] = document_id  # For backwards compatibility
-                else:
-                    # Handle per-item document_ids (create documents if any item has document_id or if chunks exist)
-                    has_any_doc_ids = any(item.get("document_id") for item in contents_dicts)
-
-                    if has_any_doc_ids or chunks:
-                        for original_doc_id, doc_contents in contents_by_doc.items():
-                            actual_doc_id = original_doc_id
-
-                            # Only create document record if:
-                            # 1. Item has explicit document_id, OR
-                            # 2. There are chunks (need document for chunk storage)
-                            should_create_doc = (original_doc_id is not None) or chunks
-
-                            if should_create_doc:
-                                if actual_doc_id is None:
-                                    # No document_id but have chunks - generate one
-                                    actual_doc_id = str(uuid.uuid4())
-
-                                # Store mapping for later use
-                                doc_id_mapping[original_doc_id] = actual_doc_id
-
-                                # Combine content for this document
-                                combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
-
-                                # Collect tags from all content items for this document and merge with document_tags
-                                all_tags = set(document_tags or [])
-                                for _, item in doc_contents:
-                                    item_tags = item.get("tags", []) or []
-                                    all_tags.update(item_tags)
-                                merged_tags = list(all_tags)
-
-                                # Extract retain params from first content item
-                                retain_params = {}
-                                if doc_contents:
-                                    first_item = doc_contents[0][1]
-                                    if first_item.get("context"):
-                                        retain_params["context"] = first_item["context"]
-                                    if first_item.get("event_date"):
-                                        retain_params["event_date"] = (
-                                            first_item["event_date"].isoformat()
-                                            if hasattr(first_item["event_date"], "isoformat")
-                                            else str(first_item["event_date"])
-                                        )
-                                    if first_item.get("metadata"):
-                                        retain_params["metadata"] = first_item["metadata"]
-
-                                await fact_storage.handle_document_tracking(
-                                    conn,
-                                    bank_id,
-                                    actual_doc_id,
-                                    combined_content,
-                                    is_first_batch,
-                                    retain_params,
-                                    merged_tags,
-                                )
-                                document_ids_added.append(actual_doc_id)
-
-                if document_ids_added:
-                    log_buffer.append(
-                        f"[2.5] Document tracking: {len(document_ids_added)} documents in {time.time() - step_start:.3f}s"
-                    )
-
-                # Store chunks and map to facts for all documents
-                step_start = time.time()
-                chunk_id_map_by_doc = {}  # Maps (doc_id, chunk_index) -> chunk_id
-
-                if chunks:
-                    # Group chunks by their source document
-                    chunks_by_doc = defaultdict(list)
-                    for chunk in chunks:
-                        # chunk.content_index tells us which content this chunk came from
-                        original_doc_id = contents_dicts[chunk.content_index].get("document_id")
-                        # Map to actual document_id (handles None -> generated UUID mapping)
-                        actual_doc_id = doc_id_mapping.get(original_doc_id, original_doc_id)
-                        if actual_doc_id is None and document_id:
-                            actual_doc_id = document_id
-                        chunks_by_doc[actual_doc_id].append(chunk)
-
-                    # Store chunks for each document
-                    for doc_id, doc_chunks in chunks_by_doc.items():
-                        chunk_id_map = await chunk_storage.store_chunks_batch(conn, bank_id, doc_id, doc_chunks)
-                        # Store mapping with document context
-                        for chunk_idx, chunk_id in chunk_id_map.items():
-                            chunk_id_map_by_doc[(doc_id, chunk_idx)] = chunk_id
-
-                    log_buffer.append(
-                        f"[3] Store chunks: {len(chunks)} chunks for {len(chunks_by_doc)} documents in {time.time() - step_start:.3f}s"
-                    )
-
-                    # Map chunk_ids and document_ids to facts
-                    for fact, processed_fact in zip(extracted_facts, processed_facts):
-                        # Get the original document_id for this fact's source content
-                        original_doc_id = contents_dicts[fact.content_index].get("document_id")
-                        # Map to actual document_id (handles None -> generated UUID mapping)
-                        actual_doc_id = doc_id_mapping.get(original_doc_id, original_doc_id)
-                        if actual_doc_id is None and document_id:
-                            actual_doc_id = document_id
-
-                        # Set document_id on the fact
-                        processed_fact.document_id = actual_doc_id
-
-                        # Map chunk_id if this fact came from a chunk
-                        if fact.chunk_index is not None:
-                            # Look up chunk_id using (doc_id, chunk_index)
-                            chunk_id = chunk_id_map_by_doc.get((actual_doc_id, fact.chunk_index))
-                            if chunk_id:
-                                processed_fact.chunk_id = chunk_id
-                else:
-                    # No chunks - still need to set document_id on facts
-                    for fact, processed_fact in zip(extracted_facts, processed_facts):
-                        original_doc_id = contents_dicts[fact.content_index].get("document_id")
-                        # Map to actual document_id (handles None -> generated UUID mapping)
-                        actual_doc_id = doc_id_mapping.get(original_doc_id, original_doc_id)
-                        if actual_doc_id is None and document_id:
-                            actual_doc_id = document_id
-                        processed_fact.document_id = actual_doc_id
-
-                non_duplicate_facts = processed_facts
-
-                # Insert facts (document_id is now stored per-fact)
-                step_start = time.time()
-                unit_ids = await fact_storage.insert_facts_batch(conn, bank_id, non_duplicate_facts)
-                log_buffer.append(f"[5] Insert facts: {len(unit_ids)} units in {time.time() - step_start:.3f}s")
-
-                # Process entities
-                step_start = time.time()
-                # Build map of content_index -> user entities for merging
-                user_entities_per_content = {
-                    idx: content.entities for idx, content in enumerate(contents) if content.entities
-                }
-                entity_links = await entity_processing.process_entities_batch(
-                    entity_resolver,
-                    conn,
-                    bank_id,
-                    unit_ids,
-                    non_duplicate_facts,
-                    log_buffer,
-                    user_entities_per_content=user_entities_per_content,
-                    entity_labels=getattr(config, "entity_labels", None),
-                )
-                log_buffer.append(f"[6] Process entities: {len(entity_links)} links in {time.time() - step_start:.3f}s")
-
-                # Create temporal links
-                step_start = time.time()
-                temporal_link_count = await link_creation.create_temporal_links_batch(conn, bank_id, unit_ids)
-                log_buffer.append(f"[7] Temporal links: {temporal_link_count} links in {time.time() - step_start:.3f}s")
-
-                # Create semantic links
-                step_start = time.time()
-                embeddings_for_links = [fact.embedding for fact in non_duplicate_facts]
-                semantic_link_count = await link_creation.create_semantic_links_batch(
-                    conn, bank_id, unit_ids, embeddings_for_links
-                )
-                log_buffer.append(f"[8] Semantic links: {semantic_link_count} links in {time.time() - step_start:.3f}s")
-
-                # Insert entity links
-                step_start = time.time()
-                if entity_links:
-                    await entity_processing.insert_entity_links_batch(conn, entity_links)
-                log_buffer.append(
-                    f"[9] Entity links: {len(entity_links) if entity_links else 0} links in {time.time() - step_start:.3f}s"
-                )
-
-                # Create causal links
-                step_start = time.time()
-                causal_link_count = await link_creation.create_causal_links_batch(conn, unit_ids, non_duplicate_facts)
-                log_buffer.append(f"[10] Causal links: {causal_link_count} links in {time.time() - step_start:.3f}s")
-
-                # Map results back to original content items
-                result_unit_ids = _map_results_to_contents(contents, extracted_facts, unit_ids)
-
-                # Transactional outbox: queue any side-effect tasks (e.g. webhook deliveries)
-                # inside the same transaction so they are atomically committed with the retain data.
-                if outbox_callback:
-                    await outbox_callback(conn)
-
-            # Flush entity stats (mention_count / last_seen) now that the transaction
-            # has committed.  Uses a fresh pool connection — no locks held.
-            await entity_resolver.flush_pending_stats()
-
-            # Log final summary
-            total_time = time.time() - start_time
-            log_buffer.append(f"{'=' * 60}")
-            log_buffer.append(f"RETAIN_BATCH COMPLETE: {len(unit_ids)} units in {total_time:.3f}s")
            if document_ids_added:
-                log_buffer.append(f"Documents: {', '.join(document_ids_added)}")
-            log_buffer.append(f"{'=' * 60}")
+                log_buffer.append(
+                    f"[2.5] Document tracking: {len(document_ids_added)} documents in {time.time() - step_start:.3f}s"
+                )

-            logger.info("\n" + "\n".join(log_buffer) + "\n")
+            # Store chunks and map to facts for all documents
+            step_start = time.time()
+            chunk_id_map_by_doc = {}  # Maps (doc_id, chunk_index) -> chunk_id

-    await retry_with_backoff(_run_db_work)
-    return result_unit_ids, usage
+            if chunks:
+                # Group chunks by their source document
+                chunks_by_doc = defaultdict(list)
+                for chunk in chunks:
+                    # chunk.content_index tells us which content this chunk came from
+                    original_doc_id = contents_dicts[chunk.content_index].get("document_id")
+                    # Map to actual document_id (handles None -> generated UUID mapping)
+                    actual_doc_id = doc_id_mapping.get(original_doc_id, original_doc_id)
+                    if actual_doc_id is None and document_id:
+                        actual_doc_id = document_id
+                    chunks_by_doc[actual_doc_id].append(chunk)
+
+                # Store chunks for each document
+                for doc_id, doc_chunks in chunks_by_doc.items():
+                    chunk_id_map = await chunk_storage.store_chunks_batch(conn, bank_id, doc_id, doc_chunks)
+                    # Store mapping with document context
+                    for chunk_idx, chunk_id in chunk_id_map.items():
+                        chunk_id_map_by_doc[(doc_id, chunk_idx)] = chunk_id
+
+                log_buffer.append(
+                    f"[3] Store chunks: {len(chunks)} chunks for {len(chunks_by_doc)} documents in {time.time() - step_start:.3f}s"
+                )
+
+                # Map chunk_ids and document_ids to facts
+                for fact, processed_fact in zip(extracted_facts, processed_facts):
+                    # Get the original document_id for this fact's source content
+                    original_doc_id = contents_dicts[fact.content_index].get("document_id")
+                    # Map to actual document_id (handles None -> generated UUID mapping)
+                    actual_doc_id = doc_id_mapping.get(original_doc_id, original_doc_id)
+                    if actual_doc_id is None and document_id:
+                        actual_doc_id = document_id
+
+                    # Set document_id on the fact
+                    processed_fact.document_id = actual_doc_id
+
+                    # Map chunk_id if this fact came from a chunk
+                    if fact.chunk_index is not None:
+                        # Look up chunk_id using (doc_id, chunk_index)
+                        chunk_id = chunk_id_map_by_doc.get((actual_doc_id, fact.chunk_index))
+                        if chunk_id:
+                            processed_fact.chunk_id = chunk_id
+            else:
+                # No chunks - still need to set document_id on facts
+                for fact, processed_fact in zip(extracted_facts, processed_facts):
+                    original_doc_id = contents_dicts[fact.content_index].get("document_id")
+                    # Map to actual document_id (handles None -> generated UUID mapping)
+                    actual_doc_id = doc_id_mapping.get(original_doc_id, original_doc_id)
+                    if actual_doc_id is None and document_id:
+                        actual_doc_id = document_id
+                    processed_fact.document_id = actual_doc_id
+
+            # Deduplication
+            step_start = time.time()
+            is_duplicate_flags = await deduplication.check_duplicates_batch(
+                conn, bank_id, processed_facts, duplicate_checker_fn
+            )
+            log_buffer.append(
+                f"[4] Deduplication: {sum(is_duplicate_flags)} duplicates in {time.time() - step_start:.3f}s"
+            )
+
+            # Filter out duplicates
+            non_duplicate_facts = deduplication.filter_duplicates(processed_facts, is_duplicate_flags)
+
+            if not non_duplicate_facts:
+                return [[] for _ in contents], usage
+
+            # Insert facts (document_id is now stored per-fact)
+            step_start = time.time()
+            unit_ids = await fact_storage.insert_facts_batch(conn, bank_id, non_duplicate_facts)
+            log_buffer.append(f"[5] Insert facts: {len(unit_ids)} units in {time.time() - step_start:.3f}s")
+
+            # Process entities
+            step_start = time.time()
+            # Build map of content_index -> user entities for merging
+            user_entities_per_content = {
+                idx: content.entities for idx, content in enumerate(contents) if content.entities
+            }
+            entity_links = await entity_processing.process_entities_batch(
+                entity_resolver,
+                conn,
+                bank_id,
+                unit_ids,
+                non_duplicate_facts,
+                log_buffer,
+                user_entities_per_content=user_entities_per_content,
+            )
+            log_buffer.append(f"[6] Process entities: {len(entity_links)} links in {time.time() - step_start:.3f}s")
+
+            # Create temporal links
+            step_start = time.time()
+            temporal_link_count = await link_creation.create_temporal_links_batch(conn, bank_id, unit_ids)
+            log_buffer.append(f"[7] Temporal links: {temporal_link_count} links in {time.time() - step_start:.3f}s")
+
+            # Create semantic links
+            step_start = time.time()
+            embeddings_for_links = [fact.embedding for fact in non_duplicate_facts]
+            semantic_link_count = await link_creation.create_semantic_links_batch(
+                conn, bank_id, unit_ids, embeddings_for_links
+            )
+            log_buffer.append(f"[8] Semantic links: {semantic_link_count} links in {time.time() - step_start:.3f}s")
+
+            # Insert entity links
+            step_start = time.time()
+            if entity_links:
+                await entity_processing.insert_entity_links_batch(conn, entity_links)
+            log_buffer.append(
+                f"[9] Entity links: {len(entity_links) if entity_links else 0} links in {time.time() - step_start:.3f}s"
+            )
+
+            # Create causal links
+            step_start = time.time()
+            causal_link_count = await link_creation.create_causal_links_batch(conn, unit_ids, non_duplicate_facts)
+            log_buffer.append(f"[10] Causal links: {causal_link_count} links in {time.time() - step_start:.3f}s")
+
+            # Map results back to original content items
+            result_unit_ids = _map_results_to_contents(contents, extracted_facts, is_duplicate_flags, unit_ids)
+
+        # Log final summary
+        total_time = time.time() - start_time
+        log_buffer.append(f"{'=' * 60}")
+        log_buffer.append(f"RETAIN_BATCH COMPLETE: {len(unit_ids)} units in {total_time:.3f}s")
+        if document_ids_added:
+            log_buffer.append(f"Documents: {', '.join(document_ids_added)}")
+        log_buffer.append(f"{'=' * 60}")
+
+        logger.info("\n" + "\n".join(log_buffer) + "\n")
+
+        return result_unit_ids, usage


 def _map_results_to_contents(
    contents: list[RetainContent],
    extracted_facts: list[ExtractedFact],
+    is_duplicate_flags: list[bool],
    unit_ids: list[str],
 ) -> list[list[str]]:
-    """Map created unit IDs back to original content items."""
-    facts_by_content: dict[int, list[int]] = {i: [] for i in range(len(contents))}
+    """
+    Map created unit IDs back to original content items.
+
+    Accounts for duplicates when mapping back.
+    """
+    result_unit_ids = []
+    filtered_idx = 0
+
+    # Group facts by content_index
+    facts_by_content = {i: [] for i in range(len(contents))}
    for i, fact in enumerate(extracted_facts):
        facts_by_content[fact.content_index].append(i)

-    result_unit_ids = []
-    unit_idx = 0
    for content_index in range(len(contents)):
        content_unit_ids = []
-        for _ in facts_by_content[content_index]:
-            content_unit_ids.append(unit_ids[unit_idx])
-            unit_idx += 1
+        for fact_idx in facts_by_content[content_index]:
+            if not is_duplicate_flags[fact_idx]:
+                content_unit_ids.append(unit_ids[filtered_idx])
+                filtered_idx += 1
        result_unit_ids.append(content_unit_ids)

    return result_unit_ids
@@ -6,8 +6,8 @@ from content input to fact storage.
 """

 from dataclasses import dataclass, field
-from datetime import datetime
-from typing import Literal, TypedDict
+from datetime import UTC, datetime
+from typing import TypedDict
 from uuid import UUID


@@ -22,21 +22,20 @@ class RetainContentDict(TypedDict, total=False):
        document_id: Document ID for this content item (optional)
        entities: User-provided entities to merge with extracted entities (optional)
        tags: Visibility scope tags for this content item (optional)
-        observation_scopes: How to scope observations for consolidation (optional).
-            "per_tag" runs one pass per individual tag; "combined" (default) runs a
-            single pass with all tags; a list[list[str]] specifies exact passes.
    """

    content: str  # Required
    context: str
-    event_date: datetime | None
+    event_date: datetime
    metadata: dict[str, str]
    document_id: str
    entities: list[dict[str, str]]  # [{"text": "...", "type": "..."}]
    tags: list[str]  # Visibility scope tags
-    observation_scopes: (
-        Literal["per_tag", "combined", "all_combinations"] | list[list[str]]
-    )  # Observation scopes for consolidation
+
+
+def _now_utc() -> datetime:
+    """Factory function for default event_date."""
+    return datetime.now(UTC)


@dataclass
@@ -49,13 +48,10 @@ class RetainContent:

    content: str
    context: str = ""
-    event_date: datetime | None = None
+    event_date: datetime = field(default_factory=_now_utc)
    metadata: dict[str, str] = field(default_factory=dict)
    entities: list[dict[str, str]] = field(default_factory=list)  # User-provided entities
    tags: list[str] = field(default_factory=list)  # Visibility scope tags
-    observation_scopes: Literal["per_tag", "combined", "all_combinations"] | list[list[str]] | None = (
-        None  # Observation scopes
-    )


@dataclass
@@ -121,9 +117,6 @@ class ExtractedFact:
    mentioned_at: datetime | None = None
    metadata: dict[str, str] = field(default_factory=dict)
    tags: list[str] = field(default_factory=list)  # Visibility scope tags
-    observation_scopes: Literal["per_tag", "combined", "all_combinations"] | list[list[str]] | None = (
-        None  # Observation scopes
-    )


@dataclass
@@ -142,7 +135,7 @@ class ProcessedFact:
    # Temporal data
    occurred_start: datetime | None
    occurred_end: datetime | None
-    mentioned_at: datetime | None
+    mentioned_at: datetime

    # Context and metadata
    context: str
@@ -172,9 +165,6 @@ class ProcessedFact:
    # Visibility scope tags
    tags: list[str] = field(default_factory=list)

-    # Observation scopes for consolidation
-    observation_scopes: Literal["per_tag", "combined", "all_combinations"] | list[list[str]] | None = None
-
    @property
    def is_duplicate(self) -> bool:
        """Check if this fact was marked as a duplicate."""
@@ -195,10 +185,12 @@ class ProcessedFact:
        Returns:
            ProcessedFact ready for storage
        """
+        from datetime import datetime
+
        # Use occurred dates only if explicitly provided by LLM
        occurred_start = extracted_fact.occurred_start
        occurred_end = extracted_fact.occurred_end
-        mentioned_at = extracted_fact.mentioned_at  # May be None when caller opted into no timestamp
+        mentioned_at = extracted_fact.mentioned_at or datetime.now(UTC)

        # Convert entity strings to EntityRef objects
        entities = [EntityRef(name=name) for name in extracted_fact.entities]
@@ -217,7 +209,6 @@ class ProcessedFact:
            chunk_id=chunk_id,
            content_index=extracted_fact.content_index,
            tags=extracted_fact.tags,
-            observation_scopes=extracted_fact.observation_scopes,
        )


@@ -162,7 +162,7 @@ class BFSGraphRetriever(GraphRetriever):
        entry_points = await conn.fetch(
            f"""
            SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, fact_type, document_id, chunk_id, tags,
+                   mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                   1 - (embedding <=> $1::vector) AS similarity
            FROM {fq_table("memory_units")}
            WHERE bank_id = $2
@@ -216,7 +216,7 @@ class BFSGraphRetriever(GraphRetriever):
                neighbors = await conn.fetch(
                    f"""
                    SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.occurred_end,
-                           mu.mentioned_at, mu.fact_type,
+                           mu.mentioned_at, mu.embedding, mu.fact_type,
                           mu.document_id, mu.chunk_id, mu.tags,
                           ml.weight, ml.link_type, ml.from_unit_id
                    FROM {fq_table("memory_links")} ml
@@ -1,28 +1,18 @@
 """
 Link Expansion graph retrieval.

-Expands from semantic/temporal seeds through three parallel, first-class signals
-stored in memory_links:
+A simple, fast graph retrieval that expands from seeds via:
+1. Entity links: Find facts sharing entities with seeds (filtered by entity frequency)
+2. Causal links: Find facts causally linked to seeds (top-k by weight)

-1. Entity links  — precomputed co-occurrence graph (created at retain time, bounded to
-                   MAX_LINKS_PER_ENTITY per entity). Score = number of distinct shared
-                   entities between the seed set and each candidate.
-2. Semantic links — precomputed kNN graph (each new fact linked to its top-5 most
-                    similar existing facts at insert time, similarity >= 0.7). Checked
-                    in both directions since the graph is not symmetric. Score = weight.
-3. Causal links  — explicit causal chains (causes/caused_by/enables/prevents).
-                   Score = weight + 1.0 (boosted as highest-quality signal).
-
-All three signals are bounded at retain time, so no LATERAL fan-out caps are needed
-at query time. Each expansion is a simple aggregation over a small result set.
-
-For non-observation fact types the three expansions are issued as a single CTE query
-(one roundtrip, one connection) with a `source` discriminator column so the Python
-merge step can apply per-signal score transformations.
+Characteristics:
+- 2-3 DB queries (seed finding + parallel entity/causal expansion)
+- Sublinear: only touches connected facts via indexes
+- No iteration, no propagation, no normalization
+- Target: <100ms
 """

 import logging
-import math
 import time

 from ..db_utils import acquire_with_retry
@@ -55,7 +45,7 @@ async def _find_semantic_seeds(
    rows = await conn.fetch(
        f"""
        SELECT id, text, context, event_date, occurred_start, occurred_end,
-               mentioned_at, fact_type, document_id, chunk_id, tags,
+               mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
               1 - (embedding <=> $1::vector) AS similarity
        FROM {fq_table("memory_units")}
        WHERE bank_id = $2
@@ -75,23 +65,27 @@ class LinkExpansionRetriever(GraphRetriever):
    """
    Graph retrieval via direct link expansion from seeds.

-    Runs three expansions through precomputed memory_links: entity co-occurrence,
-    semantic kNN, and causal chains, all bounded at retain time.
-
-    For non-observation fact types the three expansions are issued as a single CTE
-    query (one roundtrip, one connection slot) with a `source` discriminator column.
-    The Python merge step applies per-signal score transformations.
+    Expands through entity co-occurrence and causal links in a single query.
+    Fast and simple alternative to MPFP.
    """

    def __init__(
        self,
+        max_entity_frequency: int = 500,
        causal_weight_threshold: float = 0.3,
+        causal_limit_per_seed: int = 10,
    ):
        """
+        Initialize link expansion retriever.
+
        Args:
-            causal_weight_threshold: Minimum weight for causal links to follow.
+            max_entity_frequency: Skip entities appearing in more than this many facts
+            causal_weight_threshold: Minimum weight for causal links
+            causal_limit_per_seed: Max causal links to follow per seed
        """
+        self.max_entity_frequency = max_entity_frequency
        self.causal_weight_threshold = causal_weight_threshold
+        self.causal_limit_per_seed = causal_limit_per_seed

    @property
    def name(self) -> str:
@@ -116,7 +110,7 @@ class LinkExpansionRetriever(GraphRetriever):

        Args:
            pool: Database connection pool
-            query_embedding_str: Query embedding as string
+            query_embedding_str: Query embedding (unused, kept for interface)
            bank_id: Memory bank ID
            fact_type: Fact type to filter
            budget: Maximum results to return
@@ -124,7 +118,7 @@ class LinkExpansionRetriever(GraphRetriever):
            semantic_seeds: Pre-computed semantic entry points
            temporal_seeds: Pre-computed temporal entry points
            adjacency: Unused, kept for interface compatibility
-            tags: Optional list of tags for visibility filtering
+            tags: Optional list of tags for visibility filtering (OR matching)

        Returns:
            Tuple of (results, timings)
@@ -132,6 +126,8 @@ class LinkExpansionRetriever(GraphRetriever):
        start_time = time.time()
        timings = MPFPTimings(fact_type=fact_type)

+        # Use single connection for all queries to reduce pool pressure
+        # (queries are fast ~50ms each, connection acquisition is the bottleneck)
        async with acquire_with_retry(pool) as conn:
            # Find seeds if not provided
            if semantic_seeds:
@@ -154,6 +150,7 @@ class LinkExpansionRetriever(GraphRetriever):
                    f"(tags={tags}, tags_match={tags_match})"
                )

+            # Add temporal seeds if provided
            if temporal_seeds:
                all_seeds.extend(temporal_seeds)

@@ -163,61 +160,223 @@ class LinkExpansionRetriever(GraphRetriever):
            seed_ids = list({s.id for s in all_seeds})
            timings.pattern_count = len(seed_ids)

+            # Run entity and causal expansion sequentially on same connection
            query_start = time.time()

+            # For observations, traverse through source_memory_ids to find entity connections.
+            # Observations don't have direct unit_entities - they inherit entities via their
+            # source world/experience facts.
+            #
+            # Path: observation → source_memory_ids → world fact → entities →
+            #       ALL world facts with those entities → their observations (excluding seeds)
            if fact_type == "observation":
-                entity_rows, semantic_rows, causal_rows = await self._expand_observations(conn, seed_ids, budget)
+                # Debug: Check what source_memory_ids exist on seed observations
+                debug_sources = await conn.fetch(
+                    f"""
+                    SELECT id, source_memory_ids
+                    FROM {fq_table("memory_units")}
+                    WHERE id = ANY($1::uuid[])
+                    """,
+                    seed_ids,
+                )
+                source_ids_found = []
+                for row in debug_sources:
+                    if row["source_memory_ids"]:
+                        source_ids_found.extend(row["source_memory_ids"])
+                logger.debug(
+                    f"[LinkExpansion] observation graph: {len(seed_ids)} seeds, "
+                    f"{len(source_ids_found)} source_memory_ids found"
+                )
+
+                entity_rows = await conn.fetch(
+                    f"""
+                    WITH seed_sources AS (
+                        -- Get source memory IDs from seed observations
+                        SELECT DISTINCT unnest(source_memory_ids) AS source_id
+                        FROM {fq_table("memory_units")}
+                        WHERE id = ANY($1::uuid[])
+                          AND source_memory_ids IS NOT NULL
+                    ),
+                    source_entities AS (
+                        -- Get entities from those source memories (filtered by frequency)
+                        SELECT DISTINCT ue.entity_id
+                        FROM seed_sources ss
+                        JOIN {fq_table("unit_entities")} ue ON ss.source_id = ue.unit_id
+                        JOIN {fq_table("entities")} e ON ue.entity_id = e.id
+                        WHERE e.mention_count < $2
+                    ),
+                    all_connected_sources AS (
+                        -- Find ALL world facts sharing those entities (don't exclude seed sources)
+                        -- The exclusion happens at the observation level, not the source level
+                        SELECT DISTINCT other_ue.unit_id AS source_id
+                        FROM source_entities se
+                        JOIN {fq_table("unit_entities")} other_ue ON se.entity_id = other_ue.entity_id
+                    )
+                    -- Find observations derived from connected source memories
+                    -- Only exclude the actual seed observations
+                    SELECT
+                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                        mu.occurred_end, mu.mentioned_at, mu.embedding,
+                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                        COUNT(DISTINCT cs.source_id)::float AS score
+                    FROM all_connected_sources cs
+                    JOIN {fq_table("memory_units")} mu
+                        ON mu.source_memory_ids @> ARRAY[cs.source_id]
+                    WHERE mu.fact_type = 'observation'
+                      AND mu.id != ALL($1::uuid[])
+                    GROUP BY mu.id
+                    ORDER BY score DESC
+                    LIMIT $3
+                    """,
+                    seed_ids,
+                    self.max_entity_frequency,
+                    budget,
+                )
+                logger.debug(f"[LinkExpansion] observation graph: found {len(entity_rows)} connected observations")
            else:
-                entity_rows, semantic_rows, causal_rows = await self._expand_combined(conn, seed_ids, fact_type, budget)
+                # For world/experience facts, use direct entity lookup
+                entity_rows = await conn.fetch(
+                    f"""
+                    SELECT
+                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                        mu.occurred_end, mu.mentioned_at, mu.embedding,
+                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                        COUNT(*)::float AS score
+                    FROM {fq_table("unit_entities")} seed_ue
+                    JOIN {fq_table("entities")} e ON seed_ue.entity_id = e.id
+                    JOIN {fq_table("unit_entities")} other_ue ON seed_ue.entity_id = other_ue.entity_id
+                    JOIN {fq_table("memory_units")} mu ON other_ue.unit_id = mu.id
+                    WHERE seed_ue.unit_id = ANY($1::uuid[])
+                      AND e.mention_count < $2
+                      AND mu.id != ALL($1::uuid[])
+                      AND mu.fact_type = $3
+                    GROUP BY mu.id
+                    ORDER BY score DESC
+                    LIMIT $4
+                    """,
+                    seed_ids,
+                    self.max_entity_frequency,
+                    fact_type,
+                    budget,
+                )
+
+            causal_rows = await conn.fetch(
+                f"""
+                SELECT DISTINCT ON (mu.id)
+                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                    mu.occurred_end, mu.mentioned_at, mu.embedding,
+                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                    ml.weight + 1.0 AS score
+                FROM {fq_table("memory_links")} ml
+                JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
+                WHERE ml.from_unit_id = ANY($1::uuid[])
+                  AND ml.link_type IN ('causes', 'caused_by', 'enables', 'prevents')
+                  AND ml.weight >= $2
+                  AND mu.fact_type = $3
+                ORDER BY mu.id, ml.weight DESC
+                LIMIT $4
+                """,
+                seed_ids,
+                self.causal_weight_threshold,
+                fact_type,
+                budget,
+            )
+
+            # Fallback: semantic/temporal/entity links from memory_links table
+            # These are secondary to entity links (via unit_entities) and causal links
+            # Weight is halved (0.5x) to prioritize primary link types
+            # Check both directions: seeds -> others AND others -> seeds
+            fallback_rows = await conn.fetch(
+                f"""
+                WITH outgoing AS (
+                    -- Links FROM seeds TO other facts
+                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                           mu.occurred_end, mu.mentioned_at, mu.embedding,
+                           mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                           ml.weight
+                    FROM {fq_table("memory_links")} ml
+                    JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
+                    WHERE ml.from_unit_id = ANY($1::uuid[])
+                      AND ml.link_type IN ('semantic', 'temporal', 'entity')
+                      AND ml.weight >= $2
+                      AND mu.fact_type = $3
+                      AND mu.id != ALL($1::uuid[])
+                ),
+                incoming AS (
+                    -- Links FROM other facts TO seeds (reverse direction)
+                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                           mu.occurred_end, mu.mentioned_at, mu.embedding,
+                           mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                           ml.weight
+                    FROM {fq_table("memory_links")} ml
+                    JOIN {fq_table("memory_units")} mu ON ml.from_unit_id = mu.id
+                    WHERE ml.to_unit_id = ANY($1::uuid[])
+                      AND ml.link_type IN ('semantic', 'temporal', 'entity')
+                      AND ml.weight >= $2
+                      AND mu.fact_type = $3
+                      AND mu.id != ALL($1::uuid[])
+                ),
+                combined AS (
+                    SELECT * FROM outgoing
+                    UNION ALL
+                    SELECT * FROM incoming
+                )
+                SELECT DISTINCT ON (id)
+                    id, text, context, event_date, occurred_start,
+                    occurred_end, mentioned_at, embedding,
+                    fact_type, document_id, chunk_id, tags,
+                    (MAX(weight) * 0.5) AS score
+                FROM combined
+                GROUP BY id, text, context, event_date, occurred_start,
+                         occurred_end, mentioned_at, embedding,
+                         fact_type, document_id, chunk_id, tags
+                ORDER BY id, score DESC
+                LIMIT $4
+                """,
+                seed_ids,
+                self.causal_weight_threshold,
+                fact_type,
+                budget,
+            )

            timings.edge_load_time = time.time() - query_start
-            timings.db_queries = 1
-            timings.edge_count = len(entity_rows) + len(semantic_rows) + len(causal_rows)
+            timings.db_queries = 3
+            timings.edge_count = len(entity_rows) + len(causal_rows) + len(fallback_rows)

-        # Merge results with additive intra-score: entity + semantic + causal ∈ [0, 3].
-        #
-        # Entity score: tanh(count × 0.5) maps shared-entity count to [0, 1]:
-        #   1 entity → 0.46,  2 → 0.76,  3 → 0.91,  4 → 0.96  (saturates naturally)
-        # Semantic score: similarity weight, already ∈ [0.7, 1.0].
-        # Causal score:   link weight, already ∈ [0, 1].
-        #
-        # Facts appearing in multiple signals accumulate higher scores, rewarding
-        # convergent evidence. The outer RRF uses rank position from this sorted list.
-        entity_scores: dict[str, float] = {}
-        semantic_scores: dict[str, float] = {}
-        causal_scores: dict[str, float] = {}
+        # Merge results, taking max score per fact
+        # Priority: entity links (unit_entities) > causal links > fallback links
+        score_map: dict[str, float] = {}
        row_map: dict[str, dict] = {}

        for row in entity_rows:
            fact_id = str(row["id"])
-            entity_scores[fact_id] = math.tanh(row["score"] * 0.5)
+            score_map[fact_id] = max(score_map.get(fact_id, 0), row["score"])
            row_map[fact_id] = dict(row)

-        for row in semantic_rows:
-            fact_id = str(row["id"])
-            semantic_scores[fact_id] = max(semantic_scores.get(fact_id, 0.0), row["score"])
-            row_map.setdefault(fact_id, dict(row))
-
        for row in causal_rows:
            fact_id = str(row["id"])
-            causal_scores[fact_id] = max(causal_scores.get(fact_id, 0.0), row["score"])
-            row_map.setdefault(fact_id, dict(row))
+            score_map[fact_id] = max(score_map.get(fact_id, 0), row["score"])
+            if fact_id not in row_map:
+                row_map[fact_id] = dict(row)

-        all_ids = set(entity_scores) | set(semantic_scores) | set(causal_scores)
-        score_map = {
-            fid: entity_scores.get(fid, 0.0) + semantic_scores.get(fid, 0.0) + causal_scores.get(fid, 0.0)
-            for fid in all_ids
-        }
+        for row in fallback_rows:
+            fact_id = str(row["id"])
+            score_map[fact_id] = max(score_map.get(fact_id, 0), row["score"])
+            if fact_id not in row_map:
+                row_map[fact_id] = dict(row)

+        # Sort by score and limit
        sorted_ids = sorted(score_map.keys(), key=lambda x: score_map[x], reverse=True)[:budget]
        rows = [row_map[fact_id] for fact_id in sorted_ids]

+        # Convert to results
        results = []
        for row in rows:
            result = RetrievalResult.from_db_row(dict(row))
            result.activation = row["score"]
            results.append(result)

+        # Apply tags filtering (graph expansion may reach untagged memories)
        if tags:
            results = filter_results_by_tags(results, tags, match=tags_match)

@@ -230,254 +389,3 @@ class LinkExpansionRetriever(GraphRetriever):
        )

        return results, timings
-
-    async def _expand_combined(
-        self,
-        conn,
-        seed_ids: list,
-        fact_type: str,
-        budget: int,
-    ) -> tuple[list, list, list]:
-        """
-        Single-roundtrip CTE query combining entity, semantic, and causal expansions.
-
-        Uses a `source` discriminator column so the caller can apply per-signal
-        score transformations.  The three CTEs share one connection slot — important
-        for asyncpg which does not allow concurrent queries on the same connection.
-
-        Index coverage (requires migration d2e3f4a5b6c7):
-          entity:   idx_memory_links_entity_covering (from_unit_id) INCLUDE (to_unit_id, entity_id)
-                    WHERE link_type = 'entity'  → index-only scan, no heap reads
-          semantic incoming:
-                    idx_memory_links_to_type_weight (to_unit_id, link_type, weight DESC)
-                    → replaces costly BitmapAnd of two separate scans
-        """
-        ml = fq_table("memory_links")
-        mu = fq_table("memory_units")
-        all_rows = await conn.fetch(
-            f"""
-            WITH entity_expanded AS (
-                -- Entity co-occurrence: seeds → their precomputed entity-link neighbors.
-                -- Score = distinct shared entities (bounded at retain time to
-                -- MAX_LINKS_PER_ENTITY=50).  GROUP BY mu.id is sufficient because mu.id
-                -- is the primary key and functionally determines all other mu columns.
-                SELECT
-                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                    mu.occurred_end, mu.mentioned_at,
-                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                    COUNT(DISTINCT ml.entity_id)::float AS score,
-                    'entity'::text AS source
-                FROM {ml} ml
-                JOIN {mu} mu ON mu.id = ml.to_unit_id
-                WHERE ml.from_unit_id = ANY($1::uuid[])
-                  AND ml.link_type = 'entity'
-                  AND mu.fact_type = $2
-                  AND mu.id != ALL($1::uuid[])
-                GROUP BY mu.id
-                ORDER BY score DESC
-                LIMIT $3
-            ),
-            semantic_expanded AS (
-                -- Semantic kNN: both outgoing (seeds → their kNN at insert time) and
-                -- incoming (facts inserted after seeds that found seeds as kNN).
-                -- Score = max similarity weight across both directions.
-                SELECT
-                    id, text, context, event_date, occurred_start,
-                    occurred_end, mentioned_at,
-                    fact_type, document_id, chunk_id, tags,
-                    MAX(weight) AS score,
-                    'semantic'::text AS source
-                FROM (
-                    SELECT
-                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                        mu.occurred_end, mu.mentioned_at,
-                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                        ml.weight
-                    FROM {ml} ml
-                    JOIN {mu} mu ON mu.id = ml.to_unit_id
-                    WHERE ml.from_unit_id = ANY($1::uuid[])
-                      AND ml.link_type = 'semantic'
-                      AND mu.fact_type = $2
-                      AND mu.id != ALL($1::uuid[])
-                    UNION ALL
-                    SELECT
-                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                        mu.occurred_end, mu.mentioned_at,
-                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                        ml.weight
-                    FROM {ml} ml
-                    JOIN {mu} mu ON mu.id = ml.from_unit_id
-                    WHERE ml.to_unit_id = ANY($1::uuid[])
-                      AND ml.link_type = 'semantic'
-                      AND mu.fact_type = $2
-                      AND mu.id != ALL($1::uuid[])
-                ) sem_raw
-                GROUP BY id, text, context, event_date, occurred_start,
-                         occurred_end, mentioned_at,
-                         fact_type, document_id, chunk_id, tags
-                ORDER BY score DESC
-                LIMIT $3
-            ),
-            causal_expanded AS (
-                -- Causal chains: explicit causes/enables/prevents links from seeds.
-                -- DISTINCT ON handles the case where a seed has multiple causal links
-                -- to the same target; best weight wins.
-                SELECT DISTINCT ON (mu.id)
-                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                    mu.occurred_end, mu.mentioned_at,
-                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                    ml.weight AS score,
-                    'causal'::text AS source
-                FROM {ml} ml
-                JOIN {mu} mu ON ml.to_unit_id = mu.id
-                WHERE ml.from_unit_id = ANY($1::uuid[])
-                  AND ml.link_type IN ('causes', 'caused_by', 'enables', 'prevents')
-                  AND ml.weight >= $4
-                  AND mu.fact_type = $2
-                ORDER BY mu.id, ml.weight DESC
-                LIMIT $3
-            )
-            SELECT * FROM entity_expanded
-            UNION ALL
-            SELECT * FROM semantic_expanded
-            UNION ALL
-            SELECT * FROM causal_expanded
-            """,
-            seed_ids,
-            fact_type,
-            budget,
-            self.causal_weight_threshold,
-        )
-
-        entity_rows = [r for r in all_rows if r["source"] == "entity"]
-        semantic_rows = [r for r in all_rows if r["source"] == "semantic"]
-        causal_rows = [r for r in all_rows if r["source"] == "causal"]
-        return entity_rows, semantic_rows, causal_rows
-
-    async def _expand_observations(
-        self,
-        conn,
-        seed_ids: list,
-        budget: int,
-    ) -> tuple[list, list, list]:
-        """
-        Observation-specific expansion.
-
-        Observations don't have direct entity links in memory_links (they're created
-        by consolidation, not retain).  Instead, traverse source_memory_ids → world
-        facts → entities → other world facts → their observations.
-
-        Semantic and causal expansions run as a second combined CTE query.
-        """
-        source_ids_found: list = []
-        if logger.isEnabledFor(logging.DEBUG):
-            debug_rows = await conn.fetch(
-                f"""
-                SELECT id, source_memory_ids
-                FROM {fq_table("memory_units")}
-                WHERE id = ANY($1::uuid[])
-                """,
-                seed_ids,
-            )
-            for row in debug_rows:
-                if row["source_memory_ids"]:
-                    source_ids_found.extend(row["source_memory_ids"])
-            logger.debug(
-                f"[LinkExpansion] observation graph: {len(seed_ids)} seeds, "
-                f"{len(source_ids_found)} source_memory_ids found"
-            )
-
-        entity_rows = await conn.fetch(
-            f"""
-            WITH seed_sources AS (
-                SELECT DISTINCT unnest(source_memory_ids) AS source_id
-                FROM {fq_table("memory_units")}
-                WHERE id = ANY($1::uuid[])
-                  AND source_memory_ids IS NOT NULL
-            ),
-            connected_sources AS (
-                -- Mirror the non-observation entity expansion: follow pre-bounded entity
-                -- links in memory_links (capped to MAX_LINKS_PER_ENTITY=50 at retain time).
-                -- Score = number of distinct shared entities, same as the non-obs path.
-                SELECT DISTINCT ml.to_unit_id AS source_id
-                FROM seed_sources ss
-                JOIN {fq_table("memory_links")} ml ON ml.from_unit_id = ss.source_id
-                WHERE ml.link_type = 'entity'
-            ),
-            connected_array AS (
-                SELECT array_agg(source_id) AS source_ids FROM connected_sources
-            )
-            SELECT
-                mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                mu.occurred_end, mu.mentioned_at,
-                mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                (SELECT COUNT(DISTINCT s) FROM unnest(mu.source_memory_ids) s WHERE s = ANY(ca.source_ids))::float AS score
-            FROM {fq_table("memory_units")} mu, connected_array ca
-            WHERE mu.fact_type = 'observation'
-              AND mu.id != ALL($1::uuid[])
-              AND ca.source_ids IS NOT NULL
-              AND mu.source_memory_ids && ca.source_ids
-            ORDER BY score DESC
-            LIMIT $2
-            """,
-            seed_ids,
-            budget,
-        )
-        logger.debug(f"[LinkExpansion] observation graph: found {len(entity_rows)} connected observations")
-
-        # Semantic + causal for observations in one query
-        ml = fq_table("memory_links")
-        mu = fq_table("memory_units")
-        sem_causal_rows = await conn.fetch(
-            f"""
-            WITH semantic_expanded AS (
-                SELECT
-                    id, text, context, event_date, occurred_start,
-                    occurred_end, mentioned_at,
-                    fact_type, document_id, chunk_id, tags,
-                    MAX(weight) AS score,
-                    'semantic'::text AS source
-                FROM (
-                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                           mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id,
-                           mu.chunk_id, mu.tags, ml.weight
-                    FROM {ml} ml JOIN {mu} mu ON mu.id = ml.to_unit_id
-                    WHERE ml.from_unit_id = ANY($1::uuid[])
-                      AND ml.link_type = 'semantic' AND mu.fact_type = 'observation'
-                      AND mu.id != ALL($1::uuid[])
-                    UNION ALL
-                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                           mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id,
-                           mu.chunk_id, mu.tags, ml.weight
-                    FROM {ml} ml JOIN {mu} mu ON mu.id = ml.from_unit_id
-                    WHERE ml.to_unit_id = ANY($1::uuid[])
-                      AND ml.link_type = 'semantic' AND mu.fact_type = 'observation'
-                      AND mu.id != ALL($1::uuid[])
-                ) sem_raw
-                GROUP BY id, text, context, event_date, occurred_start, occurred_end,
-                         mentioned_at, fact_type, document_id, chunk_id, tags
-                ORDER BY score DESC LIMIT $2
-            ),
-            causal_expanded AS (
-                SELECT DISTINCT ON (mu.id)
-                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                    mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id,
-                    mu.chunk_id, mu.tags, ml.weight AS score, 'causal'::text AS source
-                FROM {ml} ml JOIN {mu} mu ON ml.to_unit_id = mu.id
-                WHERE ml.from_unit_id = ANY($1::uuid[])
-                  AND ml.link_type IN ('causes', 'caused_by', 'enables', 'prevents')
-                  AND ml.weight >= $3 AND mu.fact_type = 'observation'
-                ORDER BY mu.id, ml.weight DESC LIMIT $2
-            )
-            SELECT * FROM semantic_expanded
-            UNION ALL
-            SELECT * FROM causal_expanded
-            """,
-            seed_ids,
-            budget,
-            self.causal_weight_threshold,
-        )
-
-        semantic_rows = [r for r in sem_causal_rows if r["source"] == "semantic"]
-        causal_rows = [r for r in sem_causal_rows if r["source"] == "causal"]
-        return entity_rows, semantic_rows, causal_rows
@@ -449,7 +449,7 @@ async def fetch_memory_units_by_ids(
        rows = await conn.fetch(
            f"""
            SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, fact_type, document_id, chunk_id, tags
+                   mentioned_at, embedding, fact_type, document_id, chunk_id, tags
            FROM {fq_table("memory_units")}
            WHERE id = ANY($1::uuid[])
              AND fact_type = $2
@@ -2,72 +2,8 @@
 Cross-encoder neural reranking for search results.
 """

-from datetime import datetime, timezone
-
 from .types import MergedCandidate, ScoredResult

-UTC = timezone.utc
-
-# Multiplicative boost alphas for recency and temporal proximity.
-# Each signal contributes at most ±(alpha/2) relative adjustment to the base CE score,
-# so the max combined boost is (1 + alpha/2)^2 ≈ +21% and min is (1 - alpha/2)^2 ≈ -19%.
-_RECENCY_ALPHA: float = 0.2
-_TEMPORAL_ALPHA: float = 0.2
-
-
-def apply_combined_scoring(
-    scored_results: list[ScoredResult],
-    now: datetime,
-    recency_alpha: float = _RECENCY_ALPHA,
-    temporal_alpha: float = _TEMPORAL_ALPHA,
-) -> None:
-    """Apply combined scoring to a list of ScoredResults in-place.
-
-    Uses the cross-encoder score as the primary relevance signal, with recency
-    and temporal proximity applied as multiplicative boosts. This ensures the
-    influence of these secondary signals is always proportional to the base
-    relevance score, regardless of the cross-encoder model's score calibration.
-
-    Formula::
-
-        recency_boost  = 1 + recency_alpha  * (recency  - 0.5)   # in [1-α/2, 1+α/2]
-        temporal_boost = 1 + temporal_alpha * (temporal - 0.5)   # in [1-α/2, 1+α/2]
-        combined_score = cross_encoder_score_normalized * recency_boost * temporal_boost
-
-    Temporal proximity is treated as neutral (0.5) when not set by temporal retrieval,
-    so temporal_boost collapses to 1.0 for non-temporal queries.
-
-    Args:
-        scored_results: Results from the cross-encoder reranker. Mutated in place.
-        now: Current UTC datetime for recency calculation.
-        recency_alpha: Max relative recency adjustment (default 0.2 → ±10%).
-        temporal_alpha: Max relative temporal adjustment (default 0.2 → ±10%).
-    """
-    if now.tzinfo is None:
-        now = now.replace(tzinfo=UTC)
-
-    for sr in scored_results:
-        # Recency: linear decay over 365 days → [0.1, 1.0]; neutral 0.5 if no date.
-        sr.recency = 0.5
-        if sr.retrieval.occurred_start:
-            occurred = sr.retrieval.occurred_start
-            if occurred.tzinfo is None:
-                occurred = occurred.replace(tzinfo=UTC)
-            days_ago = (now - occurred).total_seconds() / 86400
-            sr.recency = max(0.1, min(1.0, 1.0 - (days_ago / 365)))
-
-        # Temporal proximity: meaningful only for temporal queries; neutral otherwise.
-        sr.temporal = sr.retrieval.temporal_proximity if sr.retrieval.temporal_proximity is not None else 0.5
-
-        # RRF: kept at 0.0 for trace continuity but excluded from scoring.
-        # RRF is batch-relative (min-max normalised) and redundant after reranking.
-        sr.rrf_normalized = 0.0
-
-        recency_boost = 1.0 + recency_alpha * (sr.recency - 0.5)
-        temporal_boost = 1.0 + temporal_alpha * (sr.temporal - 0.5)
-        sr.combined_score = sr.cross_encoder_score_normalized * recency_boost * temporal_boost
-        sr.weight = sr.combined_score
-

 class CrossEncoderReranker:
    """
@@ -127,7 +127,7 @@ async def retrieve_semantic_bm25_combined(
        results = await conn.fetch(
            f"""
            WITH semantic_ranked AS (
-                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags,
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                       1 - (embedding <=> $1::vector) AS similarity,
                       NULL::float AS bm25_score,
                       'semantic' AS source,
@@ -139,7 +139,7 @@ async def retrieve_semantic_bm25_combined(
                  AND (1 - (embedding <=> $1::vector)) >= 0.3
                  {tags_clause}
            )
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                   similarity, bm25_score, source
            FROM semantic_ranked
            WHERE rn <= $4
@@ -194,7 +194,7 @@ async def retrieve_semantic_bm25_combined(
    # Single query template with backend-specific parts injected
    query = f"""
        WITH semantic_ranked AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                   1 - (embedding <=> $1::vector) AS similarity,
                   NULL::float AS bm25_score,
                   'semantic' AS source,
@@ -207,7 +207,7 @@ async def retrieve_semantic_bm25_combined(
              {tags_clause}
        ),
        bm25_ranked AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                   NULL::float AS similarity,
                   {bm25_score_expr} AS bm25_score,
                   'bm25' AS source,
@@ -219,12 +219,12 @@ async def retrieve_semantic_bm25_combined(
              {tags_clause}
        ),
        semantic AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                   similarity, bm25_score, source
            FROM semantic_ranked WHERE rn <= $4
        ),
        bm25 AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                   similarity, bm25_score, source
            FROM bm25_ranked WHERE rn <= $4
        )
@@ -297,20 +297,13 @@ async def retrieve_temporal_combined(
    if tags:
        params.append(tags)

-    # Two-phase entry point query:
-    # Phase 1 (date_ranked): rank by date only — no embedding computation — for all units in
-    #   the temporal window. This lets the planner use date indexes for filtering.
-    # Phase 2 (sim_ranked): join back to memory_units for only the top-50-per-type candidates
-    #   and compute embedding similarity for that small set (≤ 50 × len(fact_types) rows).
-    # This avoids computing embedding distances for potentially thousands of date-range rows.
+    # Batch query: Get entry points for ALL fact types at once with window function
    entry_points = await conn.fetch(
        f"""
-        WITH date_ranked AS MATERIALIZED (
-            SELECT id, fact_type,
-                   ROW_NUMBER() OVER (
-                       PARTITION BY fact_type
-                       ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC NULLS LAST
-                   ) AS rn
+        WITH ranked_entries AS (
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                   1 - (embedding <=> $1::vector) AS similarity,
+                   ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC, embedding <=> $1::vector) AS rn
            FROM {fq_table("memory_units")}
            WHERE bank_id = $2
              AND fact_type = ANY($3)
@@ -325,20 +318,12 @@ async def retrieve_temporal_combined(
                  OR
                  (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
              )
+              AND (1 - (embedding <=> $1::vector)) >= $6
              {tags_clause}
-        ),
-        sim_ranked AS (
-            SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                   1 - (mu.embedding <=> $1::vector) AS similarity,
-                   ROW_NUMBER() OVER (PARTITION BY mu.fact_type ORDER BY mu.embedding <=> $1::vector) AS sim_rn
-            FROM date_ranked dr
-            JOIN {fq_table("memory_units")} mu ON mu.id = dr.id
-            WHERE dr.rn <= 50
-              AND (1 - (mu.embedding <=> $1::vector)) >= $6
        )
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags, similarity
-        FROM sim_ranked
-        WHERE sim_rn <= 10
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags, similarity
+        FROM ranked_entries
+        WHERE rn <= 10
        """,
        *params,
    )
@@ -402,52 +387,34 @@ async def retrieve_temporal_combined(
        frontier = list(node_scores.keys())
        budget_remaining = budget - len(ft_entry_points)
        batch_size = 20
-        # Per-source neighbor limit: lets the planner use the composite index
-        # (from_unit_id, link_type, weight DESC) with early termination, avoiding
-        # a full scan of all links from all source nodes before sorting.
-        per_source_limit = 10
-        # Safety cap on BFS iterations to prevent runaway spreading in dense graphs.
-        max_iterations = 5
-        iteration = 0

-        # Build tags clause for spreading (use param 7 since 1-6 are used)
-        spreading_tags_clause = build_tags_where_clause_simple(tags, 7, table_alias="mu.", match=tags_match)
+        # Build tags clause for spreading (use param 6 since 1-5 are used)
+        spreading_tags_clause = build_tags_where_clause_simple(tags, 6, table_alias="mu.", match=tags_match)

-        while frontier and budget_remaining > 0 and iteration < max_iterations:
-            iteration += 1
+        while frontier and budget_remaining > 0:
            batch_ids = frontier[:batch_size]
            frontier = frontier[batch_size:]

-            # $1=query_emb, $2=batch_ids, $3=fact_type, $4=threshold, $5=per_source_limit, $6=bank_id, $7=tags
-            spreading_params = [query_emb_str, batch_ids, ft, semantic_threshold, per_source_limit, bank_id]
+            spreading_params = [query_emb_str, batch_ids, ft, semantic_threshold, batch_size * 10]
            if tags:
                spreading_params.append(tags)

-            # LATERAL join: for each source node, fetch top-K neighbors by weight using
-            # the existing idx_memory_links_from_type_weight index with early-exit semantics.
-            # This avoids scanning all temporal links from all source nodes before sorting.
-            # bank_id on memory_units lets the planner use idx_memory_units_bank_fact_type.
            neighbors = await conn.fetch(
                f"""
-                SELECT src.from_unit_id, mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                       l.weight, l.link_type,
+                SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                       ml.weight, ml.link_type, ml.from_unit_id,
                       1 - (mu.embedding <=> $1::vector) AS similarity
-                FROM unnest($2::uuid[]) AS src(from_unit_id)
-                CROSS JOIN LATERAL (
-                    SELECT ml.to_unit_id, ml.weight, ml.link_type
-                    FROM {fq_table("memory_links")} ml
-                    WHERE ml.from_unit_id = src.from_unit_id
-                      AND ml.link_type IN ('temporal', 'causes', 'caused_by', 'enables', 'prevents')
-                      AND ml.weight >= 0.1
-                    ORDER BY ml.weight DESC
-                    LIMIT $5
-                ) l
-                JOIN {fq_table("memory_units")} mu ON mu.id = l.to_unit_id
-                WHERE mu.bank_id = $6
+                FROM {fq_table("memory_links")} ml
+                JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
+                WHERE ml.from_unit_id = ANY($2::uuid[])
+                  AND ml.link_type IN ('temporal', 'causes', 'caused_by', 'enables', 'prevents')
+                  AND ml.weight >= 0.1
                  AND mu.fact_type = $3
                  AND mu.embedding IS NOT NULL
                  AND (1 - (mu.embedding <=> $1::vector)) >= $4
                  {spreading_tags_clause}
+                ORDER BY ml.weight DESC
+                LIMIT $5
                """,
                *spreading_params,
            )
@@ -46,6 +46,7 @@ class RetrievalResult:
    mentioned_at: datetime | None = None
    document_id: str | None = None
    chunk_id: str | None = None
+    embedding: list[float] | None = None
    tags: list[str] | None = None  # Visibility scope tags

    # Retrieval-specific scores (only one will be set depending on retrieval method)
@@ -69,6 +70,7 @@ class RetrievalResult:
            mentioned_at=row.get("mentioned_at"),
            document_id=row.get("document_id"),
            chunk_id=row.get("chunk_id"),
+            embedding=row.get("embedding"),
            tags=row.get("tags"),
            similarity=row.get("similarity"),
            bm25_score=row.get("bm25_score"),
@@ -152,6 +154,7 @@ class ScoredResult:
            "mentioned_at": self.retrieval.mentioned_at,
            "document_id": self.retrieval.document_id,
            "chunk_id": self.retrieval.chunk_id,
+            "embedding": self.retrieval.embedding,
            "tags": self.retrieval.tags,
            "semantic_similarity": self.retrieval.similarity,
            "bm25_score": self.retrieval.bm25_score,
@@ -12,7 +12,6 @@ def create_file_storage(
    storage_type: str,
    pool_getter: Callable | None = None,
    schema: str | None = None,
-    schema_getter: Callable | None = None,
    **kwargs,
 ) -> FileStorage:
    """
@@ -21,8 +20,7 @@ def create_file_storage(
    Args:
        storage_type: "native" (PostgreSQL BYTEA) or "s3" (S3-compatible object storage)
        pool_getter: Database pool getter (required for native)
-        schema: Static database schema (for native single-tenant)
-        schema_getter: Callable returning current schema at query time (for native multi-tenant)
+        schema: Database schema (for native multi-tenant)
        **kwargs: Additional args passed to storage backend

    Returns:
@@ -34,7 +32,7 @@ def create_file_storage(
    if storage_type == "native":
        if not pool_getter:
            raise ValueError("pool_getter required for native (PostgreSQL) storage")
-        return PostgreSQLFileStorage(pool_getter=pool_getter, schema=schema, schema_getter=schema_getter)
+        return PostgreSQLFileStorage(pool_getter=pool_getter, schema=schema)
    elif storage_type == "s3":
        from ...config import get_config
        from .s3 import S3FileStorage
@@ -1,8 +1,7 @@
 """Google Cloud Storage backend using obstore."""

 import logging
-import os
-from datetime import datetime, timedelta, timezone
+from datetime import timedelta

 import obstore as obs
 from obstore.store import GCSStore
@@ -12,30 +11,6 @@ from .base import FileStorage
 logger = logging.getLogger(__name__)


-def _make_google_auth_credential_provider():
-    """Create a credential provider using google.auth (supports all credential types).
-
-    obstore's built-in credential parsing only supports service_account and
-    authorized_user JSON types.  This provider uses the google-auth library
-    which additionally handles external_account (Workload Identity Federation),
-    impersonated credentials, and metadata-server credentials.
-    """
-    import google.auth
-    import google.auth.transport.requests
-
-    credentials, _ = google.auth.default(scopes=["https://www.googleapis.com/auth/cloud-platform"])
-    request = google.auth.transport.requests.Request()
-
-    def _provide():
-        credentials.refresh(request)
-        expiry = credentials.expiry
-        if expiry and expiry.tzinfo is None:
-            expiry = expiry.replace(tzinfo=timezone.utc)
-        return {"token": credentials.token, "expires_at": expiry}
-
-    return _provide
-
-
 class GCSFileStorage(FileStorage):
    """
    Google Cloud Storage backend.
@@ -52,29 +27,8 @@ class GCSFileStorage(FileStorage):
        kwargs: dict = {}
        if service_account_key:
            kwargs["service_account_key"] = service_account_key
-        else:
-            # Use google.auth credential provider for broad credential type support
-            # (service_account, authorized_user, external_account, metadata server, etc.)
-            try:
-                kwargs["credential_provider"] = _make_google_auth_credential_provider()
-                logger.info("Using google.auth credential provider for GCS")
-            except Exception as e:
-                logger.warning(
-                    f"Failed to create google.auth credential provider, falling back to obstore defaults: {e}"
-                )

-        # Workaround for https://github.com/developmentseed/obstore/issues/605
-        # obstore's Rust layer doesn't support external_account credentials (Workload
-        # Identity Federation) and eagerly parses GOOGLE_APPLICATION_CREDENTIALS even
-        # when credential_provider is given. Per the obstore maintainer's guidance,
-        # remove env vars so the Rust code doesn't try to authenticate itself.
-        # google.auth (used by credential_provider above) has already loaded credentials.
-        gac = os.environ.pop("GOOGLE_APPLICATION_CREDENTIALS", None)
-        try:
-            self._store = GCSStore(bucket, **kwargs)
-        finally:
-            if gac is not None:
-                os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = gac
+        self._store = GCSStore(bucket, **kwargs)
        logger.info(f"Initialized GCS file storage: bucket={bucket}")

    async def store(self, file_data: bytes, key: str, metadata: dict[str, str] | None = None) -> str:
@@ -40,30 +40,16 @@ class PostgreSQLFileStorage(FileStorage):
    For production/scale, consider S3FileStorage instead.
    """

-    def __init__(
-        self,
-        pool_getter: Callable[[], "asyncpg.Pool"],
-        schema: str | None = None,
-        schema_getter: Callable[[], str] | None = None,
-    ):
+    def __init__(self, pool_getter: Callable[[], "asyncpg.Pool"], schema: str | None = None):
        """
        Initialize PostgreSQL file storage.

        Args:
            pool_getter: Function that returns asyncpg connection pool
-            schema: Static database schema (fallback for single-tenant / tests)
-            schema_getter: Callable returning current schema at query time (for multi-tenant)
+            schema: Database schema (for multi-tenant support)
        """
        self._pool_getter = pool_getter
-        self._static_schema = schema
-        self._schema_getter = schema_getter
-
-    @property
-    def _schema(self) -> str | None:
-        """Resolve schema dynamically per-request when schema_getter is provided."""
-        if self._schema_getter:
-            return self._schema_getter()
-        return self._static_schema
+        self._schema = schema

    async def store(
        self,
@@ -22,16 +22,9 @@ from hindsight_api.extensions.http import HttpExtension
 from hindsight_api.extensions.loader import load_extension
 from hindsight_api.extensions.mcp import MCPExtension
 from hindsight_api.extensions.operation_validator import (
-    # Bank Management operations
-    BankListContext,
-    BankListResult,
-    BankReadContext,
-    BankWriteContext,
    # Consolidation operation
    ConsolidateContext,
    ConsolidateResult,
-    # File Conversion
-    FileConvertResult,
    # Mental Model operations
    MentalModelGetContext,
    MentalModelGetResult,
@@ -77,16 +70,9 @@ __all__ = [
    "RetainContext",
    "RetainResult",
    "ValidationResult",
-    # Operation Validator - Bank Management
-    "BankListContext",
-    "BankListResult",
-    "BankReadContext",
-    "BankWriteContext",
    # Operation Validator - Consolidation
    "ConsolidateContext",
    "ConsolidateResult",
-    # Operation Validator - File Conversion
-    "FileConvertResult",
    # Operation Validator - Mental Model
    "MentalModelGetContext",
    "MentalModelGetResult",
@@ -96,8 +96,6 @@ class DefaultExtensionContext(ExtensionContext):

    async def run_migration(self, schema: str) -> None:
        """Run migrations for a specific schema."""
-        import asyncio
-
        from hindsight_api.config import get_config
        from hindsight_api.migrations import (
            ensure_embedding_dimension,
@@ -113,14 +111,10 @@ class DefaultExtensionContext(ExtensionContext):
            if engine_url:
                db_url = engine_url

-        # Run synchronous migration functions in a thread so the asyncio event loop
-        # remains free. This is critical for single-machine deployments where the
-        # worker runs in-process: if run_migrations() blocks the event loop, any
-        # in-flight asyncpg transactions cannot flush their COMMIT, and
-        # CREATE INDEX CONCURRENTLY inside the migration waits for those transactions
-        # forever — a deadlock.
+        run_migrations(db_url, schema=schema)
+
+        # Get config for vector extension setting
        config = get_config()
-        await asyncio.to_thread(run_migrations, db_url, schema=schema)

        # Ensure embedding column dimension matches the model's dimension
        # This is needed because migrations create columns with default dimension
@@ -129,23 +123,15 @@ class DefaultExtensionContext(ExtensionContext):
            if embeddings is not None:
                dimension = getattr(embeddings, "dimension", None)
                if dimension is not None:
-                    await asyncio.to_thread(
-                        ensure_embedding_dimension,
-                        db_url,
-                        dimension,
-                        schema=schema,
-                        vector_extension=config.vector_extension,
+                    ensure_embedding_dimension(
+                        db_url, dimension, schema=schema, vector_extension=config.vector_extension
                    )

        # Ensure vector indexes match the configured extension
-        await asyncio.to_thread(
-            ensure_vector_extension, db_url, vector_extension=config.vector_extension, schema=schema
-        )
+        ensure_vector_extension(db_url, vector_extension=config.vector_extension, schema=schema)

        # Ensure text search columns/indexes match the configured extension
-        await asyncio.to_thread(
-            ensure_text_search_extension, db_url, text_search_extension=config.text_search_extension, schema=schema
-        )
+        ensure_text_search_extension(db_url, text_search_extension=config.text_search_extension, schema=schema)

    def get_memory_engine(self) -> "MemoryEngineInterface":
        """Get the memory engine interface."""
@@ -87,15 +87,3 @@ class HttpExtension(Extension, ABC):
            ```
        """
        pass
-
-    def get_root_router(self, memory: "MemoryEngine") -> APIRouter | None:
-        """
-        Return a FastAPI router with endpoints mounted at the app root.
-
-        Unlike get_router() which is mounted at /ext/, this router is mounted
-        directly on the application root. Use for well-known endpoints or other
-        paths that must be at specific locations.
-
-        Returns None by default (no root routes). Override to provide root-level routes.
-        """
-        return None
@@ -200,44 +200,6 @@ class ConsolidateResult:
    error: str | None = None


-# =============================================================================
-# Bank Management Contexts
-# =============================================================================
-
-
-@dataclass
-class BankReadContext:
-    """Context for a bank read operation validation (pre-operation)."""
-
-    bank_id: str
-    operation: str  # "get_bank_profile", "get_bank_stats"
-    request_context: "RequestContext"
-
-
-@dataclass
-class BankWriteContext:
-    """Context for a bank write operation validation (pre-operation)."""
-
-    bank_id: str
-    operation: str  # "delete_bank", "update_bank", "update_bank_disposition", "set_bank_mission", "merge_bank_mission", "clear_observations", "clear_observations_for_memory"
-    request_context: "RequestContext"
-
-
-@dataclass
-class BankListContext:
-    """Context for filtering the bank list (post-query)."""
-
-    banks: list[dict]
-    request_context: "RequestContext"
-
-
-@dataclass
-class BankListResult:
-    """Result of filtering the bank list."""
-
-    banks: list[dict]
-
-
 # =============================================================================
 # Mental Model Contexts
 # =============================================================================
@@ -289,28 +251,6 @@ class MentalModelRefreshResult:
    error: str | None = None


-# =============================================================================
-# File Conversion Post-operation Context
-# =============================================================================
-
-
-@dataclass
-class FileConvertResult:
-    """Result context for post-file-conversion hook.
-
-    Fired after a file is converted to markdown, before the retain step.
-    """
-
-    bank_id: str
-    parser_name: str
-    filename: str
-    output_chars: int
-    output_text: str
-    request_context: "RequestContext"
-    success: bool = True
-    error: str | None = None
-
-
 class OperationValidatorExtension(Extension, ABC):
    """
    Validates and hooks into retain/recall/reflect/consolidate operations.
@@ -518,31 +458,6 @@ class OperationValidatorExtension(Extension, ABC):
        """
        pass

-    # =========================================================================
-    # File Conversion - Post-operation hook (optional - override to implement)
-    # =========================================================================
-
-    async def on_file_convert_complete(self, result: FileConvertResult) -> None:
-        """
-        Called after a file is converted to markdown (before the retain step).
-
-        Override to implement post-conversion logic such as:
-        - Billing for premium parsers (e.g., Iris)
-        - Usage tracking
-        - Audit logging
-
-        Args:
-            result: Result context containing:
-                - bank_id: Bank identifier
-                - parser_name: Name of the parser used (e.g., 'markitdown', 'iris')
-                - filename: Original filename
-                - output_chars: Character count of the converted markdown
-                - request_context: Request context with auth info
-                - success: Whether the conversion succeeded
-                - error: Error message (if failed)
-        """
-        pass
-
    # =========================================================================
    # Mental Model - Pre-operation validation hook (optional - override to implement)
    # =========================================================================
@@ -620,63 +535,3 @@ class OperationValidatorExtension(Extension, ABC):
                - error: Error message (if failed)
        """
        pass
-
-    # =========================================================================
-    # Bank Management - Validation hooks (optional - override to implement)
-    # =========================================================================
-
-    async def validate_bank_read(self, ctx: BankReadContext) -> ValidationResult:
-        """
-        Validate a bank read operation before execution.
-
-        Override to implement custom validation logic for bank reads
-        (get_bank_profile, get_bank_stats).
-
-        Args:
-            ctx: Context containing:
-                - bank_id: Bank identifier
-                - operation: Operation name
-                - request_context: Request context with auth info
-
-        Returns:
-            ValidationResult indicating whether the operation is allowed.
-        """
-        return ValidationResult.accept()
-
-    async def validate_bank_write(self, ctx: BankWriteContext) -> ValidationResult:
-        """
-        Validate a bank write operation before execution.
-
-        Override to implement custom validation logic for bank writes
-        (delete_bank, update_bank, update_bank_disposition, set_bank_mission,
-        merge_bank_mission, clear_observations, clear_observations_for_memory).
-
-        Args:
-            ctx: Context containing:
-                - bank_id: Bank identifier
-                - operation: Operation name
-                - request_context: Request context with auth info
-
-        Returns:
-            ValidationResult indicating whether the operation is allowed.
-        """
-        return ValidationResult.accept()
-
-    async def filter_bank_list(self, ctx: BankListContext) -> BankListResult:
-        """
-        Filter the bank list after querying.
-
-        Unlike validate_* methods, this is a post-query filter that narrows results
-        rather than a gate that blocks the operation.
-
-        Override to implement custom filtering (e.g., restrict to allowed banks).
-
-        Args:
-            ctx: Context containing:
-                - banks: List of bank dicts from the database
-                - request_context: Request context with auth info
-
-        Returns:
-            BankListResult with the filtered list of banks.
-        """
-        return BankListResult(banks=ctx.banks)
@@ -11,9 +11,8 @@ from hindsight_api.models import RequestContext
 class AuthenticationError(Exception):
    """Raised when authentication fails."""

-    def __init__(self, reason: str, headers: dict[str, str] | None = None):
+    def __init__(self, reason: str):
        self.reason = reason
-        self.headers = headers or {}
        super().__init__(f"Authentication failed: {reason}")


@@ -171,7 +171,6 @@ def main():
            llm_vertexai_project_id=config.llm_vertexai_project_id,
            llm_vertexai_region=config.llm_vertexai_region,
            llm_vertexai_service_account_key=config.llm_vertexai_service_account_key,
-            llm_gemini_safety_settings=config.llm_gemini_safety_settings,
            retain_llm_provider=config.retain_llm_provider,
            retain_llm_api_key=config.retain_llm_api_key,
            retain_llm_model=config.retain_llm_model,
@@ -232,15 +231,12 @@ def main():
            reranker_litellm_sdk_api_key=config.reranker_litellm_sdk_api_key,
            reranker_litellm_sdk_model=config.reranker_litellm_sdk_model,
            reranker_litellm_sdk_api_base=config.reranker_litellm_sdk_api_base,
-            reranker_zeroentropy_api_key=config.reranker_zeroentropy_api_key,
-            reranker_zeroentropy_model=config.reranker_zeroentropy_model,
            host=args.host,
            port=args.port,
            base_path=config.base_path,
            log_level=args.log_level,
            log_format=config.log_format,
            mcp_enabled=config.mcp_enabled,
-            mcp_enabled_tools=config.mcp_enabled_tools,
            enable_bank_config_api=config.enable_bank_config_api,
            graph_retriever=config.graph_retriever,
            mpfp_top_k_neighbors=config.mpfp_top_k_neighbors,
@@ -250,10 +246,8 @@ def main():
            retain_chunk_size=config.retain_chunk_size,
            retain_extract_causal_links=config.retain_extract_causal_links,
            retain_extraction_mode=config.retain_extraction_mode,
-            retain_mission=config.retain_mission,
            retain_custom_instructions=config.retain_custom_instructions,
            retain_batch_tokens=config.retain_batch_tokens,
-            retain_entity_lookup=config.retain_entity_lookup,
            retain_batch_enabled=config.retain_batch_enabled,
            retain_batch_poll_interval_seconds=config.retain_batch_poll_interval_seconds,
            file_storage_type=config.file_storage_type,
@@ -268,7 +262,6 @@ def main():
            file_storage_azure_account_name=config.file_storage_azure_account_name,
            file_storage_azure_account_key=config.file_storage_azure_account_key,
            file_parser=config.file_parser,
-            file_parser_allowlist=config.file_parser_allowlist,
            file_parser_iris_token=config.file_parser_iris_token,
            file_parser_iris_org_id=config.file_parser_iris_org_id,
            file_conversion_max_batch_size_mb=config.file_conversion_max_batch_size_mb,
@@ -276,16 +269,8 @@ def main():
            enable_file_upload_api=config.enable_file_upload_api,
            file_delete_after_retain=config.file_delete_after_retain,
            enable_observations=config.enable_observations,
-            enable_observation_history=config.enable_observation_history,
-            enable_mental_model_history=config.enable_mental_model_history,
            consolidation_batch_size=config.consolidation_batch_size,
-            consolidation_llm_batch_size=config.consolidation_llm_batch_size,
            consolidation_max_tokens=config.consolidation_max_tokens,
-            consolidation_source_facts_max_tokens=config.consolidation_source_facts_max_tokens,
-            consolidation_source_facts_max_tokens_per_observation=config.consolidation_source_facts_max_tokens_per_observation,
-            observations_mission=config.observations_mission,
-            entity_labels=config.entity_labels,
-            entities_allow_free_form=config.entities_allow_free_form,
            skip_llm_verification=config.skip_llm_verification,
            lazy_reranker=config.lazy_reranker,
            run_migrations_on_startup=config.run_migrations_on_startup,
@@ -301,21 +286,12 @@ def main():
            worker_max_slots=config.worker_max_slots,
            worker_consolidation_max_slots=config.worker_consolidation_max_slots,
            reflect_max_iterations=config.reflect_max_iterations,
-            reflect_max_context_tokens=config.reflect_max_context_tokens,
-            reflect_mission=config.reflect_mission,
-            disposition_skepticism=config.disposition_skepticism,
-            disposition_literalism=config.disposition_literalism,
-            disposition_empathy=config.disposition_empathy,
            mental_model_refresh_concurrency=config.mental_model_refresh_concurrency,
            otel_traces_enabled=config.otel_traces_enabled,
            otel_exporter_otlp_endpoint=config.otel_exporter_otlp_endpoint,
            otel_exporter_otlp_headers=config.otel_exporter_otlp_headers,
            otel_service_name=config.otel_service_name,
            otel_deployment_environment=config.otel_deployment_environment,
-            webhook_url=config.webhook_url,
-            webhook_secret=config.webhook_secret,
-            webhook_event_types=config.webhook_event_types,
-            webhook_delivery_poll_interval_seconds=config.webhook_delivery_poll_interval_seconds,
        )
    config.configure_logging()
    if not args.daemon:
@@ -394,7 +370,6 @@ def main():
        "ws": "wsproto",  # Use wsproto instead of websockets to avoid deprecation warnings
        "loop": loop_impl,  # Explicitly set event loop implementation
        "timeout_keep_alive": 30,  # Exceed aiohttp's 15s client timeout so the client always closes first
-        "timeout_graceful_shutdown": 5,  # Cap graceful shutdown at 5s; also enables force-kill on second Ctrl+C
    }

    # Add optional parameters if provided
@@ -1,14 +1,8 @@
 """
-Local MCP server entry point for use with Claude Code (HTTP transport).
+Local MCP server for use with Claude Code (stdio transport).

-This is a thin wrapper around the main hindsight-api server that pre-configures
-sensible defaults for local use (embedded PostgreSQL via pg0, warning log level).
-
-The full API runs on localhost:8888. Configure Claude Code's MCP settings:
-    claude mcp add --transport http hindsight http://localhost:8888/mcp/
-
-Or pinned to a specific bank (single-bank mode):
-    claude mcp add --transport http hindsight http://localhost:8888/mcp/default/
+This runs a fully local Hindsight instance with embedded PostgreSQL (pg0).
+No external database or server required.

 Run with:
    hindsight-local-mcp
@@ -16,24 +10,148 @@ Run with:
 Or with uvx:
    uvx hindsight-api@latest hindsight-local-mcp

+Configure in Claude Code's MCP settings:
+    {
+        "mcpServers": {
+            "hindsight": {
+                "command": "uvx",
+                "args": ["hindsight-api@latest", "hindsight-local-mcp"],
+                "env": {
+                    "HINDSIGHT_API_LLM_API_KEY": "your-openai-key"
+                }
+            }
+        }
+    }
+
 Environment variables:
    HINDSIGHT_API_LLM_API_KEY: Required. API key for LLM provider.
    HINDSIGHT_API_LLM_PROVIDER: Optional. LLM provider (default: "openai").
    HINDSIGHT_API_LLM_MODEL: Optional. LLM model (default: "gpt-4o-mini").
-    HINDSIGHT_API_DATABASE_URL: Optional. Override database URL (default: pg0://hindsight-mcp).
+    HINDSIGHT_API_MCP_LOCAL_BANK_ID: Optional. Memory bank ID (default: "mcp").
+    HINDSIGHT_API_LOG_LEVEL: Optional. Log level (default: "warning").
+    HINDSIGHT_API_MCP_INSTRUCTIONS: Optional. Additional instructions appended to both retain and recall tools.
+
+Example custom instructions (these are ADDED to the default behavior):
+    To also store assistant actions:
+        HINDSIGHT_API_MCP_INSTRUCTIONS="Also store every action you take, including tool calls, code written, and decisions made."
+
+    To also store conversation summaries:
+        HINDSIGHT_API_MCP_INSTRUCTIONS="Also store summaries of important conversations and their outcomes."
 """

+import logging
 import os
+import sys
+
+from mcp.server.fastmcp import FastMCP
+
+from hindsight_api.config import (
+    DEFAULT_MCP_LOCAL_BANK_ID,
+    DEFAULT_MCP_RECALL_DESCRIPTION,
+    DEFAULT_MCP_RETAIN_DESCRIPTION,
+    ENV_MCP_INSTRUCTIONS,
+    ENV_MCP_LOCAL_BANK_ID,
+)
+from hindsight_api.mcp_tools import MCPToolsConfig, register_mcp_tools
+
+# Configure logging - default to warning to avoid polluting stderr during MCP init
+# MCP clients interpret stderr output as errors, so we suppress INFO logs by default
+_log_level_str = os.environ.get("HINDSIGHT_API_LOG_LEVEL", "warning").lower()
+_log_level_map = {
+    "critical": logging.CRITICAL,
+    "error": logging.ERROR,
+    "warning": logging.WARNING,
+    "info": logging.INFO,
+    "debug": logging.DEBUG,
+}
+logging.basicConfig(
+    level=_log_level_map.get(_log_level_str, logging.WARNING),
+    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+    stream=sys.stderr,  # MCP uses stdout for protocol, logs go to stderr
+)
+logger = logging.getLogger(__name__)


-def main() -> None:
-    """Start the Hindsight API server with local defaults."""
-    # Set local defaults (only if not already configured by the user)
-    os.environ.setdefault("HINDSIGHT_API_DATABASE_URL", "pg0://hindsight-mcp")
+def create_local_mcp_server(bank_id: str, memory=None) -> FastMCP:
+    """
+    Create a stdio MCP server with retain/recall tools.

-    from hindsight_api.main import main as api_main
+    Args:
+        bank_id: The memory bank ID to use for all operations.
+        memory: Optional MemoryEngine instance. If not provided, creates one with pg0.

-    api_main()
+    Returns:
+        Configured FastMCP server instance.
+    """
+    # Import here to avoid slow startup if just checking --help
+    from hindsight_api import MemoryEngine
+
+    # Create memory engine with pg0 embedded database if not provided
+    if memory is None:
+        memory = MemoryEngine(db_url="pg0://hindsight-mcp")
+
+    # Get custom instructions from environment variable (appended to both tools)
+    extra_instructions = os.environ.get(ENV_MCP_INSTRUCTIONS, "")
+
+    retain_description = DEFAULT_MCP_RETAIN_DESCRIPTION
+    recall_description = DEFAULT_MCP_RECALL_DESCRIPTION
+
+    if extra_instructions:
+        retain_description = f"{DEFAULT_MCP_RETAIN_DESCRIPTION}\n\nAdditional instructions: {extra_instructions}"
+        recall_description = f"{DEFAULT_MCP_RECALL_DESCRIPTION}\n\nAdditional instructions: {extra_instructions}"
+
+    mcp = FastMCP("hindsight")
+
+    # Configure and register tools using shared module
+    config = MCPToolsConfig(
+        bank_id_resolver=lambda: bank_id,
+        include_bank_id_param=False,  # Local MCP uses fixed bank_id
+        tools={"retain", "recall"},  # Local MCP only has retain and recall
+        retain_description=retain_description,
+        recall_description=recall_description,
+        retain_fire_and_forget=True,  # Local MCP uses fire-and-forget pattern
+    )
+
+    register_mcp_tools(mcp, memory, config)
+
+    return mcp
+
+
+async def _initialize_and_run(bank_id: str):
+    """Initialize memory and run the MCP server."""
+    from hindsight_api import MemoryEngine
+
+    # Create and initialize memory engine with pg0 embedded database
+    # Note: We avoid printing to stderr during init as MCP clients show it as "errors"
+    memory = MemoryEngine(db_url="pg0://hindsight-mcp")
+    await memory.initialize()
+
+    # Create and run the server
+    mcp = create_local_mcp_server(bank_id, memory=memory)
+    await mcp.run_stdio_async()
+
+
+def main():
+    """Main entry point for the stdio MCP server."""
+    import asyncio
+
+    from hindsight_api.config import ENV_LLM_API_KEY, get_config
+
+    # Check for required environment variables
+    config = get_config()
+    if not config.llm_api_key:
+        print(f"Error: {ENV_LLM_API_KEY} environment variable is required", file=sys.stderr)
+        print("Set it in your MCP configuration or shell environment", file=sys.stderr)
+        sys.exit(1)
+
+    # Get bank ID from environment, default to "mcp"
+    bank_id = os.environ.get(ENV_MCP_LOCAL_BANK_ID, DEFAULT_MCP_LOCAL_BANK_ID)
+
+    # Note: We don't print to stderr as MCP clients display it as "error output"
+    # Use HINDSIGHT_API_LOG_LEVEL=debug for verbose startup logging
+
+    # Run the async initialization and server
+    asyncio.run(_initialize_and_run(bank_id))


 if __name__ == "__main__":
@@ -18,14 +18,12 @@ No alembic.ini required - all configuration is done programmatically.
 import hashlib
 import logging
 import os
-import threading
-import time
 from pathlib import Path

 from alembic import command
 from alembic.config import Config
 from alembic.script.revision import ResolutionError
-from sqlalchemy import Connection, create_engine, text
+from sqlalchemy import create_engine, text

 from .utils import mask_network_location

@@ -34,13 +32,6 @@ logger = logging.getLogger(__name__)
 # Advisory lock ID for migrations (arbitrary unique number)
 MIGRATION_LOCK_ID = 123456789

-# Alembic's command.upgrade() is NOT thread-safe: it uses module-level global
-# proxies (context._proxy, script) that get overwritten when two threads call
-# upgrade() concurrently.  This causes migrations to target the wrong schema
-# and crash with "relation already exists" or KeyError: 'script'.
-# Serialize all Alembic invocations with a process-level lock.
-_alembic_lock = threading.Lock()
-

 def _detect_vector_extension(conn, vector_extension: str = "pgvector") -> str:
    """
@@ -152,12 +143,9 @@ def _run_migrations_internal(database_url: str, script_location: str, schema: st
    if schema:
        alembic_cfg.set_main_option("target_schema", schema)

-    # Run migrations under a process-level lock.  Alembic uses module-level
-    # global proxies that are not thread-safe, so concurrent command.upgrade()
-    # calls from different threads corrupt each other's context.
+    # Run migrations
    try:
-        with _alembic_lock:
-            command.upgrade(alembic_cfg, "head")
+        command.upgrade(alembic_cfg, "head")
    except ResolutionError as e:
        # This happens during rolling deployments when a newer version of the code
        # has already run migrations, and this older replica doesn't have the new
@@ -232,40 +220,13 @@ def run_migrations(
        lock_id = _get_schema_lock_id(schema) if schema else MIGRATION_LOCK_ID
        schema_name = schema or "public"

-        # Use PostgreSQL advisory lock to coordinate between distributed workers.
-        #
-        # IMPORTANT: We must avoid holding an open transaction on the advisory-lock
-        # connection while CREATE INDEX CONCURRENTLY runs inside a migration.
-        # CONCURRENTLY waits for ALL active transactions to finish before the index
-        # becomes valid.  If the advisory-lock connection (or any waiting worker's
-        # connection) holds an open transaction, CONCURRENTLY deadlocks:
-        #   - migration worker waits for other workers' transactions to close
-        #   - other workers wait for the advisory lock to be released
-        #
-        # Fix:
-        #   1. Use pg_try_advisory_lock (non-blocking) in a poll loop instead of
-        #      blocking pg_advisory_lock, so we can COMMIT the transaction between
-        #      retries.  Between retries the connection holds no open transaction.
-        #   2. After acquiring the lock, COMMIT the transaction on the advisory-lock
-        #      connection itself before running migrations.  pg_advisory_lock is
-        #      session-level, so the lock survives the COMMIT.
+        # Use PostgreSQL advisory lock to coordinate between distributed workers
        engine = create_engine(database_url)
        with engine.connect() as conn:
+            # pg_advisory_lock blocks until the lock is acquired
+            # The lock is automatically released when the connection closes
            logger.debug(f"Acquiring migration advisory lock for schema '{schema_name}' (id={lock_id})...")
-            while True:
-                acquired = conn.execute(text(f"SELECT pg_try_advisory_lock({lock_id})")).scalar()
-                if acquired:
-                    break
-                # Commit the transaction so this connection holds no open snapshot
-                # while waiting.  This prevents blocking CREATE INDEX CONCURRENTLY
-                # that may be running in the migration worker.
-                conn.commit()
-                time.sleep(0.5)
-
-            # Commit AFTER acquiring the lock too.  pg_advisory_lock is session-level
-            # and survives the COMMIT, but the open transaction on this connection
-            # would otherwise block any CREATE INDEX CONCURRENTLY in the migration.
-            conn.commit()
+            conn.execute(text(f"SELECT pg_advisory_lock({lock_id})"))
            logger.debug("Migration advisory lock acquired")

            try:
@@ -386,13 +347,6 @@ def run_migrations(
                                    "Please install it with: CREATE EXTENSION vectorscale CASCADE;"
                                ) from e

-                # Commit any pending transaction on the advisory-lock connection
-                # before running migrations.  Some code paths above (e.g., the
-                # pgvector extension check) may have started a transaction via
-                # SQLAlchemy's autobegin.  If we leave it open, CREATE INDEX
-                # CONCURRENTLY inside a migration will deadlock waiting for it.
-                conn.commit()
-
                # Run migrations while holding the lock
                _run_migrations_internal(database_url, script_location, schema=schema)
            finally:
@@ -471,125 +425,6 @@ def check_migration_status(
        return None, None


-def _migrate_table_embedding_dimension(
-    conn: Connection,
-    schema_name: str,
-    table_name: str,
-    required_dimension: int,
-    vector_ext: str,
-) -> None:
-    """
-    Migrate the embedding column of a single table to the required dimension.
-
-    - If dimensions match: no action needed
-    - If dimensions differ and table is empty: ALTER COLUMN to new dimension
-    - If dimensions differ and table has data: raise error with migration guidance
-    """
-    current_dim = conn.execute(
-        text("""
-            SELECT atttypmod
-            FROM pg_attribute a
-            JOIN pg_class c ON a.attrelid = c.oid
-            JOIN pg_namespace n ON c.relnamespace = n.oid
-            WHERE n.nspname = :schema
-              AND c.relname = :table
-              AND a.attname = 'embedding'
-        """),
-        {"schema": schema_name, "table": table_name},
-    ).scalar()
-
-    if current_dim is None:
-        logger.debug(f"No embedding column found on {table_name}, skipping")
-        return
-
-    if current_dim == required_dimension:
-        logger.debug(f"Embedding dimension OK for {table_name}: {current_dim}")
-        return
-
-    logger.info(
-        f"Embedding dimension mismatch on {table_name}: database has {current_dim}, model requires {required_dimension}"
-    )
-
-    row_count = conn.execute(
-        text(f"SELECT COUNT(*) FROM {schema_name}.{table_name} WHERE embedding IS NOT NULL")
-    ).scalar()
-
-    if row_count > 0:
-        raise RuntimeError(
-            f"Cannot change embedding dimension from {current_dim} to {required_dimension}: "
-            f"{table_name} table contains {row_count} rows with embeddings. "
-            f"To change dimensions, you must either:\n"
-            f"  1. Re-embed all data: DELETE FROM {schema_name}.{table_name}; then restart\n"
-            f"  2. Use a model with {current_dim}-dimensional embeddings"
-        )
-
-    logger.info(f"Altering {table_name}.embedding column dimension from {current_dim} to {required_dimension}")
-
-    # Drop existing vector index (works for both HNSW and vchordrq)
-    conn.execute(
-        text(f"""
-            DO $$
-            DECLARE idx_name TEXT;
-            BEGIN
-                FOR idx_name IN
-                    SELECT indexname FROM pg_indexes
-                    WHERE schemaname = '{schema_name}'
-                      AND tablename = '{table_name}'
-                      AND (indexdef LIKE '%hnsw%' OR indexdef LIKE '%vchordrq%' OR indexdef LIKE '%diskann%')
-                      AND indexdef LIKE '%embedding%'
-                LOOP
-                    EXECUTE 'DROP INDEX IF EXISTS {schema_name}.' || idx_name;
-                END LOOP;
-            END $$;
-        """)
-    )
-
-    conn.execute(
-        text(f"ALTER TABLE {schema_name}.{table_name} ALTER COLUMN embedding TYPE vector({required_dimension})")
-    )
-    conn.commit()
-
-    # Recreate index with appropriate type based on detected extension
-    if vector_ext == "pgvectorscale":
-        conn.execute(
-            text(f"""
-                CREATE INDEX IF NOT EXISTS idx_{table_name}_embedding_diskann
-                ON {schema_name}.{table_name}
-                USING diskann (embedding vector_cosine_ops)
-                WITH (num_neighbors = 50)
-            """)
-        )
-        logger.info(f"Created DiskANN index on {table_name} for {required_dimension}-dimensional embeddings")
-    elif vector_ext == "vchord":
-        conn.execute(
-            text(f"""
-                CREATE INDEX IF NOT EXISTS idx_{table_name}_embedding_vchordrq
-                ON {schema_name}.{table_name}
-                USING vchordrq (embedding vector_l2_ops)
-            """)
-        )
-        logger.info(f"Created vchordrq index on {table_name} for {required_dimension}-dimensional embeddings")
-    else:  # pgvector
-        if required_dimension > 2000:
-            raise RuntimeError(
-                f"Embedding dimension {required_dimension} exceeds pgvector HNSW index limit of 2000. "
-                f"Use an embedding model with <= 2000 dimensions, or switch to a vector extension "
-                f"that supports higher dimensions (e.g., pgvectorscale/DiskANN)."
-            )
-        conn.execute(
-            text(f"""
-                CREATE INDEX IF NOT EXISTS idx_{table_name}_embedding_hnsw
-                ON {schema_name}.{table_name}
-                USING hnsw (embedding vector_cosine_ops)
-                WITH (m = 16, ef_construction = 64)
-            """)
-        )
-        logger.info(f"Created HNSW index on {table_name} for {required_dimension}-dimensional embeddings")
-    conn.commit()
-
-    logger.info(f"Successfully changed {table_name}.embedding dimension to {required_dimension}")
-
-
 def ensure_embedding_dimension(
    database_url: str,
    required_dimension: int,
@@ -597,9 +432,10 @@ def ensure_embedding_dimension(
    vector_extension: str = "pgvector",
 ) -> None:
    """
-    Ensure the embedding column dimension matches the model's dimension for all tables.
+    Ensure the embedding column dimension matches the model's dimension.

-    Checks and adjusts memory_units.embedding and mental_models.embedding:
+    This function checks the current vector column dimension in the database
+    and adjusts it if necessary:
    - If dimensions match: no action needed
    - If dimensions differ and table is empty: ALTER COLUMN to new dimension
    - If dimensions differ and table has data: raise error with migration guidance
@@ -617,7 +453,7 @@ def ensure_embedding_dimension(

    engine = create_engine(database_url)
    with engine.connect() as conn:
-        # Check if memory_units table exists (proxy for schema being initialized)
+        # Check if memory_units table exists
        table_exists = conn.execute(
            text("""
                SELECT EXISTS (
@@ -636,8 +472,111 @@ def ensure_embedding_dimension(
        vector_ext = _detect_vector_extension(conn, vector_extension)
        logger.info(f"Using vector extension: {vector_ext}")

-        _migrate_table_embedding_dimension(conn, schema_name, "memory_units", required_dimension, vector_ext)
-        _migrate_table_embedding_dimension(conn, schema_name, "mental_models", required_dimension, vector_ext)
+        # Get current column dimension from pg_attribute
+        # pgvector stores dimension in atttypmod
+        current_dim = conn.execute(
+            text("""
+                SELECT atttypmod
+                FROM pg_attribute a
+                JOIN pg_class c ON a.attrelid = c.oid
+                JOIN pg_namespace n ON c.relnamespace = n.oid
+                WHERE n.nspname = :schema
+                  AND c.relname = 'memory_units'
+                  AND a.attname = 'embedding'
+            """),
+            {"schema": schema_name},
+        ).scalar()
+
+        if current_dim is None:
+            logger.warning("Could not determine current embedding dimension, skipping check")
+            return
+
+        # pgvector stores dimension directly in atttypmod (no offset like other types)
+        current_dimension = current_dim
+
+        if current_dimension == required_dimension:
+            logger.debug(f"Embedding dimension OK: {current_dimension}")
+            return
+
+        logger.info(
+            f"Embedding dimension mismatch: database has {current_dimension}, model requires {required_dimension}"
+        )
+
+        # Check if table has data
+        row_count = conn.execute(
+            text(f"SELECT COUNT(*) FROM {schema_name}.memory_units WHERE embedding IS NOT NULL")
+        ).scalar()
+
+        if row_count > 0:
+            raise RuntimeError(
+                f"Cannot change embedding dimension from {current_dimension} to {required_dimension}: "
+                f"memory_units table contains {row_count} rows with embeddings. "
+                f"To change dimensions, you must either:\n"
+                f"  1. Re-embed all data: DELETE FROM {schema_name}.memory_units; then restart\n"
+                f"  2. Use a model with {current_dimension}-dimensional embeddings"
+            )
+
+        # Table is empty, safe to alter column
+        logger.info(f"Altering embedding column dimension from {current_dimension} to {required_dimension}")
+
+        # Drop existing vector index (works for both HNSW and vchordrq)
+        conn.execute(
+            text(f"""
+                DO $$
+                DECLARE idx_name TEXT;
+                BEGIN
+                    FOR idx_name IN
+                        SELECT indexname FROM pg_indexes
+                        WHERE schemaname = '{schema_name}'
+                          AND tablename = 'memory_units'
+                          AND (indexdef LIKE '%hnsw%' OR indexdef LIKE '%vchordrq%')
+                          AND indexdef LIKE '%embedding%'
+                    LOOP
+                        EXECUTE 'DROP INDEX IF EXISTS {schema_name}.' || idx_name;
+                    END LOOP;
+                END $$;
+            """)
+        )
+
+        # Alter the column type
+        conn.execute(
+            text(f"ALTER TABLE {schema_name}.memory_units ALTER COLUMN embedding TYPE vector({required_dimension})")
+        )
+        conn.commit()
+
+        # Recreate index with appropriate type based on detected extension
+        if vector_ext == "pgvectorscale":
+            conn.execute(
+                text(f"""
+                    CREATE INDEX IF NOT EXISTS idx_memory_units_embedding_diskann
+                    ON {schema_name}.memory_units
+                    USING diskann (embedding vector_cosine_ops)
+                    WITH (num_neighbors = 50)
+                """)
+            )
+            logger.info(f"Created DiskANN index for {required_dimension}-dimensional embeddings")
+        elif vector_ext == "vchord":
+            conn.execute(
+                text(f"""
+                    CREATE INDEX IF NOT EXISTS idx_memory_units_embedding_vchordrq
+                    ON {schema_name}.memory_units
+                    USING vchordrq (embedding vector_l2_ops)
+                """)
+            )
+            logger.info(f"Created vchordrq index for {required_dimension}-dimensional embeddings")
+        else:  # pgvector
+            conn.execute(
+                text(f"""
+                    CREATE INDEX IF NOT EXISTS idx_memory_units_embedding_hnsw
+                    ON {schema_name}.memory_units
+                    USING hnsw (embedding vector_cosine_ops)
+                    WITH (m = 16, ef_construction = 64)
+                """)
+            )
+            logger.info(f"Created HNSW index for {required_dimension}-dimensional embeddings")
+        conn.commit()
+
+        logger.info(f"Successfully changed embedding dimension to {required_dimension}")


 def ensure_vector_extension(
@@ -811,24 +750,6 @@ def ensure_vector_extension(
                    """)
                )
            else:  # pgvector
-                # Check embedding dimension — pgvector HNSW indexes only support up to 2000 dims
-                embed_dim = conn.execute(
-                    text("""
-                        SELECT atttypmod
-                        FROM pg_attribute a
-                        JOIN pg_class c ON a.attrelid = c.oid
-                        JOIN pg_namespace n ON c.relnamespace = n.oid
-                        WHERE n.nspname = :schema AND c.relname = :table_name AND a.attname = 'embedding'
-                    """),
-                    {"schema": schema_name, "table_name": table_name},
-                ).scalar()
-
-                if embed_dim and embed_dim > 2000:
-                    raise RuntimeError(
-                        f"Embedding dimension {embed_dim} on {table_name} exceeds pgvector HNSW index limit of 2000. "
-                        f"Use an embedding model with <= 2000 dimensions, or switch to a vector extension "
-                        f"that supports higher dimensions (e.g., pgvectorscale/DiskANN)."
-                    )
                logger.info(f"Creating HNSW index on {table_name}")
                conn.execute(
                    text(f"""
@@ -22,7 +22,6 @@ class RequestContext:
    tenant_id: str | None = None  # Tenant identifier (set by extension after auth)
    internal: bool = False  # True for background/internal operations (skips extension auth)
    user_initiated: bool = False  # True for async operations that originated from a user request
-    allowed_bank_ids: list[str] | None = None  # None = unrestricted (all banks)


 from pgvector.sqlalchemy import Vector
@@ -1,13 +0,0 @@
-"""Webhook system for Hindsight API event notifications."""
-
-from .manager import WebhookManager
-from .models import ConsolidationEventData, RetainEventData, WebhookConfig, WebhookEvent, WebhookEventType
-
-__all__ = [
-    "WebhookManager",
-    "WebhookConfig",
-    "WebhookEvent",
-    "WebhookEventType",
-    "ConsolidationEventData",
-    "RetainEventData",
-]
@@ -1,242 +0,0 @@
-"""Webhook manager for delivering event notifications."""
-
-import hashlib
-import hmac
-import json
-import logging
-import uuid
-from datetime import datetime, timezone
-from typing import TYPE_CHECKING
-
-import asyncpg
-
-from .models import WebhookConfig, WebhookEvent, WebhookHttpConfig
-
-if TYPE_CHECKING:
-    from hindsight_api.extensions.tenant import TenantExtension
-
-logger = logging.getLogger(__name__)
-
-# Retry delay schedule in seconds: 5 retries after the first attempt.
-# Fast early retries catch transient failures; later retries handle longer outages.
-RETRY_DELAYS = [5, 300, 1800, 7200, 18000]
-MAX_ATTEMPTS = len(RETRY_DELAYS) + 1  # first attempt + len(RETRY_DELAYS) retries
-
-
-def _fq_table(table: str, schema: str | None = None) -> str:
-    """Get fully-qualified table name with optional schema prefix."""
-    if schema:
-        return f'"{schema}".{table}'
-    return table
-
-
-def _parse_http_config(value: str | dict | None) -> WebhookHttpConfig:
-    """Parse http_config column value (JSONB returned as text or dict) into a model."""
-    if value is None:
-        return WebhookHttpConfig()
-    if isinstance(value, str):
-        return WebhookHttpConfig.model_validate_json(value)
-    return WebhookHttpConfig.model_validate(value)
-
-
-class WebhookManager:
-    """
-    Manages webhook registration and event firing.
-
-    Supports both global webhooks (configured via env vars) and per-bank
-    webhooks stored in the database. Deliveries are queued as async_operations
-    tasks (operation_type='webhook_delivery') and picked up by the worker poller.
-    """
-
-    def __init__(
-        self,
-        pool: asyncpg.Pool,
-        global_webhooks: list[WebhookConfig],
-        tenant_extension: "TenantExtension | None" = None,
-    ):
-        self._pool = pool
-        self._global_webhooks = global_webhooks
-        self._tenant_extension = tenant_extension
-
-    def _sign_payload(self, secret: str, payload_bytes: bytes) -> str:
-        """Compute HMAC-SHA256 signature for a payload."""
-        return "sha256=" + hmac.new(secret.encode(), payload_bytes, hashlib.sha256).hexdigest()
-
-    async def fire_event(self, event: WebhookEvent, schema: str | None = None) -> None:
-        """
-        Queue webhook deliveries for an event as async_operations tasks.
-
-        Loads per-bank and global webhooks, inserts pending webhook_delivery tasks for
-        any webhook whose event_types list matches the fired event type. The worker
-        poller picks these up and calls MemoryEngine._handle_webhook_delivery().
-
-        Args:
-            event: The event to deliver.
-            schema: Database schema (for multi-tenant). None = default schema.
-        """
-        webhook_table = _fq_table("webhooks", schema)
-        ops_table = _fq_table("async_operations", schema)
-        now = datetime.now(timezone.utc)
-        payload_str = event.model_dump_json()
-
-        try:
-            # Load per-bank webhooks from DB (bank-specific + global NULL rows)
-            rows = await self._pool.fetch(
-                f"""
-                SELECT id, bank_id, url, secret, event_types, enabled, http_config::text
-                FROM {webhook_table}
-                WHERE (bank_id = $1 OR bank_id IS NULL) AND enabled = true
-                """,
-                event.bank_id,
-            )
-
-            db_webhooks = [
-                WebhookConfig(
-                    id=str(row["id"]),
-                    bank_id=row["bank_id"],
-                    url=row["url"],
-                    secret=row["secret"],
-                    event_types=list(row["event_types"]) if row["event_types"] else [],
-                    enabled=row["enabled"],
-                    http_config=_parse_http_config(row["http_config"]),
-                )
-                for row in rows
-            ]
-
-            # Merge with global webhooks from env config
-            all_webhooks = self._global_webhooks + db_webhooks
-            matched = 0
-
-            for webhook in all_webhooks:
-                if not webhook.enabled:
-                    continue
-                if event.event.value not in webhook.event_types:
-                    continue
-
-                operation_id = uuid.uuid4()
-                webhook_id = webhook.id if webhook.id else None
-
-                task_payload = json.dumps(
-                    {
-                        "type": "webhook_delivery",
-                        "operation_id": str(operation_id),
-                        "bank_id": event.bank_id,
-                        "url": webhook.url,
-                        "secret": webhook.secret,
-                        "event_type": event.event.value,
-                        "payload": payload_str,
-                        "webhook_id": webhook_id,
-                        "http_config": webhook.http_config.model_dump(),
-                    }
-                )
-
-                await self._pool.execute(
-                    f"""
-                    INSERT INTO {ops_table}
-                      (operation_id, bank_id, operation_type, status, task_payload, result_metadata, created_at, updated_at)
-                    VALUES ($1, $2, 'webhook_delivery', 'pending', $3::jsonb, '{{}}'::jsonb, $4, $4)
-                    """,
-                    operation_id,
-                    event.bank_id,
-                    task_payload,
-                    now,
-                )
-                matched += 1
-
-            logger.debug(f"Fired webhook event {event.event} for bank {event.bank_id}: {matched} delivery(ies) queued")
-
-        except Exception as e:
-            logger.error(f"Failed to queue webhook deliveries for event {event.event}: {e}")
-
-    async def fire_event_with_conn(
-        self, event: WebhookEvent, conn: asyncpg.Connection, schema: str | None = None
-    ) -> None:
-        """
-        Queue webhook deliveries within an existing database connection/transaction.
-
-        Identical to fire_event() but uses the provided connection instead of acquiring
-        one from the pool. Use this to atomically insert delivery tasks in the same
-        transaction as the primary operation (transactional outbox pattern).
-
-        Args:
-            event: The event to deliver.
-            conn: Existing asyncpg connection (may be inside an active transaction).
-            schema: Database schema (for multi-tenant). None = default schema.
-        """
-        webhook_table = _fq_table("webhooks", schema)
-        ops_table = _fq_table("async_operations", schema)
-        now = datetime.now(timezone.utc)
-        payload_str = event.model_dump_json()
-
-        try:
-            rows = await conn.fetch(
-                f"""
-                SELECT id, bank_id, url, secret, event_types, enabled, http_config::text
-                FROM {webhook_table}
-                WHERE (bank_id = $1 OR bank_id IS NULL) AND enabled = true
-                """,
-                event.bank_id,
-            )
-
-            db_webhooks = [
-                WebhookConfig(
-                    id=str(row["id"]),
-                    bank_id=row["bank_id"],
-                    url=row["url"],
-                    secret=row["secret"],
-                    event_types=list(row["event_types"]) if row["event_types"] else [],
-                    enabled=row["enabled"],
-                    http_config=_parse_http_config(row["http_config"]),
-                )
-                for row in rows
-            ]
-
-            all_webhooks = self._global_webhooks + db_webhooks
-            matched = 0
-
-            for webhook in all_webhooks:
-                if not webhook.enabled:
-                    continue
-                if event.event.value not in webhook.event_types:
-                    continue
-
-                operation_id = uuid.uuid4()
-                webhook_id = webhook.id if webhook.id else None
-
-                task_payload = json.dumps(
-                    {
-                        "type": "webhook_delivery",
-                        "operation_id": str(operation_id),
-                        "bank_id": event.bank_id,
-                        "url": webhook.url,
-                        "secret": webhook.secret,
-                        "event_type": event.event.value,
-                        "payload": payload_str,
-                        "webhook_id": webhook_id,
-                        "http_config": webhook.http_config.model_dump(),
-                    }
-                )
-
-                await conn.execute(
-                    f"""
-                    INSERT INTO {ops_table}
-                      (operation_id, bank_id, operation_type, status, task_payload, result_metadata, created_at, updated_at)
-                    VALUES ($1, $2, 'webhook_delivery', 'pending', $3::jsonb, '{{}}'::jsonb, $4, $4)
-                    """,
-                    operation_id,
-                    event.bank_id,
-                    task_payload,
-                    now,
-                )
-                matched += 1
-
-            logger.debug(
-                f"Fired webhook event {event.event} for bank {event.bank_id}: {matched} delivery(ies) queued (in-transaction)"
-            )
-
-        except Exception as e:
-            logger.error(
-                f"Failed to queue webhook deliveries (in-transaction) for event {event.event}: {e}. "
-                "CRITICAL: The enclosing database transaction is now aborted and will roll back all changes."
-            )
-            raise
@@ -1,51 +0,0 @@
-"""Pydantic models for the webhook system."""
-
-from datetime import datetime
-from enum import StrEnum
-
-from pydantic import BaseModel, Field
-
-
-class WebhookEventType(StrEnum):
-    CONSOLIDATION_COMPLETED = "consolidation.completed"
-    RETAIN_COMPLETED = "retain.completed"
-
-
-class ConsolidationEventData(BaseModel):
-    observations_created: int | None = None
-    observations_updated: int | None = None
-    observations_deleted: int | None = None
-    error_message: str | None = None
-
-
-class RetainEventData(BaseModel):
-    document_id: str | None = None
-    tags: list[str] | None = None
-
-
-class WebhookEvent(BaseModel):
-    event: WebhookEventType
-    bank_id: str
-    operation_id: str
-    status: str  # "completed" or "failed"
-    timestamp: datetime
-    data: ConsolidationEventData | RetainEventData
-
-
-class WebhookHttpConfig(BaseModel):
-    """HTTP delivery configuration for a webhook."""
-
-    method: str = Field(default="POST", description="HTTP method: GET or POST")
-    timeout_seconds: int = Field(default=30, description="HTTP request timeout in seconds")
-    headers: dict[str, str] = Field(default_factory=dict, description="Custom HTTP headers")
-    params: dict[str, str] = Field(default_factory=dict, description="Custom HTTP query parameters")
-
-
-class WebhookConfig(BaseModel):
-    id: str
-    bank_id: str | None
-    url: str
-    secret: str | None
-    event_types: list[str]
-    enabled: bool
-    http_config: WebhookHttpConfig = Field(default_factory=WebhookHttpConfig)
@@ -1,9 +0,0 @@
-from datetime import datetime
-
-
-class RetryTaskAt(Exception):
-    """Raise from a task handler to schedule a retry at a specific time."""
-
-    def __init__(self, retry_at: datetime, message: str = ""):
-        self.retry_at = retry_at
-        super().__init__(message)
@@ -219,6 +219,7 @@ def main():
            worker_id=args.worker_id,
            executor=memory.execute_task,
            poll_interval_ms=args.poll_interval,
+            max_retries=args.max_retries,
            schema=schema,
            tenant_extension=tenant_extension,
            max_slots=config.worker_max_slots,
@@ -14,8 +14,6 @@ from collections.abc import Awaitable, Callable
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any

-from .exceptions import RetryTaskAt
-
 if TYPE_CHECKING:
    import asyncpg

@@ -59,6 +57,7 @@ class WorkerPoller:
        worker_id: str,
        executor: Callable[[dict[str, Any]], Awaitable[None]],
        poll_interval_ms: int = 500,
+        max_retries: int = 3,
        schema: str | None = None,
        tenant_extension: "TenantExtension | None" = None,
        max_slots: int = 10,
@@ -72,6 +71,7 @@ class WorkerPoller:
            worker_id: Unique identifier for this worker
            executor: Async function to execute tasks (typically MemoryEngine.execute_task)
            poll_interval_ms: Interval between polls when no tasks found (milliseconds)
+            max_retries: Maximum retry attempts before marking task as failed
            schema: Database schema for single-tenant support (deprecated, use tenant_extension)
            tenant_extension: Extension for dynamic multi-tenant discovery. If None, creates a
                            DefaultTenantExtension with the configured schema.
@@ -82,6 +82,7 @@ class WorkerPoller:
        self._worker_id = worker_id
        self._executor = executor
        self._poll_interval_ms = poll_interval_ms
+        self._max_retries = max_retries
        self._schema = schema
        # Always set tenant extension (use DefaultTenantExtension if none provided)
        if tenant_extension is None:
@@ -217,12 +218,11 @@ class WorkerPoller:
                # 1. Claim non-consolidation tasks (up to limit)
                non_consolidation_rows = await conn.fetch(
                    f"""
-                    SELECT operation_id, task_payload, retry_count
+                    SELECT operation_id, task_payload
                    FROM {table}
                    WHERE status = 'pending'
                      AND task_payload IS NOT NULL
                      AND operation_type != 'consolidation'
-                      AND (next_retry_at IS NULL OR next_retry_at <= NOW())
                    ORDER BY created_at
                    LIMIT $1
                    FOR UPDATE SKIP LOCKED
@@ -238,12 +238,11 @@ class WorkerPoller:
                if consolidation_limit > 0 and remaining_limit > 0:
                    consolidation_rows = await conn.fetch(
                        f"""
-                        SELECT operation_id, task_payload, retry_count
+                        SELECT operation_id, task_payload
                        FROM {table} AS pending
                        WHERE status = 'pending'
                          AND task_payload IS NOT NULL
                          AND operation_type = 'consolidation'
-                          AND (next_retry_at IS NULL OR next_retry_at <= NOW())
                          AND NOT EXISTS (
                              SELECT 1 FROM {table} AS processing
                              WHERE processing.bank_id = pending.bank_id
@@ -275,19 +274,14 @@ class WorkerPoller:
                )

                # Parse and return task payloads with schema context
-                result = []
-                for row in all_rows:
-                    task_dict = json.loads(row["task_payload"])
-                    task_dict["_retry_count"] = row["retry_count"]
-                    task_dict["_operation_id"] = str(row["operation_id"])
-                    result.append(
-                        ClaimedTask(
-                            operation_id=str(row["operation_id"]),
-                            task_dict=task_dict,
-                            schema=schema,
-                        )
+                return [
+                    ClaimedTask(
+                        operation_id=str(row["operation_id"]),
+                        task_dict=json.loads(row["task_payload"]),
+                        schema=schema,
                    )
-                return result
+                    for row in all_rows
+                ]

    async def _mark_completed(self, operation_id: str, schema: str | None):
        """Mark a task as completed."""
@@ -316,22 +310,40 @@ class WorkerPoller:
            error_message,
        )

-    async def _schedule_retry(self, operation_id: str, retry_at: "Any", error_message: str, schema: str | None):
-        """Reset task to pending with a future retry timestamp."""
+    async def _retry_or_fail(self, operation_id: str, error_message: str, schema: str | None):
+        """Increment retry count or mark as failed if max retries exceeded."""
        table = fq_table("async_operations", schema)
-        error_message = error_message[:5000] if len(error_message) > 5000 else error_message
-        await self._pool.execute(
-            f"""
-            UPDATE {table}
-            SET status = 'pending', next_retry_at = $2, worker_id = NULL, claimed_at = NULL,
-                retry_count = retry_count + 1, error_message = $3, updated_at = now()
-            WHERE operation_id = $1
-            """,
+
+        # Get current retry count
+        row = await self._pool.fetchrow(
+            f"SELECT retry_count FROM {table} WHERE operation_id = $1",
            operation_id,
-            retry_at,
-            error_message,
        )
-        logger.warning(f"Task {operation_id} scheduled for retry at {retry_at}: {error_message}")
+
+        if row is None:
+            logger.warning(f"Operation {operation_id} not found, cannot retry")
+            return
+
+        retry_count = row["retry_count"]
+
+        if retry_count >= self._max_retries:
+            # Max retries exceeded, mark as failed
+            await self._mark_failed(
+                operation_id, f"Max retries ({self._max_retries}) exceeded. Last error: {error_message}", schema
+            )
+            logger.error(f"Task {operation_id} failed after {retry_count} retries")
+        else:
+            # Increment retry and reset to pending
+            await self._pool.execute(
+                f"""
+                UPDATE {table}
+                SET status = 'pending', worker_id = NULL, claimed_at = NULL,
+                    retry_count = retry_count + 1, updated_at = now()
+                WHERE operation_id = $1
+                """,
+                operation_id,
+            )
+            logger.warning(f"Task {operation_id} failed, will retry (attempt {retry_count + 1}/{self._max_retries})")

    async def execute_task(self, task: ClaimedTask):
        """Execute a single task as a background job (fire-and-forget)."""
@@ -364,12 +376,11 @@ class WorkerPoller:
                        del self._in_flight_by_type[operation_type]

    async def _execute_task_inner(self, task: ClaimedTask):
-        """Inner task execution with retry/fail handling.
+        """Inner task execution with error handling.

-        Tasks that want to be retried raise RetryTaskAt; the poller sets next_retry_at
-        and resets status to 'pending'. All other exceptions are marked as failed immediately.
-        Non-retryable failures (e.g., file_convert_retain) are handled by the executor
-        internally — it marks the operation as failed and returns normally.
+        Note: The executor (MemoryEngine.execute_task) handles status marking internally
+        (marking operations as completed/failed and handling retries). This method should
+        NOT override those status updates.
        """
        task_type = task.task_dict.get("type", "unknown")
        bank_id = task.task_dict.get("bank_id", "unknown")
@@ -381,12 +392,11 @@ class WorkerPoller:
                task.task_dict["_schema"] = task.schema
            await self._executor(task.task_dict)
            logger.debug(f"Task {task.operation_id} execution finished")
-        except RetryTaskAt as e:
-            await self._schedule_retry(task.operation_id, e.retry_at, str(e), task.schema)
        except Exception as e:
-            logger.error(f"Task {task.operation_id} failed: {e}")
+            # The executor should handle its own errors, but if an unexpected exception
+            # propagates (e.g., from schema setup), log it as a warning
+            logger.error(f"Task {task.operation_id} raised unexpected exception: {e}")
            traceback.print_exc()
-            await self._mark_failed(task.operation_id, str(e), task.schema)

    async def recover_own_tasks(self) -> int:
        """
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"

 [project]
 name = "hindsight-api"
-version = "0.4.17"
+version = "0.4.11"
 description = "Hindsight: Agent Memory That Works Like Human Memory"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -43,8 +43,8 @@ dependencies = [
    "cohere>=5.0.0",
    "flashrank>=0.2.0",
    "litellm>=1.0.0",
-    "markitdown[pdf,docx,pptx,xlsx,xls]>=0.1.4", # File to markdown conversion
-    "obstore>=0.4.0", # S3/GCS/Azure object storage client (Rust-backed)
+    "markitdown[pdf,docx,pptx,xlsx,xls]>=0.1.4",  # File to markdown conversion
+    "obstore>=0.4.0",  # S3/GCS/Azure object storage client (Rust-backed)
    # Local ML models for embeddings/reranking - can be excluded in Docker with INCLUDE_LOCAL_MODELS=false
    "sentence-transformers>=3.3.0",
    "transformers>=4.53.0", # Security fixes for ReDoS vulnerabilities
@@ -53,16 +53,11 @@ dependencies = [
    # Transitive dependency security fixes
    "pyasn1>=0.6.2", # DoS vulnerability fix
    "urllib3>=2.6.3", # Decompression-bomb safeguards bypass fix
-    "langchain-core>=1.2.11", # Serialization injection + SSRF vulnerability fix
-    "langsmith>=0.6.3", # SSRF via tracing header injection fix
-    "protobuf>=6.33.5", # JSON recursion depth bypass fix
-    "pillow>=12.1.1", # Out-of-bounds write in PSD image loading fix
-    "cryptography>=46.0.5", # Subgroup attack vulnerability fix
+    "langchain-core>=1.2.5", # Serialization injection vulnerability fix
    "filelock>=3.20.1", # TOCTOU race condition fix
    "authlib>=1.6.6", # Account takeover vulnerability fix
    "aiohttp>=3.13.3", # Multiple DoS vulnerabilities
    "claude-agent-sdk>=0.1.27",
-    "einops>=0.8.2",
 ]

 [project.optional-dependencies]
@@ -103,7 +98,7 @@ log_cli = true
 log_cli_level = "INFO"
 log_cli_format = "%(asctime)s - %(levelname)s - %(name)s - %(message)s"
 log_cli_date_format = "%Y-%m-%d %H:%M:%S"
-addopts = "--timeout 300 -n 8 --dist loadgroup --durations=10 -v"
+addopts = "--timeout 120 -n 8 --dist loadgroup --durations=10 -v"
 asyncio_mode = "auto"
 asyncio_default_fixture_loop_scope = "function"
 log_auto_indent = true
@@ -15,9 +15,7 @@ import asyncpg
 import pytest
 import pytest_asyncio

-import hindsight_api.admin.cli as admin_cli
 from hindsight_api.admin.cli import _backup, _restore, BACKUP_TABLES
-from hindsight_api.extensions import Tenant
 from hindsight_api.migrations import run_migrations


@@ -292,196 +290,3 @@ async def test_backup_restore_preserves_all_column_types(backup_test_schema):
    finally:
        if backup_path.exists():
            backup_path.unlink()
-
-
-@pytest.mark.asyncio
-async def test_run_migration_without_schema_discovers_and_deduplicates_schemas(monkeypatch):
-    """run-db-migration without --schema should include the base schema and deduplicate tenant schemas."""
-    calls: dict[str, list] = {
-        "run_migrations": [],
-        "ensure_vector_extension": [],
-        "ensure_text_search_extension": [],
-    }
-
-    class MockTenantExtension:
-        async def list_tenants(self):
-            return [
-                Tenant(schema="public"),
-                Tenant(schema="tenant_demo"),
-                Tenant(schema="tenant_demo"),
-            ]
-
-    async def fake_resolve_database_url(db_url: str) -> str:
-        return f"resolved::{db_url}"
-
-    def fake_run_migrations(database_url: str, schema: str | None = None) -> None:
-        calls["run_migrations"].append((database_url, schema))
-
-    def fake_ensure_vector_extension(
-        database_url: str,
-        vector_extension: str = "pgvector",
-        schema: str | None = None,
-    ) -> None:
-        calls["ensure_vector_extension"].append((database_url, vector_extension, schema))
-
-    def fake_ensure_text_search_extension(
-        database_url: str,
-        text_search_extension: str = "native",
-        schema: str | None = None,
-    ) -> None:
-        calls["ensure_text_search_extension"].append((database_url, text_search_extension, schema))
-
-    monkeypatch.setenv("HINDSIGHT_API_DATABASE_URL", "postgresql://test")
-    monkeypatch.setattr(admin_cli, "load_extension", lambda *args, **kwargs: MockTenantExtension())
-    monkeypatch.setattr(admin_cli, "resolve_database_url", fake_resolve_database_url)
-
-    from hindsight_api import migrations as migrations_module
-
-    monkeypatch.setattr(migrations_module, "run_migrations", fake_run_migrations)
-    monkeypatch.setattr(migrations_module, "ensure_vector_extension", fake_ensure_vector_extension)
-    monkeypatch.setattr(migrations_module, "ensure_text_search_extension", fake_ensure_text_search_extension)
-
-    schemas = await admin_cli._run_migration("postgresql://test")
-
-    assert schemas == ["public", "tenant_demo"]
-    assert calls["run_migrations"] == [
-        ("resolved::postgresql://test", "public"),
-        ("resolved::postgresql://test", "tenant_demo"),
-    ]
-    assert calls["ensure_vector_extension"] == [
-        ("resolved::postgresql://test", "pgvector", "public"),
-        ("resolved::postgresql://test", "pgvector", "tenant_demo"),
-    ]
-    assert calls["ensure_text_search_extension"] == [
-        ("resolved::postgresql://test", "native", "public"),
-        ("resolved::postgresql://test", "native", "tenant_demo"),
-    ]
-
-
-@pytest.mark.asyncio
-async def test_run_migration_without_schema_runs_optional_post_migration_hooks(monkeypatch):
-    """Embedding dimension sync should be optional, while vector/text checks always run."""
-    monkeypatch.setenv("HINDSIGHT_API_DATABASE_URL", "postgresql://test")
-    calls: dict[str, list] = {
-        "run_migrations": [],
-        "ensure_embedding_dimension": [],
-        "ensure_vector_extension": [],
-        "ensure_text_search_extension": [],
-    }
-
-    class MockTenantExtension:
-        async def list_tenants(self):
-            return [Tenant(schema="tenant_demo")]
-
-    async def fake_resolve_database_url(db_url: str) -> str:
-        return f"resolved::{db_url}"
-
-    def fake_run_migrations(database_url: str, schema: str | None = None) -> None:
-        calls["run_migrations"].append((database_url, schema))
-
-    def fake_ensure_embedding_dimension(
-        database_url: str,
-        dimension: int,
-        schema: str | None = None,
-        vector_extension: str = "pgvector",
-    ) -> None:
-        calls["ensure_embedding_dimension"].append((database_url, dimension, schema, vector_extension))
-
-    def fake_ensure_vector_extension(
-        database_url: str,
-        vector_extension: str = "pgvector",
-        schema: str | None = None,
-    ) -> None:
-        calls["ensure_vector_extension"].append((database_url, vector_extension, schema))
-
-    def fake_ensure_text_search_extension(
-        database_url: str,
-        text_search_extension: str = "native",
-        schema: str | None = None,
-    ) -> None:
-        calls["ensure_text_search_extension"].append((database_url, text_search_extension, schema))
-
-    monkeypatch.setattr(admin_cli, "load_extension", lambda *args, **kwargs: MockTenantExtension())
-    monkeypatch.setattr(admin_cli, "resolve_database_url", fake_resolve_database_url)
-
-    from hindsight_api import migrations as migrations_module
-
-    monkeypatch.setattr(migrations_module, "run_migrations", fake_run_migrations)
-    monkeypatch.setattr(migrations_module, "ensure_embedding_dimension", fake_ensure_embedding_dimension)
-    monkeypatch.setattr(migrations_module, "ensure_vector_extension", fake_ensure_vector_extension)
-    monkeypatch.setattr(migrations_module, "ensure_text_search_extension", fake_ensure_text_search_extension)
-
-    schemas = await admin_cli._run_migration(
-        "postgresql://test",
-        base_schema="public",
-        embedding_dimension=384,
-    )
-
-    assert schemas == ["public", "tenant_demo"]
-    assert calls["run_migrations"] == [
-        ("resolved::postgresql://test", "public"),
-        ("resolved::postgresql://test", "tenant_demo"),
-    ]
-    assert calls["ensure_embedding_dimension"] == [
-        ("resolved::postgresql://test", 384, "public", "pgvector"),
-        ("resolved::postgresql://test", 384, "tenant_demo", "pgvector"),
-    ]
-    assert calls["ensure_vector_extension"] == [
-        ("resolved::postgresql://test", "pgvector", "public"),
-        ("resolved::postgresql://test", "pgvector", "tenant_demo"),
-    ]
-    assert calls["ensure_text_search_extension"] == [
-        ("resolved::postgresql://test", "native", "public"),
-        ("resolved::postgresql://test", "native", "tenant_demo"),
-    ]
-
-
-@pytest.mark.asyncio
-async def test_run_migration_with_schema_only_runs_requested_schema(monkeypatch):
-    """run-db-migration with --schema should only migrate the requested schema."""
-    monkeypatch.setenv("HINDSIGHT_API_DATABASE_URL", "postgresql://test")
-    calls: dict[str, list] = {
-        "run_migrations": [],
-        "ensure_vector_extension": [],
-        "ensure_text_search_extension": [],
-    }
-
-    class MockTenantExtension:
-        async def list_tenants(self):
-            return [Tenant(schema="tenant_demo"), Tenant(schema="tenant_other")]
-
-    async def fake_resolve_database_url(db_url: str) -> str:
-        return f"resolved::{db_url}"
-
-    def fake_run_migrations(database_url: str, schema: str | None = None) -> None:
-        calls["run_migrations"].append((database_url, schema))
-
-    def fake_ensure_vector_extension(
-        database_url: str,
-        vector_extension: str = "pgvector",
-        schema: str | None = None,
-    ) -> None:
-        calls["ensure_vector_extension"].append((database_url, vector_extension, schema))
-
-    def fake_ensure_text_search_extension(
-        database_url: str,
-        text_search_extension: str = "native",
-        schema: str | None = None,
-    ) -> None:
-        calls["ensure_text_search_extension"].append((database_url, text_search_extension, schema))
-
-    monkeypatch.setattr(admin_cli, "load_extension", lambda *args, **kwargs: MockTenantExtension())
-    monkeypatch.setattr(admin_cli, "resolve_database_url", fake_resolve_database_url)
-
-    from hindsight_api import migrations as migrations_module
-
-    monkeypatch.setattr(migrations_module, "run_migrations", fake_run_migrations)
-    monkeypatch.setattr(migrations_module, "ensure_vector_extension", fake_ensure_vector_extension)
-    monkeypatch.setattr(migrations_module, "ensure_text_search_extension", fake_ensure_text_search_extension)
-
-    schemas = await admin_cli._run_migration("postgresql://test", schema="tenant_demo")
-
-    assert schemas == ["tenant_demo"]
-    assert calls["run_migrations"] == [("resolved::postgresql://test", "tenant_demo")]
-    assert calls["ensure_vector_extension"] == [("resolved::postgresql://test", "pgvector", "tenant_demo")]
-    assert calls["ensure_text_search_extension"] == [("resolved::postgresql://test", "native", "tenant_demo")]
@@ -27,9 +27,9 @@ class TestAgentProfile:
        assert "disposition" in profile

        disposition = profile["disposition"]
-        assert disposition["skepticism"] == 3
-        assert disposition["literalism"] == 3
-        assert disposition["empathy"] == 3
+        assert disposition.skepticism == 3
+        assert disposition.literalism == 3
+        assert disposition.empathy == 3

    @pytest.mark.asyncio
    async def test_update_agent_disposition(self, memory: MemoryEngine, request_context):
@@ -37,7 +37,7 @@ class TestAgentProfile:
        bank_id = unique_agent_id("test_profile_update")

        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-        assert profile["disposition"]["skepticism"] == 3
+        assert profile["disposition"].skepticism == 3

        new_disposition = {
            "skepticism": 5,
@@ -48,9 +48,9 @@ class TestAgentProfile:

        updated_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
        disposition = updated_profile["disposition"]
-        assert disposition["skepticism"] == new_disposition["skepticism"]
-        assert disposition["literalism"] == new_disposition["literalism"]
-        assert disposition["empathy"] == new_disposition["empathy"]
+        assert disposition.skepticism == new_disposition["skepticism"]
+        assert disposition.literalism == new_disposition["literalism"]
+        assert disposition.empathy == new_disposition["empathy"]

    @pytest.mark.asyncio
    async def test_list_agents(self, memory: MemoryEngine, request_context):
@@ -104,8 +104,8 @@ class TestAgentEndpoint:

        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)

-        assert final_profile["disposition"]["skepticism"] == 4
-        assert final_profile["disposition"]["literalism"] == 5
+        assert final_profile["disposition"].skepticism == 4
+        assert final_profile["disposition"].literalism == 5


 class TestAgentDispositionIntegration:
@@ -14,7 +14,6 @@ async def test_submit_async_retain_includes_document_tags_in_task_payload():
    engine = MemoryEngine.__new__(MemoryEngine)
    engine._initialized = True
    engine._authenticate_tenant = AsyncMock()
-    engine._operation_validator = None
    engine._submit_async_operation = AsyncMock(return_value={"operation_id": "op-1"})

    # Mock the pool and connection for parent operation creation
@@ -413,6 +413,7 @@ async def test_worker_batch_recovery(memory, request_context):
            worker_id="test_worker_recovery",
            executor=memory,
            poll_interval_ms=100,
+            max_retries=3,
            schema=schema,
            tenant_extension=tenant_extension,
            max_slots=5,
@@ -1,171 +1,334 @@
 """
-Tests for combined scoring (apply_combined_scoring).
+Tests for combined scoring functionality.

-The function applies multiplicative recency/temporal boosts to the cross-encoder
-score so that the relative influence of these signals is proportional to the base
-relevance score, independent of the cross-encoder model's score calibration.
+Verifies that:
+1. RRF scores are properly normalized to [0, 1] range
+2. Combined scoring formula is applied correctly
+3. Tracer captures normalized values (not raw values)
 """
-
-from datetime import datetime, timedelta, timezone
-from unittest.mock import MagicMock
-
 import pytest
-
-from hindsight_api.engine.search.reranking import apply_combined_scoring, _RECENCY_ALPHA, _TEMPORAL_ALPHA
-from hindsight_api.engine.search.types import MergedCandidate, RetrievalResult, ScoredResult
-
-UTC = timezone.utc
-NOW = datetime(2024, 6, 1, tzinfo=UTC)
+from datetime import datetime, timezone
+from hindsight_api.engine.search.types import RetrievalResult, MergedCandidate, ScoredResult
+from hindsight_api.engine.memory_engine import Budget
+from hindsight_api import RequestContext


-def _make_result(
-    ce_norm: float,
-    occurred_start: datetime | None = None,
-    temporal_proximity: float | None = None,
-) -> ScoredResult:
-    retrieval = MagicMock(spec=RetrievalResult)
-    retrieval.occurred_start = occurred_start
-    retrieval.temporal_proximity = temporal_proximity
+class TestRRFNormalization:
+    """Test that RRF scores are properly normalized."""

-    candidate = MagicMock(spec=MergedCandidate)
-    candidate.retrieval = retrieval
-    candidate.rrf_score = 0.05
+    def test_rrf_normalized_range(self):
+        """RRF normalized values should be in [0, 1] range, not raw [0.04, 0.06]."""
+        # Simulate RRF scores like what we get from actual retrieval
+        raw_rrf_scores = [0.0607, 0.0550, 0.0480, 0.0390]

-    return ScoredResult(
-        candidate=candidate,
-        cross_encoder_score=1.0,
-        cross_encoder_score_normalized=ce_norm,
-        weight=ce_norm,
-    )
+        max_rrf = max(raw_rrf_scores)
+        min_rrf = min(raw_rrf_scores)
+        rrf_range = max_rrf - min_rrf
+
+        normalized = []
+        for score in raw_rrf_scores:
+            if rrf_range > 0:
+                norm = (score - min_rrf) / rrf_range
+            else:
+                norm = 0.5
+            normalized.append(norm)
+
+        # Verify normalized values are in [0, 1]
+        for i, norm in enumerate(normalized):
+            assert 0.0 <= norm <= 1.0, f"Normalized RRF {norm} not in [0, 1] for raw {raw_rrf_scores[i]}"
+
+        # Highest raw should be 1.0
+        assert normalized[0] == 1.0, f"Highest RRF should normalize to 1.0, got {normalized[0]}"
+
+        # Lowest raw should be 0.0
+        assert normalized[-1] == 0.0, f"Lowest RRF should normalize to 0.0, got {normalized[-1]}"
+
+    def test_rrf_all_same_scores(self):
+        """When all RRF scores are the same, normalized should be 0.5 (neutral)."""
+        raw_rrf_scores = [0.0500, 0.0500, 0.0500]
+
+        max_rrf = max(raw_rrf_scores)
+        min_rrf = min(raw_rrf_scores)
+        rrf_range = max_rrf - min_rrf
+
+        normalized = []
+        for score in raw_rrf_scores:
+            if rrf_range > 0:
+                norm = (score - min_rrf) / rrf_range
+            else:
+                norm = 0.5  # Neutral value when all same
+            normalized.append(norm)
+
+        # All should be 0.5 when scores are identical
+        for norm in normalized:
+            assert norm == 0.5, f"Expected 0.5 for identical scores, got {norm}"


-class TestBoostFormula:
-    def test_neutral_signals_leave_score_unchanged(self):
-        """recency=0.5 and temporal=0.5 both produce boost=1.0, so weight == ce."""
-        sr = _make_result(ce_norm=0.6)
-        apply_combined_scoring([sr], now=NOW)
-        assert abs(sr.weight - 0.6) < 1e-9
+class TestCombinedScoringFormula:
+    """Test that the combined scoring formula is applied correctly."""

-    def test_max_recency_boost(self):
-        """A memory from today (recency≈1.0) should boost by (1 + alpha*0.5)."""
-        sr = _make_result(ce_norm=0.5, occurred_start=NOW)
-        apply_combined_scoring([sr], now=NOW)
-        expected = 0.5 * (1.0 + _RECENCY_ALPHA * 0.5) * 1.0  # temporal neutral
-        assert abs(sr.weight - expected) < 1e-6
+    def test_combined_score_calculation(self):
+        """Verify the weighted combination: 0.6*CE + 0.2*RRF + 0.1*temporal + 0.1*recency."""
+        # Test case 1: All components at 1.0
+        ce_norm = 1.0
+        rrf_norm = 1.0
+        temporal = 1.0
+        recency = 1.0

-    def test_min_recency_penalty(self):
-        """A memory from >365 days ago (recency=0.1) should penalise score."""
-        old = NOW - timedelta(days=400)
-        sr = _make_result(ce_norm=0.5, occurred_start=old)
-        apply_combined_scoring([sr], now=NOW)
-        expected = 0.5 * (1.0 + _RECENCY_ALPHA * (0.1 - 0.5)) * 1.0
-        assert abs(sr.weight - expected) < 1e-6
+        expected = 0.6 * ce_norm + 0.2 * rrf_norm + 0.1 * temporal + 0.1 * recency
+        assert expected == 1.0, f"All 1.0 should give 1.0, got {expected}"

-    def test_max_temporal_boost(self):
-        """temporal_proximity=1.0 should boost by (1 + alpha*0.5)."""
-        sr = _make_result(ce_norm=0.5, temporal_proximity=1.0)
-        apply_combined_scoring([sr], now=NOW)
-        expected = 0.5 * 1.0 * (1.0 + _TEMPORAL_ALPHA * 0.5)  # recency neutral
-        assert abs(sr.weight - expected) < 1e-6
+        # Test case 2: All components at 0.0
+        ce_norm = 0.0
+        rrf_norm = 0.0
+        temporal = 0.0
+        recency = 0.0

-    def test_temporal_none_is_neutral(self):
-        """temporal_proximity=None must be treated as 0.5 (no boost/penalty)."""
-        sr_none = _make_result(ce_norm=0.5, temporal_proximity=None)
-        sr_half = _make_result(ce_norm=0.5, temporal_proximity=0.5)
-        apply_combined_scoring([sr_none], now=NOW)
-        apply_combined_scoring([sr_half], now=NOW)
-        assert abs(sr_none.weight - sr_half.weight) < 1e-9
+        expected = 0.6 * ce_norm + 0.2 * rrf_norm + 0.1 * temporal + 0.1 * recency
+        assert expected == 0.0, f"All 0.0 should give 0.0, got {expected}"

-    def test_both_signals_combined(self):
-        """Both boosts are applied multiplicatively."""
-        sr = _make_result(ce_norm=0.5, occurred_start=NOW, temporal_proximity=1.0)
-        apply_combined_scoring([sr], now=NOW)
-        recency_boost = 1.0 + _RECENCY_ALPHA * (1.0 - 0.5)
-        temporal_boost = 1.0 + _TEMPORAL_ALPHA * (1.0 - 0.5)
-        expected = 0.5 * recency_boost * temporal_boost
-        assert abs(sr.weight - expected) < 1e-6
+        # Test case 3: High CE, low RRF (cross-encoder finds something retrieval missed)
+        ce_norm = 0.999
+        rrf_norm = 0.0  # Lowest in set
+        temporal = 0.5
+        recency = 0.5

-    def test_boost_is_proportional_to_ce(self):
-        """The absolute boost from recency scales with the CE score."""
-        sr_high = _make_result(ce_norm=0.9, occurred_start=NOW)
-        sr_low = _make_result(ce_norm=0.3, occurred_start=NOW)
-        apply_combined_scoring([sr_high, sr_low], now=NOW)
+        expected = 0.6 * ce_norm + 0.2 * rrf_norm + 0.1 * temporal + 0.1 * recency
+        # 0.5994 + 0.0 + 0.05 + 0.05 = 0.6994
+        assert abs(expected - 0.6994) < 0.001, f"Expected ~0.6994, got {expected}"

-        # Both get the same recency boost factor — absolute gain is proportional to CE
-        boost_factor = 1.0 + _RECENCY_ALPHA * 0.5
-        assert abs(sr_high.weight - 0.9 * boost_factor) < 1e-6
-        assert abs(sr_low.weight - 0.3 * boost_factor) < 1e-6
+        # Test case 4: Medium CE, high RRF (retrieval consensus)
+        ce_norm = 0.8
+        rrf_norm = 1.0  # Highest in set
+        temporal = 0.5
+        recency = 0.5

-    def test_boost_capped(self):
-        """Max boost: recency=1.0 + temporal=1.0 gives ≤21% uplift on CE."""
-        sr = _make_result(ce_norm=1.0, occurred_start=NOW, temporal_proximity=1.0)
-        apply_combined_scoring([sr], now=NOW)
-        assert sr.weight <= 1.0 * (1 + _RECENCY_ALPHA / 2) * (1 + _TEMPORAL_ALPHA / 2) + 1e-9
+        expected = 0.6 * ce_norm + 0.2 * rrf_norm + 0.1 * temporal + 0.1 * recency
+        # 0.48 + 0.2 + 0.05 + 0.05 = 0.78
+        assert abs(expected - 0.78) < 0.001, f"Expected ~0.78, got {expected}"

-    def test_rrf_normalized_always_zero(self):
-        """RRF is excluded from scoring; rrf_normalized is set to 0.0 for trace clarity."""
-        sr = _make_result(ce_norm=0.5)
-        apply_combined_scoring([sr], now=NOW)
-        assert sr.rrf_normalized == 0.0
+    def test_rrf_contribution_is_significant(self):
+        """Verify RRF actually contributes to the final score (not negligible)."""
+        # Same CE, different RRF
+        ce_norm = 0.8
+        temporal = 0.5
+        recency = 0.5

-    def test_combined_score_equals_weight(self):
-        """combined_score and weight must stay in sync."""
-        sr = _make_result(ce_norm=0.7, occurred_start=NOW, temporal_proximity=0.8)
-        apply_combined_scoring([sr], now=NOW)
-        assert sr.combined_score == sr.weight
+        # Low RRF
+        score_low_rrf = 0.6 * ce_norm + 0.2 * 0.0 + 0.1 * temporal + 0.1 * recency

-    def test_model_calibration_independence(self):
-        """
-        A low-calibration model (low CE scores) and a high-calibration model
-        (high CE scores) should produce the same ranking for identical content.
+        # High RRF
+        score_high_rrf = 0.6 * ce_norm + 0.2 * 1.0 + 0.1 * temporal + 0.1 * recency

-        With additive scoring the recency term would dominate for low-CE models;
-        with multiplicative boosting the relative ranking is stable.
-        """
-        recent = NOW - timedelta(days=10)
-        old = NOW - timedelta(days=300)
+        # Difference should be 0.2 (20% contribution)
+        diff = score_high_rrf - score_low_rrf
+        assert abs(diff - 0.2) < 0.001, f"RRF should contribute 0.2 difference, got {diff}"

-        # High-calibration model: clear winner is #1 (more relevant, slightly older)
-        h_relevant = _make_result(ce_norm=0.85, occurred_start=old)
-        h_recent = _make_result(ce_norm=0.60, occurred_start=recent)
-        apply_combined_scoring([h_relevant, h_recent], now=NOW)
-        assert h_relevant.weight > h_recent.weight, "High-CE model: relevance should win"

-        # Low-calibration model: same relative difference, just compressed scores
-        l_relevant = _make_result(ce_norm=0.34, occurred_start=old)
-        l_recent = _make_result(ce_norm=0.24, occurred_start=recent)
-        apply_combined_scoring([l_relevant, l_recent], now=NOW)
-        assert l_relevant.weight > l_recent.weight, "Low-CE model: relevance should still win"
+@pytest.mark.asyncio
+async def test_trace_has_normalized_rrf(memory, request_context):
+    """Integration test: verify trace contains normalized RRF values, not raw."""
+    bank_id = f"test_scoring_{datetime.now(timezone.utc).timestamp()}"

-    def test_no_occurred_start_defaults_recency_neutral(self):
-        """Missing occurred_start → recency=0.5 → no boost/penalty."""
-        sr = _make_result(ce_norm=0.5, occurred_start=None)
-        apply_combined_scoring([sr], now=NOW)
-        assert sr.recency == 0.5
-        assert abs(sr.weight - 0.5) < 1e-9
+    try:
+        # Store multiple memories to ensure different RRF scores
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="Python is a programming language created by Guido van Rossum",
+            context="tech facts",
+            request_context=request_context,
+        )
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="JavaScript was created by Brendan Eich at Netscape",
+            context="tech facts",
+            request_context=request_context,
+        )
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="The Eiffel Tower is located in Paris, France",
+            context="geography facts",
+            request_context=request_context,
+        )
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="Mount Everest is the tallest mountain on Earth",
+            context="geography facts",
+            request_context=request_context,
+        )

-    def test_timezone_naive_occurred_start_handled(self):
-        """Naive datetimes in occurred_start should not raise."""
-        naive_date = datetime(2024, 1, 1)  # no tzinfo
-        sr = _make_result(ce_norm=0.5, occurred_start=naive_date)
-        apply_combined_scoring([sr], now=NOW)  # must not raise
-        assert 0.0 < sr.weight < 1.0
+        # Search with tracing
+        result = await memory.recall_async(
+            bank_id=bank_id,
+            query="programming languages",
+            fact_type=["world"],
+            budget=Budget.LOW,
+            max_tokens=1024,
+            enable_trace=True,
+            request_context=request_context,
+        )

-    def test_custom_alpha_values(self):
-        """Custom alpha parameters are respected."""
-        sr = _make_result(ce_norm=0.5, occurred_start=NOW)
-        apply_combined_scoring([sr], now=NOW, recency_alpha=0.4, temporal_alpha=0.0)
-        expected = 0.5 * (1.0 + 0.4 * 0.5) * 1.0
-        assert abs(sr.weight - expected) < 1e-6
+        assert result.trace is not None, "Trace should be present"
+        trace = result.trace

-    def test_future_event_recency_capped_at_one(self):
-        """Events in the future must not produce recency > 1.0, keeping boost within bounds."""
-        future = NOW + timedelta(days=180)
-        sr = _make_result(ce_norm=0.5, occurred_start=future)
-        apply_combined_scoring([sr], now=NOW)
-        assert sr.recency == 1.0
-        expected_max_boost = 1.0 + _RECENCY_ALPHA * 0.5
-        assert sr.weight <= 0.5 * expected_max_boost + 1e-9
+        # Check reranked results have proper score_components
+        assert "reranked" in trace, "Trace should have reranked results"
+        assert len(trace["reranked"]) > 0, "Should have reranked results"

-    def test_empty_list_is_noop(self):
-        apply_combined_scoring([], now=NOW)  # must not raise
+        has_valid_rrf = False
+        has_valid_temporal = False
+        has_valid_recency = False
+
+        for r in trace["reranked"]:
+            sc = r.get("score_components", {})
+
+            # Check RRF normalized is present and in valid range
+            if "rrf_normalized" in sc:
+                rrf_norm = sc["rrf_normalized"]
+                assert 0.0 <= rrf_norm <= 1.0, f"rrf_normalized {rrf_norm} should be in [0, 1]"
+                # Should NOT be raw RRF score (which would be ~0.04-0.06)
+                # A normalized value of exactly 0.0 or 1.0 is valid (min/max of set)
+                # But raw scores like 0.0607 should never appear as normalized
+                if rrf_norm > 0.1:  # Any value > 0.1 is likely properly normalized
+                    has_valid_rrf = True
+
+            # Check temporal is present and in valid range
+            if "temporal" in sc:
+                temporal = sc["temporal"]
+                assert 0.0 <= temporal <= 1.0, f"temporal {temporal} should be in [0, 1]"
+                has_valid_temporal = True
+
+            # Check recency is present and in valid range
+            if "recency" in sc:
+                recency = sc["recency"]
+                assert 0.0 <= recency <= 1.0, f"recency {recency} should be in [0, 1]"
+                has_valid_recency = True
+
+        # At least some results should have these components
+        # (might not have rrf > 0.1 if all scores are same, which is fine)
+        assert has_valid_temporal, "Should have temporal scores in trace"
+        assert has_valid_recency, "Should have recency scores in trace"
+
+        print("\n✓ Combined scoring trace test passed!")
+        print(f"  - Reranked results: {len(trace['reranked'])}")
+        if trace["reranked"]:
+            sc = trace["reranked"][0].get("score_components", {})
+            print(f"  - First result score components: {sc}")
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_rrf_normalized_not_raw_in_trace(memory, request_context):
+    """Verify that raw RRF scores (0.04-0.06 range) don't appear as normalized values."""
+    bank_id = f"test_rrf_raw_{datetime.now(timezone.utc).timestamp()}"
+
+    try:
+        # Store enough memories to get varied RRF scores
+        for i in range(5):
+            await memory.retain_async(
+                bank_id=bank_id,
+                content=f"Test fact number {i} about various topics",
+                context="test context",
+                request_context=request_context,
+            )
+
+        result = await memory.recall_async(
+            bank_id=bank_id,
+            query="test fact",
+            fact_type=["world"],
+            budget=Budget.LOW,
+            max_tokens=512,
+            enable_trace=True,
+            request_context=request_context,
+        )
+
+        trace = result.trace
+        assert trace is not None
+
+        # Check that rrf_normalized values are NOT in the raw range
+        raw_rrf_range = (0.01, 0.08)  # Raw RRF scores are typically in this range
+
+        for r in trace.get("reranked", []):
+            sc = r.get("score_components", {})
+
+            if "rrf_normalized" in sc and "rrf_score" in sc:
+                rrf_norm = sc["rrf_normalized"]
+                rrf_raw = sc["rrf_score"]
+
+                # Raw should be in the typical range
+                assert raw_rrf_range[0] <= rrf_raw <= raw_rrf_range[1], \
+                    f"Raw RRF {rrf_raw} should be in typical range {raw_rrf_range}"
+
+                # Normalized should either be:
+                # - 0.0 (min in set)
+                # - 1.0 (max in set)
+                # - 0.5 (all same)
+                # - Something in between (0.0 to 1.0)
+                # But NOT the same as raw (which would indicate no normalization)
+                if len(trace["reranked"]) > 1:
+                    # If we have multiple results, normalized should differ from raw
+                    # (unless by coincidence, which is very unlikely)
+                    assert rrf_norm != rrf_raw, \
+                        f"Normalized RRF ({rrf_norm}) should differ from raw ({rrf_raw})"
+
+        print("\n✓ RRF raw vs normalized test passed!")
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_combined_score_matches_components(memory, request_context):
+    """Verify the final score actually equals the weighted sum of components."""
+    bank_id = f"test_combined_{datetime.now(timezone.utc).timestamp()}"
+
+    try:
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="The quick brown fox jumps over the lazy dog",
+            context="test",
+            request_context=request_context,
+        )
+        await memory.retain_async(
+            bank_id=bank_id,
+            content="A quick test of the emergency broadcast system",
+            context="test",
+            request_context=request_context,
+        )
+
+        result = await memory.recall_async(
+            bank_id=bank_id,
+            query="quick test",
+            fact_type=["world"],
+            budget=Budget.LOW,
+            max_tokens=512,
+            enable_trace=True,
+            request_context=request_context,
+        )
+
+        trace = result.trace
+        assert trace is not None
+
+        for r in trace.get("reranked", []):
+            sc = r.get("score_components", {})
+            final_score = r.get("rerank_score", 0)
+
+            # Get components (use defaults if missing)
+            ce = sc.get("cross_encoder_score_normalized", 0)
+            rrf = sc.get("rrf_normalized", 0.5)
+            tmp = sc.get("temporal", 0.5)
+            rec = sc.get("recency", 0.5)
+
+            # Calculate expected score
+            expected = 0.6 * ce + 0.2 * rrf + 0.1 * tmp + 0.1 * rec
+
+            # Allow small floating point difference
+            assert abs(final_score - expected) < 0.01, \
+                f"Final score {final_score} doesn't match expected {expected} from components"
+
+        print("\n✓ Combined score verification test passed!")
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
@@ -5,17 +5,11 @@ Note: Consolidation runs automatically after retain via SyncTaskBackend in tests
 """

 import uuid
-from datetime import datetime, timezone
-from unittest.mock import AsyncMock, call, patch
+from unittest.mock import patch

 import pytest

-from hindsight_api.config import _get_raw_config
-from hindsight_api.engine.consolidation.consolidator import (
-    _aggregate_source_fields,
-    _find_related_observations,
-    run_consolidation_job,
-)
+from hindsight_api.engine.consolidation.consolidator import run_consolidation_job
 from hindsight_api.engine.memory_engine import MemoryEngine
 from hindsight_api.engine.reflect.tools import (
    tool_recall,
@@ -506,7 +500,6 @@ class TestConsolidationIntegration:
            content="Alex loves pizza.",
            request_context=request_context,
        )
-        await memory.wait_for_background_tasks()

        # Check we have one observation
        async with memory._pool.acquire() as conn:
@@ -525,7 +518,6 @@ class TestConsolidationIntegration:
            content="Alex hates pizza.",
            request_context=request_context,
        )
-        await memory.wait_for_background_tasks()

        # Check observations after consolidation
        async with memory._pool.acquire() as conn:
@@ -836,7 +828,6 @@ class TestConsolidationTagRouting:
            content="Pizza is a popular Italian food.",
            request_context=request_context,
        )
-        await memory.wait_for_background_tasks()

        # Check untagged observation exists
        async with memory._pool.acquire() as conn:
@@ -858,7 +849,6 @@ class TestConsolidationTagRouting:
        await self._retain_with_tags(
            memory, bank_id, "Pizza originated in Naples.", ["history"], request_context
        )
-        await memory.wait_for_background_tasks()

        # Check - global observation should be updated OR new scoped observation created
        async with memory._pool.acquire() as conn:
@@ -911,7 +901,6 @@ class TestConsolidationTagRouting:
            "Alice recommends the Thai restaurant on Main Street.",
            ["alice"], request_context
        )
-        await memory.wait_for_background_tasks()

        # Check Alice's observation exists with correct tags
        async with memory._pool.acquire() as conn:
@@ -930,7 +919,6 @@ class TestConsolidationTagRouting:
            "Bob visited the Thai restaurant on Main Street and loved it.",
            ["bob"], request_context
        )
-        await memory.wait_for_background_tasks()

        # Check observations
        async with memory._pool.acquire() as conn:
@@ -943,19 +931,22 @@ class TestConsolidationTagRouting:
                bank_id,
            )

-            # Note: some LLMs may or may not consolidate cross-scope facts.
-            # Just verify structural correctness of any observations that exist.
+            # Should have multiple observations (alice's, bob's, potentially global)
+            assert len(obs_after) >= 2, (
+                f"Expected at least 2 observations for different scopes, got {len(obs_after)}"
+            )

-            # If observations were created, ensure alice and bob are not merged into same observation
-            # (cross-scope merging should not produce an observation with both tags)
-            if obs_after:
-                observations_with_both = [
-                    o for o in obs_after
-                    if o["tags"] and "alice" in o["tags"] and "bob" in o["tags"]
-                ]
-                assert len(observations_with_both) == 0, (
-                    "Should not merge different scopes into one observation with both tags"
-                )
+            # Check we have observations with different tags (alice, bob, or untagged)
+            tag_sets = [frozenset(o["tags"] or []) for o in obs_after]
+
+            # Should NOT merge alice and bob into same observation
+            observations_with_both = [
+                o for o in obs_after
+                if o["tags"] and "alice" in o["tags"] and "bob" in o["tags"]
+            ]
+            assert len(observations_with_both) == 0, (
+                "Should not merge different scopes into one observation with both tags"
+            )

        # Cleanup
        await memory.delete_bank(bank_id, request_context=request_context)
@@ -1032,7 +1023,6 @@ class TestConsolidationTagRouting:
            "Alice works on machine learning projects.",
            ["alice"], request_context
        )
-        await memory.wait_for_background_tasks()

        # Retain untagged memory on same topic
        await memory.retain_async(
@@ -1040,7 +1030,6 @@ class TestConsolidationTagRouting:
            content="Machine learning involves training neural networks.",
            request_context=request_context,
        )
-        await memory.wait_for_background_tasks()

        # Check observations
        async with memory._pool.acquire() as conn:
@@ -1053,10 +1042,11 @@ class TestConsolidationTagRouting:
                bank_id,
            )

+            # Should have at least one observation
+            assert len(observations) >= 1, "Expected at least one observation"
+
            # Either alice's observation was updated OR a global observation was created
-            # This is valid LLM behavior - just verify no errors and structure is correct.
-            # Note: with some LLMs, a single simple fact may not generate an observation,
-            # so we don't assert a minimum count - just verify structural correctness if any exist.
+            # This is valid LLM behavior - just verify no errors and structure is correct
            for obs in observations:
                assert obs["text"], "Observation should have text"

@@ -1441,20 +1431,22 @@ class TestObservationDrillDown:

        assert result["count"] > 0, "Expected at least one observation"

-        # Verify source_fact_ids is present (MemoryFact field name for source memories)
+        # Verify source_memory_ids and proof_count are present
        obs = result["observations"][0]
-        assert "source_fact_ids" in obs, "Observation should have source_fact_ids"
+        assert "source_memory_ids" in obs, "Observation should have source_memory_ids"
+        assert "proof_count" in obs, "Observation should have proof_count"
+        assert obs["proof_count"] >= 1, "proof_count should be at least 1"

-        # If source_fact_ids exist, verify they can be used with expand
-        if obs["source_fact_ids"]:
-            assert len(obs["source_fact_ids"]) >= 1, "Should have at least one source memory"
+        # If source_memory_ids exist, verify they can be used with expand
+        if obs["source_memory_ids"]:
+            assert len(obs["source_memory_ids"]) >= 1, "Should have at least one source memory"

            # Use expand tool to get source memory details
            async with memory._pool.acquire() as conn:
                expand_result = await tool_expand(
                    conn=conn,
                    bank_id=bank_id,
-                    memory_ids=obs["source_fact_ids"][:2],  # Take first 2
+                    memory_ids=obs["source_memory_ids"][:2],  # Take first 2
                    depth="chunk",
                )

@@ -1721,10 +1713,11 @@ class TestHierarchicalRetrieval:
            query="What was the quarterly revenue?",
            request_context=request_context,
            max_tokens=2048,
+            max_results=10,
        )

        # Should have raw facts with specific numbers
-        assert len(recall_result["memories"]) >= 1, "Recall should find the raw facts"
+        assert recall_result["count"] >= 1, "Recall should find the raw facts"

        # Check that we get the actual numbers from the original memories
        all_memory_text = " ".join([m["text"] for m in recall_result["memories"]])
@@ -1937,7 +1930,9 @@ class TestMentalModelRefreshAfterConsolidation:
        )

        # Wait for consolidation to create observations
-        await memory.wait_for_background_tasks()
+        import asyncio
+
+        await asyncio.sleep(2)

        # Get graph data filtered by observation type only
        graph_data = await memory.get_graph_data(
@@ -1955,26 +1950,12 @@ class TestMentalModelRefreshAfterConsolidation:
        for row in graph_data["table_rows"]:
            assert row["fact_type"] == "observation", f"All nodes should be observations, got {row['fact_type']}"

-        # Edges are inherited from source memories when multiple observations exist.
-        # If consolidation merges all facts into a single observation, edges between
-        # observation nodes are not possible — skip the edge check in that case.
-        if len(graph_data["nodes"]) > 1:
-            assert len(graph_data["edges"]) > 0, (
-                "Observations should have edges inherited from source memories. "
-                f"Found {len(graph_data['edges'])} edges among {len(graph_data['nodes'])} nodes"
-            )
-            # Verify edge types are valid
-            valid_link_types = {"semantic", "temporal", "entity"}
-            for edge in graph_data["edges"]:
-                link_type = edge["data"]["linkType"]
-                assert link_type in valid_link_types, f"Invalid link type: {link_type}"
-            # Verify all edges connect visible observation nodes
-            visible_node_ids = {row["id"] for row in graph_data["table_rows"]}
-            for edge in graph_data["edges"]:
-                source_id = edge["data"]["source"]
-                target_id = edge["data"]["target"]
-                assert source_id in visible_node_ids, f"Edge source {source_id[:8]} not in visible nodes"
-                assert target_id in visible_node_ids, f"Edge target {target_id[:8]} not in visible nodes"
+        # Should have edges (inherited from source memories)
+        # Even though we're only showing observations, they should inherit links from their sources
+        assert len(graph_data["edges"]) > 0, (
+            "Observations should have edges inherited from source memories. "
+            f"Found {len(graph_data['edges'])} edges"
+        )

        # Should have entities (inherited from source memories)
        observations_with_entities = [
@@ -1991,510 +1972,19 @@ class TestMentalModelRefreshAfterConsolidation:
            f"Expected to find Alice, Bob, or Google in entities, got: {all_entities}"
        )

+        # Verify edge types are valid
+        valid_link_types = {"semantic", "temporal", "entity"}
+        for edge in graph_data["edges"]:
+            link_type = edge["data"]["linkType"]
+            assert link_type in valid_link_types, f"Invalid link type: {link_type}"
+
+        # Verify all edges connect visible observation nodes
+        visible_node_ids = {row["id"] for row in graph_data["table_rows"]}
+        for edge in graph_data["edges"]:
+            source_id = edge["data"]["source"]
+            target_id = edge["data"]["target"]
+            assert source_id in visible_node_ids, f"Edge source {source_id[:8]} not in visible nodes"
+            assert target_id in visible_node_ids, f"Edge target {target_id[:8]} not in visible nodes"
+
        # Cleanup
        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-def test_consolidation_prompt_default():
-    """Test that the default consolidation prompt contains the built-in mission and processing rules."""
-    from hindsight_api.engine.consolidation.prompts import build_batch_consolidation_prompt
-
-    prompt = build_batch_consolidation_prompt()
-    assert "temporal markers" in prompt
-    assert "RESOLVE REFERENCES" in prompt
-    assert "{facts_text}" in prompt
-    assert "{observations_text}" in prompt
-
-
-def test_consolidation_prompt_observations_mission():
-    """Test that observations_mission replaces the default mission but keeps processing rules."""
-    from hindsight_api.engine.consolidation.prompts import build_batch_consolidation_prompt
-
-    spec = "Observations are weekly summaries of sprint outcomes and team dynamics."
-    prompt = build_batch_consolidation_prompt(observations_mission=spec)
-
-    # Spec is injected
-    assert spec in prompt
-    # Processing rules and output format always remain
-    assert "RESOLVE REFERENCES" in prompt
-    assert "creates" in prompt
-    assert "updates" in prompt
-    assert "{facts_text}" in prompt
-    assert "{observations_text}" in prompt
-
-    # Renders cleanly
-    rendered = prompt.format(facts_text="Alice fixed a bug.", observations_text="[]")
-    assert "{facts_text}" not in rendered
-    assert spec in rendered
-
-
-def test_observations_mission_config():
-    """Test that observations_mission is loaded from env and exposed as configurable."""
-    import os
-
-    from hindsight_api.config import HindsightConfig, _get_raw_config, clear_config_cache
-
-    original = os.getenv("HINDSIGHT_API_OBSERVATIONS_MISSION")
-    try:
-        os.environ["HINDSIGHT_API_OBSERVATIONS_MISSION"] = "Weekly sprint summaries only."
-        clear_config_cache()
-        config = _get_raw_config()
-        assert config.observations_mission == "Weekly sprint summaries only."
-        assert "observations_mission" in HindsightConfig.get_configurable_fields()
-    finally:
-        if original is None:
-            os.environ.pop("HINDSIGHT_API_OBSERVATIONS_MISSION", None)
-        else:
-            os.environ["HINDSIGHT_API_OBSERVATIONS_MISSION"] = original
-        clear_config_cache()
-
-
-@pytest.mark.asyncio
-async def test_consolidation_with_observations_mission(memory: "MemoryEngine", request_context):
-    """Test that observations_mission is used during consolidation without errors."""
-    import os
-
-    from hindsight_api.config import _get_raw_config, clear_config_cache
-
-    original = os.getenv("HINDSIGHT_API_OBSERVATIONS_MISSION")
-    try:
-        os.environ["HINDSIGHT_API_OBSERVATIONS_MISSION"] = (
-            "Observations are summaries of programming language usage patterns."
-        )
-        clear_config_cache()
-        config = _get_raw_config()
-
-        bank_id = f"test-obs-spec-{uuid.uuid4().hex[:8]}"
-        original_global_config = memory._config_resolver._global_config
-        memory._config_resolver._global_config = config
-
-        try:
-            await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
-            await memory.retain_async(
-                bank_id=bank_id,
-                content="Alice uses Python for data analysis and loves its simplicity.",
-                request_context=request_context,
-            )
-            async with memory._pool.acquire() as conn:
-                observations = await conn.fetch(
-                    "SELECT id, text, fact_type FROM memory_units WHERE bank_id = $1 AND fact_type = 'observation'",
-                    bank_id,
-                )
-            assert isinstance(observations, list)
-        finally:
-            memory._config_resolver._global_config = original_global_config
-            await memory.delete_bank(bank_id, request_context=request_context)
-    finally:
-        if original is None:
-            os.environ.pop("HINDSIGHT_API_OBSERVATIONS_MISSION", None)
-        else:
-            os.environ["HINDSIGHT_API_OBSERVATIONS_MISSION"] = original
-        clear_config_cache()
-
-
-
-@pytest.mark.asyncio
-async def test_observation_scopes_explicit_multi_pass(memory: MemoryEngine, request_context):
-    """Test that observation_scopes with an explicit list triggers separate consolidation passes.
-
-    A single memory stored with observation_scopes=[["user:alice"], ["teacher:ben"]]
-    must produce:
-      - At least one observation with tags containing ONLY "user:alice" (not "teacher:ben")
-      - At least one observation with tags containing ONLY "teacher:ben" (not "user:alice")
-
-    The two tag scopes must remain isolated — no observation should carry both tags,
-    which would indicate the scopes were incorrectly merged.
-    """
-    bank_id = f"test-obs-scopes-explicit-{uuid.uuid4().hex[:8]}"
-
-    await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
-
-    # Retain a memory with two explicit observation scopes
-    await memory.retain_batch_async(
-        bank_id=bank_id,
-        contents=[
-            {
-                "content": "Alice, a student, worked hard in the lesson with teacher Ben.",
-                "observation_scopes": [["user:alice"], ["teacher:ben"]],
-            }
-        ],
-        request_context=request_context,
-    )
-
-    async with memory._pool.acquire() as conn:
-        observations = await conn.fetch(
-            """
-            SELECT id, text, tags
-            FROM memory_units
-            WHERE bank_id = $1 AND fact_type = 'observation'
-            ORDER BY created_at
-            """,
-            bank_id,
-        )
-
-    try:
-        # Must have at least 2 observations (one per tag scope)
-        assert len(observations) >= 2, (
-            f"Expected at least 2 observations (one per tag scope), got {len(observations)}: "
-            + str([dict(o) for o in observations])
-        )
-
-        tag_sets = [set(obs["tags"] or []) for obs in observations]
-
-        # There must be at least one observation scoped to user:alice only
-        alice_only = [ts for ts in tag_sets if "user:alice" in ts and "teacher:ben" not in ts]
-        assert alice_only, (
-            f"Expected an observation scoped to 'user:alice' only, got tag sets: {tag_sets}"
-        )
-
-        # There must be at least one observation scoped to teacher:ben only
-        ben_only = [ts for ts in tag_sets if "teacher:ben" in ts and "user:alice" not in ts]
-        assert ben_only, (
-            f"Expected an observation scoped to 'teacher:ben' only, got tag sets: {tag_sets}"
-        )
-
-        # No observation should carry both tags (scopes must not be merged)
-        both = [ts for ts in tag_sets if "user:alice" in ts and "teacher:ben" in ts]
-        assert not both, (
-            f"Found observation(s) with both tags — scopes were incorrectly merged: {both}"
-        )
-    finally:
-        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-@pytest.mark.asyncio
-async def test_observation_scopes_per_tag(memory: MemoryEngine, request_context):
-    """Test that observation_scopes='per_tag' derives one pass per individual tag.
-
-    A memory with tags=["user:alice", "teacher:ben"] and observation_scopes="per_tag"
-    must produce isolated observations — one scoped to "user:alice" and one to "teacher:ben".
-    """
-    bank_id = f"test-obs-scopes-pertag-{uuid.uuid4().hex[:8]}"
-
-    await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
-
-    await memory.retain_batch_async(
-        bank_id=bank_id,
-        contents=[
-            {
-                "content": "Alice, a student, worked hard in the lesson with teacher Ben.",
-                "tags": ["user:alice", "teacher:ben"],
-                "observation_scopes": "per_tag",
-            }
-        ],
-        request_context=request_context,
-    )
-
-    async with memory._pool.acquire() as conn:
-        observations = await conn.fetch(
-            """
-            SELECT id, text, tags
-            FROM memory_units
-            WHERE bank_id = $1 AND fact_type = 'observation'
-            ORDER BY created_at
-            """,
-            bank_id,
-        )
-
-    try:
-        assert len(observations) >= 2, (
-            f"Expected at least 2 observations (one per tag), got {len(observations)}: "
-            + str([dict(o) for o in observations])
-        )
-
-        tag_sets = [set(obs["tags"] or []) for obs in observations]
-
-        alice_only = [ts for ts in tag_sets if "user:alice" in ts and "teacher:ben" not in ts]
-        assert alice_only, f"Expected an observation scoped to 'user:alice' only, got: {tag_sets}"
-
-        ben_only = [ts for ts in tag_sets if "teacher:ben" in ts and "user:alice" not in ts]
-        assert ben_only, f"Expected an observation scoped to 'teacher:ben' only, got: {tag_sets}"
-    finally:
-        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-@pytest.mark.asyncio
-async def test_observation_scopes_combined(memory: MemoryEngine, request_context):
-    """Test that observation_scopes='combined' produces a single observation with all tags.
-
-    A memory with tags=["user:alice", "teacher:ben"] and observation_scopes="combined"
-    must produce at least one observation that carries both tags together, and no
-    observation scoped to only one of them.
-    """
-    bank_id = f"test-obs-scopes-combined-{uuid.uuid4().hex[:8]}"
-
-    await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
-
-    await memory.retain_batch_async(
-        bank_id=bank_id,
-        contents=[
-            {
-                "content": "Alice, a student, worked hard in the lesson with teacher Ben.",
-                "tags": ["user:alice", "teacher:ben"],
-                "observation_scopes": "combined",
-            }
-        ],
-        request_context=request_context,
-    )
-
-    async with memory._pool.acquire() as conn:
-        observations = await conn.fetch(
-            """
-            SELECT id, text, tags
-            FROM memory_units
-            WHERE bank_id = $1 AND fact_type = 'observation'
-            ORDER BY created_at
-            """,
-            bank_id,
-        )
-
-    try:
-        assert len(observations) >= 1, (
-            "Expected at least 1 observation, got 0"
-        )
-
-        tag_sets = [set(obs["tags"] or []) for obs in observations]
-
-        # All observations must carry both tags (combined scope)
-        combined = [ts for ts in tag_sets if "user:alice" in ts and "teacher:ben" in ts]
-        assert combined, f"Expected at least one observation with both tags, got: {tag_sets}"
-
-        # No observation should be scoped to only one tag
-        alice_only = [ts for ts in tag_sets if "user:alice" in ts and "teacher:ben" not in ts]
-        assert not alice_only, f"Expected no alice-only observation in combined mode, got: {tag_sets}"
-
-        ben_only = [ts for ts in tag_sets if "teacher:ben" in ts and "user:alice" not in ts]
-        assert not ben_only, f"Expected no ben-only observation in combined mode, got: {tag_sets}"
-    finally:
-        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-@pytest.mark.asyncio
-async def test_observation_scopes_all_combinations(memory: MemoryEngine, request_context):
-    """Test that observation_scopes='all_combinations' generates passes for every tag subset.
-
-    A memory with tags=["user:alice", "teacher:ben"] and observation_scopes="all_combinations"
-    must produce observations covering all subsets: ["user:alice"], ["teacher:ben"], and
-    ["user:alice", "teacher:ben"].
-    """
-    bank_id = f"test-obs-scopes-allcombos-{uuid.uuid4().hex[:8]}"
-
-    await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
-
-    await memory.retain_batch_async(
-        bank_id=bank_id,
-        contents=[
-            {
-                "content": "Alice, a student, worked hard in the lesson with teacher Ben.",
-                "tags": ["user:alice", "teacher:ben"],
-                "observation_scopes": "all_combinations",
-            }
-        ],
-        request_context=request_context,
-    )
-
-    async with memory._pool.acquire() as conn:
-        observations = await conn.fetch(
-            """
-            SELECT id, text, tags
-            FROM memory_units
-            WHERE bank_id = $1 AND fact_type = 'observation'
-            ORDER BY created_at
-            """,
-            bank_id,
-        )
-
-    try:
-        # With 2 tags there are 3 subsets: {alice}, {ben}, {alice, ben}
-        assert len(observations) >= 3, (
-            f"Expected at least 3 observations (one per subset), got {len(observations)}: "
-            + str([dict(o) for o in observations])
-        )
-
-        tag_sets = [set(obs["tags"] or []) for obs in observations]
-
-        alice_only = [ts for ts in tag_sets if "user:alice" in ts and "teacher:ben" not in ts]
-        assert alice_only, f"Expected an observation scoped to 'user:alice' only, got: {tag_sets}"
-
-        ben_only = [ts for ts in tag_sets if "teacher:ben" in ts and "user:alice" not in ts]
-        assert ben_only, f"Expected an observation scoped to 'teacher:ben' only, got: {tag_sets}"
-
-        combined = [ts for ts in tag_sets if "user:alice" in ts and "teacher:ben" in ts]
-        assert combined, f"Expected an observation scoped to both tags, got: {tag_sets}"
-    finally:
-        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-def _dt(year: int, month: int, day: int) -> datetime:
-    return datetime(year, month, day, tzinfo=timezone.utc)
-
-
-class TestAggregateSourceFields:
-    """Unit tests for _aggregate_source_fields – no database required."""
-
-    def test_all_none_temporal_fields_stay_none(self):
-        """When source memories carry no temporal data, all fields must remain None."""
-        source_mems = [
-            {"tags": ["t1"], "event_date": None, "occurred_start": None, "occurred_end": None, "mentioned_at": None},
-            {"tags": ["t1"], "event_date": None, "occurred_start": None, "occurred_end": None, "mentioned_at": None},
-        ]
-        agg = _aggregate_source_fields(source_mems)
-        assert agg.event_date is None
-        assert agg.occurred_start is None
-        assert agg.occurred_end is None
-        assert agg.mentioned_at is None
-
-    def test_temporal_fields_aggregated_correctly(self):
-        """occurred_start and event_date are minimised; occurred_end and mentioned_at are maximised."""
-        early = _dt(2023, 1, 1)
-        late = _dt(2024, 6, 15)
-        source_mems = [
-            {
-                "tags": [],
-                "event_date": late,
-                "occurred_start": late,
-                "occurred_end": early,
-                "mentioned_at": early,
-            },
-            {
-                "tags": [],
-                "event_date": early,
-                "occurred_start": early,
-                "occurred_end": late,
-                "mentioned_at": late,
-            },
-        ]
-        agg = _aggregate_source_fields(source_mems)
-        assert agg.event_date == early
-        assert agg.occurred_start == early
-        assert agg.occurred_end == late
-        assert agg.mentioned_at == late
-
-    def test_partial_temporal_fields_ignored_when_none(self):
-        """None values in individual sources do not corrupt the min/max from sources that do have dates."""
-        d = _dt(2023, 3, 10)
-        source_mems = [
-            {"tags": [], "event_date": None, "occurred_start": None, "occurred_end": None, "mentioned_at": None},
-            {"tags": [], "event_date": d, "occurred_start": d, "occurred_end": d, "mentioned_at": d},
-        ]
-        agg = _aggregate_source_fields(source_mems)
-        assert agg.event_date == d
-        assert agg.occurred_start == d
-        assert agg.occurred_end == d
-        assert agg.mentioned_at == d
-
-    def test_tags_inherited_from_first_source_memory(self):
-        """Tags default to those of the first source memory (batch invariant)."""
-        source_mems = [
-            {"tags": ["user:alice"], "event_date": None, "occurred_start": None, "occurred_end": None, "mentioned_at": None},
-            {"tags": ["user:alice"], "event_date": None, "occurred_start": None, "occurred_end": None, "mentioned_at": None},
-        ]
-        agg = _aggregate_source_fields(source_mems)
-        assert agg.tags == ["user:alice"]
-
-    def test_tags_override_takes_precedence(self):
-        """Explicit tags parameter overrides the source-memory tags."""
-        source_mems = [
-            {"tags": ["user:alice"], "event_date": None, "occurred_start": None, "occurred_end": None, "mentioned_at": None},
-        ]
-        agg = _aggregate_source_fields(source_mems, tags=["scope:override"])
-        assert agg.tags == ["scope:override"]
-
-    def test_empty_tags_override_is_respected(self):
-        """An explicit empty list override must not fall back to source tags."""
-        source_mems = [
-            {"tags": ["user:alice"], "event_date": None, "occurred_start": None, "occurred_end": None, "mentioned_at": None},
-        ]
-        agg = _aggregate_source_fields(source_mems, tags=[])
-        assert agg.tags == []
-
-    def test_single_source_memory(self):
-        """Single-source aggregation should just pass through that memory's fields."""
-        d = _dt(2024, 11, 5)
-        source_mems = [
-            {"tags": ["x"], "event_date": d, "occurred_start": d, "occurred_end": d, "mentioned_at": d},
-        ]
-        agg = _aggregate_source_fields(source_mems)
-        assert agg.event_date == d
-        assert agg.occurred_start == d
-        assert agg.occurred_end == d
-        assert agg.mentioned_at == d
-        assert agg.tags == ["x"]
-
-
-class TestConsolidationSourceFactsConfig:
-    """Tests that consolidation uses the source_facts token config when calling recall."""
-
-    @pytest.fixture(autouse=True)
-    def enable_observations(self):
-        config = _get_raw_config()
-        original = config.enable_observations
-        config.enable_observations = True
-        yield
-        config.enable_observations = original
-
-    @pytest.mark.asyncio
-    async def test_consolidation_passes_source_facts_max_tokens_to_recall(
-        self, memory: MemoryEngine, request_context
-    ):
-        """consolidation_source_facts_max_tokens from config is forwarded to recall_async."""
-        bank_id = f"test-sf-config-total-{uuid.uuid4().hex[:8]}"
-        await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
-
-        raw = _get_raw_config()
-        fake_config = type(raw)(**{
-            **{f: getattr(raw, f) for f in raw.__dataclass_fields__},
-            "consolidation_source_facts_max_tokens": 999,
-            "consolidation_source_facts_max_tokens_per_observation": -1,
-        })
-
-        try:
-            with (
-                patch.object(memory._config_resolver, "resolve_full_config", return_value=fake_config),
-                patch.object(memory, "recall_async", wraps=memory.recall_async) as mock_recall,
-            ):
-                await _find_related_observations(
-                    memory_engine=memory,
-                    bank_id=bank_id,
-                    query="test query",
-                    request_context=request_context,
-                )
-                assert mock_recall.called
-                _, kwargs = mock_recall.call_args
-                assert kwargs.get("max_source_facts_tokens") == 999
-                assert kwargs.get("max_source_facts_tokens_per_observation") == -1
-        finally:
-            await memory.delete_bank(bank_id, request_context=request_context)
-
-    @pytest.mark.asyncio
-    async def test_consolidation_passes_source_facts_per_obs_tokens_to_recall(
-        self, memory: MemoryEngine, request_context
-    ):
-        """consolidation_source_facts_max_tokens_per_observation from config is forwarded to recall_async."""
-        bank_id = f"test-sf-config-per-obs-{uuid.uuid4().hex[:8]}"
-        await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
-
-        raw = _get_raw_config()
-        fake_config = type(raw)(**{
-            **{f: getattr(raw, f) for f in raw.__dataclass_fields__},
-            "consolidation_source_facts_max_tokens": -1,
-            "consolidation_source_facts_max_tokens_per_observation": 128,
-        })
-
-        try:
-            with (
-                patch.object(memory._config_resolver, "resolve_full_config", return_value=fake_config),
-                patch.object(memory, "recall_async", wraps=memory.recall_async) as mock_recall,
-            ):
-                await _find_related_observations(
-                    memory_engine=memory,
-                    bank_id=bank_id,
-                    query="test query",
-                    request_context=request_context,
-                )
-                assert mock_recall.called
-                _, kwargs = mock_recall.call_args
-                assert kwargs.get("max_source_facts_tokens") == -1
-                assert kwargs.get("max_source_facts_tokens_per_observation") == 128
-        finally:
-            await memory.delete_bank(bank_id, request_context=request_context)
@@ -15,7 +15,7 @@ import pytest
 from sqlalchemy import create_engine, text

 from hindsight_api import MemoryEngine, RequestContext
-from hindsight_api.engine.cross_encoder import CohereCrossEncoder, LocalSTCrossEncoder, ZeroEntropyCrossEncoder
+from hindsight_api.engine.cross_encoder import CohereCrossEncoder, LocalSTCrossEncoder
 from hindsight_api.engine.embeddings import CohereEmbeddings, LocalSTEmbeddings, OpenAIEmbeddings
 from hindsight_api.engine.query_analyzer import DateparserQueryAnalyzer
 from hindsight_api.engine.task_backend import SyncTaskBackend
@@ -75,7 +75,7 @@ def drop_schema(db_url: str, schema_name: str):
        conn.commit()


-def get_column_dimension(db_url: str, schema: str = "public", table: str = "memory_units") -> int | None:
+def get_column_dimension(db_url: str, schema: str = "public") -> int | None:
    """Get the current embedding column dimension from the database."""
    engine = create_engine(db_url)
    with engine.connect() as conn:
@@ -86,10 +86,10 @@ def get_column_dimension(db_url: str, schema: str = "public", table: str = "memo
                JOIN pg_class c ON a.attrelid = c.oid
                JOIN pg_namespace n ON c.relnamespace = n.oid
                WHERE n.nspname = :schema
-                  AND c.relname = :table
+                  AND c.relname = 'memory_units'
                  AND a.attname = 'embedding'
            """),
-            {"schema": schema, "table": table},
+            {"schema": schema},
        ).scalar()
        return result

@@ -98,7 +98,9 @@ def get_row_count(db_url: str, schema: str = "public") -> int:
    """Get the number of rows with embeddings in memory_units."""
    engine = create_engine(db_url)
    with engine.connect() as conn:
-        return conn.execute(text(f"SELECT COUNT(*) FROM {schema}.memory_units WHERE embedding IS NOT NULL")).scalar()
+        return conn.execute(
+            text(f"SELECT COUNT(*) FROM {schema}.memory_units WHERE embedding IS NOT NULL")
+        ).scalar()


 def insert_test_embedding(db_url: str, schema: str, dimension: int):
@@ -125,38 +127,6 @@ def clear_embeddings(db_url: str, schema: str):
        conn.commit()


-def insert_test_mental_model_embedding(db_url: str, schema: str, dimension: int):
-    """Insert a test mental model row with a dummy embedding."""
-    engine = create_engine(db_url)
-    embedding = [0.1] * dimension
-    embedding_str = "[" + ",".join(str(x) for x in embedding) + "]"
-
-    with engine.connect() as conn:
-        # Ensure test bank exists
-        conn.execute(
-            text(f"""
-                INSERT INTO {schema}.banks (bank_id, name)
-                VALUES ('test-bank-mm', 'Test Bank')
-                ON CONFLICT (bank_id) DO NOTHING
-            """)
-        )
-        conn.execute(
-            text(f"""
-                INSERT INTO {schema}.mental_models (bank_id, name, source_query, content, embedding)
-                VALUES ('test-bank-mm', 'test model', 'test query', 'test content', '{embedding_str}'::vector)
-            """)
-        )
-        conn.commit()
-
-
-def clear_mental_model_embeddings(db_url: str, schema: str):
-    """Clear all rows from mental_models."""
-    engine = create_engine(db_url)
-    with engine.connect() as conn:
-        conn.execute(text(f"DELETE FROM {schema}.mental_models"))
-        conn.commit()
-
-
 # =============================================================================
 # Embedding Dimension Tests (Local Embeddings)
 # =============================================================================
@@ -231,49 +201,6 @@ class TestEmbeddingDimension:
        # Cleanup
        clear_embeddings(db_url, schema)

-    def test_mental_models_dimension_matches_no_change(self, dimension_test_schema):
-        """When mental_models dimension matches, no changes should be made."""
-        db_url, schema = dimension_test_schema
-
-        initial_dim = get_column_dimension(db_url, schema, table="mental_models")
-        assert initial_dim == 384, f"Expected 384, got {initial_dim}"
-
-        ensure_embedding_dimension(db_url, 384, schema=schema)
-
-        assert get_column_dimension(db_url, schema, table="mental_models") == 384
-
-    def test_mental_models_dimension_change_empty_table(self, dimension_test_schema):
-        """When mental_models is empty, dimension can be changed."""
-        db_url, schema = dimension_test_schema
-
-        clear_mental_model_embeddings(db_url, schema)
-
-        ensure_embedding_dimension(db_url, 768, schema=schema)
-
-        assert get_column_dimension(db_url, schema, table="mental_models") == 768
-
-        # Change back for other tests
-        ensure_embedding_dimension(db_url, 384, schema=schema)
-        assert get_column_dimension(db_url, schema, table="mental_models") == 384
-
-    def test_mental_models_dimension_change_blocked_with_data(self, dimension_test_schema):
-        """When mental_models has data, dimension change should be blocked."""
-        db_url, schema = dimension_test_schema
-
-        clear_mental_model_embeddings(db_url, schema)
-        insert_test_mental_model_embedding(db_url, schema, 384)
-
-        with pytest.raises(RuntimeError) as exc_info:
-            ensure_embedding_dimension(db_url, 768, schema=schema)
-
-        assert "Cannot change embedding dimension" in str(exc_info.value)
-        assert "mental_models" in str(exc_info.value)
-
-        assert get_column_dimension(db_url, schema, table="mental_models") == 384
-
-        # Cleanup
-        clear_mental_model_embeddings(db_url, schema)
-
    def test_local_embeddings_dimension_detection(self, embeddings):
        """Test that LocalSTEmbeddings correctly detects dimension."""
        # Initialize embeddings if not already done
@@ -683,59 +610,3 @@ class TestCohereIntegration:
                    await memory.close()
            except Exception:
                pass
-
-
-# =============================================================================
-# ZeroEntropy Reranker Tests
-# =============================================================================
-
-
-def has_zeroentropy_api_key() -> bool:
-    """Check if ZeroEntropy API key is available."""
-    return bool(os.environ.get("ZEROENTROPY_API_KEY"))
-
-
-def get_zeroentropy_api_key() -> str:
-    """Get ZeroEntropy API key from environment."""
-    return os.environ.get("ZEROENTROPY_API_KEY", "")
-
-
-@pytest.fixture(scope="module")
-def zeroentropy_cross_encoder():
-    """Create ZeroEntropy cross-encoder instance."""
-    if not has_zeroentropy_api_key():
-        pytest.skip("ZeroEntropy API key not available (set ZEROENTROPY_API_KEY)")
-
-    cross_encoder = ZeroEntropyCrossEncoder(
-        api_key=get_zeroentropy_api_key(),
-        model="zerank-2",
-    )
-    loop = asyncio.new_event_loop()
-    try:
-        loop.run_until_complete(cross_encoder.initialize())
-    finally:
-        loop.close()
-    return cross_encoder
-
-
-class TestZeroEntropyCrossEncoder:
-    """Tests for ZeroEntropy cross-encoder/reranker."""
-
-    def test_zeroentropy_cross_encoder_initialization(self, zeroentropy_cross_encoder):
-        """Test that ZeroEntropy cross-encoder initializes correctly."""
-        assert zeroentropy_cross_encoder.provider_name == "zeroentropy"
-
-    @pytest.mark.asyncio
-    async def test_zeroentropy_cross_encoder_predict(self, zeroentropy_cross_encoder):
-        """Test that ZeroEntropy cross-encoder can score pairs."""
-        pairs = [
-            ("What is the capital of France?", "Paris is the capital of France."),
-            ("What is the capital of France?", "The Eiffel Tower is in Paris."),
-            ("What is the capital of France?", "Python is a programming language."),
-        ]
-        scores = await zeroentropy_cross_encoder.predict(pairs)
-
-        assert len(scores) == 3
-        assert all(isinstance(s, float) for s in scores)
-        # The first result should be most relevant
-        assert scores[0] > scores[2], "Direct answer should score higher than unrelated text"
@@ -2,13 +2,9 @@
 Tests for document tracking and upsert functionality.
 """
 import logging
-from datetime import datetime, timezone
-from unittest.mock import patch
-
 import pytest
-
+from datetime import datetime, timezone
 from hindsight_api import RequestContext
-from hindsight_api.engine.response_models import TokenUsage


@pytest.mark.asyncio
@@ -315,52 +311,3 @@ async def test_document_persisted_with_zero_facts_async_submit(memory, request_c

    finally:
        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-@pytest.mark.asyncio
-async def test_document_stored_without_chunks_when_zero_facts(memory_no_llm_verify, request_context):
-    """
-    Regression test: when 0 facts are extracted from chunked content, the document row
-    must be stored but no chunk rows should be written.
-    """
-    bank_id = f"test_zero_facts_no_chunks_{datetime.now(timezone.utc).timestamp()}"
-    document_id = "doc-zero-facts-chunked"
-
-    # Content large enough to exceed default retain_chunk_size (3000 chars) so chunking is triggered
-    content = "Alice works at Google. " * 200  # ~4600 chars
-
-    async def mock_llm_zero_facts(*args, **kwargs):
-        response = {"facts": []}
-        if kwargs.get("return_usage", False):
-            return response, TokenUsage(input_tokens=10, output_tokens=2)
-        return response
-
-    try:
-        with patch("hindsight_api.engine.llm_wrapper.LLMProvider.call", new=mock_llm_zero_facts):
-            units = await memory_no_llm_verify.retain_async(
-                bank_id=bank_id,
-                content=content,
-                document_id=document_id,
-                request_context=request_context,
-            )
-
-        assert units == [], "Should return no memory units when LLM extracts zero facts"
-
-        # Document row must exist
-        doc = await memory_no_llm_verify.get_document(document_id, bank_id, request_context=request_context)
-        assert doc is not None, "Document row must be stored even when zero facts are extracted"
-        assert doc["id"] == document_id
-        assert doc["memory_unit_count"] == 0
-
-        # No chunk rows should be stored
-        pool = await memory_no_llm_verify._get_pool()
-        async with pool.acquire() as conn:
-            chunk_count = await conn.fetchval(
-                "SELECT COUNT(*) FROM chunks WHERE document_id = $1 AND bank_id = $2",
-                document_id,
-                bank_id,
-            )
-        assert chunk_count == 0, "No chunk rows should be stored when zero facts are extracted"
-
-    finally:
-        await memory_no_llm_verify.delete_bank(bank_id, request_context=request_context)
@@ -535,9 +535,8 @@ class TestOperationHooksParameters:
            request_context=ctx,
        )

-        # Use >= 1 since consolidation may trigger internal recall calls when observations are enabled
-        assert len(validator.pre_recall_calls) >= 1
-        assert len(validator.post_recall_calls) >= 1
+        assert len(validator.pre_recall_calls) == 1
+        assert len(validator.post_recall_calls) == 1


 class TestTenantExtension:
@@ -1,61 +0,0 @@
-"""
-Unit tests for metadata inclusion in fact extraction LLM prompt.
-"""
-from datetime import datetime
-
-from hindsight_api.engine.retain.fact_extraction import _build_user_message
-
-
-def test_build_user_message_includes_metadata():
-    """Metadata key-value pairs should appear in the user message."""
-    event_date = datetime(2024, 6, 15, 12, 0, 0)
-    metadata = {"title": "Q2 Planning Doc", "source": "confluence", "author": "Alice"}
-
-    msg = _build_user_message(
-        chunk="Some content.",
-        chunk_index=0,
-        total_chunks=1,
-        event_date=event_date,
-        context="planning meeting",
-        metadata=metadata,
-    )
-
-    assert "title" in msg
-    assert "Q2 Planning Doc" in msg
-    assert "source" in msg
-    assert "confluence" in msg
-    assert "author" in msg
-    assert "Alice" in msg
-
-
-def test_build_user_message_no_metadata():
-    """When metadata is empty, the message should still be valid and not include a metadata section."""
-    event_date = datetime(2024, 6, 15, 12, 0, 0)
-
-    msg = _build_user_message(
-        chunk="Some content.",
-        chunk_index=0,
-        total_chunks=1,
-        event_date=event_date,
-        context="planning meeting",
-        metadata={},
-    )
-
-    assert "Some content." in msg
-    assert "Metadata:" not in msg
-
-
-def test_build_user_message_without_metadata_arg():
-    """Calling without metadata (default) should behave the same as empty metadata."""
-    event_date = datetime(2024, 6, 15, 12, 0, 0)
-
-    msg = _build_user_message(
-        chunk="Some content.",
-        chunk_index=0,
-        total_chunks=1,
-        event_date=event_date,
-        context="none",
-    )
-
-    assert "Some content." in msg
-    assert "Metadata:" not in msg
@@ -1,141 +0,0 @@
-"""
-Unit tests for fact extraction retry logic.
-
-Tests the fix for the TypeError when LLM returns invalid JSON across all retries.
-Previously, `raise last_error` would raise None (TypeError) because last_error was
-only set in the BadRequestError handler, not when the LLM returned non-dict JSON.
-"""
-
-from datetime import datetime, timezone
-from unittest.mock import AsyncMock, MagicMock, patch
-
-import pytest
-
-
-def _make_config(llm_max_retries: int = 3, retain_llm_max_retries: int | None = None):
-    """Build a minimal HindsightConfig for fact extraction tests."""
-    from hindsight_api.config import HindsightConfig
-
-    cfg = MagicMock(spec=HindsightConfig)
-    cfg.retain_llm_max_retries = retain_llm_max_retries
-    cfg.llm_max_retries = llm_max_retries
-    cfg.retain_llm_initial_backoff = None
-    cfg.llm_initial_backoff = 0.0
-    cfg.retain_llm_max_backoff = None
-    cfg.llm_max_backoff = 0.0
-    cfg.retain_max_completion_tokens = 8192
-    cfg.retain_extraction_mode = "concise"
-    cfg.retain_extract_causal_links = False
-    cfg.retain_mission = None
-    return cfg
-
-
-def _make_llm_config(mock_response):
-    """Build a mock LLMProvider that returns the given response."""
-    from hindsight_api.engine.llm_wrapper import LLMProvider
-
-    llm = MagicMock(spec=LLMProvider)
-    llm.provider = "mock"
-    token_usage = MagicMock()
-    token_usage.__add__ = lambda self, other: self
-    llm.call = AsyncMock(return_value=(mock_response, token_usage))
-    return llm
-
-
-@pytest.mark.asyncio
-async def test_non_dict_json_all_retries_returns_empty():
-    """
-    When LLM returns non-dict JSON on every attempt, extraction should return []
-    without raising TypeError ('exceptions must derive from BaseException').
-
-    This was the bug: the loop ran range(2) times (hardcoded), but comparisons
-    used config.llm_max_retries (default 10). On the last loop iteration (attempt=1),
-    `attempt < 10 - 1` was True, so the code called `continue`, the loop
-    exhausted, and `raise last_error` raised None → TypeError.
-    """
-    from hindsight_api.engine.retain.fact_extraction import _extract_facts_from_chunk
-
-    # llm_max_retries=3 ensures the bug triggers with the old code (3 != 2 hardcoded)
-    config = _make_config(llm_max_retries=3, retain_llm_max_retries=None)
-
-    # Mock: always returns a list (non-dict), which is invalid
-    llm_config = _make_llm_config(mock_response=[{"invalid": "response"}])
-
-    with patch(
-        "hindsight_api.engine.retain.fact_extraction._build_extraction_prompt_and_schema",
-        return_value=("system prompt", MagicMock()),
-    ):
-        facts, usage = await _extract_facts_from_chunk(
-            chunk="Alice visited Paris in 2023.",
-            chunk_index=0,
-            total_chunks=1,
-            event_date=datetime(2023, 1, 1, tzinfo=timezone.utc),
-            context="travel notes",
-            llm_config=llm_config,
-            config=config,
-            agent_name="test-agent",
-        )
-
-    assert facts == []
-
-
-@pytest.mark.asyncio
-async def test_non_dict_json_with_default_max_retries_returns_empty():
-    """
-    Same scenario with the default llm_max_retries=10 (matching real default config).
-    The old code ran range(2) but checked against 10, always continuing until
-    the loop exhausted, then raised None → TypeError.
-    """
-    from hindsight_api.engine.retain.fact_extraction import _extract_facts_from_chunk
-
-    config = _make_config(llm_max_retries=10, retain_llm_max_retries=None)
-    llm_config = _make_llm_config(mock_response="not a dict at all")
-
-    with patch(
-        "hindsight_api.engine.retain.fact_extraction._build_extraction_prompt_and_schema",
-        return_value=("system prompt", MagicMock()),
-    ):
-        facts, usage = await _extract_facts_from_chunk(
-            chunk="Some text.",
-            chunk_index=0,
-            total_chunks=1,
-            event_date=datetime(2023, 6, 1, tzinfo=timezone.utc),
-            context="",
-            llm_config=llm_config,
-            config=config,
-            agent_name="agent",
-        )
-
-    assert facts == []
-
-
-@pytest.mark.asyncio
-async def test_retain_llm_max_retries_overrides_global():
-    """
-    When retain_llm_max_retries is set, it should be used for the loop range
-    and all comparisons (no shadowing bug).
-    """
-    from hindsight_api.engine.retain.fact_extraction import _extract_facts_from_chunk
-
-    # retain_llm_max_retries=5 should override llm_max_retries=10
-    config = _make_config(llm_max_retries=10, retain_llm_max_retries=5)
-    llm_config = _make_llm_config(mock_response=42)  # non-dict: integer
-
-    with patch(
-        "hindsight_api.engine.retain.fact_extraction._build_extraction_prompt_and_schema",
-        return_value=("system prompt", MagicMock()),
-    ):
-        facts, usage = await _extract_facts_from_chunk(
-            chunk="Bob likes Python.",
-            chunk_index=0,
-            total_chunks=1,
-            event_date=datetime(2024, 1, 1, tzinfo=timezone.utc),
-            context="",
-            llm_config=llm_config,
-            config=config,
-            agent_name="agent",
-        )
-
-    assert facts == []
-    # Verify it retried exactly retain_llm_max_retries times
-    assert llm_config.call.call_count == 5
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Nicolò Boschi	79f683cdf9	fix	2026-02-18 13:44:37 +01:00
Nicolò Boschi	ce74b1fc56	feat: add iris as file parser	2026-02-18 12:06:04 +01:00