test: add client tests for ReflectResponse parsing

Added comprehensive tests in hindsight-clients/python/tests to verify: - v0.4.0+ format with empty based_on object - v0.4.0+ format with null based_on - v0.4.0+ format with populated facts - v0.3.0 format (list) correctly fails validation - Missing based_on field handling These tests document the v0.3.0 -> v0.4.0 breaking change where based_on changed from list to object.
fix: remove client imports from API test
2026-02-12 10:06:20 +01:00 · 2026-02-12 10:05:08 +01:00 · 2026-02-11 17:51:09 +01:00
630 changed files with 15915 additions and 84748 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
@@ -31,22 +31,10 @@ HINDSIGHT_API_HOST=0.0.0.0
 HINDSIGHT_API_PORT=8888
 HINDSIGHT_API_LOG_LEVEL=info

-# Base Path / Reverse Proxy Support (Optional)
-# Set these when deploying behind a reverse proxy with path-based routing
-# Example: To deploy at example.com/hindsight/, set both to "/hindsight"
-# HINDSIGHT_API_BASE_PATH=/hindsight
-# NEXT_PUBLIC_BASE_PATH=/hindsight
-
 # Database (Optional - uses embedded pg0 by default)
 # HINDSIGHT_API_DATABASE_URL=postgresql://user:pass@host:5432/db
 # HINDSIGHT_API_DATABASE_SCHEMA=public  # PostgreSQL schema name (default: public)

-# Vector Extension (Optional - uses pgvector by default)
-# Options: "pgvector" (default), "vchord", "pgvectorscale" (DiskANN)
-# HINDSIGHT_API_VECTOR_EXTENSION=pgvector
-# For Azure PostgreSQL with DiskANN:
-# HINDSIGHT_API_VECTOR_EXTENSION=pgvectorscale  # Auto-detects pg_diskann on Azure
-
 # Embeddings Configuration (Optional - uses local by default)
 # Provider: "local" (default) or "tei" (HuggingFace Text Embeddings Inference)
 # HINDSIGHT_API_EMBEDDINGS_PROVIDER=local
@@ -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
@@ -46,10 +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
-
    # 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
@@ -81,12 +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
-
    # Upload artifacts for GitHub release
    - name: Upload artifacts
      uses: actions/upload-artifact@v4
@@ -98,7 +88,6 @@ jobs:
          hindsight/dist/*
          hindsight-integrations/litellm/dist/*
          hindsight-embed/dist/*
-          hindsight-integrations/crewai/dist/*
        retention-days: 1

  release-typescript-client:
@@ -248,55 +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
@@ -328,14 +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
@@ -550,7 +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

@@ -585,12 +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:
@@ -636,8 +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
@@ -689,9 +571,9 @@ jobs:
  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)
@@ -700,12 +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:
@@ -737,46 +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
@@ -793,225 +648,12 @@ jobs:
        echo "=== API Server Logs ==="
        cat /tmp/api-server.log || echo "No API server log found"

-  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_URL: http://localhost:8888
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Setup GCP credentials
-      run: |
-        printf '%s' '${{ 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 Go
-      uses: actions/setup-go@v5
-      with:
-        go-version: '1.23'
-        cache-dependency-path: hindsight-clients/go/go.sum
-
-    - 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: 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_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: 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
-
-    - name: Show API server logs
-      if: always()
-      run: |
-        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
@@ -1019,12 +661,6 @@ 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" >> $GITHUB_ENV
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -1072,22 +708,21 @@ jobs:
      run: |
        cat > .env << EOF
        HINDSIGHT_API_LLM_PROVIDER=${{ 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
@@ -1104,35 +739,6 @@ 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

@@ -1165,21 +771,15 @@ jobs:
  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:
@@ -1215,25 +815,19 @@ 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:
@@ -1270,9 +864,9 @@ jobs:
    runs-on: ubuntu-latest
    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
@@ -1280,12 +874,6 @@ 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: Download CLI artifact
      uses: actions/download-artifact@v4
      with:
@@ -1328,57 +916,55 @@ jobs:
        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
          sleep 1
        done

+    - name: Run Python doc examples
+      working-directory: ./hindsight-clients/python
+      run: |
+        for f in ../../hindsight-docs/examples/api/*.py; do
+          echo "Running $f..."
+          uv run python "$f"
+        done
+
+    - name: Run Node.js doc examples
+      run: |
+        for f in hindsight-docs/examples/api/*.mjs; do
+          echo "Running $f..."
+          node "$f"
+        done
+
    - name: Configure CLI
      run: hindsight configure --api-url http://localhost:8888

-    - name: Run all doc examples
-      run: ./scripts/test-doc-examples.sh
+    - name: Run CLI doc examples
+      run: |
+        for f in hindsight-docs/examples/api/*.sh; do
+          echo "Running $f..."
+          bash "$f"
+        done

    - name: Show API server logs
      if: always()
@@ -1389,9 +975,9 @@ 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

@@ -1400,12 +986,6 @@ 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

@@ -46,7 +46,6 @@ hindsight-docs/static/llms-full.txt
 hindsight-dev/benchmarks/locomo/results/
 hindsight-dev/benchmarks/longmemeval/results/
 hindsight-dev/benchmarks/consolidation/results/
-hindsight-dev/benchmarks/perf/results/
 benchmarks/results/
 hindsight-cli/target
 hindsight-clients/rust/target
@@ -57,15 +57,8 @@ cd hindsight-control-plane && npm run dev

 ### Benchmarks
 ```bash
-# Accuracy benchmarks
 ./scripts/benchmarks/run-longmemeval.sh
 ./scripts/benchmarks/run-locomo.sh
-
-# Performance benchmarks
-./scripts/benchmarks/run-consolidation.sh
-./scripts/benchmarks/run-retain-perf.sh --document <path>  # Requires API server running
-
-# Results viewer
 ./scripts/benchmarks/start-visualizer.sh  # View results at localhost:8001
 ```

@@ -245,61 +238,26 @@ def process(data: UserData) -> str:

 ### Adding New API Configuration Flags

-Configuration follows a hierarchical system: **Global (env vars) → Tenant (via extension) → Bank (database)**.
-
-Fields must be categorized as either **hierarchical** (can be overridden per-tenant/bank) or **static** (server-level only).
-
-#### Adding a New Configuration Field
+When adding a new environment variable configuration:

 1. **config.py** (`hindsight-api/hindsight_api/config.py`):
-   - Add `ENV_*` constant for the environment variable name (e.g., `ENV_MY_SETTING = "HINDSIGHT_API_MY_SETTING"`)
+   - Add `ENV_*` constant for the environment variable name
   - Add `DEFAULT_*` constant for the default value
-   - Add field to `HindsightConfig` dataclass with type annotation
-   - **Mark as hierarchical or static** by adding to `_HIERARCHICAL_FIELDS` set (hierarchical) or leaving it out (static)
+   - Add field to `HindsightConfig` dataclass
   - Add initialization in `from_env()` method

-   ```python
-   # Hierarchical field (can be overridden per-bank)
-   _HIERARCHICAL_FIELDS = {
-       ...,
-       "my_setting",  # Add here for hierarchical
-   }
-
-   # Static field - just don't add to _HIERARCHICAL_FIELDS
-   ```
-
 2. **main.py** (`hindsight-api/hindsight_api/main.py`):
   - Add field to the manual `HindsightConfig()` constructor call (search for "CLI override")

-3. **Use hierarchical config in MemoryEngine**:
-   ```python
-   # Config is resolved automatically per bank via ConfigResolver
-   config_dict = await self._config_resolver.get_bank_config(bank_id, context)
-   value = config_dict["my_setting"]
-   ```
-
-4. **Use static config** (non-hierarchical):
+3. **Use the config** in code:
   ```python
   from ...config import get_config
   config = get_config()
-   value = config.my_static_field
+   value = config.your_new_field
   ```

-5. **Documentation** (`hindsight-docs/docs/developer/configuration.md`):
+4. **Documentation** (`hindsight-docs/docs/developer/configuration.md`):
   - Add to appropriate section table with Variable, Description, Default
-   - Mark if it's hierarchical (can be overridden per-bank)
-
-#### Hierarchical vs Static Guidelines
-
-**Hierarchical** (per-bank overridable):
- LLM settings (provider, model, API key, base URL)
- Operation-specific settings (retain mode, chunk size, etc.)
- Feature flags that vary by customer/bank
-
-**Static** (server-level only):
- Infrastructure settings (database URL, port, host)
- Global limits (max concurrent operations)
- System-wide feature flags

 ## Environment Setup

@@ -317,10 +275,9 @@ 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)
@@ -2,7 +2,7 @@

 ![Hindsight Banner](./hindsight-docs/static/img/hindsight-github-banner.png)

-[Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://ui.hindsight.vectorize.io/signup)
+[Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://vectorize.io/hindsight/cloud)

 [![CI](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml)
 [![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
@@ -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="" />
@@ -1,96 +0,0 @@
-# Nginx Reverse Proxy with Custom Base Path
-
-Deploy Hindsight API under `/hindsight` (or any custom path) using Nginx reverse proxy.
-
-## Quick Start (Published Image - API Only)
-
-```bash
-docker-compose up
-```
-
- **API:** http://localhost:8080/hindsight/docs
- **Control Plane:** http://localhost:9999 (direct access, not proxied)
-
-## Full Stack with Custom Base Path (Requires Build)
-
-**Important:** You cannot rebuild from the published image with build args. You must build from source.
-
-### Build from Source with Custom Base Path
-
-1. **Clone the repository** (if you haven't):
-```bash
-git clone https://github.com/vectorize-io/hindsight.git
-cd hindsight
-```
-
-2. **Build with base path**:
-```bash
-docker build \
-  --build-arg NEXT_PUBLIC_BASE_PATH=/hindsight \
-  -f docker/standalone/Dockerfile \
-  -t hindsight:custom \
-  .
-```
-
-3. **Update docker-compose.yml** to use your built image:
-```yaml
-services:
-  hindsight:
-    image: hindsight:custom  # ← Change this
-    environment:
-      HINDSIGHT_API_BASE_PATH: /hindsight
-      NEXT_PUBLIC_BASE_PATH: /hindsight
-```
-
-4. **Update nginx.conf** to handle Control Plane routes (see below)
-
-5. **Run**:
-```bash
-docker-compose up
-```
-
-### Required nginx.conf for Full Stack
-
-Replace the current `nginx.conf` with this to proxy both API and Control Plane:
-
-```nginx
-events { worker_connections 1024; }
-
-http {
-    include /etc/nginx/mime.types;
-    default_type application/octet-stream;
-
-    upstream hindsight_api { server hindsight:8888; }
-    upstream hindsight_cp { server hindsight:9999; }
-
-    server {
-        listen 80;
-
-        # API
-        location ~ ^/hindsight/(docs|openapi\.json|health|metrics|v1|mcp) {
-            proxy_pass http://hindsight_api;
-            proxy_set_header Host $http_host;
-        }
-
-        # Control Plane static files
-        location ~ ^/hindsight/_next/ {
-            proxy_pass http://hindsight_cp;
-            proxy_set_header Host $http_host;
-        }
-
-        # Control Plane UI
-        location /hindsight {
-            proxy_pass http://hindsight_cp;
-            proxy_set_header Host $http_host;
-        }
-
-        location = / { return 301 /hindsight; }
-    }
-}
-```
-
-### Why Build is Required
-
-Next.js requires `basePath` at **build time**. The published image was built without a custom base path, so you must rebuild from source with the `NEXT_PUBLIC_BASE_PATH` build arg to deploy the Control Plane under a subpath.
-
-The API works without rebuild because `HINDSIGHT_API_BASE_PATH` is a runtime environment variable.
@@ -1,88 +0,0 @@
-# Hindsight API deployment with Nginx reverse proxy (API-only)
-#
-# This example deploys Hindsight API under the path /hindsight with:
-# - Hindsight standalone image (API + Control Plane + embedded pg0)
-# - Nginx reverse proxy (API only)
-#
-# Quick Start:
-#   docker-compose -f docker/docker-compose/nginx/docker-compose.yml up
-#
-# Access:
-#   API (via nginx):         http://localhost:8080/hindsight/docs
-#   Control Plane (direct):  http://localhost:9999
-#
-# For full stack deployment (API + Control Plane both under /hindsight):
-#   See README.md in this directory for instructions on building with basePath.
-#
-# Note: This configuration uses the published image (no build required).
-#       Control Plane is served directly because Next.js basePath requires
-#       build-time configuration. See README.md for the full stack option.
-
-services:
-  # Hindsight (API + Control Plane + embedded pg0)
-  hindsight:
-    image: ghcr.io/vectorize-io/hindsight:latest
-    ports:
-      - "9999:9999"  # Control Plane (direct access, not proxied)
-    environment:
-      # API base path for reverse proxy
-      HINDSIGHT_API_BASE_PATH: /hindsight
-
-      # LLM configuration
-      # Using mock provider for testing (no API key needed)
-      # For production, set OPENAI_API_KEY or ANTHROPIC_API_KEY and use a real provider
-      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-mock}
-      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-not-needed-for-mock}
-      HINDSIGHT_API_LLM_MODEL: ${HINDSIGHT_API_LLM_MODEL:-mock-model}
-
-      # Production examples (uncomment and set appropriate API key):
-      # HINDSIGHT_API_LLM_PROVIDER: openai
-      # HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY}
-      # HINDSIGHT_API_LLM_MODEL: gpt-4o-mini
-
-      # HINDSIGHT_API_LLM_PROVIDER: anthropic
-      # HINDSIGHT_API_LLM_API_KEY: ${ANTHROPIC_API_KEY}
-      # HINDSIGHT_API_LLM_MODEL: claude-sonnet-4-20250514
-
-      # Server config
-      HINDSIGHT_API_HOST: 0.0.0.0
-      HINDSIGHT_API_PORT: 8888
-      HINDSIGHT_API_LOG_LEVEL: info
-
-      # Control Plane config
-      HINDSIGHT_CP_DATAPLANE_API_URL: http://localhost:8888
-    volumes:
-      # Persist embedded pg0 database
-      - hindsight_data:/app/data
-    # Note: Ports not exposed - access via Nginx at localhost:8080/hindsight/
-    # To debug directly, uncomment these ports:
-    # ports:
-    #   - "8888:8888"  # API
-    #   - "9999:9999"  # Control Plane
-    healthcheck:
-      test: ["CMD", "curl", "-f", "http://localhost:8888/hindsight/health"]
-      interval: 10s
-      timeout: 5s
-      retries: 3
-      start_period: 30s
-    networks:
-      - hindsight
-
-  # Nginx reverse proxy
-  nginx:
-    image: nginx:alpine
-    ports:
-      - "8080:80"
-    volumes:
-      - ./nginx.conf:/etc/nginx/nginx.conf:ro
-    depends_on:
-      hindsight:
-        condition: service_healthy
-    networks:
-      - hindsight
-
-volumes:
-  hindsight_data:
-
-networks:
-  hindsight:
@@ -1,40 +0,0 @@
-# Nginx configuration for API-only reverse proxy
-# Control Plane accessed directly (not through nginx)
-
-events {
-    worker_connections 1024;
-}
-
-http {
-    include /etc/nginx/mime.types;
-    default_type application/octet-stream;
-
-    # Logging
-    access_log /var/log/nginx/access.log;
-    error_log /var/log/nginx/error.log;
-
-    # Upstream - Hindsight API
-    upstream hindsight_api {
-        server hindsight:8888;
-    }
-
-    server {
-        listen 80;
-        server_name _;
-
-        # API endpoints - forward with /hindsight prefix
-        location /hindsight/ {
-            proxy_pass http://hindsight_api;
-
-            proxy_set_header Host $http_host;
-            proxy_set_header X-Real-IP $remote_addr;
-            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
-            proxy_set_header X-Forwarded-Proto $scheme;
-        }
-
-        # Redirect root to API docs
-        location = / {
-            return 301 /hindsight/docs;
-        }
-    }
-}
@@ -1,32 +0,0 @@
-# PostgreSQL with pgvector and pg_textsearch extensions
-# Note: pg_textsearch requires PostgreSQL 17+
-FROM postgres:17
-
-# Install build dependencies
-RUN apt-get update && apt-get install -y \
-    build-essential \
-    git \
-    postgresql-server-dev-17 \
-    libpq-dev \
-    && rm -rf /var/lib/apt/lists/*
-
-# Install pgvector
-RUN cd /tmp && \
-    git clone --branch v0.8.0 https://github.com/pgvector/pgvector.git && \
-    cd pgvector && \
-    make && \
-    make install
-
-# Install pg_textsearch
-RUN cd /tmp && \
-    git clone https://github.com/timescale/pg_textsearch.git && \
-    cd pg_textsearch && \
-    make && \
-    make install
-
-# Clean up source files and build dependencies
-RUN rm -rf /tmp/pgvector /tmp/pg_textsearch && \
-    apt-get purge -y --auto-remove build-essential git postgresql-server-dev-17
-
-# Ensure extensions are preloaded
-RUN echo "shared_preload_libraries = 'pg_textsearch'" >> /usr/share/postgresql/postgresql.conf.sample
@@ -1,91 +0,0 @@
-name: hindsight
-# Docker Compose file for Hindsight with PostgreSQL and Timescale pg_textsearch
-# docker compose -f docker/docker-compose/pg_textsearch/docker-compose.yaml down && sleep 2 && docker compose -f docker/docker-compose/pg_textsearch/docker-compose.yaml up -d
-# Make sure to set the required environment variables before running:
-# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
-# - Configure LLM provider variables as needed (see below in the hindsight service)
-#
-# Usage:
-# docker compose up -d
-#
-# Optional environment variables with defaults:
-# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
-# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
-# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
-
-services:
-  db:
-    # Use custom PostgreSQL image with pgvector and pg_textsearch extensions
-    build:
-      context: .
-      dockerfile: Dockerfile
-    container_name: hindsight-db
-    restart: always
-    # Expose PostgreSQL port
-    ports:
-      - "5437:5432"
-    environment:
-      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
-      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:-hindsight_password}
-      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
-    volumes:
-      - pg_data:/var/lib/postgresql/data
-    networks:
-      - hindsight-net
-
-  pg-textsearch-init:
-    build:
-      context: .
-      dockerfile: Dockerfile
-    depends_on:
-      - db
-    environment:
-      - PGPASSWORD=${HINDSIGHT_DB_PASSWORD:-hindsight_password}
-    command: >
-      bash -c "
-        echo 'Waiting for PostgreSQL to be ready...';
-        until pg_isready -h hindsight-db -p 5432 -U hindsight_user; do
-          echo 'PostgreSQL is unavailable - sleeping';
-          sleep 2;
-        done;
-        echo 'PostgreSQL is ready - creating hindsight_db database';
-        psql -h hindsight-db -p 5432 -U hindsight_user -c 'CREATE DATABASE hindsight_db;' 2>/dev/null || echo 'Database already exists';
-        echo 'Creating extensions in hindsight_db database';
-        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vector CASCADE;';
-        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE;';
-        echo 'Database and extensions created successfully';
-      "
-    restart: "no"
-    networks:
-      - hindsight-net
-
-  hindsight:
-    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
-    container_name: hindsight-app
-    ports:
-      - "8888:8888"
-      - "9999:9999"
-    environment:
-      # LLM Configuration
-      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-openai}
-      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-your-api-key}
-
-      # Database Configuration
-      HINDSIGHT_API_DATABASE_URL: postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:-hindsight_password}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
-
-      # Vector and Text Search Extensions
-      HINDSIGHT_API_VECTOR_EXTENSION: pgvector
-      HINDSIGHT_API_TEXT_SEARCH_EXTENSION: pg_textsearch
-
-    depends_on:
-      - db
-    networks:
-      - hindsight-net
-
-
-networks:
-  hindsight-net:
-    driver: bridge
-
-volumes:
-  pg_data:
@@ -1,83 +0,0 @@
-# Docker Compose file for Hindsight with S3 file storage (SeaweedFS)
-#
-# SeaweedFS (Apache 2.0) provides an S3-compatible object storage backend
-# for storing uploaded files instead of PostgreSQL BYTEA storage.
-#
-# Make sure to set the required environment variables before running:
-# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
-# - Configure LLM provider variables as needed (see below in the hindsight service)
-#
-# Usage:
-# docker compose up -d
-#
-# Optional environment variables with defaults:
-# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
-# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
-# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
-# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
-# - SEAWEEDFS_S3_ACCESS_KEY: S3 access key (default: hindsight_s3_key)
-# - SEAWEEDFS_S3_SECRET_KEY: S3 secret key (default: hindsight_s3_secret)
-
-services:
-  db:
-    image: pgvector/pgvector:pg${HINDSIGHT_DB_VERSION:-18}
-    container_name: hindsight-db
-    restart: always
-    environment:
-      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
-      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}
-      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
-    volumes:
-      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
-    networks:
-      - hindsight-net
-
-  seaweedfs:
-    image: chrislusf/seaweedfs:latest
-    container_name: hindsight-seaweedfs
-    restart: always
-    # Single-node mode: master + volume + filer + S3 gateway all in one process
-    command: >
-      server
-      -s3
-      -s3.port=8333
-      -s3.config=/etc/seaweedfs/s3.json
-      -ip.bind=0.0.0.0
-    volumes:
-      - seaweedfs_data:/data
-      - ./s3.json:/etc/seaweedfs/s3.json:ro
-    # Expose S3 API port (uncomment to access from host)
-    # ports:
-    #   - "8333:8333"
-    networks:
-      - hindsight-net
-
-  hindsight:
-    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
-    container_name: hindsight-app
-    ports:
-      - "8888:8888"
-      - "9999:9999"
-    environment:
-      - HINDSIGHT_API_LLM_API_KEY=${OPENAI_API_KEY?Please set the OPENAI_API_KEY env variable}
-      - HINDSIGHT_API_DATABASE_URL=postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
-      # S3 file storage configuration (SeaweedFS)
-      - HINDSIGHT_API_FILE_STORAGE_TYPE=s3
-      - HINDSIGHT_API_FILE_STORAGE_S3_BUCKET=hindsight
-      - HINDSIGHT_API_FILE_STORAGE_S3_ENDPOINT=http://seaweedfs:8333
-      - HINDSIGHT_API_FILE_STORAGE_S3_REGION=us-east-1
-      - HINDSIGHT_API_FILE_STORAGE_S3_ACCESS_KEY_ID=${SEAWEEDFS_S3_ACCESS_KEY:-hindsight_s3_key}
-      - HINDSIGHT_API_FILE_STORAGE_S3_SECRET_ACCESS_KEY=${SEAWEEDFS_S3_SECRET_KEY:-hindsight_s3_secret}
-    depends_on:
-      - db
-      - seaweedfs
-    networks:
-      - hindsight-net
-
-networks:
-  hindsight-net:
-    driver: bridge
-
-volumes:
-  pg_data:
-  seaweedfs_data:
@@ -1,19 +0,0 @@
-{
-  "identities": [
-    {
-      "name": "hindsight",
-      "credentials": [
-        {
-          "accessKey": "hindsight_s3_key",
-          "secretKey": "hindsight_s3_secret"
-        }
-      ],
-      "actions": [
-        "Admin",
-        "Read",
-        "Write",
-        "List"
-      ]
-    }
-  ]
-}
@@ -1,16 +0,0 @@
-# Git
-.git
-.gitignore
-.gitattributes
-
-# Docker
-docker-compose.yaml
-.dockerignore
-
-# Documentation
-README.md
-*.md
-
-# Environment
-.env
-.env.example
@@ -1,25 +0,0 @@
-# PostgreSQL Configuration
-HINDSIGHT_DB_USER=hindsight_user
-HINDSIGHT_DB_PASSWORD=change-me-to-secure-password
-HINDSIGHT_DB_NAME=hindsight_db
-
-# Hindsight Version
-HINDSIGHT_VERSION=latest
-
-# LLM Configuration
-HINDSIGHT_API_LLM_PROVIDER=openai
-OPENAI_API_KEY=your-openai-api-key-here
-
-# Alternative LLM providers (uncomment and configure as needed):
-# HINDSIGHT_API_LLM_PROVIDER=anthropic
-# ANTHROPIC_API_KEY=your-anthropic-api-key
-
-# HINDSIGHT_API_LLM_PROVIDER=gemini
-# GEMINI_API_KEY=your-gemini-api-key
-
-# HINDSIGHT_API_LLM_PROVIDER=groq
-# GROQ_API_KEY=your-groq-api-key
-
-# Vector and Text Search (already configured in docker-compose.yaml)
-# HINDSIGHT_API_VECTOR_EXTENSION=pgvectorscale
-# HINDSIGHT_API_TEXT_SEARCH_EXTENSION=pg_textsearch
@@ -1,55 +0,0 @@
-# PostgreSQL with pgvector, pgvectorscale, and pg_textsearch extensions
-# All three extensions from Timescale/pgvector for high-performance vector and text search
-# Note: Requires PostgreSQL 16+
-FROM postgres:17
-
-# Install build dependencies and Rust toolchain
-RUN apt-get update && apt-get install -y \
-    build-essential \
-    git \
-    postgresql-server-dev-17 \
-    libpq-dev \
-    cmake \
-    curl \
-    pkg-config \
-    libssl-dev \
-    && rm -rf /var/lib/apt/lists/*
-
-# Install Rust toolchain (required for pgvectorscale)
-RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
-ENV PATH="/root/.cargo/bin:${PATH}"
-
-# Install pgvector (required by pgvectorscale)
-RUN cd /tmp && \
-    git clone --branch v0.8.0 https://github.com/pgvector/pgvector.git && \
-    cd pgvector && \
-    make && \
-    make install && \
-    rm -rf /tmp/pgvector
-
-# Install cargo-pgrx (PostgreSQL extension framework for Rust)
-RUN cargo install cargo-pgrx --version 0.12.5 --locked && \
-    cargo pgrx init --pg17 /usr/bin/pg_config
-
-# Install pgvectorscale (DiskANN index support)
-RUN cd /tmp && \
-    git clone --branch 0.5.1 https://github.com/timescale/pgvectorscale.git && \
-    cd pgvectorscale/pgvectorscale && \
-    cargo pgrx install --release && \
-    rm -rf /tmp/pgvectorscale
-
-# Install pg_textsearch (BM25 text search)
-RUN cd /tmp && \
-    git clone https://github.com/timescale/pg_textsearch.git && \
-    cd pg_textsearch && \
-    make && \
-    make install && \
-    rm -rf /tmp/pg_textsearch
-
-# Clean up build dependencies (keep runtime dependencies)
-RUN apt-get purge -y --auto-remove git cmake curl && \
-    rm -rf /root/.cargo/registry /root/.cargo/git
-
-# Ensure extensions are preloaded (pg_textsearch requires preloading)
-RUN echo "shared_preload_libraries = 'pg_textsearch'" >> /usr/share/postgresql/postgresql.conf.sample
-
@@ -1,101 +0,0 @@
-# Hindsight with Timescale Extensions
-
-This Docker Compose setup provides a complete Hindsight deployment with **Timescale extensions**:
- **pgvectorscale** - DiskANN algorithm for disk-based scalable vector search  
- **pg_textsearch** - High-performance BM25 text search
-
-Both extensions are from [Timescale](https://github.com/timescale) and provide production-grade performance.
-
-## Prerequisites
-
- Docker and Docker Compose installed
- OpenAI API key (or another LLM provider)
-
-## Quick Start
-
-```bash
-# Set environment variables
-export HINDSIGHT_DB_PASSWORD="your-secure-password"
-export OPENAI_API_KEY="your-openai-api-key"
-
-# Build and start
-docker compose -f docker/docker-compose/timescale/docker-compose.yaml up -d --build
-
-# Check logs
-
-docker compose -f docker/docker-compose/timescale/docker-compose.yaml logs -f
-```
-
-**Access:**
- API: http://localhost:8888
- Control Plane: http://localhost:9999
-
-## Stop and Clean Up
-
-```bash
-# Stop services
-docker compose -f docker/docker-compose/timescale/docker-compose.yaml down
-
-# Remove volumes (deletes all data)
-docker compose -f docker/docker-compose/timescale/docker-compose.yaml down -v
-```
-
-## Configuration
-
-### Environment Variables
-
-| Variable | Description | Default |
-|----------|-------------|---------|
-| `HINDSIGHT_DB_PASSWORD` | PostgreSQL password | `hindsight_password` |
-| `HINDSIGHT_DB_USER` | PostgreSQL username | `hindsight_user` |
-| `HINDSIGHT_DB_NAME` | Database name | `hindsight_db` |
-| `HINDSIGHT_VERSION` | Hindsight Docker image version | `latest` |
-| `OPENAI_API_KEY` | OpenAI API key | (required) |
-| `HINDSIGHT_API_LLM_PROVIDER` | LLM provider | `openai` |
-
-### Why Timescale Extensions?
-
-**pgvectorscale (DiskANN):**
- 28x lower p95 latency vs dedicated vector databases
- 16x higher query throughput at 99% recall
- 60-75% cost reduction (disk is cheaper than RAM)
- Best for large datasets (10M+ vectors)
-
-**pg_textsearch (BM25):**
- High-performance keyword retrieval
- Native BM25 ranking algorithm
- Optimized for full-text search
-
-## Troubleshooting
-
-### Extensions not installed
-
-Check if extensions are available:
-
-```bash
-docker exec -it hindsight-db-timescale psql -U hindsight_user -d hindsight_db -c "\dx"
-```
-
-You should see:
- `vector` (pgvector)
- `vectorscale` (pgvectorscale/DiskANN)
- `pg_textsearch` (BM25 search)
-
-### Build fails
-
-If the Docker build fails during pgvectorscale compilation:
-
-1. Ensure you have sufficient memory (recommended: 4GB+)
-2. Check Docker build logs for Rust compilation errors
-3. Try building with more resources: `docker compose build --no-cache --memory 4g`
-
-### Port conflicts
-
-If port 5438 is already in use, modify the `ports` section in docker-compose.yaml.
-
-## Learn More
-
- [pgvectorscale GitHub](https://github.com/timescale/pgvectorscale)
- [pg_textsearch GitHub](https://github.com/timescale/pg_textsearch)
- [HNSW vs DiskANN](https://www.tigerdata.com/learn/hnsw-vs-diskann)
- [Hindsight Documentation](https://hindsight.dev)
@@ -1,108 +0,0 @@
-name: hindsight
-# Docker Compose file for Hindsight with Timescale extensions
-# - pgvectorscale: DiskANN vector search (disk-based, scalable)
-# - pg_textsearch: BM25 text search (high-performance keyword retrieval)
-#
-# Quick start:
-# docker compose -f docker/docker-compose/timescale/docker-compose.yaml up -d --build
-#
-# Required environment variables:
-# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
-# - OPENAI_API_KEY (or configure another LLM provider)
-#
-# Optional environment variables with defaults:
-# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
-# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
-# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
-
-services:
-  db:
-    # Custom PostgreSQL image with Timescale extensions (pgvectorscale + pg_textsearch)
-    build:
-      context: .
-      dockerfile: Dockerfile
-    container_name: hindsight-db-timescale
-    restart: always
-    # Expose PostgreSQL port (using 5438 to avoid conflicts with other setups)
-    ports:
-      - "5438:5432"
-    environment:
-      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
-      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:-hindsight_password}
-      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
-    volumes:
-      - pg_data:/var/lib/postgresql/data
-    networks:
-      - hindsight-net
-    # Health check to ensure database is ready
-    healthcheck:
-      test: ["CMD-SHELL", "pg_isready -U hindsight_user"]
-      interval: 5s
-      timeout: 5s
-      retries: 5
-
-  timescale-init:
-    build:
-      context: .
-      dockerfile: Dockerfile
-    depends_on:
-      db:
-        condition: service_healthy
-    environment:
-      - PGPASSWORD=${HINDSIGHT_DB_PASSWORD:-hindsight_password}
-    command: >
-      bash -c "
-        echo 'PostgreSQL is ready - creating hindsight_db database';
-        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -c 'CREATE DATABASE hindsight_db;' 2>/dev/null || echo 'Database already exists';
-        echo 'Installing Timescale extensions...';
-        echo '1/3: Installing pgvector (required by pgvectorscale)...';
-        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vector CASCADE;';
-        echo '2/3: Installing pgvectorscale (DiskANN vector search)...';
-        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vectorscale CASCADE;';
-        echo '3/3: Installing pg_textsearch (BM25 text search)...';
-        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE;';
-        echo '';
-        echo '✅ Timescale extensions installed successfully';
-        echo '';
-        echo 'Installed extensions:';
-        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c \"\\dx\" | grep -E '(vector|vectorscale|pg_textsearch)';
-      "
-    restart: "no"
-    networks:
-      - hindsight-net
-
-  hindsight:
-    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
-    container_name: hindsight-app-timescale
-    ports:
-      - "8888:8888"
-      - "9999:9999"
-    environment:
-      # LLM Configuration
-      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-openai}
-      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-your-api-key}
-
-      # Database Configuration
-      HINDSIGHT_API_DATABASE_URL: postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:-hindsight_password}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
-
-      # Timescale Extensions
-      # pgvectorscale: DiskANN algorithm for disk-based scalable vector search
-      HINDSIGHT_API_VECTOR_EXTENSION: pgvectorscale
-      # pg_textsearch: High-performance BM25 text search
-      HINDSIGHT_API_TEXT_SEARCH_EXTENSION: pg_textsearch
-
-    depends_on:
-      db:
-        condition: service_healthy
-      timescale-init:
-        condition: service_completed_successfully
-    networks:
-      - hindsight-net
-
-
-networks:
-  hindsight-net:
-    driver: bridge
-
-volumes:
-  pg_data:
@@ -1,93 +0,0 @@
-name: hindsight
-# Docker Compose file for Hindsight with PostgreSQL and vectorchord
-# docker compose -f docker/docker-compose/docker-compose.yaml down && sleep 2 && docker compose -f docker/docker-compose/docker-compose.yaml up -d
-# Make sure to set the required environment variables before running:
-# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
-# - Configure LLM provider variables as needed (see below in the hindsight service)
-#
-# Usage:
-# docker compose up -d
-#
-# Optional environment variables with defaults:
-# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
-# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
-# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
-# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
-
-services:
-  db:
-    # Use a PostgreSQL-Image with vectorchord extension pre-installed
-    image: tensorchord/vchord-suite:pg${HINDSIGHT_DB_VERSION:-18-latest}
-    container_name: hindsight-db
-    restart: always
-    # Expose PostgreSQL port
-    ports:
-      - "5436:5432"
-    environment:
-      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
-      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:-hindsight_password}
-      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
-    volumes:
-      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
-    networks:
-      - hindsight-net
-
-  vectorchord-init:
-    image: tensorchord/vchord-suite:pg18-latest
-    #container_name: vectorchord-init
-    depends_on:
-      - db
-    environment:
-      - PGPASSWORD=${HINDSIGHT_DB_PASSWORD:-hindsight_password}
-    command: >
-      bash -c "
-        echo 'Waiting for PostgreSQL to be ready...';
-        until pg_isready -h hindsight-db -p 5432 -U hindsight_user; do
-          echo 'PostgreSQL is unavailable - sleeping';
-          sleep 2;
-        done;
-        echo 'PostgreSQL is ready - creating hindsight_db database';
-        psql -h hindsight-db -p 5432 -U hindsight_user -c 'CREATE DATABASE hindsight_db;' 2>/dev/null || echo 'Database already exists';
-        echo 'Creating extensions in hindsight_db database';
-        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vchord CASCADE;';
-        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS pg_tokenizer CASCADE;';
-        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE;';
-        echo 'Creating llmlingua2 tokenizer';
-        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c \"SELECT create_tokenizer('llmlingua2', \\$\\$ model = \\\"llmlingua2\\\" \\$\\$);\" 2>/dev/null || echo 'Tokenizer already exists or creation skipped';
-        echo 'Database and extensions created successfully';
-      "
-    restart: "no"
-    networks:
-      - hindsight-net
-
-  hindsight:
-    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
-    container_name: hindsight-app
-    ports:
-      - "8888:8888"
-      - "9999:9999"
-    environment:
-      # LLM Configuration (uses OpenAI for testing vchord)
-      # LLM configuration
-      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-openai}
-      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-your-api-key}
-
-      # Database Configuration
-      HINDSIGHT_API_DATABASE_URL: postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:-hindsight_password}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
-
-      # Vector and Text Search Extensions
-      HINDSIGHT_API_VECTOR_EXTENSION: vchord
-      HINDSIGHT_API_TEXT_SEARCH_EXTENSION: vchord
-
-    depends_on:
-      - db
-    networks:
-      - hindsight-net
-
-
-networks:
-  hindsight-net:
-    driver: bridge
-
-volumes:
-  pg_data:
@@ -112,10 +112,6 @@ RUN rm -f package-lock.json && sed -i '/"@vectorize-io\/hindsight-client":/d' pa
 # Copy built SDK directly into node_modules (more reliable than npm link in Docker)
 COPY --from=sdk-builder /app/hindsight-clients/typescript ./node_modules/@vectorize-io/hindsight-client

-# Accept base path as build argument for reverse proxy deployments
-# Usage: docker build --build-arg NEXT_PUBLIC_BASE_PATH=/hindsight ...
-ARG NEXT_PUBLIC_BASE_PATH=""
-
 # Build Control Plane - run next build first, then custom standalone copy
 # (The build:standalone script expects a specific path structure that differs in Docker)
 RUN npm exec -- next build
@@ -170,11 +166,6 @@ 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 +317,6 @@ 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)
@@ -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: 0.4.14
-appVersion: "0.4.14"
+version: 0.4.10
+appVersion: "0.4.10"
 keywords:
  - ai
  - memory
@@ -46,4 +46,4 @@ __all__ = [
    "RemoteTEICrossEncoder",
    "LLMConfig",
 ]
-__version__ = "0.4.14"
+__version__ = "0.4.10"
@@ -6,7 +6,6 @@ Create Date: 2025-11-27 11:54:19.228030

 """

-import os
 from collections.abc import Sequence

 import sqlalchemy as sa
@@ -22,96 +21,6 @@ branch_labels: str | Sequence[str] | None = None
 depends_on: str | Sequence[str] | None = None


-def _detect_vector_extension() -> str:
-    """
-    Detect or validate vector extension: 'pgvector', 'vchord', or 'pgvectorscale'.
-    Respects HINDSIGHT_API_VECTOR_EXTENSION env var if set.
-    """
-    conn = op.get_bind()
-    vector_extension = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
-
-    # Validate configured extension is installed
-    if vector_extension == "pgvectorscale":
-        # pgvectorscale/DiskANN requires pgvector
-        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
-        if not pgvector_check:
-            raise RuntimeError(
-                "DiskANN requires pgvector. Install with: CREATE EXTENSION vector; then vectorscale or pg_diskann CASCADE;"
-            )
-        # Check for either vectorscale (open source) or pg_diskann (Azure)
-        vectorscale_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vectorscale'")).scalar()
-        pg_diskann_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_diskann'")).scalar()
-
-        if vectorscale_check:
-            return "pgvectorscale"
-        elif pg_diskann_check:
-            return "pg_diskann"
-        else:
-            raise RuntimeError(
-                "Configured vector extension 'pgvectorscale' not found. Install either:\n"
-                "  - pgvectorscale: CREATE EXTENSION vectorscale CASCADE;\n"
-                "  - pg_diskann (Azure): CREATE EXTENSION pg_diskann CASCADE;"
-            )
-    elif vector_extension == "vchord":
-        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
-        if not vchord_check:
-            raise RuntimeError(
-                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
-            )
-        return "vchord"
-    elif vector_extension == "pgvector":
-        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
-        if not pgvector_check:
-            raise RuntimeError(
-                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
-            )
-        return "pgvector"
-    else:
-        raise ValueError(
-            f"Invalid HINDSIGHT_API_VECTOR_EXTENSION: {vector_extension}. Must be 'pgvector', 'vchord', or 'pgvectorscale'"
-        )
-
-
-def _detect_text_search_extension() -> str:
-    """
-    Detect or validate text search extension: 'native', 'vchord', or 'pg_textsearch'.
-    Respects HINDSIGHT_API_TEXT_SEARCH_EXTENSION env var.
-    Creates the extension if needed.
-    """
-    text_search_extension = os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
-
-    if text_search_extension == "vchord":
-        # Create vchord_bm25 extension if not exists
-        try:
-            op.execute("CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE")
-        except Exception:
-            # Extension might already exist or user lacks permissions - verify it exists
-            conn = op.get_bind()
-            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord_bm25'")).fetchone()
-            if not result:
-                # Extension truly doesn't exist - re-raise the error
-                raise
-        return "vchord"
-    elif text_search_extension == "pg_textsearch":
-        # Create pg_textsearch extension if not exists
-        try:
-            op.execute("CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE")
-        except Exception:
-            # Extension might already exist or user lacks permissions - verify it exists
-            conn = op.get_bind()
-            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_textsearch'")).fetchone()
-            if not result:
-                # Extension truly doesn't exist - re-raise the error
-                raise
-        return "pg_textsearch"
-    elif text_search_extension == "native":
-        return "native"
-    else:
-        raise ValueError(
-            f"Invalid HINDSIGHT_API_TEXT_SEARCH_EXTENSION: {text_search_extension}. Must be 'native', 'vchord', or 'pg_textsearch'"
-        )
-
-
 def upgrade() -> None:
    """Upgrade schema - create all tables from scratch."""

@@ -257,29 +166,11 @@ def upgrade() -> None:
    )

    # Add search_vector column for full-text search
-    # Type depends on configured text search backend
-    text_search_ext = _detect_text_search_extension()
-
-    if text_search_ext == "vchord":
-        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT)
-        # Note: vchord_bm25 extension creates types in bm25_catalog schema
-        op.execute("""
-            ALTER TABLE memory_units
-            ADD COLUMN search_vector bm25_catalog.bm25vector
-        """)
-    elif text_search_ext == "pg_textsearch":
-        # Timescale pg_textsearch: dummy TEXT column for consistency (indexes operate on base columns directly)
-        op.execute("""
-            ALTER TABLE memory_units
-            ADD COLUMN search_vector TEXT
-        """)
-    else:  # native
-        # Native PostgreSQL: tsvector with automatic generation
-        op.execute("""
-            ALTER TABLE memory_units
-            ADD COLUMN search_vector tsvector
-            GENERATED ALWAYS AS (to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))) STORED
-        """)
+    op.execute("""
+        ALTER TABLE memory_units
+        ADD COLUMN search_vector tsvector
+        GENERATED ALWAYS AS (to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))) STORED
+    """)

    op.create_index("idx_memory_units_bank_id", "memory_units", ["bank_id"])
    op.create_index("idx_memory_units_document_id", "memory_units", ["document_id"])
@@ -309,61 +200,19 @@ def upgrade() -> None:
        ["bank_id", sa.text("event_date DESC")],
        postgresql_where=sa.text("fact_type = 'observation'"),
    )
-    # Create vector index - conditional based on available extension
-    vector_ext = _detect_vector_extension()
+    op.create_index(
+        "idx_memory_units_embedding",
+        "memory_units",
+        ["embedding"],
+        postgresql_using="hnsw",
+        postgresql_ops={"embedding": "vector_cosine_ops"},
+    )

-    if vector_ext == "pgvectorscale":
-        # Use DiskANN index for pgvectorscale (disk-based, scalable)
-        op.execute("""
-            CREATE INDEX idx_memory_units_embedding ON memory_units
-            USING diskann (embedding vector_cosine_ops)
-            WITH (num_neighbors = 50)
-        """)
-    elif vector_ext == "pg_diskann":
-        # Use DiskANN index for pg_diskann (Azure)
-        op.execute("""
-            CREATE INDEX idx_memory_units_embedding ON memory_units
-            USING diskann (embedding vector_cosine_ops)
-            WITH (max_neighbors = 50)
-        """)
-    elif vector_ext == "vchord":
-        # Use vchordrq index for vchord (supports high-dimensional embeddings)
-        op.execute("""
-            CREATE INDEX idx_memory_units_embedding ON memory_units
-            USING vchordrq (embedding vector_l2_ops)
-        """)
-    else:  # pgvector
-        # Use HNSW index for pgvector
-        op.create_index(
-            "idx_memory_units_embedding",
-            "memory_units",
-            ["embedding"],
-            postgresql_using="hnsw",
-            postgresql_ops={"embedding": "vector_cosine_ops"},
-        )
-
-    # Create full-text search index on search_vector
-    # Index type depends on text search backend
-    if text_search_ext == "vchord":
-        # VectorChord BM25 index
-        op.execute("""
-            CREATE INDEX idx_memory_units_text_search ON memory_units
-            USING bm25 (search_vector bm25_catalog.bm25_ops)
-        """)
-    elif text_search_ext == "pg_textsearch":
-        # Timescale pg_textsearch BM25 index on text column
-        # Note: pg_textsearch doesn't support expressions, so we index the main text column
-        op.execute("""
-            CREATE INDEX idx_memory_units_text_search ON memory_units
-            USING bm25(text)
-            WITH (text_config='english')
-        """)
-    else:  # native
-        # Native PostgreSQL GIN index
-        op.execute("""
-            CREATE INDEX idx_memory_units_text_search ON memory_units
-            USING gin(search_vector)
-        """)
+    # Create BM25 full-text search index on search_vector
+    op.execute("""
+        CREATE INDEX idx_memory_units_text_search ON memory_units
+        USING gin(search_vector)
+    """)

    op.execute("""
        CREATE MATERIALIZED VIEW memory_units_bm25 AS
@@ -1,70 +0,0 @@
-"""Add file_storage table for BYTEA-based file storage
-
-Revision ID: a1b2c3d4e5f6
-Revises: y0t1u2v3w4x5
-Create Date: 2026-02-16
-
-Creates a dedicated table for storing uploaded files using BYTEA.
-This provides zero-config file storage that "just works" for development
-and small deployments. For production/scale, use S3-compatible storage.
-
-Files are stored in a separate table to avoid bloating the documents table.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "a1b2c3d4e5f6"
-down_revision: str | Sequence[str] | None = "y0t1u2v3w4x5"
-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:
-    """Create file_storage table for BYTEA storage."""
-    schema = _get_schema_prefix()
-
-    # Create file_storage table (minimal: just key + data)
-    op.execute(
-        f"""
-        CREATE TABLE {schema}file_storage (
-            storage_key TEXT PRIMARY KEY,
-            data BYTEA NOT NULL
-        )
-    """
-    )
-
-    # Add file tracking columns to documents table
-    op.execute(
-        f"""
-        ALTER TABLE {schema}documents
-        ADD COLUMN IF NOT EXISTS file_storage_key TEXT,
-        ADD COLUMN IF NOT EXISTS file_original_name TEXT,
-        ADD COLUMN IF NOT EXISTS file_content_type TEXT
-    """
-    )
-
-
-def downgrade() -> None:
-    """Remove file_storage table and related columns."""
-    schema = _get_schema_prefix()
-
-    # Drop columns from documents table
-    op.execute(
-        f"""
-        ALTER TABLE {schema}documents
-        DROP COLUMN IF EXISTS file_storage_key,
-        DROP COLUMN IF EXISTS file_original_name,
-        DROP COLUMN IF EXISTS file_content_type
-    """
-    )
-
-    # Drop file_storage table
-    op.execute(f"DROP TABLE IF EXISTS {schema}file_storage")
@@ -10,11 +10,9 @@ This migration:
 3. Adds consolidation tracking columns to the 'banks' table
 """

-import os
 from collections.abc import Sequence

 from alembic import context, op
-from sqlalchemy import text

 # revision identifiers, used by Alembic.
 revision: str = "n9i0j1k2l3m4"
@@ -29,106 +27,10 @@ def _get_schema_prefix() -> str:
    return f'"{schema}".' if schema else ""


-def _detect_vector_extension() -> str:
-    """
-    Detect or validate vector extension: 'pgvector', 'vchord', or 'pgvectorscale'.
-    Respects HINDSIGHT_API_VECTOR_EXTENSION env var if set.
-    """
-    conn = op.get_bind()
-    vector_extension = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
-
-    # Validate configured extension is installed
-    if vector_extension == "pgvectorscale":
-        # pgvectorscale/DiskANN requires pgvector
-        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
-        if not pgvector_check:
-            raise RuntimeError(
-                "DiskANN requires pgvector. Install with: CREATE EXTENSION vector; then vectorscale or pg_diskann CASCADE;"
-            )
-        # Check for either vectorscale (open source) or pg_diskann (Azure)
-        vectorscale_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vectorscale'")).scalar()
-        pg_diskann_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_diskann'")).scalar()
-
-        if vectorscale_check:
-            return "pgvectorscale"
-        elif pg_diskann_check:
-            return "pg_diskann"
-        else:
-            raise RuntimeError(
-                "Configured vector extension 'pgvectorscale' not found. Install either:\n"
-                "  - pgvectorscale: CREATE EXTENSION vectorscale CASCADE;\n"
-                "  - pg_diskann (Azure): CREATE EXTENSION pg_diskann CASCADE;"
-            )
-    elif vector_extension == "vchord":
-        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
-        if not vchord_check:
-            raise RuntimeError(
-                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
-            )
-        return "vchord"
-    elif vector_extension == "pgvector":
-        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
-        if not pgvector_check:
-            raise RuntimeError(
-                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
-            )
-        return "pgvector"
-    else:
-        raise ValueError(
-            f"Invalid HINDSIGHT_API_VECTOR_EXTENSION: {vector_extension}. Must be 'pgvector', 'vchord', or 'pgvectorscale'"
-        )
-
-
-def _detect_text_search_extension() -> str:
-    """
-    Detect or validate text search extension: 'native', 'vchord', or 'pg_textsearch'.
-    Respects HINDSIGHT_API_TEXT_SEARCH_EXTENSION env var.
-    Creates the extension if needed.
-    """
-    text_search_extension = os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
-
-    if text_search_extension == "vchord":
-        # Create vchord_bm25 extension if not exists
-        try:
-            op.execute("CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE")
-        except Exception:
-            # Extension might already exist or user lacks permissions - verify it exists
-            conn = op.get_bind()
-            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord_bm25'")).fetchone()
-            if not result:
-                # Extension truly doesn't exist - re-raise the error
-                raise
-        return "vchord"
-    elif text_search_extension == "pg_textsearch":
-        # Create pg_textsearch extension if not exists
-        try:
-            op.execute("CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE")
-        except Exception:
-            # Extension might already exist or user lacks permissions - verify it exists
-            conn = op.get_bind()
-            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_textsearch'")).fetchone()
-            if not result:
-                # Extension truly doesn't exist - re-raise the error
-                raise
-        return "pg_textsearch"
-    elif text_search_extension == "native":
-        return "native"
-    else:
-        raise ValueError(
-            f"Invalid HINDSIGHT_API_TEXT_SEARCH_EXTENSION: {text_search_extension}. Must be 'native', 'vchord', or 'pg_textsearch'"
-        )
-
-
 def upgrade() -> None:
    """Create learnings and pinned_reflections tables."""
    schema = _get_schema_prefix()

-    # Detect which vector extension is available
-    vector_ext = _detect_vector_extension()
-
-    # Detect which text search extension to use
-    text_search_ext = _detect_text_search_extension()
-
    # 1. Create learnings table
    op.execute(f"""
        CREATE TABLE {schema}learnings (
@@ -155,60 +57,18 @@ def upgrade() -> None:

    # Indexes for learnings
    op.execute(f"CREATE INDEX idx_learnings_bank_id ON {schema}learnings(bank_id)")
-
-    # Create vector index based on detected extension
-    if vector_ext == "pgvectorscale":
-        op.execute(f"""
-            CREATE INDEX idx_learnings_embedding ON {schema}learnings
-            USING diskann (embedding vector_cosine_ops)
-            WITH (num_neighbors = 50)
-        """)
-    elif vector_ext == "pg_diskann":
-        op.execute(f"""
-            CREATE INDEX idx_learnings_embedding ON {schema}learnings
-            USING diskann (embedding vector_cosine_ops)
-            WITH (max_neighbors = 50)
-        """)
-    elif vector_ext == "vchord":
-        op.execute(f"""
-            CREATE INDEX idx_learnings_embedding ON {schema}learnings
-            USING vchordrq (embedding vector_l2_ops)
-        """)
-    else:  # pgvector
-        op.execute(f"""
-            CREATE INDEX idx_learnings_embedding ON {schema}learnings
-            USING hnsw (embedding vector_cosine_ops)
-        """)
-
+    op.execute(f"""
+        CREATE INDEX idx_learnings_embedding ON {schema}learnings
+        USING hnsw (embedding vector_cosine_ops)
+    """)
    op.execute(f"CREATE INDEX idx_learnings_tags ON {schema}learnings USING GIN(tags)")

    # Full-text search for learnings
-    if text_search_ext == "vchord":
-        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT)
-        # Note: vchord_bm25 extension creates types in bm25_catalog schema
-        op.execute(f"""
-            ALTER TABLE {schema}learnings ADD COLUMN search_vector bm25_catalog.bm25vector
-        """)
-        op.execute(f"""
-            CREATE INDEX idx_learnings_text_search ON {schema}learnings
-            USING bm25 (search_vector bm25_catalog.bm25_ops)
-        """)
-    elif text_search_ext == "pg_textsearch":
-        # Timescale pg_textsearch: dummy TEXT column for consistency (indexes operate on base columns directly)
-        op.execute(f"""
-            ALTER TABLE {schema}learnings ADD COLUMN search_vector TEXT
-        """)
-        op.execute(f"""
-            CREATE INDEX idx_learnings_text_search ON {schema}learnings
-            USING bm25(text) WITH (text_config='english')
-        """)
-    else:  # native
-        # Native PostgreSQL: tsvector with automatic generation
-        op.execute(f"""
-            ALTER TABLE {schema}learnings ADD COLUMN search_vector tsvector
-            GENERATED ALWAYS AS (to_tsvector('english', text)) STORED
-        """)
-        op.execute(f"CREATE INDEX idx_learnings_text_search ON {schema}learnings USING gin(search_vector)")
+    op.execute(f"""
+        ALTER TABLE {schema}learnings ADD COLUMN search_vector tsvector
+        GENERATED ALWAYS AS (to_tsvector('english', text)) STORED
+    """)
+    op.execute(f"CREATE INDEX idx_learnings_text_search ON {schema}learnings USING gin(search_vector)")

    # 2. Create pinned_reflections table
    op.execute(f"""
@@ -234,64 +94,21 @@ def upgrade() -> None:

    # Indexes for pinned_reflections
    op.execute(f"CREATE INDEX idx_pinned_reflections_bank_id ON {schema}pinned_reflections(bank_id)")
-
-    # Create vector index based on detected extension
-    if vector_ext == "pgvectorscale":
-        op.execute(f"""
-            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
-            USING diskann (embedding vector_cosine_ops)
-            WITH (num_neighbors = 50)
-        """)
-    elif vector_ext == "pg_diskann":
-        op.execute(f"""
-            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
-            USING diskann (embedding vector_cosine_ops)
-            WITH (max_neighbors = 50)
-        """)
-    elif vector_ext == "vchord":
-        op.execute(f"""
-            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
-            USING vchordrq (embedding vector_l2_ops)
-        """)
-    else:  # pgvector
-        op.execute(f"""
-            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
-            USING hnsw (embedding vector_cosine_ops)
-        """)
-
+    op.execute(f"""
+        CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
+        USING hnsw (embedding vector_cosine_ops)
+    """)
    op.execute(f"CREATE INDEX idx_pinned_reflections_tags ON {schema}pinned_reflections USING GIN(tags)")

    # Full-text search for pinned_reflections
-    if text_search_ext == "vchord":
-        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT/UPDATE)
-        # Note: vchord_bm25 extension creates types in bm25_catalog schema
-        op.execute(f"""
-            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector bm25_catalog.bm25vector
-        """)
-        op.execute(f"""
-            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
-            USING bm25 (search_vector bm25_catalog.bm25_ops)
-        """)
-    elif text_search_ext == "pg_textsearch":
-        # Timescale pg_textsearch: dummy TEXT column for consistency (indexes operate on base columns directly)
-        op.execute(f"""
-            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector TEXT
-        """)
-        op.execute(f"""
-            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
-            USING bm25(content)
-            WITH (text_config='english')
-        """)
-    else:  # native
-        # Native PostgreSQL: tsvector with automatic generation
-        op.execute(f"""
-            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector tsvector
-            GENERATED ALWAYS AS (to_tsvector('english', COALESCE(name, '') || ' ' || content)) STORED
-        """)
-        op.execute(f"""
-            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
-            USING gin(search_vector)
-        """)
+    op.execute(f"""
+        ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector tsvector
+        GENERATED ALWAYS AS (to_tsvector('english', COALESCE(name, '') || ' ' || content)) STORED
+    """)
+    op.execute(f"""
+        CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
+        USING gin(search_vector)
+    """)

    # 3. Add consolidation tracking columns to banks table
    op.execute(f"""
@@ -1,64 +0,0 @@
-"""Add config JSONB column to banks table for hierarchical configuration
-
-Revision ID: x9s0t1u2v3w4
-Revises: w8r9s0t1u2v3
-Create Date: 2026-02-09
-
-This migration adds a `config` JSONB column to the banks table to support
-per-bank configuration overrides. This enables hierarchical configuration where:
- Global config is loaded from environment variables
- Tenant config is provided via TenantExtension
- Bank config overrides are stored in banks.config JSONB column
-
-The config column stores overrides for hierarchical fields (LLM settings,
-retention parameters, retrieval settings, etc.) in Python field name format.
-"""
-
-from collections.abc import Sequence
-
-import sqlalchemy as sa
-from alembic import context, op
-from sqlalchemy.dialects.postgresql import JSONB
-
-revision: str = "x9s0t1u2v3w4"
-down_revision: str | Sequence[str] | None = "w8r9s0t1u2v3"
-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:
-    """Add config JSONB column to banks table with GIN index."""
-    schema = _get_schema_prefix()
-
-    # Add config column to banks table
-    op.execute(f"""
-        ALTER TABLE {schema}banks
-        ADD COLUMN config JSONB NOT NULL DEFAULT '{{}}'::jsonb
-    """)
-
-    # Add GIN index for efficient JSONB queries
-    op.execute(f"""
-        CREATE INDEX idx_banks_config
-        ON {schema}banks
-        USING gin(config)
-    """)
-
-
-def downgrade() -> None:
-    """Remove config column and index from banks table."""
-    schema = _get_schema_prefix()
-
-    # Drop index first
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_banks_config")
-
-    # Drop column
-    op.execute(f"""
-        ALTER TABLE {schema}banks
-        DROP COLUMN IF EXISTS config
-    """)
@@ -1,49 +0,0 @@
-"""Add GIN index on async_operations.result_metadata for parent_operation_id queries
-
-Revision ID: y0t1u2v3w4x5
-Revises: x9s0t1u2v3w4
-Create Date: 2026-02-13
-
-This migration adds a GIN index on the result_metadata JSONB column in the
-async_operations table to support efficient queries for child operations by
-parent_operation_id.
-
-The index enables fast lookups when querying for child operations:
-  SELECT * FROM async_operations
-  WHERE result_metadata::jsonb @> '{"parent_operation_id": "uuid"}'::jsonb
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "y0t1u2v3w4x5"
-down_revision: str | Sequence[str] | None = "x9s0t1u2v3w4"
-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:
-    """Add GIN index on result_metadata for efficient parent_operation_id queries."""
-    schema = _get_schema_prefix()
-
-    # Add GIN index for JSONB containment queries (@> operator)
-    op.execute(f"""
-        CREATE INDEX idx_async_operations_result_metadata
-        ON {schema}async_operations
-        USING gin(result_metadata)
-    """)
-
-
-def downgrade() -> None:
-    """Remove GIN index on result_metadata."""
-    schema = _get_schema_prefix()
-
-    # Drop index
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_result_metadata")
@@ -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)
@@ -176,39 +114,9 @@ def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
        logger.info(f"Loading MCP extension: {mcp_extension.__class__.__name__}")
        mcp_extension.register_tools(mcp, memory)

-    # Make all tools tolerant of extra arguments from LLMs (e.g., "explanation")
-    _make_tools_tolerant(mcp)
-
    return mcp


-def _make_tools_tolerant(mcp: FastMCP) -> None:
-    """Wrap all tool run methods to strip unknown arguments before validation.
-
-    LLMs frequently add extra fields like "explanation" or "reasoning" to tool calls.
-    FastMCP's Pydantic TypeAdapter rejects these with "Unexpected keyword argument".
-    This wraps each tool's run() to filter arguments to only known parameters.
-    """
-    try:
-        for name, tool in mcp._tool_manager._tools.items():
-            if hasattr(tool, "parameters") and tool.parameters:
-                allowed = set(tool.parameters.get("properties", {}).keys())
-                original_run = tool.run
-
-                async def _tolerant_run(arguments, _allowed=allowed, _orig=original_run):
-                    extra_keys = set(arguments.keys()) - _allowed
-                    if extra_keys:
-                        logger.debug(f"Stripping unknown arguments from tool call: {extra_keys}")
-                        arguments = {k: v for k, v in arguments.items() if k in _allowed}
-                    return await _orig(arguments)
-
-                # FunctionTool is a Pydantic model with extra='forbid', so use
-                # object.__setattr__ to bypass Pydantic's setter validation.
-                object.__setattr__(tool, "run", _tolerant_run)
-    except (AttributeError, KeyError) as e:
-        logger.warning(f"Could not make tools tolerant of extra arguments: {e}")
-
-
 class MCPMiddleware:
    """ASGI middleware that intercepts MCP requests and routes to appropriate MCP server.

@@ -234,11 +142,6 @@ class MCPMiddleware:
       - No bank management tools (list_banks, create_bank)
       - Recommended for agent isolation

-    Bank ID resolution priority:
-        1. URL path (e.g., /mcp/{bank_id}/) → single-bank mode
-        2. X-Bank-Id header → multi-bank mode
-        3. HINDSIGHT_MCP_BANK_ID env var → multi-bank mode (default: "default")
-
    Examples:
        # Single-bank mode (recommended for agent isolation)
        claude mcp add --transport http my-agent http://localhost:8888/mcp/my-agent-bank/ \\
@@ -273,9 +176,9 @@ 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."""
@@ -339,25 +242,20 @@ class MCPMiddleware:
            _current_schema.set(tenant_context.schema_name) if tenant_context and tenant_context.schema_name else None
        )

-        # Resolve bank_id: path takes priority over header.
-        # Path = user's explicit connection endpoint (e.g., /mcp/my-bank/).
-        # X-Bank-Id header = per-request override for multi-bank mode only.
-        bank_id = None
+        # Try to get bank_id from header first (for Claude Code compatibility)
+        bank_id = self._get_header(scope, "X-Bank-Id")
        bank_id_from_path = False
-        new_path = path

-        # First, try to extract from path: /{bank_id}/...
-        if path.startswith("/") and len(path) > 1:
+        # If no header, try to extract from path: /{bank_id}/...
+        new_path = path
+        if not bank_id and path.startswith("/") and len(path) > 1:
            parts = path[1:].split("/", 1)
            if parts[0]:
+                # First segment looks like a bank_id
                bank_id = parts[0]
                bank_id_from_path = True
                new_path = "/" + parts[1] if len(parts) > 1 else "/"

-        # If no path-based bank_id, try X-Bank-Id header (multi-bank mode)
-        if not bank_id:
-            bank_id = self._get_header(scope, "X-Bank-Id")
-
        # Fall back to default bank_id
        if not bank_id:
            bank_id = DEFAULT_BANK_ID
@@ -441,9 +339,9 @@ 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
@@ -86,8 +86,6 @@ def print_startup_info(
    reranker_provider: str,
    mcp_enabled: bool = False,
    version: str | None = None,
-    vector_extension: str | None = None,
-    text_search_extension: str | None = None,
 ):
    """Print styled startup information."""
    print(color_start("Starting Hindsight API..."))
@@ -98,8 +96,6 @@ def print_startup_info(
    print(f"  {dim('LLM:')} {color(f'{llm_provider} / {llm_model}', 0.6)}")
    print(f"  {dim('Embeddings:')} {color(embeddings_provider, 0.8)}")
    print(f"  {dim('Reranker:')} {color(reranker_provider, 1.0)}")
-    extensions = f"{vector_extension or 'default'} (vector) / {text_search_extension or 'default'} (text)"
-    print(f"  {dim('Extensions:')} {color(extensions, 0.4)}")
    if mcp_enabled:
        print(f"  {dim('MCP:')} {color_end('enabled at /mcp')}")
    print()
@@ -8,9 +8,8 @@ import json
 import logging
 import os
 import sys
-from dataclasses import dataclass, field, fields
+from dataclasses import dataclass
 from datetime import datetime, timezone
-from typing import Any

 from dotenv import find_dotenv, load_dotenv

@@ -19,103 +18,6 @@ load_dotenv(find_dotenv(usecwd=True), override=True)

 logger = logging.getLogger(__name__)

-
-class ConfigFieldAccessError(AttributeError):
-    """Raised when trying to access a bank-configurable field from global config."""
-
-    pass
-
-
-class StaticConfigProxy:
-    """
-    Proxy that wraps HindsightConfig and only allows access to static (non-configurable) fields.
-
-    Raises ConfigFieldAccessError when trying to access configurable fields that vary per-bank.
-    Forces developers to use get_resolved_config(bank_id, context) for bank-specific settings.
-    """
-
-    def __init__(self, config: "HindsightConfig"):
-        object.__setattr__(self, "_config", config)
-        object.__setattr__(self, "_configurable_fields", HindsightConfig.get_configurable_fields())
-
-    def __getattribute__(self, name: str):
-        if name.startswith("_"):
-            return object.__getattribute__(self, name)
-
-        configurable_fields = object.__getattribute__(self, "_configurable_fields")
-        if name in configurable_fields:
-            raise ConfigFieldAccessError(
-                f"Field '{name}' is bank-configurable and cannot be accessed from global config. "
-                f"Use ConfigResolver.resolve_full_config(bank_id, context) to get bank-specific config. "
-                f"This prevents accidentally using global defaults when bank-specific overrides exist."
-            )
-
-        config = object.__getattribute__(self, "_config")
-        return getattr(config, name)
-
-    def __setattr__(self, name: str, value):
-        raise AttributeError("Config is read-only. Modifications must go through ConfigResolver.")
-
-
-# Configuration field markers for hierarchical configuration
-def hierarchical(default_value):
-    """
-    Mark a config field as hierarchical (can be overridden per-tenant/bank).
-
-    Hierarchical fields can be customized at the tenant or bank level via database
-    configuration. Examples: LLM settings, retention parameters, retrieval settings.
-    """
-    return field(default=default_value, metadata={"hierarchical": True})
-
-
-def static(default_value):
-    """
-    Mark a config field as static (server-level only, cannot be overridden).
-
-    Static fields are infrastructure-level settings that affect the entire server
-    and cannot vary per tenant or bank. Examples: database URL, API port, worker settings.
-    """
-    return field(default=default_value, metadata={"hierarchical": False})
-
-
-# Configuration key normalization utilities
-def normalize_config_key(key: str) -> str:
-    """
-    Convert environment variable format to Python field name format.
-
-    Examples:
-        HINDSIGHT_API_LLM_PROVIDER -> llm_provider
-        LLM_MODEL -> llm_model
-        llm_model -> llm_model (already normalized)
-
-    Args:
-        key: Environment variable name or Python field name
-
-    Returns:
-        Normalized Python field name (lowercase snake_case)
-    """
-    if key.startswith("HINDSIGHT_API_"):
-        key = key[len("HINDSIGHT_API_") :]
-    return key.lower()
-
-
-def normalize_config_dict(config: dict[str, Any]) -> dict[str, Any]:
-    """
-    Normalize all keys in a config dict to Python field names.
-
-    Allows users to provide config overrides in either format:
-    - Python field format: {"llm_provider": "openai"}
-    - Env var format: {"HINDSIGHT_API_LLM_PROVIDER": "openai"}
-
-    Args:
-        config: Dict with env var or Python field names as keys
-
-    Returns:
-        Dict with all keys normalized to Python field names
-    """
-    return {normalize_config_key(k): v for k, v in config.items()}
-
-
 # Environment variable names
 ENV_DATABASE_URL = "HINDSIGHT_API_DATABASE_URL"
 ENV_DATABASE_SCHEMA = "HINDSIGHT_API_DATABASE_SCHEMA"
@@ -129,11 +31,6 @@ ENV_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_LLM_INITIAL_BACKOFF"
 ENV_LLM_MAX_BACKOFF = "HINDSIGHT_API_LLM_MAX_BACKOFF"
 ENV_LLM_TIMEOUT = "HINDSIGHT_API_LLM_TIMEOUT"
 ENV_LLM_GROQ_SERVICE_TIER = "HINDSIGHT_API_LLM_GROQ_SERVICE_TIER"
-ENV_LLM_OPENAI_SERVICE_TIER = "HINDSIGHT_API_LLM_OPENAI_SERVICE_TIER"
-
-# Defaults for service tiers
-DEFAULT_LLM_GROQ_SERVICE_TIER = "auto"  # "on_demand", "flex", or "auto"
-DEFAULT_LLM_OPENAI_SERVICE_TIER = None  # None (default) or "flex" (50% cheaper)

 # Per-operation LLM configuration (optional, falls back to global LLM config)
 ENV_RETAIN_LLM_PROVIDER = "HINDSIGHT_API_RETAIN_LLM_PROVIDER"
@@ -194,14 +91,6 @@ ENV_RERANKER_LITELLM_API_BASE = "HINDSIGHT_API_RERANKER_LITELLM_API_BASE"
 ENV_RERANKER_LITELLM_API_KEY = "HINDSIGHT_API_RERANKER_LITELLM_API_KEY"
 ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"

-# LiteLLM SDK configuration (direct API access, no proxy needed)
-ENV_EMBEDDINGS_LITELLM_SDK_API_KEY = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_API_KEY"
-ENV_EMBEDDINGS_LITELLM_SDK_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_MODEL"
-ENV_EMBEDDINGS_LITELLM_SDK_API_BASE = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_API_BASE"
-ENV_RERANKER_LITELLM_SDK_API_KEY = "HINDSIGHT_API_RERANKER_LITELLM_SDK_API_KEY"
-ENV_RERANKER_LITELLM_SDK_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_SDK_MODEL"
-ENV_RERANKER_LITELLM_SDK_API_BASE = "HINDSIGHT_API_RERANKER_LITELLM_SDK_API_BASE"
-
 # Deprecated: Legacy shared LiteLLM config (for backward compatibility)
 ENV_LITELLM_API_BASE = "HINDSIGHT_API_LITELLM_API_BASE"
 ENV_LITELLM_API_KEY = "HINDSIGHT_API_LITELLM_API_KEY"
@@ -218,26 +107,18 @@ 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"
-
 ENV_HOST = "HINDSIGHT_API_HOST"
 ENV_PORT = "HINDSIGHT_API_PORT"
-ENV_BASE_PATH = "HINDSIGHT_API_BASE_PATH"
 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
@@ -257,38 +138,12 @@ 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_BATCH_ENABLED = "HINDSIGHT_API_RETAIN_BATCH_ENABLED"
-ENV_RETAIN_BATCH_POLL_INTERVAL_SECONDS = "HINDSIGHT_API_RETAIN_BATCH_POLL_INTERVAL_SECONDS"
-
-# File storage configuration
-ENV_FILE_STORAGE_TYPE = "HINDSIGHT_API_FILE_STORAGE_TYPE"
-ENV_FILE_STORAGE_S3_BUCKET = "HINDSIGHT_API_FILE_STORAGE_S3_BUCKET"
-ENV_FILE_STORAGE_S3_REGION = "HINDSIGHT_API_FILE_STORAGE_S3_REGION"
-ENV_FILE_STORAGE_S3_ENDPOINT = "HINDSIGHT_API_FILE_STORAGE_S3_ENDPOINT"
-ENV_FILE_STORAGE_S3_ACCESS_KEY_ID = "HINDSIGHT_API_FILE_STORAGE_S3_ACCESS_KEY_ID"
-ENV_FILE_STORAGE_S3_SECRET_ACCESS_KEY = "HINDSIGHT_API_FILE_STORAGE_S3_SECRET_ACCESS_KEY"
-ENV_FILE_STORAGE_GCS_BUCKET = "HINDSIGHT_API_FILE_STORAGE_GCS_BUCKET"
-ENV_FILE_STORAGE_GCS_SERVICE_ACCOUNT_KEY = "HINDSIGHT_API_FILE_STORAGE_GCS_SERVICE_ACCOUNT_KEY"
-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_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"
-ENV_FILE_CONVERSION_MAX_BATCH_SIZE = "HINDSIGHT_API_FILE_CONVERSION_MAX_BATCH_SIZE"
-ENV_ENABLE_FILE_UPLOAD_API = "HINDSIGHT_API_ENABLE_FILE_UPLOAD_API"
-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_OBSERVATIONS_MISSION = "HINDSIGHT_API_OBSERVATIONS_MISSION"

 # Optimization flags
 ENV_SKIP_LLM_VERIFICATION = "HINDSIGHT_API_SKIP_LLM_VERIFICATION"
@@ -314,12 +169,6 @@ 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_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"
@@ -328,18 +177,18 @@ 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
@@ -374,36 +223,22 @@ 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"
-
-# Text search extension (native PostgreSQL, vchord BM25, or Timescale pg_textsearch)
-DEFAULT_TEXT_SEARCH_EXTENSION = "native"  # Options: "native", "vchord", "pg_textsearch"
-
 # LiteLLM defaults
 DEFAULT_LITELLM_API_BASE = "http://localhost:4000"
 DEFAULT_EMBEDDINGS_LITELLM_MODEL = "text-embedding-3-small"
 DEFAULT_RERANKER_LITELLM_MODEL = "cohere/rerank-english-v3.0"

-# LiteLLM SDK defaults
-DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL = "cohere/embed-english-v3.0"
-DEFAULT_RERANKER_LITELLM_SDK_MODEL = "cohere/rerank-english-v3.0"
-
 DEFAULT_HOST = "0.0.0.0"
 DEFAULT_PORT = 8888
-DEFAULT_BASE_PATH = ""  # Empty string = root path
 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_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
@@ -412,26 +247,12 @@ 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_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"  # 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
-DEFAULT_FILE_DELETE_AFTER_RETAIN = True  # Delete file bytes after retain (saves storage)

 # Observations defaults (consolidated knowledge from facts)
 DEFAULT_ENABLE_OBSERVATIONS = True  # Observations 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_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
@@ -454,11 +275,6 @@ 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

-# 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
 DEFAULT_OTEL_SERVICE_NAME = "hindsight-api"
@@ -542,8 +358,6 @@ class HindsightConfig:
    # Database
    database_url: str
    database_schema: str
-    vector_extension: str  # "pgvector" or "vchord"
-    text_search_extension: str  # "native" or "vchord"

    # LLM (default, used as fallback for per-operation config)
    llm_provider: str
@@ -555,8 +369,6 @@ class HindsightConfig:
    llm_initial_backoff: float
    llm_max_backoff: float
    llm_timeout: float
-    llm_groq_service_tier: str  # Groq: "on_demand", "flex", or "auto"
-    llm_openai_service_tier: str | None  # OpenAI: None (default) or "flex" (50% cheaper)

    # Vertex AI configuration
    llm_vertexai_project_id: str | None
@@ -607,9 +419,6 @@ class HindsightConfig:
    embeddings_litellm_api_base: str
    embeddings_litellm_api_key: str | None
    embeddings_litellm_model: str
-    embeddings_litellm_sdk_api_key: str | None
-    embeddings_litellm_sdk_model: str
-    embeddings_litellm_sdk_api_base: str | None

    # Reranker
    reranker_provider: str
@@ -627,21 +436,13 @@ class HindsightConfig:
    reranker_litellm_api_base: str
    reranker_litellm_api_key: str | None
    reranker_litellm_model: str
-    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
    port: int
-    base_path: str
    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
    graph_retriever: str
@@ -655,46 +456,12 @@ 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
-
-    # File storage (static - server-level only)
-    file_storage_type: str  # "native" (PostgreSQL) or "s3" (S3-compatible)
-    file_storage_s3_bucket: str | None  # S3 bucket name (required for s3 storage)
-    file_storage_s3_region: str | None  # S3 region (optional, uses SDK default)
-    file_storage_s3_endpoint: str | None  # S3 endpoint URL (for MinIO, R2, etc.)
-    file_storage_s3_access_key_id: str | None  # S3 access key (optional, uses env/IAM)
-    file_storage_s3_secret_access_key: str | None  # S3 secret key (optional, uses env/IAM)
-    file_storage_gcs_bucket: str | None  # GCS bucket name (required for gcs storage)
-    file_storage_gcs_service_account_key: str | None  # GCS service account key JSON (optional, uses ADC)
-    file_storage_azure_container: str | None  # Azure container name (required for azure storage)
-    file_storage_azure_account_name: str | None  # Azure storage account name
-    file_storage_azure_account_key: str | None  # Azure storage account key
-    file_parser: str  # File parser to use (e.g., "markitdown", "iris")
-    file_parser_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)
-    file_conversion_max_batch_size: int  # Max files per request
-    enable_file_upload_api: bool
-    file_delete_after_retain: bool

    # Observations settings (consolidated knowledge from facts)
    enable_observations: bool
    consolidation_batch_size: int
-    consolidation_llm_batch_size: int
    consolidation_max_tokens: int
-    observations_mission: str | None
-
-    # 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
@@ -728,130 +495,8 @@ class HindsightConfig:
    otel_service_name: str
    otel_deployment_environment: str

-    # Class-level sets for configuration categorization
-
-    # CREDENTIAL_FIELDS: Never exposed via API, never configurable per-tenant/bank
-    _CREDENTIAL_FIELDS = {
-        # API Keys
-        "llm_api_key",
-        "retain_llm_api_key",
-        "reflect_llm_api_key",
-        "consolidation_llm_api_key",
-        # Base URLs (could expose infrastructure)
-        "llm_base_url",
-        "retain_llm_base_url",
-        "reflect_llm_base_url",
-        "consolidation_llm_base_url",
-        "embeddings_tei_base_url",
-        "reranker_tei_base_url",
-        "reranker_cohere_base_url",
-        # Service Account Keys
-        "llm_vertexai_service_account_key",
-        # File storage credentials
-        "file_storage_s3_access_key_id",
-        "file_storage_s3_secret_access_key",
-        "file_storage_gcs_service_account_key",
-        "file_storage_azure_account_key",
-        # File parser credentials
-        "file_parser_iris_token",
-    }
-
-    # CONFIGURABLE_FIELDS: Safe behavioral settings that can be customized per-tenant/bank
-    # 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",
-        # Consolidation settings
-        "enable_observations",
-        "observations_mission",
-        # Reflect settings
-        "reflect_mission",
-        # Disposition settings
-        "disposition_skepticism",
-        "disposition_literalism",
-        "disposition_empathy",
-    }
-
-    @property
-    def file_conversion_max_batch_size_bytes(self) -> int:
-        """Get maximum total batch size in bytes."""
-        return self.file_conversion_max_batch_size_mb * 1024 * 1024
-
-    @classmethod
-    def get_configurable_fields(cls) -> set[str]:
-        """
-        Get set of field names that are configurable per-tenant/bank via API.
-
-        Configurable fields are manually tagged behavioral settings that are safe
-        to expose and modify (e.g., retain_chunk_size, custom_instructions).
-        Excludes credentials, infrastructure config, and provider/model selection.
-
-        Returns:
-            Set of configurable field names
-        """
-        return cls._CONFIGURABLE_FIELDS.copy()
-
-    @classmethod
-    def get_credential_fields(cls) -> set[str]:
-        """
-        Get set of field names that are credentials (NEVER exposed via API).
-
-        Credential fields include API keys, base URLs, and service account keys.
-        These must never be returned in API responses or accepted in updates.
-
-        Returns:
-            Set of credential field names
-        """
-        return cls._CREDENTIAL_FIELDS.copy()
-
-    @classmethod
-    def get_hierarchical_fields(cls) -> set[str]:
-        """
-        DEPRECATED: Use get_configurable_fields() instead.
-
-        Kept for backward compatibility during migration.
-        """
-        return cls.get_configurable_fields()
-
-    @classmethod
-    def get_static_fields(cls) -> set[str]:
-        """
-        Get set of field names that are static (server-level only).
-
-        Static fields are infrastructure-level settings that cannot vary
-        per tenant or bank. These include database config, API port, worker settings, etc.
-        Also includes credential fields which are never configurable.
-
-        Returns:
-            Set of static field names
-        """
-        # Get all field names from dataclass
-        all_fields = {f.name for f in fields(cls)}
-        # Static fields = all fields - configurable fields
-        return all_fields - cls._CONFIGURABLE_FIELDS
-
    def validate(self) -> None:
        """Validate configuration values and raise errors for invalid combinations."""
-        # Validate vector_extension
-        valid_extensions = ("pgvector", "vchord", "pgvectorscale")
-        if self.vector_extension not in valid_extensions:
-            raise ValueError(
-                f"Invalid vector_extension: {self.vector_extension}. Must be one of: {', '.join(valid_extensions)}"
-            )
-
-        # Validate text_search_extension
-        valid_text_search = ("native", "vchord", "pg_textsearch")
-        if self.text_search_extension not in valid_text_search:
-            raise ValueError(
-                f"Invalid text_search_extension: {self.text_search_extension}. Must be one of: {', '.join(valid_text_search)}"
-            )
-
        # RETAIN_MAX_COMPLETION_TOKENS must be greater than RETAIN_CHUNK_SIZE
        # to ensure the LLM has enough output capacity to extract facts from chunks
        if self.retain_max_completion_tokens <= self.retain_chunk_size:
@@ -877,8 +522,6 @@ class HindsightConfig:
            # Database
            database_url=os.getenv(ENV_DATABASE_URL, DEFAULT_DATABASE_URL),
            database_schema=os.getenv(ENV_DATABASE_SCHEMA, DEFAULT_DATABASE_SCHEMA),
-            vector_extension=os.getenv(ENV_VECTOR_EXTENSION, DEFAULT_VECTOR_EXTENSION).lower(),
-            text_search_extension=os.getenv(ENV_TEXT_SEARCH_EXTENSION, DEFAULT_TEXT_SEARCH_EXTENSION).lower(),
            # LLM
            llm_provider=llm_provider,
            llm_api_key=os.getenv(ENV_LLM_API_KEY),
@@ -889,8 +532,6 @@ class HindsightConfig:
            llm_initial_backoff=float(os.getenv(ENV_LLM_INITIAL_BACKOFF, str(DEFAULT_LLM_INITIAL_BACKOFF))),
            llm_max_backoff=float(os.getenv(ENV_LLM_MAX_BACKOFF, str(DEFAULT_LLM_MAX_BACKOFF))),
            llm_timeout=float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT))),
-            llm_groq_service_tier=os.getenv(ENV_LLM_GROQ_SERVICE_TIER, DEFAULT_LLM_GROQ_SERVICE_TIER),
-            llm_openai_service_tier=os.getenv(ENV_LLM_OPENAI_SERVICE_TIER, DEFAULT_LLM_OPENAI_SERVICE_TIER),
            # Vertex AI
            llm_vertexai_project_id=os.getenv(ENV_LLM_VERTEXAI_PROJECT_ID) or DEFAULT_LLM_VERTEXAI_PROJECT_ID,
            llm_vertexai_region=os.getenv(ENV_LLM_VERTEXAI_REGION, DEFAULT_LLM_VERTEXAI_REGION),
@@ -989,12 +630,6 @@ class HindsightConfig:
            or os.getenv(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE),
            embeddings_litellm_api_key=os.getenv(ENV_EMBEDDINGS_LITELLM_API_KEY) or os.getenv(ENV_LITELLM_API_KEY),
            embeddings_litellm_model=os.getenv(ENV_EMBEDDINGS_LITELLM_MODEL, DEFAULT_EMBEDDINGS_LITELLM_MODEL),
-            # LiteLLM SDK embeddings (direct API access)
-            embeddings_litellm_sdk_api_key=os.getenv(ENV_EMBEDDINGS_LITELLM_SDK_API_KEY),
-            embeddings_litellm_sdk_model=os.getenv(
-                ENV_EMBEDDINGS_LITELLM_SDK_MODEL, DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL
-            ),
-            embeddings_litellm_sdk_api_base=os.getenv(ENV_EMBEDDINGS_LITELLM_SDK_API_BASE) or None,
            # Reranker
            reranker_provider=os.getenv(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER),
            reranker_local_model=os.getenv(ENV_RERANKER_LOCAL_MODEL, DEFAULT_RERANKER_LOCAL_MODEL),
@@ -1024,25 +659,12 @@ class HindsightConfig:
            or os.getenv(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE),
            reranker_litellm_api_key=os.getenv(ENV_RERANKER_LITELLM_API_KEY) or os.getenv(ENV_LITELLM_API_KEY),
            reranker_litellm_model=os.getenv(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL),
-            # LiteLLM SDK reranker (direct API access)
-            reranker_litellm_sdk_api_key=os.getenv(ENV_RERANKER_LITELLM_SDK_API_KEY),
-            reranker_litellm_sdk_model=os.getenv(ENV_RERANKER_LITELLM_SDK_MODEL, DEFAULT_RERANKER_LITELLM_SDK_MODEL),
-            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)),
-            base_path=os.getenv(ENV_BASE_PATH, DEFAULT_BASE_PATH),
            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
            graph_retriever=os.getenv(ENV_GRAPH_RETRIEVER, DEFAULT_GRAPH_RETRIEVER),
            mpfp_top_k_neighbors=int(os.getenv(ENV_MPFP_TOP_K_NEIGHBORS, str(DEFAULT_MPFP_TOP_K_NEIGHBORS))),
@@ -1068,53 +690,15 @@ 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_batch_enabled=os.getenv(ENV_RETAIN_BATCH_ENABLED, str(DEFAULT_RETAIN_BATCH_ENABLED)).lower()
-            == "true",
-            retain_batch_poll_interval_seconds=int(
-                os.getenv(ENV_RETAIN_BATCH_POLL_INTERVAL_SECONDS, str(DEFAULT_RETAIN_BATCH_POLL_INTERVAL_SECONDS))
-            ),
-            # File storage
-            file_storage_type=os.getenv(ENV_FILE_STORAGE_TYPE, DEFAULT_FILE_STORAGE_TYPE),
-            file_storage_s3_bucket=os.getenv(ENV_FILE_STORAGE_S3_BUCKET) or None,
-            file_storage_s3_region=os.getenv(ENV_FILE_STORAGE_S3_REGION) or None,
-            file_storage_s3_endpoint=os.getenv(ENV_FILE_STORAGE_S3_ENDPOINT) or None,
-            file_storage_s3_access_key_id=os.getenv(ENV_FILE_STORAGE_S3_ACCESS_KEY_ID) or None,
-            file_storage_s3_secret_access_key=os.getenv(ENV_FILE_STORAGE_S3_SECRET_ACCESS_KEY) or None,
-            file_storage_gcs_bucket=os.getenv(ENV_FILE_STORAGE_GCS_BUCKET) or None,
-            file_storage_gcs_service_account_key=os.getenv(ENV_FILE_STORAGE_GCS_SERVICE_ACCOUNT_KEY) or None,
-            file_storage_azure_container=os.getenv(ENV_FILE_STORAGE_AZURE_CONTAINER) or None,
-            file_storage_azure_account_name=os.getenv(ENV_FILE_STORAGE_AZURE_ACCOUNT_NAME) or None,
-            file_storage_azure_account_key=os.getenv(ENV_FILE_STORAGE_AZURE_ACCOUNT_KEY) or None,
-            file_parser=os.getenv(ENV_FILE_PARSER, DEFAULT_FILE_PARSER),
-            file_parser_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(
-                os.getenv(ENV_FILE_CONVERSION_MAX_BATCH_SIZE_MB, str(DEFAULT_FILE_CONVERSION_MAX_BATCH_SIZE_MB))
-            ),
-            file_conversion_max_batch_size=int(
-                os.getenv(ENV_FILE_CONVERSION_MAX_BATCH_SIZE, str(DEFAULT_FILE_CONVERSION_MAX_BATCH_SIZE))
-            ),
-            enable_file_upload_api=os.getenv(ENV_ENABLE_FILE_UPLOAD_API, str(DEFAULT_ENABLE_FILE_UPLOAD_API)).lower()
-            == "true",
-            file_delete_after_retain=os.getenv(
-                ENV_FILE_DELETE_AFTER_RETAIN, str(DEFAULT_FILE_DELETE_AFTER_RETAIN)
-            ).lower()
-            == "true",
            # Observations settings (consolidated knowledge from facts)
            enable_observations=os.getenv(ENV_ENABLE_OBSERVATIONS, str(DEFAULT_ENABLE_OBSERVATIONS)).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))
            ),
-            observations_mission=os.getenv(ENV_OBSERVATIONS_MISSION) or DEFAULT_OBSERVATIONS_MISSION,
            # Database migrations
            run_migrations_on_startup=os.getenv(ENV_RUN_MIGRATIONS_ON_STARTUP, "true").lower() == "true",
            # Database connection pool
@@ -1134,17 +718,6 @@ class HindsightConfig:
            ),
            # Reflect agent settings
            reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
-            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"),
@@ -1232,35 +805,8 @@ class HindsightConfig:
 _config_cache: HindsightConfig | None = None


-def get_config() -> StaticConfigProxy:
-    """
-    Get global configuration with ONLY static (non-configurable) fields accessible.
-
-    This returns a proxy that prevents access to bank-configurable fields
-    (like enable_observations, retain_chunk_size, etc.).
-
-    For bank-specific configuration, use:
-        config_resolver.resolve_full_config(bank_id, context)
-
-    This design prevents accidentally using global defaults when bank-specific
-    overrides exist.
-
-    Returns:
-        StaticConfigProxy that only exposes static infrastructure fields
-
-    Raises:
-        ConfigFieldAccessError: If you try to access a bank-configurable field
-    """
-    return StaticConfigProxy(_get_raw_config())
-
-
-def _get_raw_config() -> HindsightConfig:
-    """
-    Get raw config (internal use only).
-
-    INTERNAL USE ONLY. Do not use this directly in application code.
-    Use get_config() for static fields or ConfigResolver.resolve_full_config() for bank-specific config.
-    """
+def get_config() -> HindsightConfig:
+    """Get the cached configuration, loading from environment on first call."""
    global _config_cache
    if _config_cache is None:
        _config_cache = HindsightConfig.from_env()
@@ -1,275 +0,0 @@
-"""
-Configuration resolution with hierarchical overrides.
-
-Resolves config values through the hierarchy:
-  Global (env vars) → Tenant config (via extension) → Bank config (database)
-
-Config values are resolved on every request to ensure consistency across
-multiple API servers.
-"""
-
-import json
-import logging
-from dataclasses import asdict
-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
-
-logger = logging.getLogger(__name__)
-
-
-class ConfigResolver:
-    """Resolves hierarchical configuration with tenant/bank overrides."""
-
-    def __init__(self, pool: asyncpg.Pool, tenant_extension: TenantExtension | None = None):
-        """
-        Initialize config resolver.
-
-        Args:
-            pool: Database connection pool
-            tenant_extension: Optional tenant extension for tenant-level config and permissions
-        """
-        self.pool = pool
-        self.tenant_extension = tenant_extension
-        self._global_config = _get_raw_config()
-        self._configurable_fields = HindsightConfig.get_configurable_fields()
-        self._credential_fields = HindsightConfig.get_credential_fields()
-
-    async def resolve_full_config(self, bank_id: str, context: RequestContext | None = None) -> HindsightConfig:
-        """
-        Resolve full HindsightConfig for a bank with hierarchical overrides applied.
-
-        This is for INTERNAL USE ONLY. Returns the complete config object with all fields
-        including credentials and static fields. Use get_bank_config() for API responses.
-
-        Resolution order:
-        1. Global config (from environment variables)
-        2. Tenant config overrides (from TenantExtension.get_tenant_config())
-        3. Bank config overrides (from banks.config JSONB)
-
-        Args:
-            bank_id: Bank identifier
-            context: Request context for tenant config resolution
-
-        Returns:
-            Complete HindsightConfig with hierarchical overrides applied
-        """
-        # Start with global config (all fields)
-        config_dict = asdict(self._global_config)
-
-        # Load tenant config overrides (if tenant extension available)
-        if self.tenant_extension and context:
-            try:
-                tenant_overrides = await self.tenant_extension.get_tenant_config(context)
-                if tenant_overrides:
-                    # Normalize keys and filter to configurable fields only
-                    normalized_tenant = normalize_config_dict(tenant_overrides)
-                    configurable_tenant = {k: v for k, v in normalized_tenant.items() if k in self._configurable_fields}
-                    config_dict.update(configurable_tenant)
-                    logger.debug(
-                        f"Applied tenant config overrides for bank {bank_id}: {list(configurable_tenant.keys())}"
-                    )
-            except Exception as e:
-                logger.warning(f"Failed to load tenant config for bank {bank_id}: {e}")
-
-        # Load bank config overrides
-        bank_overrides = await self._load_bank_config(bank_id)
-        if bank_overrides:
-            config_dict.update(bank_overrides)
-            logger.debug(f"Applied bank config overrides for bank {bank_id}: {list(bank_overrides.keys())}")
-
-        # Return full config object (dataclass doesn't have __init__ that accepts kwargs, so we update the object)
-        # Create a new config instance by copying the global config and updating fields
-        resolved_config = HindsightConfig(**config_dict)
-        return resolved_config
-
-    async def get_bank_config(self, bank_id: str, context: RequestContext | None = None) -> dict[str, Any]:
-        """
-        Get fully resolved config for a bank (filtered by permissions).
-
-        Resolution order:
-        1. Global config (from environment variables)
-        2. Tenant config overrides (from TenantExtension.get_tenant_config())
-        3. Bank config overrides (from banks.config JSONB)
-
-        Note: Config is resolved on every call (not cached) to ensure consistency
-        across multiple API servers.
-
-        SECURITY:
-        - Only returns configurable fields (excludes static/infrastructure fields)
-        - Filters out ALL credential fields (API keys, base URLs, etc.)
-        - Further filtered by tenant/bank permissions if extension provides them
-
-        Args:
-            bank_id: Bank identifier
-            context: Request context for tenant config resolution and permissions
-
-        Returns:
-            Dict of allowed configurable fields only (never includes credentials or static fields)
-        """
-        # Resolve full config with all hierarchical overrides
-        resolved_config = await self.resolve_full_config(bank_id, context)
-        config_dict = asdict(resolved_config)
-
-        # SECURITY: Filter to only configurable fields (exclude static/infrastructure)
-        filtered = {k: v for k, v in config_dict.items() if k in self._configurable_fields}
-
-        # SECURITY: Remove ALL credential fields (API keys, base URLs, etc.)
-        filtered = {k: v for k, v in filtered.items() if k not in self._credential_fields}
-
-        # PERMISSIONS: Further filter based on tenant/bank permissions
-        if self.tenant_extension and context:
-            try:
-                allowed_fields = await self.tenant_extension.get_allowed_config_fields(context, bank_id)
-                if allowed_fields is not None:  # None means "allow all"
-                    filtered = {k: v for k, v in filtered.items() if k in allowed_fields}
-                    logger.debug(
-                        f"Applied permission filter for bank {bank_id}: allowed={len(allowed_fields)} fields, "
-                        f"returned={len(filtered)} fields"
-                    )
-            except Exception as e:
-                logger.warning(f"Failed to load permissions for bank {bank_id}: {e}")
-
-        return filtered
-
-    async def _load_bank_config(self, bank_id: str) -> dict[str, Any]:
-        """
-        Load bank config overrides from banks.config JSONB column.
-
-        Args:
-            bank_id: Bank identifier
-
-        Returns:
-            Dict of config overrides (only configurable fields, normalized keys)
-        """
-        try:
-            async with self.pool.acquire() as conn:
-                row = await conn.fetchrow(
-                    f"""
-                    SELECT config FROM {fq_table("banks")} WHERE bank_id = $1
-                    """,
-                    bank_id,
-                )
-
-                if row and row["config"]:
-                    config_data = row["config"]
-
-                    # Handle case where JSONB is returned as JSON string
-                    if isinstance(config_data, str):
-                        config_data = json.loads(config_data)
-
-                    # Normalize keys (handle both env var format and Python field format)
-                    normalized = normalize_config_dict(config_data)
-
-                    # Only return overrides for configurable fields
-                    return {k: v for k, v in normalized.items() if k in self._configurable_fields}
-        except Exception as e:
-            logger.error(f"Failed to load bank config for {bank_id}: {e}")
-
-        return {}
-
-    async def update_bank_config(
-        self, bank_id: str, updates: dict[str, Any], context: RequestContext | None = None
-    ) -> None:
-        """
-        Update bank configuration overrides (with permission checking).
-
-        Args:
-            bank_id: Bank identifier
-            updates: Dict of config field names to new values.
-                    Keys can be in env var format (HINDSIGHT_API_LLM_PROVIDER)
-                    or Python field format (llm_provider).
-                    Only configurable fields are allowed.
-            context: Request context for permission checking
-
-        Raises:
-            ValueError: If attempting to override invalid/disallowed fields
-        """
-        # Normalize keys
-        normalized_updates = normalize_config_dict(updates)
-
-        # SECURITY: Reject credential fields explicitly
-        credential_attempts = set(normalized_updates.keys()) & self._credential_fields
-        if credential_attempts:
-            raise ValueError(
-                f"Cannot set credential fields via API: {sorted(credential_attempts)}. "
-                f"Credentials (API keys, base URLs) must be set at server level only."
-            )
-
-        # Validate all fields are configurable
-        invalid_fields = set(normalized_updates.keys()) - self._configurable_fields
-        if invalid_fields:
-            static_fields = HindsightConfig.get_static_fields()
-            invalid_static = invalid_fields & static_fields
-            if invalid_static:
-                raise ValueError(
-                    f"Cannot override static (server-level) fields: {sorted(invalid_static)}. "
-                    f"Only configurable fields can be overridden per-bank. "
-                    f"Configurable fields include: {sorted(list(self._configurable_fields)[:10])}... "
-                    f"(total: {len(self._configurable_fields)} fields)"
-                )
-            else:
-                raise ValueError(
-                    f"Unknown configuration fields: {sorted(invalid_fields)}. "
-                    f"Valid configurable fields: {sorted(list(self._configurable_fields)[:10])}..."
-                )
-
-        # PERMISSIONS: Check tenant/bank permissions
-        if self.tenant_extension and context:
-            try:
-                allowed_fields = await self.tenant_extension.get_allowed_config_fields(context, bank_id)
-                if allowed_fields is not None:  # None means "allow all"
-                    disallowed = set(normalized_updates.keys()) - allowed_fields
-                    if disallowed:
-                        raise ValueError(
-                            f"Not allowed to modify fields: {sorted(disallowed)}. "
-                            f"Your permissions allow: {sorted(list(allowed_fields)[:10])}..."
-                            if allowed_fields
-                            else "Not allowed to modify fields: {sorted(disallowed)}. "
-                            "Your permissions do not allow any config modifications."
-                        )
-            except ValueError:
-                raise  # Re-raise permission errors
-            except Exception as e:
-                logger.warning(f"Failed to check permissions for bank {bank_id}: {e}")
-                # Continue without permission check (fail open for backward compatibility)
-
-        # Merge with existing config (JSONB || operator)
-        async with self.pool.acquire() as conn:
-            await conn.execute(
-                f"""
-                UPDATE {fq_table("banks")}
-                SET config = config || $1::jsonb,
-                    updated_at = now()
-                WHERE bank_id = $2
-                """,
-                json.dumps(normalized_updates),
-                bank_id,
-            )
-
-        logger.info(f"Updated bank config for {bank_id}: {list(normalized_updates.keys())}")
-
-    async def reset_bank_config(self, bank_id: str) -> None:
-        """
-        Reset bank configuration to defaults (remove all overrides).
-
-        Args:
-            bank_id: Bank identifier
-        """
-        async with self.pool.acquire() as conn:
-            await conn.execute(
-                f"""
-                UPDATE {fq_table("banks")}
-                SET config = '{{}}'::jsonb,
-                    updated_at = now()
-                WHERE bank_id = $1
-                """,
-                bank_id,
-            )
-
-        logger.info(f"Reset bank config for {bank_id} to defaults")
@@ -1,66 +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 (no create/update needed)"""
+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 (showing the required UUID format for all IDs):
-{{"creates": [{{"text": "Alice lives in Berlin", "source_fact_ids": ["a1b2c3d4-e5f6-7890-abcd-ef1234567890", "b2c3d4e5-f6a7-8901-bcde-f12345678901"]}}],
-  "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"}}]}}
+Return [] if fact contains no durable knowledge.

-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"""
@@ -21,7 +21,6 @@ from ..config import (
    DEFAULT_RERANKER_FLASHRANK_CACHE_DIR,
    DEFAULT_RERANKER_FLASHRANK_MODEL,
    DEFAULT_RERANKER_LITELLM_MODEL,
-    DEFAULT_RERANKER_LITELLM_SDK_MODEL,
    DEFAULT_RERANKER_LOCAL_FORCE_CPU,
    DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT,
    DEFAULT_RERANKER_LOCAL_MODEL,
@@ -29,12 +28,10 @@ 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,
    ENV_RERANKER_FLASHRANK_MODEL,
-    ENV_RERANKER_LITELLM_SDK_API_KEY,
    ENV_RERANKER_LOCAL_FORCE_CPU,
    ENV_RERANKER_LOCAL_MAX_CONCURRENT,
    ENV_RERANKER_LOCAL_MODEL,
@@ -43,7 +40,6 @@ 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 +554,6 @@ 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/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.
@@ -930,126 +828,6 @@ class LiteLLMCrossEncoder(CrossEncoderModel):
        return all_scores


-class LiteLLMSDKCrossEncoder(CrossEncoderModel):
-    """
-    LiteLLM SDK cross-encoder for direct API integration.
-
-    Supports reranking via LiteLLM SDK without requiring a proxy server.
-    Supported providers: Cohere, DeepInfra, Together AI, HuggingFace, Jina AI, Voyage AI, AWS Bedrock.
-
-    Example model names:
-    - cohere/rerank-english-v3.0
-    - deepinfra/Qwen3-reranker-8B
-    - together_ai/Salesforce/Llama-Rank-V1
-    - huggingface/BAAI/bge-reranker-v2-m3
-    """
-
-    def __init__(
-        self,
-        api_key: str,
-        model: str = DEFAULT_RERANKER_LITELLM_SDK_MODEL,
-        api_base: str | None = None,
-        timeout: float = 60.0,
-    ):
-        """
-        Initialize LiteLLM SDK cross-encoder client.
-
-        Args:
-            api_key: API key for the reranking provider
-            model: Model name with provider prefix (e.g., "deepinfra/Qwen3-reranker-8B")
-            api_base: Custom base URL for API (optional)
-            timeout: Request timeout in seconds (default: 60.0)
-        """
-        self.api_key = api_key
-        self.model = model
-        self.api_base = api_base
-        self.timeout = timeout
-        self._initialized = False
-        self._litellm = None  # Will be set during initialization
-
-    @property
-    def provider_name(self) -> str:
-        return "litellm-sdk"
-
-    async def initialize(self) -> None:
-        """Initialize the LiteLLM SDK client."""
-        if self._initialized:
-            return
-
-        try:
-            import litellm
-
-            self._litellm = litellm  # Store reference
-        except ImportError:
-            raise ImportError("litellm is required for LiteLLMSDKCrossEncoder. Install it with: pip install litellm")
-
-        api_base_msg = f" at {self.api_base}" if self.api_base else ""
-        logger.info(f"Reranker: initializing LiteLLM SDK provider with model {self.model}{api_base_msg}")
-
-        self._initialized = True
-        logger.info("Reranker: LiteLLM SDK provider initialized")
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """
-        Score query-document pairs using the LiteLLM SDK.
-
-        Args:
-            pairs: List of (query, document) tuples to score
-
-        Returns:
-            List of relevance scores
-        """
-        if not self._initialized:
-            raise RuntimeError("Reranker not initialized. Call initialize() first.")
-
-        if not pairs:
-            return []
-
-        # Group pairs by query for efficient batching
-        # LiteLLM rerank expects one query with multiple documents
-        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]
-
-            # Build kwargs for rerank call
-            rerank_kwargs = {
-                "model": self.model,
-                "query": query,
-                "documents": texts,
-                "api_key": self.api_key,
-            }
-            if self.api_base:
-                rerank_kwargs["api_base"] = self.api_base
-
-            response = await self._litellm.arerank(**rerank_kwargs)
-
-            # Map scores back to original positions
-            # Response format: RerankResponse with results list
-            # Each result is a TypedDict with "index" and "relevance_score"
-            if hasattr(response, "results") and response.results:
-                for result in response.results:
-                    # Results are TypedDicts, use dict-style access
-                    original_idx = result["index"]
-                    score = result.get("relevance_score", result.get("score", 0.0))
-                    all_scores[indices[original_idx]] = score
-            elif isinstance(response, list):
-                # Direct list of scores (unlikely but defensive)
-                for i, score in enumerate(response):
-                    all_scores[indices[i]] = score
-            else:
-                logger.warning(f"Unexpected response format from LiteLLM rerank: {type(response)}")
-
-        return all_scores
-
-
 def create_cross_encoder_from_env() -> CrossEncoderModel:
    """
    Create a CrossEncoderModel instance based on configuration.
@@ -1099,30 +877,9 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
            api_key=config.reranker_litellm_api_key,
            model=config.reranker_litellm_model,
        )
-    elif provider == "litellm-sdk":
-        api_key = config.reranker_litellm_sdk_api_key
-        if not api_key:
-            raise ValueError(
-                f"{ENV_RERANKER_LITELLM_SDK_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'litellm-sdk'"
-            )
-        return LiteLLMSDKCrossEncoder(
-            api_key=api_key,
-            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', 'rrf'"
        )
@@ -19,7 +19,6 @@ import httpx
 from ..config import (
    DEFAULT_EMBEDDINGS_COHERE_MODEL,
    DEFAULT_EMBEDDINGS_LITELLM_MODEL,
-    DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL,
    DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU,
    DEFAULT_EMBEDDINGS_LOCAL_MODEL,
    DEFAULT_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
@@ -27,7 +26,6 @@ from ..config import (
    DEFAULT_EMBEDDINGS_PROVIDER,
    DEFAULT_LITELLM_API_BASE,
    ENV_EMBEDDINGS_COHERE_API_KEY,
-    ENV_EMBEDDINGS_LITELLM_SDK_API_KEY,
    ENV_EMBEDDINGS_LOCAL_FORCE_CPU,
    ENV_EMBEDDINGS_LOCAL_MODEL,
    ENV_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
@@ -722,150 +720,6 @@ class LiteLLMEmbeddings(Embeddings):
        return all_embeddings


-class LiteLLMSDKEmbeddings(Embeddings):
-    """
-    LiteLLM SDK embeddings for direct API integration.
-
-    Supports embeddings via LiteLLM SDK without requiring a proxy server.
-    Supported providers: Cohere, OpenAI, Azure OpenAI, HuggingFace, Voyage AI, Together AI, etc.
-
-    Example model names:
-    - cohere/embed-english-v3.0
-    - openai/text-embedding-3-small
-    - together_ai/togethercomputer/m2-bert-80M-8k-retrieval
-    - voyage/voyage-2
-    """
-
-    def __init__(
-        self,
-        api_key: str,
-        model: str = DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL,
-        api_base: str | None = None,
-        batch_size: int = 100,
-        timeout: float = 60.0,
-    ):
-        """
-        Initialize LiteLLM SDK embeddings client.
-
-        Args:
-            api_key: API key for the embedding provider
-            model: Model name with provider prefix (e.g., "cohere/embed-english-v3.0")
-            api_base: Custom base URL for API (optional)
-            batch_size: Maximum batch size for embedding requests (default: 100)
-            timeout: Request timeout in seconds (default: 60.0)
-        """
-        self.api_key = api_key
-        self.model = model
-        self.api_base = api_base
-        self.batch_size = batch_size
-        self.timeout = timeout
-        self._litellm = None  # Will be set during initialization
-        self._dimension: int | None = None
-
-    @property
-    def provider_name(self) -> str:
-        return "litellm-sdk"
-
-    @property
-    def dimension(self) -> int:
-        if self._dimension is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-        return self._dimension
-
-    async def initialize(self) -> None:
-        """Initialize the LiteLLM SDK client and detect dimension."""
-        if self._litellm is not None:
-            return
-
-        try:
-            import litellm
-
-            self._litellm = litellm  # Store reference
-        except ImportError:
-            raise ImportError("litellm is required for LiteLLMSDKEmbeddings. Install it with: pip install litellm")
-
-        api_base_msg = f" at {self.api_base}" if self.api_base else ""
-        logger.info(f"Embeddings: initializing LiteLLM SDK provider with model {self.model}{api_base_msg}")
-
-        # Do a test embedding to detect dimension
-        try:
-            # Build kwargs for embedding call
-            embed_kwargs = {
-                "model": self.model,
-                "input": ["test"],
-                "api_key": self.api_key,
-                "encoding_format": "float",
-            }
-            if self.api_base:
-                embed_kwargs["api_base"] = self.api_base
-
-            # Use async embedding method (standard in litellm)
-            response = await self._litellm.aembedding(**embed_kwargs)
-
-            # Extract dimension from response
-            if response.data and len(response.data) > 0:
-                self._dimension = len(response.data[0]["embedding"])
-            else:
-                raise RuntimeError(f"Unable to detect embedding dimension for model {self.model}")
-
-        except Exception as e:
-            raise RuntimeError(f"Failed to initialize LiteLLM SDK embeddings: {e}")
-
-        logger.info(f"Embeddings: LiteLLM SDK provider initialized (model: {self.model}, dim: {self._dimension})")
-
-    def encode(self, texts: list[str]) -> list[list[float]]:
-        """
-        Generate embeddings using the LiteLLM SDK.
-
-        Args:
-            texts: List of text strings to encode
-
-        Returns:
-            List of embedding vectors (one per input text)
-        """
-        if self._litellm is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-
-        if not texts:
-            return []
-
-        all_embeddings = []
-
-        # Process in batches
-        for i in range(0, len(texts), self.batch_size):
-            batch = texts[i : i + self.batch_size]
-
-            try:
-                # Build kwargs for embedding call
-                embed_kwargs = {
-                    "model": self.model,
-                    "input": batch,
-                    "api_key": self.api_key,
-                    "encoding_format": "float",
-                }
-                if self.api_base:
-                    embed_kwargs["api_base"] = self.api_base
-
-                # Use sync embedding (litellm doesn't have async in thread-safe way)
-                response = self._litellm.embedding(**embed_kwargs)
-
-                # Extract embeddings from response
-                # Sort by index to ensure correct order
-                batch_embeddings = sorted(response.data, key=lambda x: x.get("index", 0))
-                all_embeddings.extend([e["embedding"] for e in batch_embeddings])
-
-            except Exception as e:
-                import traceback
-
-                logger.error(
-                    f"Error in LiteLLM embedding for batch starting at index {i}: {e}\n"
-                    f"Traceback: {traceback.format_exc()}"
-                )
-                raise
-
-        return all_embeddings
-
-
 def create_embeddings_from_env() -> Embeddings:
    """
    Create an Embeddings instance based on configuration.
@@ -917,19 +771,7 @@ def create_embeddings_from_env() -> Embeddings:
            api_key=config.embeddings_litellm_api_key,
            model=config.embeddings_litellm_model,
        )
-    elif provider == "litellm-sdk":
-        api_key = config.embeddings_litellm_sdk_api_key
-        if not api_key:
-            raise ValueError(
-                f"{ENV_EMBEDDINGS_LITELLM_SDK_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'litellm-sdk'"
-            )
-        return LiteLLMSDKEmbeddings(
-            api_key=api_key,
-            model=config.embeddings_litellm_sdk_model,
-            api_base=config.embeddings_litellm_sdk_api_base,
-        )
    else:
        raise ValueError(
-            f"Unknown embeddings provider: {provider}. "
-            f"Supported: 'local', 'tei', 'openai', 'cohere', 'litellm', 'litellm-sdk'"
+            f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai', 'cohere', 'litellm'"
        )
@@ -48,7 +48,6 @@ class MemoryEngineInterface(ABC):
        contents: list[dict[str, Any]],
        *,
        request_context: "RequestContext",
-        document_tags: list[str] | None = None,
    ) -> dict[str, Any]:
        """
        Retain a batch of memory items.
@@ -56,9 +55,8 @@ class MemoryEngineInterface(ABC):
        Args:
            bank_id: The memory bank ID.
            contents: List of content dicts with 'content', optional 'event_date',
-                     'context', 'metadata', 'document_id', and per-item 'tags'.
+                     'context', 'metadata', 'document_id'.
            request_context: Request context for authentication.
-            document_tags: Optional tags applied to all items in the batch.

        Returns:
            Dict with processing results.
@@ -563,7 +561,6 @@ class MemoryEngineInterface(ABC):
        contents: list[dict[str, Any]],
        *,
        request_context: "RequestContext",
-        document_tags: list[str] | None = None,
    ) -> dict[str, Any]:
        """
        Submit a batch retain operation to run asynchronously.
@@ -572,7 +569,6 @@ class MemoryEngineInterface(ABC):
            bank_id: The memory bank ID.
            contents: List of content dicts to retain.
            request_context: Request context for authentication.
-            document_tags: Optional tags applied to all items in the async batch.

        Returns:
            Dict with operation_id and items_count.
@@ -128,67 +128,6 @@ class LLMInterface(ABC):
        """
        pass

-    async def supports_batch_api(self) -> bool:
-        """
-        Check if this provider supports batch API operations.
-
-        Returns:
-            True if provider supports submit_batch/get_batch_status/retrieve_batch_results
-        """
-        return False
-
-    async def submit_batch(
-        self,
-        requests: list[dict[str, Any]],
-        endpoint: str = "/v1/chat/completions",
-        completion_window: str = "24h",
-    ) -> dict[str, Any]:
-        """
-        Submit a batch of requests to the provider's batch API.
-
-        Args:
-            requests: List of request dicts in JSONL format (custom_id, method, url, body)
-            endpoint: API endpoint for the batch (e.g., "/v1/chat/completions")
-            completion_window: Completion window (e.g., "24h")
-
-        Returns:
-            Dict with batch metadata: {"batch_id": str, "status": str, ...}
-
-        Raises:
-            NotImplementedError: If provider doesn't support batch API
-        """
-        raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
-
-    async def get_batch_status(self, batch_id: str) -> dict[str, Any]:
-        """
-        Get the status of a batch job.
-
-        Args:
-            batch_id: Batch identifier returned from submit_batch
-
-        Returns:
-            Dict with status info: {"batch_id": str, "status": str, "completed_at": str, ...}
-
-        Raises:
-            NotImplementedError: If provider doesn't support batch API
-        """
-        raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
-
-    async def retrieve_batch_results(self, batch_id: str) -> list[dict[str, Any]]:
-        """
-        Retrieve completed batch results.
-
-        Args:
-            batch_id: Batch identifier returned from submit_batch
-
-        Returns:
-            List of result dicts (one per request, matched by custom_id)
-
-        Raises:
-            NotImplementedError: If provider doesn't support batch API
-        """
-        raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
-
    @abstractmethod
    async def cleanup(self) -> None:
        """Clean up resources (close connections, etc.)."""
@@ -60,59 +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,
@@ -120,7 +67,6 @@ def create_llm_provider(
    model: str,
    reasoning_effort: str,
    groq_service_tier: str | None = None,
-    openai_service_tier: str | None = None,
    vertexai_project_id: str | None = None,
    vertexai_region: str | None = None,
    vertexai_credentials: Any = None,
@@ -134,8 +80,7 @@ def create_llm_provider(
        base_url: Base URL for the API.
        model: Model name.
        reasoning_effort: Reasoning effort level for supported providers.
-        groq_service_tier: Groq service tier (for Groq provider) - "on_demand", "flex", or "auto".
-        openai_service_tier: OpenAI service tier (for OpenAI provider) - None (default) or "flex" (50% cheaper).
+        groq_service_tier: Groq service tier (for Groq provider).
        vertexai_project_id: Vertex AI project ID (for VertexAI provider).
        vertexai_region: Vertex AI region (for VertexAI provider).
        vertexai_credentials: Vertex AI credentials object (for VertexAI provider).
@@ -211,7 +156,6 @@ def create_llm_provider(
            model=model,
            reasoning_effort=reasoning_effort,
            groq_service_tier=groq_service_tier,
-            openai_service_tier=openai_service_tier,
        )

    else:
@@ -233,7 +177,6 @@ class LLMProvider:
        model: str,
        reasoning_effort: str = "low",
        groq_service_tier: str | None = None,
-        openai_service_tier: str | None = None,
    ):
        """
        Initialize LLM provider.
@@ -244,17 +187,15 @@ class LLMProvider:
            base_url: Base URL for the API.
            model: Model name.
            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.
+            groq_service_tier: Groq service tier ("on_demand", "flex", "auto"). Default: None (uses Groq's default).
        """
        self.provider = provider.lower()
        self.api_key = api_key
        self.base_url = base_url
        self.model = model
        self.reasoning_effort = reasoning_effort
-        # Service tiers from hierarchical config (not env vars)
-        self.groq_service_tier = groq_service_tier
-        self.openai_service_tier = openai_service_tier
+        # Default to 'auto' for best performance, users can override to 'on_demand' for free tier
+        self.groq_service_tier = groq_service_tier or os.getenv(ENV_LLM_GROQ_SERVICE_TIER, "auto")

        # Validate provider
        valid_providers = [
@@ -331,7 +272,6 @@ class LLMProvider:
            model=self.model,
            reasoning_effort=self.reasoning_effort,
            groq_service_tier=self.groq_service_tier,
-            openai_service_tier=self.openai_service_tier,
            vertexai_project_id=vertexai_project_id,
            vertexai_region=vertexai_region,
            vertexai_credentials=vertexai_credentials,
@@ -605,9 +545,8 @@ 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)"
            )
@@ -623,9 +562,8 @@ 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)"
@@ -642,9 +580,8 @@ 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)"
@@ -1,69 +0,0 @@
-"""
-Typed metadata models for async operations.
-
-These dataclasses define the structure of result_metadata for different operation types.
-The metadata is exposed in the API for debugging purposes and may change without notice.
-"""
-
-from dataclasses import asdict, dataclass
-from typing import Any
-
-
-@dataclass
-class BatchRetainParentMetadata:
-    """Metadata for parent batch_retain operations (when split into sub-batches)."""
-
-    items_count: int
-    total_tokens: int
-    num_sub_batches: int
-    is_parent: bool = True
-
-    def to_dict(self) -> dict[str, Any]:
-        """Convert to dict for JSON serialization."""
-        return asdict(self)
-
-
-@dataclass
-class BatchRetainChildMetadata:
-    """Metadata for child batch_retain operations (individual sub-batches)."""
-
-    items_count: int
-    parent_operation_id: str
-    sub_batch_index: int
-    total_sub_batches: int
-
-    def to_dict(self) -> dict[str, Any]:
-        """Convert to dict for JSON serialization."""
-        return asdict(self)
-
-
-@dataclass
-class RetainMetadata:
-    """Metadata for regular retain operations (non-batched, deprecated async path)."""
-
-    items_count: int
-
-    def to_dict(self) -> dict[str, Any]:
-        """Convert to dict for JSON serialization."""
-        return asdict(self)
-
-
-@dataclass
-class ConsolidationMetadata:
-    """Metadata for consolidation operations."""
-
-    # Currently empty, but structure for future fields
-    def to_dict(self) -> dict[str, Any]:
-        """Convert to dict for JSON serialization."""
-        return asdict(self)
-
-
-@dataclass
-class RefreshMentalModelMetadata:
-    """Metadata for mental model refresh operations."""
-
-    mental_model_id: str
-
-    def to_dict(self) -> dict[str, Any]:
-        """Convert to dict for JSON serialization."""
-        return asdict(self)
@@ -1,62 +0,0 @@
-"""File parser implementations."""
-
-from .base import FileParser, UnsupportedFileTypeError
-from .iris import IrisParser
-from .markitdown import MarkitdownParser
-
-__all__ = ["FileParser", "UnsupportedFileTypeError", "IrisParser", "MarkitdownParser", "FileParserRegistry"]
-
-
-class FileParserRegistry:
-    """Registry for file parsers with auto-detection."""
-
-    def __init__(self):
-        """Initialize empty parser registry."""
-        self._parsers: dict[str, FileParser] = {}
-
-    def register(self, parser: FileParser):
-        """
-        Register a parser.
-
-        Args:
-            parser: FileParser instance
-        """
-        self._parsers[parser.name()] = parser
-
-    def get_parser(
-        self,
-        name: str | None,
-        filename: str,
-        content_type: str | None = None,
-    ) -> FileParser:
-        """
-        Get parser by name or auto-detect.
-
-        Args:
-            name: Parser name (e.g., "markitdown") or None for auto-detect
-            filename: File name for auto-detection
-            content_type: MIME type (optional)
-
-        Returns:
-            FileParser instance
-
-        Raises:
-            ValueError: If no suitable parser found
-        """
-        if name:
-            # Explicit parser requested — return it directly, let the parser
-            # raise UnsupportedFileTypeError from convert() if needed
-            if name not in self._parsers:
-                raise ValueError(f"Parser '{name}' not found. Available: {list(self._parsers.keys())}")
-            return self._parsers[name]
-
-        # Auto-detect parser
-        for parser in self._parsers.values():
-            if parser.supports(filename, content_type):
-                return parser
-
-        raise ValueError(f"No parser found for {filename}. Available parsers: {list(self._parsers.keys())}")
-
-    def list_parsers(self) -> list[str]:
-        """Get list of registered parser names."""
-        return list(self._parsers.keys())
@@ -1,58 +0,0 @@
-"""Abstract base class for file parsers."""
-
-from abc import ABC, abstractmethod
-
-
-class UnsupportedFileTypeError(Exception):
-    """Raised by a parser when it does not support the given file type."""
-
-    pass
-
-
-class FileParser(ABC):
-    """Abstract base for file to markdown parsers."""
-
-    @abstractmethod
-    async def convert(self, file_data: bytes, filename: str) -> str:
-        """
-        Parse file to markdown.
-
-        Args:
-            file_data: Raw file bytes
-            filename: Original filename (used for format detection)
-
-        Returns:
-            Markdown content as string
-
-        Raises:
-            UnsupportedFileTypeError: If the file type is not supported by this parser
-            RuntimeError: If parsing fails for another reason
-        """
-        pass
-
-    def supports(self, filename: str, content_type: str | None = None) -> bool:
-        """
-        Check if parser supports this file type.
-
-        Override this for local/static extension-based filtering.
-        Parsers that delegate to a remote service should leave this as True
-        and raise UnsupportedFileTypeError from convert() instead.
-
-        Args:
-            filename: File name (used for extension check)
-            content_type: MIME type (optional)
-
-        Returns:
-            True if this parser can handle the file (default: True)
-        """
-        return True
-
-    @abstractmethod
-    def name(self) -> str:
-        """
-        Get parser name.
-
-        Returns:
-            Parser name (e.g., "markitdown")
-        """
-        pass
@@ -1,137 +0,0 @@
-"""Iris parser implementation using the Vectorize Iris HTTP API."""
-
-import asyncio
-import logging
-import mimetypes
-import time
-
-import httpx
-
-from .base import FileParser, UnsupportedFileTypeError
-
-logger = logging.getLogger(__name__)
-
-_IRIS_BASE_URL = "https://api.vectorize.io/v1"
-_DEFAULT_POLL_INTERVAL = 2.0  # seconds
-_DEFAULT_TIMEOUT = 300.0  # seconds
-
-
-class IrisParser(FileParser):
-    """
-    Iris file parser using the Vectorize Iris cloud extraction service.
-
-    Uploads files to the Vectorize Iris API, starts an extraction job,
-    and polls until the text is ready. The API determines which file types
-    are supported — UnsupportedFileTypeError is raised if the file is rejected.
-
-    Authentication:
-        Requires HINDSIGHT_API_FILE_PARSER_IRIS_TOKEN and
-        HINDSIGHT_API_FILE_PARSER_IRIS_ORG_ID environment variables,
-        or pass them explicitly via the constructor.
-    """
-
-    def __init__(
-        self,
-        token: str,
-        org_id: str,
-        poll_interval: float = _DEFAULT_POLL_INTERVAL,
-        timeout: float = _DEFAULT_TIMEOUT,
-    ):
-        """
-        Initialize iris parser.
-
-        Args:
-            token: Vectorize API token
-            org_id: Vectorize organization ID
-            poll_interval: Seconds between status poll requests (default: 2)
-            timeout: Maximum seconds to wait for extraction (default: 300)
-        """
-        self._token = token
-        self._org_id = org_id
-        self._poll_interval = poll_interval
-        self._timeout = timeout
-        self._auth_headers = {"Authorization": f"Bearer {token}"}
-
-    async def convert(self, file_data: bytes, filename: str) -> str:
-        """
-        Parse file to text using the Vectorize Iris API.
-
-        Raises:
-            UnsupportedFileTypeError: If the Iris API rejects the file type (4xx)
-            RuntimeError: If extraction fails for another reason
-        """
-        content_type = mimetypes.guess_type(filename)[0] or "application/octet-stream"
-
-        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",
-                headers=self._auth_headers,
-                json={"name": filename, "contentType": content_type},
-            )
-            _raise_for_status(init_resp, filename, "file upload init")
-            init_data = init_resp.json()
-            file_id: str = init_data["fileId"]
-            upload_url: str = init_data["uploadUrl"]
-
-            # Step 2: Upload the file bytes to the presigned URL (no auth header)
-            upload_resp = await client.put(
-                upload_url,
-                content=file_data,
-                headers={"Content-Type": content_type},
-            )
-            _raise_for_status(upload_resp, filename, "file upload")
-
-            # Step 3: Start extraction
-            extract_resp = await client.post(
-                f"{_IRIS_BASE_URL}/org/{self._org_id}/extraction",
-                headers=self._auth_headers,
-                json={"fileId": file_id},
-            )
-            _raise_for_status(extract_resp, filename, "start extraction")
-            extraction_id: str = extract_resp.json()["extractionId"]
-
-            # Step 4: Poll until ready or timeout
-            deadline = time.monotonic() + self._timeout
-            while True:
-                status_resp = await client.get(
-                    f"{_IRIS_BASE_URL}/org/{self._org_id}/extraction/{extraction_id}",
-                    headers=self._auth_headers,
-                )
-                _raise_for_status(status_resp, filename, "poll extraction status")
-                status_data = status_resp.json()
-
-                if status_data.get("ready"):
-                    data = status_data.get("data", {})
-                    if not data.get("success"):
-                        error = data.get("error", "unknown error")
-                        raise RuntimeError(f"Iris extraction failed for '{filename}': {error}")
-                    text = data.get("text")
-                    if not text:
-                        raise RuntimeError(f"No content extracted from '{filename}'")
-                    return text
-
-                if time.monotonic() >= deadline:
-                    raise RuntimeError(f"Iris extraction timed out after {self._timeout}s for '{filename}'")
-
-                await asyncio.sleep(self._poll_interval)
-
-    def name(self) -> str:
-        """Get parser name."""
-        return "iris"
-
-
-def _raise_for_status(response: httpx.Response, filename: str, step: str) -> None:
-    """
-    Raise an appropriate error including the response body on HTTP errors.
-
-    Raises UnsupportedFileTypeError for 4xx responses (file rejected by the API),
-    RuntimeError for other HTTP errors.
-    """
-    if not response.is_error:
-        return
-    body = response.text or "<empty>"
-    msg = f"Iris API error during {step} for '{filename}': {response.status_code} {response.reason_phrase} — {body}"
-    if response.is_client_error:
-        raise UnsupportedFileTypeError(msg)
-    raise RuntimeError(msg)
@@ -1,109 +0,0 @@
-"""Markitdown parser implementation."""
-
-import asyncio
-import logging
-import tempfile
-from pathlib import Path
-
-from .base import FileParser
-
-logger = logging.getLogger(__name__)
-
-
-class MarkitdownParser(FileParser):
-    """
-    Markitdown file parser.
-
-    Uses Microsoft's markitdown library to convert various file formats
-    to markdown including PDF, Office docs, images (via OCR), audio, HTML.
-
-    Supported formats:
-    - PDF (.pdf)
-    - Word (.docx, .doc)
-    - PowerPoint (.pptx, .ppt)
-    - Excel (.xlsx, .xls)
-    - Images (.jpg, .jpeg, .png) - with OCR
-    - HTML (.html, .htm)
-    - Text (.txt, .md)
-    - Audio (.mp3, .wav) - with transcription
-    """
-
-    def __init__(self):
-        """Initialize markitdown parser."""
-        # Lazy import to avoid requiring markitdown for all users
-        try:
-            from markitdown import MarkItDown
-
-            self._markitdown = MarkItDown()
-        except ImportError as e:
-            raise ImportError(
-                "markitdown package is required for file parsing. Install with: pip install markitdown"
-            ) from e
-
-    async def convert(self, file_data: bytes, filename: str) -> str:
-        """Parse file to markdown using markitdown."""
-        # markitdown is synchronous, so we run it in executor to avoid blocking
-        loop = asyncio.get_event_loop()
-        return await loop.run_in_executor(None, self._convert_sync, file_data, filename)
-
-    def _convert_sync(self, file_data: bytes, filename: str) -> str:
-        """Synchronous parsing (runs in thread pool)."""
-        # Write to temp file (markitdown requires file path)
-        with tempfile.NamedTemporaryFile(suffix=Path(filename).suffix, delete=False) as tmp:
-            tmp.write(file_data)
-            tmp_path = tmp.name
-
-        try:
-            # Parse using markitdown
-            result = self._markitdown.convert(tmp_path)
-
-            if not result or not result.text_content:
-                raise RuntimeError(f"No content extracted from '{filename}'")
-
-            return result.text_content
-
-        except Exception as e:
-            logger.error(f"Markitdown parsing failed for {filename}: {e}")
-            raise RuntimeError(f"Failed to parse '{filename}': {e}") from e
-
-        finally:
-            # Clean up temp file
-            try:
-                Path(tmp_path).unlink()
-            except Exception:
-                pass
-
-    def supports(self, filename: str, content_type: str | None = None) -> bool:
-        """Check if markitdown supports this file type."""
-        # Supported extensions (from markitdown docs)
-        supported_extensions = {
-            # Documents
-            ".pdf",
-            ".docx",
-            ".doc",
-            ".pptx",
-            ".ppt",
-            ".xlsx",
-            ".xls",
-            # Images (with OCR)
-            ".jpg",
-            ".jpeg",
-            ".png",
-            # Web
-            ".html",
-            ".htm",
-            # Text
-            ".txt",
-            ".md",
-            ".csv",
-            # Audio (with transcription)
-            ".mp3",
-            ".wav",
-        }
-
-        ext = Path(filename).suffix.lower()
-        return ext in supported_extensions
-
-    def name(self) -> str:
-        """Get parser name."""
-        return "markitdown"
@@ -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:
@@ -18,7 +18,6 @@ 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

@@ -222,13 +221,10 @@ class GeminiLLM(LLMInterface):

        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
@@ -251,7 +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:
@@ -409,57 +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:
@@ -467,40 +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)
-
        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
@@ -16,7 +16,6 @@ Features:
 """

 import asyncio
-import io
 import json
 import logging
 import os
@@ -97,9 +96,8 @@ class OpenAICompatibleLLM(LLMInterface):
        if self.provider in ("openai", "groq") and not self.api_key:
            raise ValueError(f"API key is required for {self.provider}")

-        # Service tier configuration (from config, not env vars)
-        self.groq_service_tier = groq_service_tier
-        self.openai_service_tier = kwargs.get("openai_service_tier")
+        # Groq service tier configuration
+        self.groq_service_tier = groq_service_tier or os.getenv("HINDSIGHT_API_LLM_GROQ_SERVICE_TIER", "auto")

        # Get timeout config
        self.timeout = timeout or float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT)))
@@ -784,140 +782,6 @@ class OpenAICompatibleLLM(LLMInterface):
            raise last_exception
        raise RuntimeError("Ollama call failed after all retries")

-    async def supports_batch_api(self) -> bool:
-        """Check if this provider supports batch API operations."""
-        # Only OpenAI and Groq support batch API
-        return self.provider in ("openai", "groq")
-
-    async def submit_batch(
-        self,
-        requests: list[dict[str, Any]],
-        endpoint: str = "/v1/chat/completions",
-        completion_window: str = "24h",
-    ) -> dict[str, Any]:
-        """
-        Submit a batch of requests to OpenAI/Groq Batch API.
-
-        Args:
-            requests: List of request dicts with custom_id, method, url, body
-            endpoint: API endpoint (e.g., "/v1/chat/completions")
-            completion_window: Completion window (e.g., "24h")
-
-        Returns:
-            Dict with batch metadata including batch_id
-
-        Raises:
-            NotImplementedError: If provider doesn't support batch API
-        """
-        if not await self.supports_batch_api():
-            raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
-
-        logger.info(f"Submitting batch with {len(requests)} requests to {self.provider}")
-
-        # Format requests as JSONL
-        jsonl_content = "\n".join(json.dumps(req) for req in requests)
-
-        # Upload file to provider (wrap in BytesIO with filename)
-        file_bytes = io.BytesIO(jsonl_content.encode("utf-8"))
-        file_bytes.name = "batch_input.jsonl"  # OpenAI SDK needs a filename
-
-        file_response = await self._client.files.create(
-            file=file_bytes,
-            purpose="batch",
-        )
-
-        logger.debug(f"Uploaded batch file: {file_response.id}")
-
-        # Create batch
-        batch_response = await self._client.batches.create(
-            input_file_id=file_response.id,
-            endpoint=endpoint,
-            completion_window=completion_window,
-        )
-
-        logger.info(f"Batch submitted: {batch_response.id}, status={batch_response.status}")
-
-        return {
-            "batch_id": batch_response.id,
-            "status": batch_response.status,
-            "input_file_id": file_response.id,
-            "created_at": batch_response.created_at,
-            "request_count": len(requests),
-        }
-
-    async def get_batch_status(self, batch_id: str) -> dict[str, Any]:
-        """
-        Get the status of a batch job.
-
-        Args:
-            batch_id: Batch identifier
-
-        Returns:
-            Dict with status info (batch_id, status, completed_at, etc.)
-        """
-        if not await self.supports_batch_api():
-            raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
-
-        batch = await self._client.batches.retrieve(batch_id)
-
-        result = {
-            "batch_id": batch.id,
-            "status": batch.status,
-            "created_at": batch.created_at,
-            "request_counts": {
-                "total": batch.request_counts.total if batch.request_counts else 0,
-                "completed": batch.request_counts.completed if batch.request_counts else 0,
-                "failed": batch.request_counts.failed if batch.request_counts else 0,
-            },
-        }
-
-        if batch.completed_at:
-            result["completed_at"] = batch.completed_at
-        if batch.output_file_id:
-            result["output_file_id"] = batch.output_file_id
-        if batch.error_file_id:
-            result["error_file_id"] = batch.error_file_id
-        if batch.errors:
-            result["errors"] = batch.errors
-
-        return result
-
-    async def retrieve_batch_results(self, batch_id: str) -> list[dict[str, Any]]:
-        """
-        Retrieve completed batch results.
-
-        Args:
-            batch_id: Batch identifier
-
-        Returns:
-            List of result dicts (one per request, matched by custom_id)
-        """
-        if not await self.supports_batch_api():
-            raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
-
-        # Get batch status
-        batch = await self._client.batches.retrieve(batch_id)
-
-        if batch.status != "completed":
-            raise ValueError(f"Batch {batch_id} is not completed yet (status: {batch.status})")
-
-        if not batch.output_file_id:
-            raise ValueError(f"Batch {batch_id} has no output file")
-
-        # Download results file
-        logger.debug(f"Downloading results for batch {batch_id} from file {batch.output_file_id}")
-        file_content = await self._client.files.content(batch.output_file_id)
-
-        # Parse JSONL results
-        results = []
-        for line in file_content.text.strip().split("\n"):
-            if line:
-                results.append(json.loads(line))
-
-        logger.info(f"Retrieved {len(results)} results for batch {batch_id}")
-
-        return results
-
    async def cleanup(self) -> None:
        """Clean up resources (close OpenAI client connections)."""
        if hasattr(self, "_client") and self._client:
@@ -20,18 +20,26 @@ 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:
@@ -266,7 +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,
@@ -382,7 +390,6 @@ 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

@@ -436,32 +443,14 @@ async def run_reflect_agent(
        # 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(
@@ -475,14 +464,13 @@ 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 (but cap consecutive errors to avoid long hangs)
+            # 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)
            )
-            if 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)
            llm_start = time.time()
@@ -819,9 +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
@@ -857,7 +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."""
@@ -929,7 +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."""
@@ -955,8 +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", [])
@@ -984,9 +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")
@@ -12,20 +12,57 @@ from typing import Any


 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 ""
@@ -132,12 +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",
@@ -174,7 +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",
                "",
            ]
@@ -192,7 +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",
                "",
            ]
@@ -270,7 +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",
            ]
@@ -286,7 +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.",
        ]
    )

@@ -482,6 +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):
@@ -26,6 +26,8 @@ def _infer_temporal_date(fact_text: str, event_date: datetime) -> str | None:
    This is a fallback for when the LLM fails to extract temporal information
    from relative time expressions like "last night", "yesterday", etc.
    """
+    import re
+
    fact_lower = fact_text.lower()

    # Map relative time expressions to day offsets
@@ -438,9 +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
@@ -479,9 +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

@@ -519,7 +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):
@@ -547,16 +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
 )
@@ -565,7 +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
@@ -690,117 +695,6 @@ Example: "Lost job → couldn't pay rent → moved apartment"
 - Fact 2: Moved apartment, causal_relations: [{target_index: 1, relation_type: "caused_by"}]"""


-def _build_extraction_prompt_and_schema(config) -> tuple[str, type]:
-    """
-    Build extraction prompt and response schema based on config.
-
-    Returns:
-        Tuple of (prompt, response_schema)
-    """
-    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,
-            )
-        else:
-            base_prompt = CUSTOM_FACT_EXTRACTION_PROMPT
-            prompt = base_prompt.format(
-                retain_mission_section=retain_mission_section,
-                custom_instructions=config.retain_custom_instructions,
-            )
-    elif extraction_mode == "verbose":
-        prompt = VERBOSE_FACT_EXTRACTION_PROMPT
-    else:
-        base_prompt = CONCISE_FACT_EXTRACTION_PROMPT
-        prompt = base_prompt.format(
-            retain_mission_section=retain_mission_section,
-        )
-
-    # Add causal relationships section if enabled
-    if extract_causal_links:
-        prompt = prompt + CAUSAL_RELATIONSHIPS_SECTION
-        response_schema = FactExtractionResponseVerbose if extraction_mode == "verbose" else FactExtractionResponse
-    else:
-        response_schema = FactExtractionResponseNoCausal
-
-    return prompt, response_schema
-
-
-def _build_user_message(
-    chunk: str,
-    chunk_index: int,
-    total_chunks: int,
-    event_date: datetime,
-    context: str,
-    metadata: dict[str, str] | None = None,
-) -> 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"
-    event_date = parse_datetime_flexible(event_date)
-    event_date_formatted = event_date.strftime("%A, %B %d, %Y")
-
-    metadata_section = ""
-    if metadata:
-        metadata_lines = "\n".join(f"  {k}: {v}" for k, v in metadata.items())
-        metadata_section = f"\nMetadata:\n{metadata_lines}"
-
-    return f"""Extract facts from the following text chunk.
-
-Chunk: {chunk_index + 1}/{total_chunks}
-Event Date: {event_date_formatted} ({event_date.isoformat()})
-Context: {sanitized_context}{metadata_section}
-
-Text:
-{sanitized_chunk}"""
-
-
-def _build_request_body(llm_config, config, prompt: str, user_message: str, response_schema: type) -> dict:
-    """Build request body for LLM API call."""
-    request_body = {
-        "model": llm_config.model,
-        "messages": [{"role": "system", "content": prompt}, {"role": "user", "content": user_message}],
-        "temperature": 0.1,
-    }
-
-    # Add max_completion_tokens if configured
-    if config.retain_max_completion_tokens:
-        request_body["max_completion_tokens"] = config.retain_max_completion_tokens
-
-    # Add service_tier for OpenAI Flex Processing
-    if llm_config.provider == "openai" and llm_config._provider_impl.openai_service_tier:
-        request_body["service_tier"] = llm_config._provider_impl.openai_service_tier
-
-    # Add response_format (JSON schema)
-    if hasattr(response_schema, "model_json_schema"):
-        schema = response_schema.model_json_schema()
-        request_body["response_format"] = {
-            "type": "json_schema",
-            "json_schema": {"name": "facts", "schema": schema},
-        }
-
-    return request_body
-
-
 async def _extract_facts_from_chunk(
    chunk: str,
    chunk_index: int,
@@ -808,9 +702,7 @@ async def _extract_facts_from_chunk(
    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).
@@ -824,20 +716,73 @@ async def _extract_facts_from_chunk(

    logger = logging.getLogger(__name__)

-    # Build prompt and schema using helper function
-    prompt, response_schema = _build_extraction_prompt_and_schema(config)
+    # Determine which fact types to extract
+    # Note: We use "assistant" in the prompt but convert to "bank" for storage
+    fact_types_instruction = "Extract ONLY 'world' and 'assistant' type facts."

    # Check config for extraction mode and causal link extraction
+    config = get_config()
    extraction_mode = config.retain_extraction_mode
    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)
+    # Select base prompt based on extraction mode
+    if extraction_mode == "custom":
+        # Custom mode: inject user-provided guidelines
+        if not config.retain_custom_instructions:
+            logger.warning(
+                "extraction_mode='custom' but HINDSIGHT_API_RETAIN_CUSTOM_INSTRUCTIONS not set. "
+                "Falling back to 'concise' mode."
+            )
+            base_prompt = CONCISE_FACT_EXTRACTION_PROMPT
+            prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
+        else:
+            base_prompt = CUSTOM_FACT_EXTRACTION_PROMPT
+            prompt = base_prompt.format(
+                fact_types_instruction=fact_types_instruction,
+                custom_instructions=config.retain_custom_instructions,
+            )
+    elif extraction_mode == "verbose":
+        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(fact_types_instruction=fact_types_instruction)
+
+    # Build the full prompt with or without causal relationships section
+    # Select appropriate response schema based on extraction mode and causal links
+    if extract_causal_links:
+        prompt = prompt + CAUSAL_RELATIONSHIPS_SECTION
+        if extraction_mode == "verbose":
+            response_schema = FactExtractionResponseVerbose
+        else:
+            response_schema = FactExtractionResponse
+    else:
+        response_schema = FactExtractionResponseNoCausal

    # Retry logic for JSON validation errors
    max_retries = 2
    last_error = None

+    # Sanitize input text to prevent Unicode encoding errors (e.g., unpaired surrogates)
+    sanitized_chunk = _sanitize_text(chunk)
+    sanitized_context = _sanitize_text(context) if context else "none"
+
+    # Build user message with metadata and chunk content in a clear format
+    # Format event_date with day of week for better temporal reasoning
+    # Handle both datetime objects and ISO string formats (from deserialized async tasks)
+    from .orchestrator import parse_datetime_flexible
+
+    event_date = parse_datetime_flexible(event_date)
+    event_date_formatted = event_date.strftime("%A, %B %d, %Y")  # e.g., "Monday, June 10, 2024"
+    user_message = f"""Extract facts from the following text chunk.
+
+Chunk: {chunk_index + 1}/{total_chunks}
+Event Date: {event_date_formatted} ({event_date.isoformat()})
+Context: {sanitized_context}
+
+Text:
+{sanitized_chunk}"""
+
    usage = TokenUsage()  # Track cumulative usage across retries
    for attempt in range(max_retries):
        try:
@@ -1110,9 +1055,7 @@ async def _extract_facts_with_auto_split(
    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.
@@ -1127,9 +1070,7 @@ async def _extract_facts_with_auto_split(
        event_date: Reference date for temporal information
        context: Context about the conversation/document
        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)
@@ -1147,9 +1088,7 @@ async def _extract_facts_with_auto_split(
            event_date=event_date,
            context=context,
            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
@@ -1193,9 +1132,7 @@ async def _extract_facts_with_auto_split(
                event_date=event_date,
                context=context,
                llm_config=llm_config,
-                config=config,
                agent_name=agent_name,
-                metadata=metadata,
            ),
            _extract_facts_with_auto_split(
                chunk=second_half,
@@ -1204,9 +1141,7 @@ async def _extract_facts_with_auto_split(
                event_date=event_date,
                context=context,
                llm_config=llm_config,
-                config=config,
                agent_name=agent_name,
-                metadata=metadata,
            ),
        ]

@@ -1229,9 +1164,7 @@ async def extract_facts_from_text(
    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.
@@ -1245,11 +1178,9 @@ async def extract_facts_from_text(
    Args:
        text: Input text (conversation, article, etc.)
        event_date: Reference date for resolving relative times
+        context: Context about the conversation/document
        llm_config: LLM configuration to use
        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:
@@ -1257,6 +1188,7 @@ async def extract_facts_from_text(
        - chunks: List of tuples (chunk_text, fact_count) for each chunk
        - usage: Aggregated token usage across all LLM calls
    """
+    config = get_config()
    chunks = chunk_text(text, max_chars=config.retain_chunk_size)

    # Log chunk count before starting LLM requests
@@ -1275,9 +1207,7 @@ async def extract_facts_from_text(
            event_date=event_date,
            context=context,
            llm_config=llm_config,
-            config=config,
            agent_name=agent_name,
-            metadata=metadata,
        )
        for i, chunk in enumerate(chunks)
    ]
@@ -1304,425 +1234,12 @@ 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
-
-
-async def extract_facts_from_contents_batch_api(
-    contents: list[RetainContent],
-    llm_config,
-    agent_name: str,
-    config,
-    pool=None,
-    operation_id: str | None = None,
-    schema: str | None = None,
-) -> tuple[list[ExtractedFactType], list[ChunkMetadata], TokenUsage]:
-    """
-    Extract facts using LLM Batch API (OpenAI/Groq).
-
-    Submits all chunks as a single batch, polls until complete, then processes results.
-    Only called when config.retain_batch_enabled=True.
-
-    Args:
-        contents: List of RetainContent objects to process
-        llm_config: LLM configuration with batch API support
-        agent_name: Name of the agent
-        config: Resolved HindsightConfig for this bank
-        pool: Database connection pool (for storing batch state)
-        operation_id: Async operation ID (for crash recovery)
-        schema: Database schema (for multi-tenant support)
-
-    Returns:
-        Tuple of (extracted_facts, chunks_metadata, usage)
-    """
-    if not contents:
-        return [], [], TokenUsage()
-
-    logger.info(f"Using Batch API for fact extraction ({len(contents)} contents)")
-
-    # Check config for extraction mode and causal link extraction (used throughout)
-    extraction_mode = config.retain_extraction_mode
-    extract_causal_links = config.retain_extract_causal_links
-
-    # Check if provider supports batch API
-    if not await llm_config._provider_impl.supports_batch_api():
-        logger.warning(f"Batch API not supported for provider {llm_config.provider}, falling back to sync mode")
-        return await extract_facts_from_contents(contents, llm_config, agent_name, config, pool, operation_id, schema)
-
-    # Check if we're resuming an existing batch (crash recovery)
-    batch_id = None
-    if operation_id and pool:
-        from ..task_backend import fq_table
-
-        table = fq_table("async_operations", schema)
-        row = await pool.fetchrow(
-            f"SELECT result_metadata FROM {table} WHERE operation_id = $1",
-            operation_id,
-        )
-
-        if row and row["result_metadata"]:
-            metadata = row["result_metadata"]
-            if isinstance(metadata, str):
-                metadata = json.loads(metadata)
-            batch_id = metadata.get("batch_id")
-
-            if batch_id:
-                logger.info(f"Resuming existing batch: batch_id={batch_id} (crash recovery)")
-
-    # Step 1: Chunk all contents and build batch requests (skip if resuming)
-    all_chunks_info = []  # List of (chunk_text, content_index, chunk_index_in_content, event_date, context)
-    batch_requests = []
-
-    # Build prompt and schema once (same for all chunks)
-    prompt, response_schema = _build_extraction_prompt_and_schema(config)
-
-    for content_index, item in enumerate(contents):
-        chunks = chunk_text(item.content, max_chars=config.retain_chunk_size)
-
-        for chunk_index_in_content, chunk in enumerate(chunks):
-            all_chunks_info.append((chunk, content_index, chunk_index_in_content, item.event_date, item.context))
-
-            # Build batch request for this chunk
-            custom_id = f"chunk_{len(all_chunks_info) - 1}"  # Global chunk index
-
-            # 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
-            )
-
-            # Build request body using helper function
-            request_body = _build_request_body(llm_config, config, prompt, user_message, response_schema)
-
-            batch_requests.append(
-                {"custom_id": custom_id, "method": "POST", "url": "/v1/chat/completions", "body": request_body}
-            )
-
-    if not batch_requests and not batch_id:  # No requests and not resuming
-        return [], [], TokenUsage()
-
-    # Step 2: Submit batch (skip if resuming)
-    if not batch_id:
-        logger.info(f"Submitting batch with {len(batch_requests)} chunk requests")
-
-        batch_metadata = await llm_config._provider_impl.submit_batch(batch_requests)
-        batch_id = batch_metadata["batch_id"]
-
-        logger.info(f"Batch submitted: {batch_id}, polling every {config.retain_batch_poll_interval_seconds}s")
-
-        # CRITICAL: Store minimal batch state in operation metadata for crash recovery
-        # This allows resuming polling if worker restarts
-        if operation_id and pool:
-            batch_state = {
-                "batch_id": batch_id,
-                "batch_provider": llm_config.provider,
-                "chunk_count": len(batch_requests),
-            }
-
-            # Update operation result_metadata
-            from ..task_backend import fq_table
-
-            table = fq_table("async_operations", schema)
-            await pool.execute(
-                f"""
-                UPDATE {table}
-                SET result_metadata = result_metadata || $1::jsonb, updated_at = now()
-                WHERE operation_id = $2
-                """,
-                json.dumps(batch_state),
-                operation_id,
-            )
-            logger.info(f"Stored batch state for operation {operation_id} (crash recovery enabled)")
-    else:
-        logger.info(f"Resuming polling for existing batch: {batch_id}")
-
-    # Step 3: Poll until complete
-    import time
-
-    start_time = time.time()
-    while True:
-        status_info = await llm_config._provider_impl.get_batch_status(batch_id)
-        status = status_info["status"]
-
-        elapsed = time.time() - start_time
-        logger.info(
-            f"Batch {batch_id}: status={status}, "
-            f"completed={status_info['request_counts']['completed']}/{status_info['request_counts']['total']}, "
-            f"elapsed={elapsed:.0f}s"
-        )
-
-        if status == "completed":
-            break
-        elif status in ("failed", "expired", "cancelled"):
-            error_msg = status_info.get("errors", "Unknown error")
-            raise RuntimeError(f"Batch {batch_id} failed with status {status}: {error_msg}")
-
-        # Wait before polling again
-        await asyncio.sleep(config.retain_batch_poll_interval_seconds)
-
-    logger.info(f"Batch {batch_id} completed in {elapsed:.0f}s, retrieving results")
-
-    # Step 4: Retrieve results
-    batch_results = await llm_config._provider_impl.retrieve_batch_results(batch_id)
-
-    # Map results by custom_id
-    results_by_id = {result["custom_id"]: result for result in batch_results}
-
-    # Step 5: Parse results into facts (same as sync mode)
-    all_facts_from_llm = []
-    chunks_metadata = []
-    total_usage = TokenUsage()
-
-    for chunk_idx, (chunk_content, content_index, chunk_index_in_content, event_date, context) in enumerate(
-        all_chunks_info
-    ):
-        custom_id = f"chunk_{chunk_idx}"
-        result = results_by_id.get(custom_id)
-
-        if not result:
-            logger.warning(f"Missing result for {custom_id}, skipping")
-            chunks_metadata.append(
-                ChunkMetadata(
-                    chunk_text=chunk_content, fact_count=0, content_index=content_index, chunk_index=chunk_idx
-                )
-            )
-            continue
-
-        # Check for errors
-        if result.get("error"):
-            logger.error(f"Error in {custom_id}: {result['error']}")
-            chunks_metadata.append(
-                ChunkMetadata(
-                    chunk_text=chunk_content, fact_count=0, content_index=content_index, chunk_index=chunk_idx
-                )
-            )
-            continue
-
-        # Extract response
-        response_body = result.get("response", {}).get("body", {})
-        choices = response_body.get("choices", [])
-
-        if not choices:
-            logger.warning(f"No choices in response for {custom_id}")
-            chunks_metadata.append(
-                ChunkMetadata(
-                    chunk_text=chunk_content, fact_count=0, content_index=content_index, chunk_index=chunk_idx
-                )
-            )
-            continue
-
-        # Parse JSON content
-        message = choices[0].get("message", {})
-        content_str = message.get("content", "{}")
-
-        try:
-            extraction_response_json = json.loads(content_str)
-        except json.JSONDecodeError as e:
-            logger.error(f"Failed to parse JSON for {custom_id}: {e}")
-            chunks_metadata.append(
-                ChunkMetadata(
-                    chunk_text=chunk_content, fact_count=0, content_index=content_index, chunk_index=chunk_idx
-                )
-            )
-            continue
-
-        # Parse facts (reuse existing logic from _extract_facts_from_chunk)
-        raw_facts = extraction_response_json.get("facts", [])
-        chunk_facts = []
-
-        for i, llm_fact in enumerate(raw_facts):
-            if not isinstance(llm_fact, dict):
-                continue
-
-            def get_value(field_name):
-                value = llm_fact.get(field_name)
-                if value and value != "" and value != [] and value != {} and str(value).upper() != "N/A":
-                    return value
-                return None
-
-            what = get_value("what")
-            if not what:
-                what = get_value("factual_core")
-            if not what:
-                continue
-
-            when = get_value("when")
-            who = get_value("who")
-            why = get_value("why")
-
-            # Critical field: fact_type
-            original_fact_type = llm_fact.get("fact_type")
-            fact_type = original_fact_type
-
-            # Convert "assistant" → "experience"
-            if fact_type == "assistant":
-                fact_type = "experience"
-
-            # Validate fact_type
-            if fact_type not in ["world", "experience", "opinion"]:
-                fact_kind = llm_fact.get("fact_kind")
-                if fact_kind == "assistant":
-                    fact_type = "experience"
-                elif fact_kind in ["world", "experience", "opinion"]:
-                    fact_type = fact_kind
-                else:
-                    fact_type = "world"
-
-            # Build combined fact text
-            combined_parts = [what]
-            if when:
-                combined_parts.append(f"When: {when}")
-            if who:
-                combined_parts.append(f"Involving: {who}")
-            if why:
-                combined_parts.append(why)
-            combined_text = " | ".join(combined_parts)
-
-            # Temporal fields
-            fact_data = {}
-            fact_kind = llm_fact.get("fact_kind", "conversation")
-            if fact_kind not in ["conversation", "event", "other"]:
-                fact_kind = "conversation"
-
-            if fact_kind == "event":
-                occurred_start = get_value("occurred_start")
-                occurred_end = get_value("occurred_end")
-
-                if not occurred_start:
-                    fact_data["occurred_start"] = _infer_temporal_date(combined_text, event_date)
-                else:
-                    fact_data["occurred_start"] = occurred_start
-
-                if occurred_end:
-                    fact_data["occurred_end"] = occurred_end
-                elif fact_data.get("occurred_start"):
-                    fact_data["occurred_end"] = fact_data["occurred_start"]
-
-            # Entities
-            entities = get_value("entities")
-            if entities:
-                validated_entities = []
-                for ent in entities:
-                    if isinstance(ent, str):
-                        validated_entities.append(Entity(text=ent))
-                    elif isinstance(ent, dict) and "text" in ent:
-                        try:
-                            validated_entities.append(Entity.model_validate(ent))
-                        except Exception:
-                            pass
-                if validated_entities:
-                    fact_data["entities"] = validated_entities
-
-            # Causal relations
-            if extract_causal_links:
-                validated_relations = []
-                causal_relations_raw = get_value("causal_relations")
-                if causal_relations_raw:
-                    for rel in causal_relations_raw:
-                        if not isinstance(rel, dict):
-                            continue
-                        target_idx = rel.get("target_index")
-                        relation_type = rel.get("relation_type")
-                        strength = rel.get("strength", 1.0)
-
-                        if target_idx is None or relation_type is None:
-                            continue
-                        if target_idx < 0 or target_idx >= i:
-                            continue
-
-                        try:
-                            validated_relations.append(
-                                CausalRelation(
-                                    target_fact_index=target_idx, relation_type=relation_type, strength=strength
-                                )
-                            )
-                        except Exception:
-                            pass
-
-                if validated_relations:
-                    fact_data["causal_relations"] = validated_relations
-
-            # Always set mentioned_at
-            fact_data["mentioned_at"] = event_date.isoformat()
-
-            try:
-                fact = Fact(fact=combined_text, fact_type=fact_type, **fact_data)
-                chunk_facts.append(fact)
-            except Exception as e:
-                logger.error(f"Failed to create Fact model for fact {i}: {e}")
-                continue
-
-        all_facts_from_llm.extend(chunk_facts)
-        chunks_metadata.append(
-            ChunkMetadata(
-                chunk_text=chunk_content,
-                fact_count=len(chunk_facts),
-                content_index=content_index,
-                chunk_index=chunk_idx,
-            )
-        )
-
-        # Track token usage
-        usage_data = response_body.get("usage", {})
-        if usage_data:
-            total_usage = total_usage + TokenUsage(
-                input_tokens=usage_data.get("prompt_tokens", 0),
-                output_tokens=usage_data.get("completion_tokens", 0),
-                total_tokens=usage_data.get("total_tokens", 0),
-            )
-
-    # Step 6: Convert to ExtractedFact objects with proper chunk mapping
-    # Group facts by chunk
-    facts_by_chunk = []  # List of (chunk_metadata, [facts])
-    fact_start_idx = 0
-
-    for chunk_meta in chunks_metadata:
-        chunk_facts = all_facts_from_llm[fact_start_idx : fact_start_idx + chunk_meta.fact_count]
-        facts_by_chunk.append((chunk_meta, chunk_facts))
-        fact_start_idx += chunk_meta.fact_count
-
-    # Now convert to ExtractedFactType
-    extracted_facts = []
-    global_fact_idx = 0
-
-    for chunk_meta, chunk_facts in facts_by_chunk:
-        content = contents[chunk_meta.content_index]
-
-        for fact_from_llm in chunk_facts:
-            extracted_fact = ExtractedFactType(
-                fact_text=fact_from_llm.fact,
-                fact_type=fact_from_llm.fact_type,
-                entities=[e.text for e in (fact_from_llm.entities or [])],
-                occurred_start=_parse_datetime(fact_from_llm.occurred_start) if fact_from_llm.occurred_start else None,
-                occurred_end=_parse_datetime(fact_from_llm.occurred_end) if fact_from_llm.occurred_end else None,
-                causal_relations=_convert_causal_relations(fact_from_llm.causal_relations or [], global_fact_idx),
-                content_index=chunk_meta.content_index,
-                chunk_index=chunk_meta.chunk_index,
-                context=content.context,
-                mentioned_at=content.event_date,
-                metadata=content.metadata,
-                tags=content.tags,
-                observation_scopes=content.observation_scopes,
-            )
-
-            extracted_facts.append(extracted_fact)
-            global_fact_idx += 1
-
-    # Step 7: Add temporal offsets
-    _add_temporal_offsets(extracted_facts, contents)
-
-    logger.info(f"Batch API extracted {len(extracted_facts)} facts from {len(all_chunks_info)} chunks")
-
-    return extracted_facts, chunks_metadata, total_usage
+# Each fact gets 10 seconds offset to preserve ordering within a document
+SECONDS_PER_FACT = 10


 async def extract_facts_from_contents(
-    contents: list[RetainContent],
-    llm_config,
-    agent_name: str,
-    config,
-    pool=None,
-    operation_id: str | None = None,
-    schema: str | None = None,
+    contents: list[RetainContent], llm_config, agent_name: str
 ) -> tuple[list[ExtractedFactType], list[ChunkMetadata], TokenUsage]:
    """
    Extract facts from multiple content items in parallel.
@@ -1733,16 +1250,10 @@ async def extract_facts_from_contents(
    3. Adds time offsets to preserve fact ordering within each content
    4. Returns typed ExtractedFact and ChunkMetadata objects

-    Routes to batch API mode if config.retain_batch_enabled=True.
-
    Args:
        contents: List of RetainContent objects to process
        llm_config: LLM configuration for fact extraction
        agent_name: Name of the agent (for agent-related fact detection)
-        config: Resolved HindsightConfig for this bank
-        pool: Database connection pool (passed to batch API for state storage)
-        operation_id: Async operation ID (passed to batch API for crash recovery)
-        schema: Database schema (passed to batch API for multi-tenant support)

    Returns:
        Tuple of (extracted_facts, chunks_metadata, usage)
@@ -1750,12 +1261,6 @@ async def extract_facts_from_contents(
    if not contents:
        return [], [], TokenUsage()

-    # Route to batch API if enabled
-    if config.retain_batch_enabled:
-        return await extract_facts_from_contents_batch_api(
-            contents, llm_config, agent_name, config, pool, operation_id, schema
-        )
-
    # Step 1: Create parallel fact extraction tasks
    fact_extraction_tasks = []
    for item in contents:
@@ -1767,8 +1272,6 @@ async def extract_facts_from_contents(
            context=item.context,
            llm_config=llm_config,
            agent_name=agent_name,
-            config=config,
-            metadata=item.metadata or None,
        )
        fact_extraction_tasks.append(task)

@@ -1832,7 +1335,6 @@ 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)
@@ -7,7 +7,6 @@ Handles insertion of facts into the database.
 import json
 import logging

-from ...config import get_config
 from ..memory_engine import fq_table
 from .fact_extraction import _sanitize_text
 from .types import ProcessedFact
@@ -47,7 +46,6 @@ async def insert_facts_batch(
    chunk_ids = []
    document_ids = []
    tags_list = []
-    observation_scopes_list = []

    for fact in facts:
        fact_texts.append(_sanitize_text(fact.fact_text))
@@ -69,72 +67,31 @@ 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
-        )

    # Batch insert all facts
    # Note: tags are passed as JSON strings and converted back to varchar[] via jsonb_array_elements_text + array_agg
-    # Query varies based on text search backend
-    config = get_config()
-    if config.text_search_extension == "vchord":
-        # VectorChord: manually tokenize and insert 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[]
-                ) 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)
-            )
-            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, search_vector)
-            SELECT
-                $1,
-                text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                context, fact_type, confidence_score, metadata, chunk_id, document_id,
-                COALESCE(
-                    (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
-                    '{{}}'::varchar[]
-                ),
-                observation_scopes_json,
-                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, 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[]
-                ) 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)
-            )
-            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)
-            SELECT
-                $1,
-                text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                context, fact_type, confidence_score, metadata, chunk_id, document_id,
-                COALESCE(
-                    (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
-                    '{{}}'::varchar[]
-                ),
-                observation_scopes_json
-            FROM input_data
-            RETURNING id
-        """
-
    results = await conn.fetch(
-        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[]
+            ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                   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)
+        SELECT
+            $1,
+            text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+            context, fact_type, confidence_score, metadata, chunk_id, document_id,
+            COALESCE(
+                (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
+                '{{}}'::varchar[]
+            )
+        FROM input_data
+        RETURNING id
+        """,
        bank_id,
        fact_texts,
        embeddings,
@@ -149,7 +106,6 @@ async def insert_facts_batch(
        chunk_ids,
        document_ids,
        tags_list,
-        observation_scopes_list,
    )

    unit_ids = [str(row["id"]) for row in results]
@@ -76,14 +76,11 @@ async def retain_batch(
    duplicate_checker_fn,
    bank_id: str,
    contents_dicts: list[RetainContentDict],
-    config,
    document_id: str | None = None,
    is_first_batch: bool = True,
    fact_type_override: str | None = None,
    confidence_score: float | None = None,
    document_tags: list[str] | None = None,
-    operation_id: str | None = None,
-    schema: str | None = None,
 ) -> tuple[list[list[str]], TokenUsage]:
    """
    Process a batch of content through the retain pipeline.
@@ -97,7 +94,6 @@ async def retain_batch(
        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
        document_id: Optional document ID
        is_first_batch: Whether this is the first batch
        fact_type_override: Override fact type for all facts
@@ -142,37 +138,25 @@ async def retain_batch(
            metadata=item.get("metadata", {}),
            entities=item.get("entities", []),
            tags=merged_tags,
-            observation_scopes=item.get("observation_scopes"),
        )
        contents.append(content)

    # Step 1: Extract facts from all contents
    step_start = time.time()

-    extracted_facts, chunks, usage = await fact_extraction.extract_facts_from_contents(
-        contents, llm_config, agent_name, config, pool, operation_id, schema
-    )
+    extracted_facts, chunks, usage = await fact_extraction.extract_facts_from_contents(contents, llm_config, agent_name)
    log_buffer.append(
        f"[1] Extract facts: {len(extracted_facts)} facts, {len(chunks)} chunks from {len(contents)} contents in {time.time() - step_start:.3f}s"
    )

    if not extracted_facts:
-        # Still need to create document if document_id was provided 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():
                await fact_storage.ensure_bank_exists(conn, bank_id)

-                # 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))
-
+                # 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 [])
@@ -197,57 +181,45 @@ 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

@@ -7,7 +7,7 @@ from content input to fact storage.

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


@@ -22,9 +22,6 @@ 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
@@ -34,9 +31,6 @@ class RetainContentDict(TypedDict, total=False):
    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:
@@ -58,9 +52,6 @@ class RetainContent:
    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
@@ -126,9 +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
@@ -177,9 +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."""
@@ -224,7 +209,6 @@ class ProcessedFact:
            chunk_id=chunk_id,
            content_index=extracted_fact.content_index,
            tags=extracted_fact.tags,
-            observation_scopes=extracted_fact.observation_scopes,
        )


@@ -13,10 +13,12 @@ from .reranking import CrossEncoderReranker
 from .retrieval import (
    ParallelRetrievalResult,
    get_default_graph_retriever,
+    retrieve_parallel,
    set_default_graph_retriever,
 )

 __all__ = [
+    "retrieve_parallel",
    "get_default_graph_retriever",
    "set_default_graph_retriever",
    "ParallelRetrievalResult",
@@ -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
@@ -45,7 +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
@@ -216,7 +216,7 @@ class LinkExpansionRetriever(GraphRetriever):
                    -- 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.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
@@ -239,7 +239,7 @@ class LinkExpansionRetriever(GraphRetriever):
                    f"""
                    SELECT
                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                        mu.occurred_end, mu.mentioned_at,
+                        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
@@ -264,7 +264,7 @@ class LinkExpansionRetriever(GraphRetriever):
                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.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
@@ -291,7 +291,7 @@ class LinkExpansionRetriever(GraphRetriever):
                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.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
@@ -305,7 +305,7 @@ class LinkExpansionRetriever(GraphRetriever):
                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.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
@@ -323,12 +323,12 @@ class LinkExpansionRetriever(GraphRetriever):
                )
                SELECT DISTINCT ON (id)
                    id, text, context, event_date, occurred_start,
-                    occurred_end, mentioned_at,
+                    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,
+                         occurred_end, mentioned_at, embedding,
                         fact_type, document_id, chunk_id, tags
                ORDER BY id, score DESC
                LIMIT $4
@@ -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
@@ -85,6 +85,116 @@ def set_default_graph_retriever(retriever: GraphRetriever) -> None:
    _default_graph_retriever = retriever


+async def retrieve_semantic(
+    conn,
+    query_emb_str: str,
+    bank_id: str,
+    fact_type: str,
+    limit: int,
+    tags: list[str] | None = None,
+) -> list[RetrievalResult]:
+    """
+    Semantic retrieval via vector similarity.
+
+    Args:
+        conn: Database connection
+        query_emb_str: Query embedding as string
+        agent_id: bank ID
+        fact_type: Fact type to filter
+        limit: Maximum results to return
+        tags: Optional list of tags for visibility filtering (OR matching)
+
+    Returns:
+        List of RetrievalResult objects
+    """
+    from .tags import TagsMatch, build_tags_where_clause_simple
+
+    tags_clause = build_tags_where_clause_simple(tags, 5)
+    params = [query_emb_str, bank_id, fact_type, limit]
+    if tags:
+        params.append(tags)
+
+    results = await conn.fetch(
+        f"""
+        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
+        FROM {fq_table("memory_units")}
+        WHERE bank_id = $2
+          AND embedding IS NOT NULL
+          AND fact_type = $3
+          AND (1 - (embedding <=> $1::vector)) >= 0.3
+          {tags_clause}
+        ORDER BY embedding <=> $1::vector
+        LIMIT $4
+        """,
+        *params,
+    )
+    return [RetrievalResult.from_db_row(dict(r)) for r in results]
+
+
+async def retrieve_bm25(
+    conn,
+    query_text: str,
+    bank_id: str,
+    fact_type: str,
+    limit: int,
+    tags: list[str] | None = None,
+) -> list[RetrievalResult]:
+    """
+    BM25 keyword retrieval via full-text search.
+
+    Args:
+        conn: Database connection
+        query_text: Query text
+        agent_id: bank ID
+        fact_type: Fact type to filter
+        limit: Maximum results to return
+        tags: Optional list of tags for visibility filtering (OR matching)
+
+    Returns:
+        List of RetrievalResult objects
+    """
+    import re
+
+    from .tags import TagsMatch, build_tags_where_clause_simple
+
+    # Sanitize query text: remove special characters that have meaning in tsquery
+    # Keep only alphanumeric characters and spaces
+    sanitized_text = re.sub(r"[^\w\s]", " ", query_text.lower())
+
+    # Split and filter empty strings
+    tokens = [token for token in sanitized_text.split() if token]
+
+    if not tokens:
+        # If no valid tokens, return empty results
+        return []
+
+    # Convert query to tsquery using OR for more flexible matching
+    # This prevents empty results when some terms are missing
+    query_tsquery = " | ".join(tokens)
+
+    tags_clause = build_tags_where_clause_simple(tags, 5)
+    params = [query_tsquery, bank_id, fact_type, limit]
+    if tags:
+        params.append(tags)
+
+    results = await conn.fetch(
+        f"""
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+               ts_rank_cd(search_vector, to_tsquery('english', $1)) AS bm25_score
+        FROM {fq_table("memory_units")}
+        WHERE bank_id = $2
+          AND fact_type = $3
+          AND search_vector @@ to_tsquery('english', $1)
+          {tags_clause}
+        ORDER BY bm25_score DESC
+        LIMIT $4
+        """,
+        *params,
+    )
+    return [RetrievalResult.from_db_row(dict(r)) for r in results]
+
+
 async def retrieve_semantic_bm25_combined(
    conn,
    query_emb_str: str,
@@ -127,7 +237,7 @@ 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 +249,7 @@ 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
@@ -158,43 +268,20 @@ async def retrieve_semantic_bm25_combined(
                result_dict[ft][0].append(RetrievalResult.from_db_row(row))
        return result_dict

-    # Build BM25 query based on text search backend
-    config = get_config()
+    query_tsquery = " | ".join(tokens)

    # Build tags clause - param 6 if tags provided
    tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
-
-    # Build backend-specific BM25 parts
-    if config.text_search_extension == "vchord":
-        # VectorChord BM25: use <&> operator with to_bm25query and tokenize
-        # Note: VectorChord scores are negative (higher = better, so -1 > -10)
-        bm25_score_expr = "search_vector <&> to_bm25query('idx_memory_units_text_search', tokenize($5, 'llmlingua2'))"
-        bm25_order_by = f"{bm25_score_expr} DESC"
-        bm25_where_filter = ""  # No additional WHERE filter for vchord
-        params = [query_emb_str, bank_id, fact_types, limit, query_text]  # Pass raw query_text for tokenization
-    elif config.text_search_extension == "pg_textsearch":
-        # Timescale pg_textsearch: use <@> operator with to_bm25query
-        # Note: pg_textsearch scores are negative (lower/more negative = better, so -10 > -1)
-        # We negate the score to maintain API consistency (higher = better)
-        bm25_score_expr = "-(text <@> to_bm25query($5, 'idx_memory_units_text_search'))"
-        bm25_order_by = "text <@> to_bm25query($5, 'idx_memory_units_text_search') ASC"
-        bm25_where_filter = ""  # No additional WHERE filter for pg_textsearch
-        params = [query_emb_str, bank_id, fact_types, limit, query_text]
-    else:  # native
-        # Native PostgreSQL: use ts_rank_cd with to_tsquery
-        query_tsquery = " | ".join(tokens)
-        bm25_score_expr = "ts_rank_cd(search_vector, to_tsquery('english', $5))"
-        bm25_order_by = f"{bm25_score_expr} DESC"
-        bm25_where_filter = "AND search_vector @@ to_tsquery('english', $5)"
-        params = [query_emb_str, bank_id, fact_types, limit, query_tsquery]
-
+    params = [query_emb_str, bank_id, fact_types, limit, query_tsquery]
    if tags:
        params.append(tags)

-    # Single query template with backend-specific parts injected
-    query = f"""
+    # Combined CTE query for both semantic and BM25 across all fact types
+    # Uses window functions to limit per fact_type per method
+    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,
@@ -207,35 +294,33 @@ 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,
+                   ts_rank_cd(search_vector, to_tsquery('english', $5)) AS bm25_score,
                   'bm25' AS source,
-                   ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY {bm25_order_by}) AS rn
+                   ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY ts_rank_cd(search_vector, to_tsquery('english', $5)) DESC) AS rn
            FROM {fq_table("memory_units")}
            WHERE bank_id = $2
              AND fact_type = ANY($3)
-              {bm25_where_filter}
+              AND search_vector @@ to_tsquery('english', $5)
              {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
        )
        SELECT * FROM semantic
        UNION ALL
        SELECT * FROM bm25
-    """
-
-    # Combined CTE query for both semantic and BM25 across all fact types
-    # Uses window functions to limit per fact_type per method
-    results = await conn.fetch(query, *params)
+        """,
+        *params,
+    )

    # Group results by fact_type and source
    result_dict: dict[str, tuple[list[RetrievalResult], list[RetrievalResult]]] = {ft: ([], []) for ft in fact_types}
@@ -301,7 +386,7 @@ async def retrieve_temporal_combined(
    entry_points = await conn.fetch(
        f"""
        WITH ranked_entries 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,
                   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")}
@@ -321,7 +406,7 @@ async def retrieve_temporal_combined(
              AND (1 - (embedding <=> $1::vector)) >= $6
              {tags_clause}
        )
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags, similarity
+        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
        """,
@@ -401,7 +486,7 @@ async def retrieve_temporal_combined(

            neighbors = await conn.fetch(
                f"""
-                SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                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 {fq_table("memory_links")} ml
@@ -476,6 +561,623 @@ async def retrieve_temporal_combined(
    return results_by_ft


+async def retrieve_temporal(
+    conn,
+    query_emb_str: str,
+    bank_id: str,
+    fact_type: str,
+    start_date: datetime,
+    end_date: datetime,
+    budget: int,
+    semantic_threshold: float = 0.1,
+    tags: list[str] | None = None,
+) -> list[RetrievalResult]:
+    """
+    Temporal retrieval with spreading activation.
+
+    Strategy:
+    1. Find entry points (facts in date range with semantic relevance)
+    2. Spread through temporal links to related facts
+    3. Score by temporal proximity + semantic similarity + link weight
+
+    Args:
+        conn: Database connection
+        query_emb_str: Query embedding as string
+        agent_id: bank ID
+        fact_type: Fact type to filter
+        start_date: Start of time range
+        end_date: End of time range
+        budget: Node budget for spreading
+        semantic_threshold: Minimum semantic similarity to include
+        tags: Optional list of tags for visibility filtering (OR matching)
+
+    Returns:
+        List of RetrievalResult objects with temporal scores
+    """
+
+    # Ensure start_date and end_date are timezone-aware (UTC) to match database datetimes
+    if start_date.tzinfo is None:
+        start_date = start_date.replace(tzinfo=UTC)
+    if end_date.tzinfo is None:
+        end_date = end_date.replace(tzinfo=UTC)
+
+    from .tags import TagsMatch, build_tags_where_clause_simple
+
+    tags_clause = build_tags_where_clause_simple(tags, 7)
+    params = [query_emb_str, bank_id, fact_type, start_date, end_date, semantic_threshold]
+    if tags:
+        params.append(tags)
+
+    entry_points = await conn.fetch(
+        f"""
+        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
+        FROM {fq_table("memory_units")}
+        WHERE bank_id = $2
+          AND fact_type = $3
+          AND embedding IS NOT NULL
+          AND (
+              -- Match if occurred range overlaps with query range
+              (occurred_start IS NOT NULL AND occurred_end IS NOT NULL
+               AND occurred_start <= $5 AND occurred_end >= $4)
+              OR
+              -- Match if mentioned_at falls within query range
+              (mentioned_at IS NOT NULL AND mentioned_at BETWEEN $4 AND $5)
+              OR
+              -- Match if any occurred date is set and overlaps (even if only start or end is set)
+              (occurred_start IS NOT NULL AND occurred_start BETWEEN $4 AND $5)
+              OR
+              (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
+          )
+          AND (1 - (embedding <=> $1::vector)) >= $6
+          {tags_clause}
+        ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC, (embedding <=> $1::vector) ASC
+        LIMIT 10
+        """,
+        *params,
+    )
+
+    if not entry_points:
+        return []
+
+    # Calculate temporal scores for entry points
+    total_days = (end_date - start_date).total_seconds() / 86400
+    mid_date = start_date + (end_date - start_date) / 2  # Calculate once for all comparisons
+    results = []
+    visited = set()
+
+    for ep in entry_points:
+        unit_id = str(ep["id"])
+        visited.add(unit_id)
+
+        # Calculate temporal proximity using the most relevant date
+        # Priority: occurred_start/end (event time) > mentioned_at (mention time)
+        best_date = None
+        if ep["occurred_start"] is not None and ep["occurred_end"] is not None:
+            # Use midpoint of occurred range
+            best_date = ep["occurred_start"] + (ep["occurred_end"] - ep["occurred_start"]) / 2
+        elif ep["occurred_start"] is not None:
+            best_date = ep["occurred_start"]
+        elif ep["occurred_end"] is not None:
+            best_date = ep["occurred_end"]
+        elif ep["mentioned_at"] is not None:
+            best_date = ep["mentioned_at"]
+
+        # Temporal proximity score (closer to range center = higher score)
+        if best_date:
+            days_from_mid = abs((best_date - mid_date).total_seconds() / 86400)
+            temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
+        else:
+            temporal_proximity = 0.5  # Fallback if no dates (shouldn't happen due to WHERE clause)
+
+        # Create RetrievalResult with temporal scores
+        ep_result = RetrievalResult.from_db_row(dict(ep))
+        ep_result.temporal_score = temporal_proximity
+        ep_result.temporal_proximity = temporal_proximity
+        results.append(ep_result)
+
+    # Spread through temporal links using BATCHED neighbor fetching
+    # Map node_id -> (semantic_sim, temporal_score) for propagation
+    node_scores = {str(ep["id"]): (ep["similarity"], 1.0) for ep in entry_points}
+    frontier = list(node_scores.keys())  # Current batch of nodes to expand
+    budget_remaining = budget - len(entry_points)
+    batch_size = 20  # Process this many nodes per DB query
+
+    while frontier and budget_remaining > 0:
+        # Take a batch from frontier
+        batch_ids = frontier[:batch_size]
+        frontier = frontier[batch_size:]
+
+        # Batch fetch all neighbors for this batch of nodes
+        neighbors = await conn.fetch(
+            f"""
+            SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id,
+                   ml.weight, ml.link_type, ml.from_unit_id,
+                   1 - (mu.embedding <=> $1::vector) AS similarity
+            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
+            ORDER BY ml.weight DESC
+            LIMIT $5
+            """,
+            query_emb_str,
+            batch_ids,
+            fact_type,
+            semantic_threshold,
+            batch_size * 10,  # Allow up to 10 neighbors per node in batch
+        )
+
+        for n in neighbors:
+            neighbor_id = str(n["id"])
+            if neighbor_id in visited:
+                continue
+
+            visited.add(neighbor_id)
+            budget_remaining -= 1
+
+            # Get parent's scores for propagation
+            parent_id = str(n["from_unit_id"])
+            _, parent_temporal_score = node_scores.get(parent_id, (0.5, 0.5))
+
+            # Calculate temporal score for neighbor using best available date
+            neighbor_best_date = None
+            if n["occurred_start"] is not None and n["occurred_end"] is not None:
+                neighbor_best_date = n["occurred_start"] + (n["occurred_end"] - n["occurred_start"]) / 2
+            elif n["occurred_start"] is not None:
+                neighbor_best_date = n["occurred_start"]
+            elif n["occurred_end"] is not None:
+                neighbor_best_date = n["occurred_end"]
+            elif n["mentioned_at"] is not None:
+                neighbor_best_date = n["mentioned_at"]
+
+            if neighbor_best_date:
+                days_from_mid = abs((neighbor_best_date - mid_date).total_seconds() / 86400)
+                neighbor_temporal_proximity = (
+                    1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
+                )
+            else:
+                neighbor_temporal_proximity = 0.3  # Lower score if no temporal data
+
+            # Boost causal links (same as graph retrieval)
+            link_type = n["link_type"]
+            if link_type in ("causes", "caused_by"):
+                causal_boost = 2.0
+            elif link_type in ("enables", "prevents"):
+                causal_boost = 1.5
+            else:
+                causal_boost = 1.0
+
+            # Propagate temporal score through links (decay, with causal boost)
+            propagated_temporal = parent_temporal_score * n["weight"] * causal_boost * 0.7
+
+            # Combined temporal score
+            combined_temporal = max(neighbor_temporal_proximity, propagated_temporal)
+
+            # Create RetrievalResult with temporal scores
+            neighbor_result = RetrievalResult.from_db_row(dict(n))
+            neighbor_result.temporal_score = combined_temporal
+            neighbor_result.temporal_proximity = neighbor_temporal_proximity
+            results.append(neighbor_result)
+
+            # Track scores for propagation and add to frontier
+            if budget_remaining > 0 and combined_temporal > 0.2:
+                node_scores[neighbor_id] = (n["similarity"], combined_temporal)
+                frontier.append(neighbor_id)
+
+            if budget_remaining <= 0:
+                break
+
+    return results
+
+
+async def retrieve_parallel(
+    pool,
+    query_text: str,
+    query_embedding_str: str,
+    bank_id: str,
+    fact_type: str,
+    thinking_budget: int,
+    question_date: datetime | None = None,
+    query_analyzer: Optional["QueryAnalyzer"] = None,
+    graph_retriever: GraphRetriever | None = None,
+    temporal_constraint: tuple | None = None,  # Pre-extracted temporal constraint
+    tags: list[str] | None = None,  # Visibility scope tags for filtering
+) -> ParallelRetrievalResult:
+    """
+    Run 3-way or 4-way parallel retrieval (adds temporal if detected).
+
+    Args:
+        pool: Database connection pool
+        query_text: Query text
+        query_embedding_str: Query embedding as string
+        bank_id: Bank ID
+        fact_type: Fact type to filter
+        thinking_budget: Budget for graph traversal and retrieval limits
+        question_date: Optional date when question was asked (for temporal filtering)
+        query_analyzer: Query analyzer to use (defaults to TransformerQueryAnalyzer)
+        graph_retriever: Graph retrieval strategy (defaults to configured retriever)
+        temporal_constraint: Pre-extracted temporal constraint (optional)
+        tags: Optional list of tags for visibility filtering (OR matching)
+
+    Returns:
+        ParallelRetrievalResult with semantic, bm25, graph, temporal results and timings
+    """
+    retriever = graph_retriever or get_default_graph_retriever()
+
+    # Use optimized parallel path for MPFP and LinkExpansion (runs all methods truly in parallel)
+    # BFS uses legacy path that extracts temporal constraint upfront
+    if retriever.name in ("mpfp", "link_expansion"):
+        return await _retrieve_parallel_mpfp(
+            pool,
+            query_text,
+            query_embedding_str,
+            bank_id,
+            fact_type,
+            thinking_budget,
+            temporal_constraint,
+            retriever,
+            question_date,
+            query_analyzer,
+            tags=tags,
+        )
+    else:
+        # For BFS, extract temporal constraint upfront (legacy path)
+        if temporal_constraint is None:
+            from .temporal_extraction import extract_temporal_constraint
+
+            temporal_constraint = extract_temporal_constraint(
+                query_text, reference_date=question_date, analyzer=query_analyzer
+            )
+        return await _retrieve_parallel_bfs(
+            pool,
+            query_text,
+            query_embedding_str,
+            bank_id,
+            fact_type,
+            thinking_budget,
+            temporal_constraint,
+            retriever,
+            tags=tags,
+        )
+
+
+@dataclass
+class _TimedResult:
+    """Internal result with timing."""
+
+    results: list[RetrievalResult]
+    time: float
+    conn_wait: float = 0.0  # Connection acquisition wait time
+
+
+async def _retrieve_parallel_mpfp(
+    pool,
+    query_text: str,
+    query_embedding_str: str,
+    bank_id: str,
+    fact_type: str,
+    thinking_budget: int,
+    temporal_constraint: tuple | None,
+    retriever: GraphRetriever,
+    question_date: datetime | None = None,
+    query_analyzer=None,
+    tags: list[str] | None = None,
+) -> ParallelRetrievalResult:
+    """
+    MPFP retrieval with true parallelization.
+
+    All methods run independently in parallel:
+    - Semantic: vector similarity search
+    - BM25: keyword search
+    - Graph: MPFP traversal (does its own semantic seeds internally)
+    - Temporal: date extraction (if needed) + date-range search
+
+    Temporal extraction runs IN PARALLEL with other retrievals, so even if
+    dateparser is slow, it doesn't block semantic/BM25/graph.
+    """
+    import time
+
+    async def run_semantic() -> _TimedResult:
+        """Independent semantic retrieval."""
+        start = time.time()
+        acquire_start = time.time()
+        async with acquire_with_retry(pool) as conn:
+            conn_wait = time.time() - acquire_start
+            results = await retrieve_semantic(
+                conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget, tags=tags
+            )
+        return _TimedResult(results, time.time() - start, conn_wait)
+
+    async def run_bm25() -> _TimedResult:
+        """Independent BM25 retrieval."""
+        start = time.time()
+        acquire_start = time.time()
+        async with acquire_with_retry(pool) as conn:
+            conn_wait = time.time() - acquire_start
+            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget, tags=tags)
+        return _TimedResult(results, time.time() - start, conn_wait)
+
+    async def run_graph() -> tuple[list[RetrievalResult], float, MPFPTimings | None]:
+        """Independent graph retrieval - does its own semantic seeds."""
+        start = time.time()
+
+        # MPFP does its own semantic seeds via _find_semantic_seeds
+        # Note: temporal_seeds not used here to avoid dependency on temporal extraction
+        results, mpfp_timing = await retriever.retrieve(
+            pool=pool,
+            query_embedding_str=query_embedding_str,
+            bank_id=bank_id,
+            fact_type=fact_type,
+            budget=thinking_budget,
+            query_text=query_text,
+            semantic_seeds=None,  # Let MPFP find its own seeds
+            temporal_seeds=None,  # Don't wait for temporal extraction
+            tags=tags,
+        )
+        return results, time.time() - start, mpfp_timing
+
+    @dataclass
+    class _TemporalWithConstraint:
+        """Temporal results with the extracted constraint."""
+
+        results: list[RetrievalResult]
+        time: float
+        constraint: tuple | None
+        extraction_time: float  # Time spent in query analyzer (dateparser)
+        conn_wait: float = 0.0  # Connection acquisition wait time
+
+    async def run_temporal_with_extraction() -> _TemporalWithConstraint:
+        """
+        Extract temporal constraint AND run temporal retrieval.
+
+        This runs in parallel with semantic/BM25/graph, so dateparser
+        latency doesn't block other retrievals.
+        """
+        start = time.time()
+
+        # Use pre-provided constraint if available
+        tc = temporal_constraint
+        extraction_time = 0.0
+
+        # Otherwise extract from query (this is the potentially slow dateparser call)
+        if tc is None:
+            from .temporal_extraction import extract_temporal_constraint
+
+            extraction_start = time.time()
+            tc = extract_temporal_constraint(query_text, reference_date=question_date, analyzer=query_analyzer)
+            extraction_time = time.time() - extraction_start
+
+        # If no temporal constraint found, return empty (but still report extraction time)
+        if tc is None:
+            return _TemporalWithConstraint([], time.time() - start, None, extraction_time, 0.0)
+
+        # Run temporal retrieval with the extracted constraint
+        tc_start, tc_end = tc
+        acquire_start = time.time()
+        async with acquire_with_retry(pool) as conn:
+            conn_wait = time.time() - acquire_start
+            results = await retrieve_temporal(
+                conn,
+                query_embedding_str,
+                bank_id,
+                fact_type,
+                tc_start,
+                tc_end,
+                budget=thinking_budget,
+                semantic_threshold=0.1,
+            )
+        return _TemporalWithConstraint(results, time.time() - start, tc, extraction_time, conn_wait)
+
+    # Run ALL methods in parallel (including temporal extraction!)
+    semantic_result, bm25_result, graph_result, temporal_result = await asyncio.gather(
+        run_semantic(),
+        run_bm25(),
+        run_graph(),
+        run_temporal_with_extraction(),
+    )
+    graph_results, graph_time, mpfp_timing = graph_result
+
+    # Compute max connection wait across all methods (graph handles its own connections)
+    max_conn_wait = max(semantic_result.conn_wait, bm25_result.conn_wait, temporal_result.conn_wait)
+
+    return ParallelRetrievalResult(
+        semantic=semantic_result.results,
+        bm25=bm25_result.results,
+        graph=graph_results,
+        temporal=temporal_result.results if temporal_result.results else None,
+        timings={
+            "semantic": semantic_result.time,
+            "bm25": bm25_result.time,
+            "graph": graph_time,
+            "temporal": temporal_result.time,
+            "temporal_extraction": temporal_result.extraction_time,
+        },
+        temporal_constraint=temporal_result.constraint,
+        mpfp_timings=[mpfp_timing] if mpfp_timing else [],
+        max_conn_wait=max_conn_wait,
+    )
+
+
+async def _get_temporal_entry_points(
+    conn,
+    query_embedding_str: str,
+    bank_id: str,
+    fact_type: str,
+    start_date: datetime,
+    end_date: datetime,
+    limit: int = 20,
+    semantic_threshold: float = 0.1,
+) -> list[RetrievalResult]:
+    """Get temporal entry points (facts in date range with semantic relevance)."""
+
+    if start_date.tzinfo is None:
+        start_date = start_date.replace(tzinfo=UTC)
+    if end_date.tzinfo is None:
+        end_date = end_date.replace(tzinfo=UTC)
+
+    rows = await conn.fetch(
+        f"""
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at,
+               embedding, fact_type, document_id, chunk_id,
+               1 - (embedding <=> $1::vector) AS similarity
+        FROM {fq_table("memory_units")}
+        WHERE bank_id = $2
+          AND fact_type = $3
+          AND embedding IS NOT NULL
+          AND (
+              (occurred_start IS NOT NULL AND occurred_end IS NOT NULL
+               AND occurred_start <= $5 AND occurred_end >= $4)
+              OR (mentioned_at IS NOT NULL AND mentioned_at BETWEEN $4 AND $5)
+              OR (occurred_start IS NOT NULL AND occurred_start BETWEEN $4 AND $5)
+              OR (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
+          )
+          AND (1 - (embedding <=> $1::vector)) >= $6
+        ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC,
+                 (embedding <=> $1::vector) ASC
+        LIMIT $7
+        """,
+        query_embedding_str,
+        bank_id,
+        fact_type,
+        start_date,
+        end_date,
+        semantic_threshold,
+        limit,
+    )
+
+    results = []
+    total_days = max((end_date - start_date).total_seconds() / 86400, 1)
+    mid_date = start_date + (end_date - start_date) / 2
+
+    for row in rows:
+        result = RetrievalResult.from_db_row(dict(row))
+
+        # Calculate temporal proximity score
+        best_date = None
+        if row["occurred_start"] and row["occurred_end"]:
+            best_date = row["occurred_start"] + (row["occurred_end"] - row["occurred_start"]) / 2
+        elif row["occurred_start"]:
+            best_date = row["occurred_start"]
+        elif row["occurred_end"]:
+            best_date = row["occurred_end"]
+        elif row["mentioned_at"]:
+            best_date = row["mentioned_at"]
+
+        if best_date:
+            days_from_mid = abs((best_date - mid_date).total_seconds() / 86400)
+            result.temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0)
+        else:
+            result.temporal_proximity = 0.5
+
+        result.temporal_score = result.temporal_proximity
+        results.append(result)
+
+    return results
+
+
+async def _retrieve_parallel_bfs(
+    pool,
+    query_text: str,
+    query_embedding_str: str,
+    bank_id: str,
+    fact_type: str,
+    thinking_budget: int,
+    temporal_constraint: tuple | None,
+    retriever: GraphRetriever,
+    tags: list[str] | None = None,
+) -> ParallelRetrievalResult:
+    """BFS retrieval: all methods run in parallel (original behavior)."""
+    import time
+
+    async def run_semantic() -> _TimedResult:
+        start = time.time()
+        async with acquire_with_retry(pool) as conn:
+            results = await retrieve_semantic(
+                conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget, tags=tags
+            )
+        return _TimedResult(results, time.time() - start)
+
+    async def run_bm25() -> _TimedResult:
+        start = time.time()
+        async with acquire_with_retry(pool) as conn:
+            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget, tags=tags)
+        return _TimedResult(results, time.time() - start)
+
+    async def run_graph() -> _TimedResult:
+        start = time.time()
+        results, _ = await retriever.retrieve(
+            pool=pool,
+            query_embedding_str=query_embedding_str,
+            bank_id=bank_id,
+            fact_type=fact_type,
+            budget=thinking_budget,
+            query_text=query_text,
+            tags=tags,
+        )
+        return _TimedResult(results, time.time() - start)
+
+    async def run_temporal(tc_start, tc_end) -> _TimedResult:
+        start = time.time()
+        async with acquire_with_retry(pool) as conn:
+            results = await retrieve_temporal(
+                conn,
+                query_embedding_str,
+                bank_id,
+                fact_type,
+                tc_start,
+                tc_end,
+                budget=thinking_budget,
+                semantic_threshold=0.1,
+                tags=tags,
+            )
+        return _TimedResult(results, time.time() - start)
+
+    if temporal_constraint:
+        tc_start, tc_end = temporal_constraint
+        semantic_r, bm25_r, graph_r, temporal_r = await asyncio.gather(
+            run_semantic(),
+            run_bm25(),
+            run_graph(),
+            run_temporal(tc_start, tc_end),
+        )
+        return ParallelRetrievalResult(
+            semantic=semantic_r.results,
+            bm25=bm25_r.results,
+            graph=graph_r.results,
+            temporal=temporal_r.results,
+            timings={
+                "semantic": semantic_r.time,
+                "bm25": bm25_r.time,
+                "graph": graph_r.time,
+                "temporal": temporal_r.time,
+            },
+            temporal_constraint=temporal_constraint,
+        )
+    else:
+        semantic_r, bm25_r, graph_r = await asyncio.gather(
+            run_semantic(),
+            run_bm25(),
+            run_graph(),
+        )
+        return ParallelRetrievalResult(
+            semantic=semantic_r.results,
+            bm25=bm25_r.results,
+            graph=graph_r.results,
+            temporal=None,
+            timings={
+                "semantic": semantic_r.time,
+                "bm25": bm25_r.time,
+                "graph": graph_r.time,
+            },
+            temporal_constraint=None,
+        )
+
+
 async def retrieve_all_fact_types_parallel(
    pool,
    query_text: str,
@@ -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,
@@ -1,79 +0,0 @@
-"""File storage backends for uploaded files."""
-
-from collections.abc import Callable
-
-from .base import FileStorage
-from .postgresql import PostgreSQLFileStorage
-
-__all__ = ["FileStorage", "PostgreSQLFileStorage", "create_file_storage"]
-
-
-def create_file_storage(
-    storage_type: str,
-    pool_getter: Callable | None = None,
-    schema: str | None = None,
-    schema_getter: Callable | None = None,
-    **kwargs,
-) -> FileStorage:
-    """
-    Create file storage backend based on configuration.
-
-    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)
-        **kwargs: Additional args passed to storage backend
-
-    Returns:
-        FileStorage instance
-
-    Raises:
-        ValueError: If storage_type is unknown or required args are missing
-    """
-    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)
-    elif storage_type == "s3":
-        from ...config import get_config
-        from .s3 import S3FileStorage
-
-        config = get_config()
-        bucket = config.file_storage_s3_bucket
-        if not bucket:
-            raise ValueError("HINDSIGHT_API_FILE_STORAGE_S3_BUCKET is required for S3 storage")
-        return S3FileStorage(
-            bucket=bucket,
-            region=config.file_storage_s3_region,
-            endpoint=config.file_storage_s3_endpoint,
-            access_key_id=config.file_storage_s3_access_key_id,
-            secret_access_key=config.file_storage_s3_secret_access_key,
-        )
-    elif storage_type == "gcs":
-        from ...config import get_config
-        from .gcs import GCSFileStorage
-
-        config = get_config()
-        bucket = config.file_storage_gcs_bucket
-        if not bucket:
-            raise ValueError("HINDSIGHT_API_FILE_STORAGE_GCS_BUCKET is required for GCS storage")
-        return GCSFileStorage(
-            bucket=bucket,
-            service_account_key=config.file_storage_gcs_service_account_key,
-        )
-    elif storage_type == "azure":
-        from ...config import get_config
-        from .azure import AzureFileStorage
-
-        config = get_config()
-        container = config.file_storage_azure_container
-        if not container:
-            raise ValueError("HINDSIGHT_API_FILE_STORAGE_AZURE_CONTAINER is required for Azure storage")
-        return AzureFileStorage(
-            container_name=container,
-            account_name=config.file_storage_azure_account_name,
-            account_key=config.file_storage_azure_account_key,
-        )
-    else:
-        raise ValueError(f"Unknown storage type: {storage_type}. Supported: 'native', 's3', 'gcs', 'azure'.")
@@ -1,62 +0,0 @@
-"""Azure Blob Storage backend using obstore."""
-
-import logging
-from datetime import timedelta
-
-import obstore as obs
-from obstore.store import AzureStore
-
-from .base import FileStorage
-
-logger = logging.getLogger(__name__)
-
-
-class AzureFileStorage(FileStorage):
-    """
-    Azure Blob Storage backend.
-
-    Uses obstore (Rust-backed) for high-throughput async access to Azure Blob Storage.
-    Supports account key, SAS token, and default Azure credentials.
-    """
-
-    def __init__(
-        self,
-        container_name: str,
-        account_name: str | None = None,
-        account_key: str | None = None,
-    ):
-        kwargs: dict = {}
-        if account_name:
-            kwargs["account_name"] = account_name
-        if account_key:
-            kwargs["account_key"] = account_key
-
-        self._store = AzureStore(container_name, **kwargs)
-        logger.info(f"Initialized Azure file storage: container={container_name}, account={account_name}")
-
-    async def store(self, file_data: bytes, key: str, metadata: dict[str, str] | None = None) -> str:
-        await obs.put_async(self._store, key, file_data)
-        logger.debug(f"Stored file {key} ({len(file_data)} bytes) in Azure")
-        return key
-
-    async def retrieve(self, key: str) -> bytes:
-        try:
-            response = await obs.get_async(self._store, key)
-            return await response.bytes_async()
-        except Exception as e:
-            if "not found" in str(e).lower() or "BlobNotFound" in str(e):
-                raise FileNotFoundError(f"File not found: {key}") from e
-            raise
-
-    async def delete(self, key: str) -> None:
-        await obs.delete_async(self._store, key)
-
-    async def exists(self, key: str) -> bool:
-        try:
-            await obs.head_async(self._store, key)
-            return True
-        except Exception:
-            return False
-
-    async def get_download_url(self, key: str, expires_in: int = 3600) -> str:
-        return await obs.sign_async(self._store, "GET", key, timedelta(seconds=expires_in))
@@ -1,83 +0,0 @@
-"""Abstract base class for file storage backends."""
-
-from abc import ABC, abstractmethod
-
-
-class FileStorage(ABC):
-    """Abstract base for file storage backends."""
-
-    @abstractmethod
-    async def store(
-        self,
-        file_data: bytes,
-        key: str,
-        metadata: dict[str, str] | None = None,
-    ) -> str:
-        """
-        Store file and return storage key.
-
-        Args:
-            file_data: Raw file bytes
-            key: Storage key (e.g., "banks/{bank_id}/files/{file_id}.pdf")
-            metadata: Optional metadata to store with file
-
-        Returns:
-            Storage key that can be used to retrieve the file
-        """
-        pass
-
-    @abstractmethod
-    async def retrieve(self, key: str) -> bytes:
-        """
-        Retrieve file by storage key.
-
-        Args:
-            key: Storage key
-
-        Returns:
-            File data as bytes
-
-        Raises:
-            FileNotFoundError: If file does not exist
-        """
-        pass
-
-    @abstractmethod
-    async def delete(self, key: str) -> None:
-        """
-        Delete file by storage key.
-
-        Args:
-            key: Storage key
-        """
-        pass
-
-    @abstractmethod
-    async def exists(self, key: str) -> bool:
-        """
-        Check if file exists.
-
-        Args:
-            key: Storage key
-
-        Returns:
-            True if file exists, False otherwise
-        """
-        pass
-
-    @abstractmethod
-    async def get_download_url(self, key: str, expires_in: int = 3600) -> str:
-        """
-        Get a URL for downloading the file.
-
-        For PostgreSQL storage, this might be a relative API path.
-        For S3, this would be a pre-signed URL.
-
-        Args:
-            key: Storage key
-            expires_in: Expiration time in seconds (may be ignored for some backends)
-
-        Returns:
-            Download URL or path
-        """
-        pass
@@ -1,59 +0,0 @@
-"""Google Cloud Storage backend using obstore."""
-
-import logging
-from datetime import timedelta
-
-import obstore as obs
-from obstore.store import GCSStore
-
-from .base import FileStorage
-
-logger = logging.getLogger(__name__)
-
-
-class GCSFileStorage(FileStorage):
-    """
-    Google Cloud Storage backend.
-
-    Uses obstore (Rust-backed) for high-throughput async access to GCS.
-    Supports Application Default Credentials, service account keys, and explicit credentials.
-    """
-
-    def __init__(
-        self,
-        bucket: str,
-        service_account_key: str | None = None,
-    ):
-        kwargs: dict = {}
-        if service_account_key:
-            kwargs["service_account_key"] = service_account_key
-
-        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:
-        await obs.put_async(self._store, key, file_data)
-        logger.debug(f"Stored file {key} ({len(file_data)} bytes) in GCS")
-        return key
-
-    async def retrieve(self, key: str) -> bytes:
-        try:
-            response = await obs.get_async(self._store, key)
-            return await response.bytes_async()
-        except Exception as e:
-            if "not found" in str(e).lower():
-                raise FileNotFoundError(f"File not found: {key}") from e
-            raise
-
-    async def delete(self, key: str) -> None:
-        await obs.delete_async(self._store, key)
-
-    async def exists(self, key: str) -> bool:
-        try:
-            await obs.head_async(self._store, key)
-            return True
-        except Exception:
-            return False
-
-    async def get_download_url(self, key: str, expires_in: int = 3600) -> str:
-        return await obs.sign_async(self._store, "GET", key, timedelta(seconds=expires_in))
@@ -1,153 +0,0 @@
-"""PostgreSQL BYTEA-based file storage (default, zero-config)."""
-
-import logging
-from collections.abc import Callable
-from typing import TYPE_CHECKING
-
-if TYPE_CHECKING:
-    import asyncpg
-
-from .base import FileStorage
-
-logger = logging.getLogger(__name__)
-
-
-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
-
-
-class PostgreSQLFileStorage(FileStorage):
-    """
-    PostgreSQL BYTEA-based file storage.
-
-    Stores files directly in PostgreSQL using BYTEA columns.
-    This is the default storage backend - zero configuration required!
-
-    Pros:
-    - Works out of the box (no external dependencies)
-    - Transactional consistency with database
-    - Simple backups (included in pg_dump)
-    - Good performance for <10MB files
-
-    Cons:
-    - Database bloat for large/many files
-    - Not ideal for distributed deployments
-    - Higher cost than object storage at scale
-
-    For production/scale, consider S3FileStorage instead.
-    """
-
-    def __init__(
-        self,
-        pool_getter: Callable[[], "asyncpg.Pool"],
-        schema: str | None = None,
-        schema_getter: Callable[[], 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)
-        """
-        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
-
-    async def store(
-        self,
-        file_data: bytes,
-        key: str,
-        metadata: dict[str, str] | None = None,
-    ) -> str:
-        """Store file in PostgreSQL."""
-        pool = self._pool_getter()
-
-        async with pool.acquire() as conn:
-            await conn.execute(
-                f"""
-                INSERT INTO {fq_table("file_storage", self._schema)}
-                (storage_key, data)
-                VALUES ($1, $2)
-                ON CONFLICT (storage_key) DO UPDATE SET
-                    data = EXCLUDED.data
-                """,
-                key,
-                file_data,
-            )
-
-        logger.debug(f"Stored file {key} ({len(file_data)} bytes) in PostgreSQL")
-        return key
-
-    async def retrieve(self, key: str) -> bytes:
-        """Retrieve file from PostgreSQL."""
-        pool = self._pool_getter()
-
-        async with pool.acquire() as conn:
-            row = await conn.fetchrow(
-                f"""
-                SELECT data FROM {fq_table("file_storage", self._schema)}
-                WHERE storage_key = $1
-                """,
-                key,
-            )
-
-            if not row:
-                raise FileNotFoundError(f"File not found: {key}")
-
-            return bytes(row["data"])
-
-    async def delete(self, key: str) -> None:
-        """Delete file from PostgreSQL."""
-        pool = self._pool_getter()
-
-        async with pool.acquire() as conn:
-            result = await conn.execute(
-                f"""
-                DELETE FROM {fq_table("file_storage", self._schema)}
-                WHERE storage_key = $1
-                """,
-                key,
-            )
-
-            # Check if anything was deleted
-            if result == "DELETE 0":
-                logger.warning(f"Attempted to delete non-existent file: {key}")
-
-    async def exists(self, key: str) -> bool:
-        """Check if file exists in PostgreSQL."""
-        pool = self._pool_getter()
-
-        async with pool.acquire() as conn:
-            row = await conn.fetchrow(
-                f"""
-                SELECT 1 FROM {fq_table("file_storage", self._schema)}
-                WHERE storage_key = $1
-                """,
-                key,
-            )
-
-            return row is not None
-
-    async def get_download_url(self, key: str, expires_in: int = 3600) -> str:
-        """
-        Get download URL for PostgreSQL-stored file.
-
-        Returns an API endpoint path (not a pre-signed URL since the file
-        is stored in the database). The expires_in parameter is ignored
-        for PostgreSQL storage.
-        """
-        # Return API path for download endpoint
-        # (expires_in ignored for database storage - auth handled at API level)
-        return f"/v1/default/files/download/{key}"
@@ -1,71 +0,0 @@
-"""S3 object storage backend using obstore."""
-
-import logging
-from datetime import timedelta
-
-import obstore as obs
-from obstore.store import S3Store
-
-from .base import FileStorage
-
-logger = logging.getLogger(__name__)
-
-
-class S3FileStorage(FileStorage):
-    """
-    S3-compatible object storage backend.
-
-    Uses obstore (Rust-backed) for high-throughput async access to
-    Amazon S3, MinIO, Cloudflare R2, and other S3-compliant APIs.
-    """
-
-    def __init__(
-        self,
-        bucket: str,
-        region: str | None = None,
-        endpoint: str | None = None,
-        access_key_id: str | None = None,
-        secret_access_key: str | None = None,
-    ):
-        kwargs: dict = {}
-        if region:
-            kwargs["region"] = region
-        if endpoint:
-            kwargs["endpoint"] = endpoint
-            # Allow plain HTTP for local S3-compatible services (MinIO, LocalStack, etc.)
-            if endpoint.startswith("http://"):
-                kwargs["allow_http"] = True
-        if access_key_id:
-            kwargs["access_key_id"] = access_key_id
-        if secret_access_key:
-            kwargs["secret_access_key"] = secret_access_key
-
-        self._store = S3Store(bucket, **kwargs)
-        logger.info(f"Initialized S3 file storage: bucket={bucket}, region={region}, endpoint={endpoint}")
-
-    async def store(self, file_data: bytes, key: str, metadata: dict[str, str] | None = None) -> str:
-        await obs.put_async(self._store, key, file_data)
-        logger.debug(f"Stored file {key} ({len(file_data)} bytes) in S3")
-        return key
-
-    async def retrieve(self, key: str) -> bytes:
-        try:
-            response = await obs.get_async(self._store, key)
-            return await response.bytes_async()
-        except Exception as e:
-            if "not found" in str(e).lower() or "NoSuchKey" in str(e):
-                raise FileNotFoundError(f"File not found: {key}") from e
-            raise
-
-    async def delete(self, key: str) -> None:
-        await obs.delete_async(self._store, key)
-
-    async def exists(self, key: str) -> bool:
-        try:
-            await obs.head_async(self._store, key)
-            return True
-        except Exception:
-            return False
-
-    async def get_download_url(self, key: str, expires_in: int = 3600) -> str:
-        return await obs.sign_async(self._store, "GET", key, timedelta(seconds=expires_in))
@@ -19,7 +19,6 @@ async def extract_facts(
    context: str = "",
    llm_config: "LLMConfig" = None,
    agent_name: str = None,
-    config=None,
 ) -> tuple[list["Fact"], list[tuple[str, int]]]:
    """
    Extract semantic facts from text using LLM.
@@ -36,7 +35,6 @@ async def extract_facts(
        context: Context about the conversation/document
        llm_config: LLM configuration to use
        agent_name: Optional agent name to help identify agent-related facts
-        config: HindsightConfig to use (defaults to global config if not provided)

    Returns:
        Tuple of (facts, chunks) where:
@@ -49,19 +47,12 @@ async def extract_facts(
    if not text or not text.strip():
        return [], []

-    # Use provided config or fall back to global config
-    if config is None:
-        from ..config import _get_raw_config
-
-        config = _get_raw_config()
-
    facts, chunks, _ = await extract_facts_from_text(
        text,
        event_date,
+        context=context,
        llm_config=llm_config,
        agent_name=agent_name,
-        config=config,
-        context=context,
    )

    if not facts:
@@ -22,11 +22,6 @@ 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,
@@ -75,11 +70,6 @@ __all__ = [
    "RetainContext",
    "RetainResult",
    "ValidationResult",
-    # Operation Validator - Bank Management
-    "BankListContext",
-    "BankListResult",
-    "BankReadContext",
-    "BankWriteContext",
    # Operation Validator - Consolidation
    "ConsolidateContext",
    "ConsolidateResult",
@@ -96,13 +96,7 @@ class DefaultExtensionContext(ExtensionContext):

    async def run_migration(self, schema: str) -> None:
        """Run migrations for a specific schema."""
-        from hindsight_api.config import get_config
-        from hindsight_api.migrations import (
-            ensure_embedding_dimension,
-            ensure_text_search_extension,
-            ensure_vector_extension,
-            run_migrations,
-        )
+        from hindsight_api.migrations import ensure_embedding_dimension, run_migrations

        # Prefer getting URL from memory engine (handles pg0 case where URL is set after init)
        db_url = self._database_url
@@ -113,9 +107,6 @@ class DefaultExtensionContext(ExtensionContext):

        run_migrations(db_url, schema=schema)

-        # Get config for vector extension setting
-        config = get_config()
-
        # Ensure embedding column dimension matches the model's dimension
        # This is needed because migrations create columns with default dimension
        if self._memory_engine is not None:
@@ -123,15 +114,7 @@ class DefaultExtensionContext(ExtensionContext):
            if embeddings is not None:
                dimension = getattr(embeddings, "dimension", None)
                if dimension is not None:
-                    ensure_embedding_dimension(
-                        db_url, dimension, schema=schema, vector_extension=config.vector_extension
-                    )
-
-        # Ensure vector indexes match the configured extension
-        ensure_vector_extension(db_url, vector_extension=config.vector_extension, schema=schema)
-
-        # Ensure text search columns/indexes match the configured extension
-        ensure_text_search_extension(db_url, text_search_extension=config.text_search_extension, schema=schema)
+                    ensure_embedding_dimension(db_url, dimension, schema=schema)

    def get_memory_engine(self) -> "MemoryEngineInterface":
        """Get the memory engine interface."""
@@ -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
 # =============================================================================
@@ -573,63 +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)
@@ -2,7 +2,6 @@

 from abc import ABC, abstractmethod
 from dataclasses import dataclass
-from typing import Any

 from hindsight_api.extensions.base import Extension
 from hindsight_api.models import RequestContext
@@ -89,54 +88,6 @@ class TenantExtension(Extension, ABC):
        """
        ...

-    async def get_tenant_config(self, context: RequestContext) -> dict[str, Any]:
-        """
-        Get tenant-specific configuration overrides.
-
-        This method is called during hierarchical configuration resolution to get
-        tenant-level config overrides. The returned dict should contain Python field
-        names (lowercase snake_case) as keys, not environment variable names.
-
-        Example:
-            {"llm_model": "gpt-4", "retain_extraction_mode": "verbose"}
-
-        The default implementation returns an empty dict (no tenant-specific config).
-        Override this method in custom extensions to provide tenant-specific configuration.
-
-        Args:
-            context: The request context containing tenant information.
-
-        Returns:
-            Dict of config field names to values (only configurable fields).
-            Empty dict if no tenant-specific config.
-        """
-        return {}
-
-    async def get_allowed_config_fields(self, context: RequestContext, bank_id: str) -> set[str] | None:
-        """
-        Get set of config fields that this tenant/bank is allowed to modify.
-
-        This method controls which configurable fields can be modified via the bank config API.
-        It enables fine-grained permission control per tenant or per bank.
-
-        Examples:
-            - Return None: Allow all configurable fields (default)
-            - Return {"retain_chunk_size", "retain_custom_instructions"}: Allow only these fields
-            - Return set(): Allow no modifications (read-only)
-
-        The default implementation returns None (all configurable fields allowed).
-        Override this method in custom extensions to implement custom permission logic.
-
-        Args:
-            context: The request context containing tenant information.
-            bank_id: The bank identifier for per-bank permissions.
-
-        Returns:
-            Set of allowed field names, or None to allow all configurable fields.
-            Returned fields must be a subset of HindsightConfig.get_configurable_fields().
-        """
-        return None
-
    async def authenticate_mcp(self, context: RequestContext) -> TenantContext:
        """
        Authenticate MCP requests.
@@ -23,7 +23,7 @@ import uvicorn
 from . import MemoryEngine, __version__
 from .api import create_app
 from .banner import print_banner
-from .config import DEFAULT_WORKERS, ENV_WORKERS, HindsightConfig, _get_raw_config
+from .config import DEFAULT_WORKERS, ENV_WORKERS, HindsightConfig, get_config
 from .daemon import (
    DEFAULT_DAEMON_PORT,
    DEFAULT_IDLE_TIMEOUT,
@@ -68,7 +68,7 @@ def main():
    global _memory

    # Load configuration from environment (for CLI args defaults)
-    config = _get_raw_config()
+    config = get_config()

    parser = argparse.ArgumentParser(
        prog="hindsight-api",
@@ -155,8 +155,6 @@ def main():
        config = HindsightConfig(
            database_url=config.database_url,
            database_schema=config.database_schema,
-            vector_extension=config.vector_extension,
-            text_search_extension=config.text_search_extension,
            llm_provider=config.llm_provider,
            llm_api_key=config.llm_api_key,
            llm_model=config.llm_model,
@@ -166,8 +164,6 @@ def main():
            llm_initial_backoff=config.llm_initial_backoff,
            llm_max_backoff=config.llm_max_backoff,
            llm_timeout=config.llm_timeout,
-            llm_groq_service_tier=config.llm_groq_service_tier,
-            llm_openai_service_tier=config.llm_openai_service_tier,
            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,
@@ -210,9 +206,6 @@ def main():
            embeddings_litellm_api_base=config.embeddings_litellm_api_base,
            embeddings_litellm_api_key=config.embeddings_litellm_api_key,
            embeddings_litellm_model=config.embeddings_litellm_model,
-            embeddings_litellm_sdk_api_key=config.embeddings_litellm_sdk_api_key,
-            embeddings_litellm_sdk_model=config.embeddings_litellm_sdk_model,
-            embeddings_litellm_sdk_api_base=config.embeddings_litellm_sdk_api_base,
            reranker_provider=config.reranker_provider,
            reranker_local_model=config.reranker_local_model,
            reranker_local_force_cpu=config.reranker_local_force_cpu,
@@ -228,19 +221,11 @@ def main():
            reranker_litellm_api_base=config.reranker_litellm_api_base,
            reranker_litellm_api_key=config.reranker_litellm_api_key,
            reranker_litellm_model=config.reranker_litellm_model,
-            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,
            recall_max_concurrent=config.recall_max_concurrent,
@@ -249,34 +234,10 @@ 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_batch_enabled=config.retain_batch_enabled,
-            retain_batch_poll_interval_seconds=config.retain_batch_poll_interval_seconds,
-            file_storage_type=config.file_storage_type,
-            file_storage_s3_bucket=config.file_storage_s3_bucket,
-            file_storage_s3_region=config.file_storage_s3_region,
-            file_storage_s3_endpoint=config.file_storage_s3_endpoint,
-            file_storage_s3_access_key_id=config.file_storage_s3_access_key_id,
-            file_storage_s3_secret_access_key=config.file_storage_s3_secret_access_key,
-            file_storage_gcs_bucket=config.file_storage_gcs_bucket,
-            file_storage_gcs_service_account_key=config.file_storage_gcs_service_account_key,
-            file_storage_azure_container=config.file_storage_azure_container,
-            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_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,
-            file_conversion_max_batch_size=config.file_conversion_max_batch_size,
-            enable_file_upload_api=config.enable_file_upload_api,
-            file_delete_after_retain=config.file_delete_after_retain,
            enable_observations=config.enable_observations,
            consolidation_batch_size=config.consolidation_batch_size,
-            consolidation_llm_batch_size=config.consolidation_llm_batch_size,
            consolidation_max_tokens=config.consolidation_max_tokens,
-            observations_mission=config.observations_mission,
            skip_llm_verification=config.skip_llm_verification,
            lazy_reranker=config.lazy_reranker,
            run_migrations_on_startup=config.run_migrations_on_startup,
@@ -292,10 +253,6 @@ 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_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,
@@ -379,7 +336,6 @@ def main():
        "proxy_headers": args.proxy_headers,
        "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
    }

    # Add optional parameters if provided
@@ -408,8 +364,6 @@ def main():
            reranker_provider=config.reranker_provider,
            mcp_enabled=config.mcp_enabled,
            version=__version__,
-            vector_extension=config.vector_extension,
-            text_search_extension=config.text_search_extension,
        )

    # Start idle checker in daemon mode
@@ -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__":
@@ -33,69 +33,6 @@ logger = logging.getLogger(__name__)
 MIGRATION_LOCK_ID = 123456789


-def _detect_vector_extension(conn, vector_extension: str = "pgvector") -> str:
-    """
-    Validate vector extension: 'pgvector', 'vchord', or 'pgvectorscale'.
-
-    Args:
-        conn: SQLAlchemy connection object
-        vector_extension: Configured extension ("pgvector", "vchord", or "pgvectorscale")
-
-    Returns:
-        "pgvector", "vchord", "pgvectorscale", or "pg_diskann"
-
-    Raises:
-        RuntimeError: If configured extension is not installed
-    """
-    # Verify the configured extension is installed
-    if vector_extension == "pgvectorscale":
-        # pgvectorscale/DiskANN requires pgvector to be installed first
-        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
-        if not pgvector_check:
-            raise RuntimeError(
-                "DiskANN (pgvectorscale/pg_diskann) requires pgvector to be installed. "
-                "Install it with: CREATE EXTENSION vector; then CREATE EXTENSION vectorscale CASCADE; (or pg_diskann on Azure)"
-            )
-
-        # Check for either vectorscale (open source) or pg_diskann (Azure)
-        vectorscale_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vectorscale'")).scalar()
-        pg_diskann_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_diskann'")).scalar()
-
-        if vectorscale_check:
-            logger.debug("Using vector extension: pgvectorscale (DiskANN)")
-            return "pgvectorscale"
-        elif pg_diskann_check:
-            logger.debug("Using vector extension: pg_diskann (Azure DiskANN)")
-            return "pg_diskann"  # Return distinct name for parameter handling
-        else:
-            raise RuntimeError(
-                "Configured vector extension 'pgvectorscale' not found. "
-                "Install either:\n"
-                "  - pgvectorscale (open source): CREATE EXTENSION vectorscale CASCADE;\n"
-                "  - pg_diskann (Azure): CREATE EXTENSION pg_diskann CASCADE;"
-            )
-    elif vector_extension == "vchord":
-        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
-        if not vchord_check:
-            raise RuntimeError(
-                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
-            )
-        logger.debug("Using configured vector extension: vchord")
-        return "vchord"
-    elif vector_extension == "pgvector":
-        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
-        if not pgvector_check:
-            raise RuntimeError(
-                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
-            )
-        logger.debug("Using configured vector extension: pgvector")
-        return "pgvector"
-    else:
-        raise ValueError(
-            f"Invalid vector_extension: {vector_extension}. Must be 'pgvector', 'vchord', or 'pgvectorscale'"
-        )
-
-
 def _get_schema_lock_id(schema: str) -> int:
    """
    Generate a unique advisory lock ID for a schema.
@@ -305,48 +242,6 @@ def run_migrations(
                                "Please install it with: CREATE EXTENSION vector;"
                            ) from e

-                # If using pgvectorscale, ensure vectorscale extension is also installed
-                vector_extension = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
-                if vector_extension == "pgvectorscale":
-                    logger.debug("Checking pgvectorscale (vectorscale) extension availability...")
-
-                    vectorscale_check = conn.execute(
-                        text("SELECT 1 FROM pg_extension WHERE extname = 'vectorscale'")
-                    ).scalar()
-
-                    if vectorscale_check:
-                        logger.info("pgvectorscale extension already installed")
-                    else:
-                        # Extension doesn't exist - try to install
-                        logger.info("pgvectorscale extension not found, attempting to install...")
-                        try:
-                            conn.execute(text("CREATE EXTENSION vectorscale CASCADE"))
-                            conn.commit()
-                            logger.info("pgvectorscale extension installed successfully")
-                        except Exception as e:
-                            # Installation failed - check one more time in case another process installed it
-                            conn.rollback()
-                            vectorscale_recheck = conn.execute(
-                                text("SELECT 1 FROM pg_extension WHERE extname = 'vectorscale'")
-                            ).fetchone()
-
-                            if vectorscale_recheck:
-                                logger.warning(
-                                    "Could not install pgvectorscale extension (permission denied?), "
-                                    "but extension exists. Continuing..."
-                                )
-                            else:
-                                # Extension truly doesn't exist and we can't install it
-                                logger.error(
-                                    f"pgvectorscale extension is not installed and cannot be installed: {e}. "
-                                    f"Please ensure pgvectorscale is installed by a database administrator. "
-                                    f"See: https://github.com/timescale/pgvectorscale#installation"
-                                )
-                                raise RuntimeError(
-                                    "pgvectorscale extension is required but not installed. "
-                                    "Please install it with: CREATE EXTENSION vectorscale CASCADE;"
-                                ) from e
-
                # Run migrations while holding the lock
                _run_migrations_internal(database_url, script_location, schema=schema)
            finally:
@@ -429,7 +324,6 @@ def ensure_embedding_dimension(
    database_url: str,
    required_dimension: int,
    schema: str | None = None,
-    vector_extension: str = "pgvector",
 ) -> None:
    """
    Ensure the embedding column dimension matches the model's dimension.
@@ -444,7 +338,6 @@ def ensure_embedding_dimension(
        database_url: SQLAlchemy database URL
        required_dimension: The embedding dimension required by the model
        schema: Target PostgreSQL schema name (None for public)
-        vector_extension: Configured vector extension ("pgvector" or "vchord")

    Raises:
        RuntimeError: If dimension mismatch with existing data
@@ -468,10 +361,6 @@ def ensure_embedding_dimension(
            logger.debug(f"memory_units table does not exist in schema '{schema_name}', skipping dimension check")
            return

-        # Detect which vector extension is available
-        vector_ext = _detect_vector_extension(conn, vector_extension)
-        logger.info(f"Using vector extension: {vector_ext}")
-
        # Get current column dimension from pg_attribute
        # pgvector stores dimension in atttypmod
        current_dim = conn.execute(
@@ -519,7 +408,8 @@ def ensure_embedding_dimension(
        # 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)
+        # Drop the HNSW index on embedding column if it exists
+        # Only drop indexes that use 'hnsw' and reference the 'embedding' column
        conn.execute(
            text(f"""
                DO $$
@@ -529,7 +419,7 @@ def ensure_embedding_dimension(
                        SELECT indexname FROM pg_indexes
                        WHERE schemaname = '{schema_name}'
                          AND tablename = 'memory_units'
-                          AND (indexdef LIKE '%hnsw%' OR indexdef LIKE '%vchordrq%')
+                          AND indexdef LIKE '%hnsw%'
                          AND indexdef LIKE '%embedding%'
                    LOOP
                        EXECUTE 'DROP INDEX IF EXISTS {schema_name}.' || idx_name;
@@ -544,476 +434,15 @@ def ensure_embedding_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
-            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_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")
+        # Recreate the HNSW index
+        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)
+            """)
+        )
        conn.commit()

        logger.info(f"Successfully changed embedding dimension to {required_dimension}")
-
-
-def ensure_vector_extension(
-    database_url: str,
-    vector_extension: str = "pgvector",
-    schema: str | None = None,
-) -> None:
-    """
-    Ensure the vector indexes match the configured vector extension.
-
-    This function checks the current vector index type in the database
-    and adjusts it if necessary:
-    - If index type matches configured extension: no action needed
-    - If they differ and tables are empty: drop old indexes, recreate with new type
-    - If they differ and tables have data: raise error with migration guidance
-
-    Args:
-        database_url: SQLAlchemy database URL
-        vector_extension: Configured vector extension ("pgvector" or "vchord")
-        schema: Target PostgreSQL schema name (None for public)
-
-    Raises:
-        RuntimeError: If extension mismatch with existing data
-    """
-    schema_name = schema or "public"
-
-    engine = create_engine(database_url)
-    with engine.connect() as conn:
-        # Detect which vector extension should be used
-        target_ext = _detect_vector_extension(conn, vector_extension)
-        logger.info(f"Target vector extension: {target_ext}")
-
-        # Tables with vector indexes to check
-        tables_to_check = [
-            ("memory_units", "idx_memory_units_embedding"),
-            ("learnings", "idx_learnings_embedding"),
-            ("pinned_reflections", "idx_pinned_reflections_embedding"),
-        ]
-
-        # Determine target index type
-        if target_ext in ("pgvectorscale", "pg_diskann"):
-            target_index_type = "diskann"
-        elif target_ext == "vchord":
-            target_index_type = "vchordrq"
-        else:
-            target_index_type = "hnsw"
-
-        mismatched_tables = []
-        tables_with_data = []
-
-        for table_name, index_name in tables_to_check:
-            # Check if table exists
-            table_exists = conn.execute(
-                text("""
-                    SELECT EXISTS (
-                        SELECT 1 FROM information_schema.tables
-                        WHERE table_schema = :schema AND table_name = :table_name
-                    )
-                """),
-                {"schema": schema_name, "table_name": table_name},
-            ).scalar()
-
-            if not table_exists:
-                logger.debug(f"Table {table_name} does not exist in schema '{schema_name}', skipping")
-                continue
-
-            # Check current index type by querying pg_indexes
-            current_index_info = conn.execute(
-                text("""
-                    SELECT indexdef
-                    FROM pg_indexes
-                    WHERE schemaname = :schema
-                      AND tablename = :table_name
-                      AND indexname LIKE :index_pattern
-                """),
-                {"schema": schema_name, "table_name": table_name, "index_pattern": "%embedding%"},
-            ).fetchone()
-
-            if not current_index_info:
-                logger.warning(f"No embedding index found for {table_name}, will create it")
-                mismatched_tables.append((table_name, index_name, None))
-                continue
-
-            indexdef = current_index_info[0].lower()
-            if "diskann" in indexdef:
-                current_index_type = "diskann"
-            elif "vchordrq" in indexdef:
-                current_index_type = "vchordrq"
-            elif "hnsw" in indexdef:
-                current_index_type = "hnsw"
-            else:
-                logger.warning(f"Unknown index type for {table_name}: {indexdef}")
-                continue
-
-            # Check if index type matches target
-            if current_index_type != target_index_type:
-                logger.info(
-                    f"Index type mismatch on {table_name}: current={current_index_type}, target={target_index_type}"
-                )
-                mismatched_tables.append((table_name, index_name, current_index_type))
-
-                # Check if table has data
-                row_count = conn.execute(
-                    text(f"SELECT COUNT(*) FROM {schema_name}.{table_name} WHERE embedding IS NOT NULL")
-                ).scalar()
-
-                if row_count > 0:
-                    tables_with_data.append((table_name, row_count))
-            else:
-                logger.debug(f"Index type OK for {table_name}: {current_index_type}")
-
-        # If no mismatches, we're done
-        if not mismatched_tables:
-            logger.debug(f"All vector indexes match configured extension: {target_ext}")
-            return
-
-        # If there's data in any mismatched table, raise error
-        if tables_with_data:
-            table_list = ", ".join([f"{table}({count} rows)" for table, count in tables_with_data])
-            # Map index type back to extension name for error message
-            current_ext_name = {"diskann": "pgvectorscale", "vchordrq": "vchord", "hnsw": "pgvector"}.get(
-                current_index_type, current_index_type
-            )
-
-            raise RuntimeError(
-                f"Cannot change vector extension from {current_index_type} to {target_index_type}: "
-                f"the following tables contain data: {table_list}. "
-                f"To change vector extension, you must either:\n"
-                f"  1. Re-embed all data: DELETE FROM {schema_name}.memory_units; "
-                f"DELETE FROM {schema_name}.learnings; DELETE FROM {schema_name}.pinned_reflections; then restart\n"
-                f"  2. Use the current vector extension (set HINDSIGHT_API_VECTOR_EXTENSION='{current_ext_name}')"
-            )
-
-        # Tables are empty, safe to recreate indexes
-        logger.info(f"Recreating vector indexes for {target_ext}")
-
-        for table_name, index_name, current_type in mismatched_tables:
-            # Drop existing index if it exists
-            if current_type:
-                logger.info(f"Dropping {current_type} index on {table_name}")
-                conn.execute(text(f"DROP INDEX IF EXISTS {schema_name}.{index_name}"))
-
-            # Create new index with appropriate type
-            if target_ext == "pgvectorscale":
-                logger.info(f"Creating DiskANN index on {table_name} (pgvectorscale)")
-                conn.execute(
-                    text(f"""
-                        CREATE INDEX IF NOT EXISTS {index_name}
-                        ON {schema_name}.{table_name}
-                        USING diskann (embedding vector_cosine_ops)
-                        WITH (num_neighbors = 50)
-                    """)
-                )
-            elif target_ext == "pg_diskann":
-                logger.info(f"Creating DiskANN index on {table_name} (pg_diskann/Azure)")
-                conn.execute(
-                    text(f"""
-                        CREATE INDEX IF NOT EXISTS {index_name}
-                        ON {schema_name}.{table_name}
-                        USING diskann (embedding vector_cosine_ops)
-                        WITH (max_neighbors = 50)
-                    """)
-                )
-            elif target_ext == "vchord":
-                logger.info(f"Creating vchordrq index on {table_name}")
-                conn.execute(
-                    text(f"""
-                        CREATE INDEX IF NOT EXISTS {index_name}
-                        ON {schema_name}.{table_name}
-                        USING vchordrq (embedding vector_l2_ops)
-                    """)
-                )
-            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"""
-                        CREATE INDEX IF NOT EXISTS {index_name}
-                        ON {schema_name}.{table_name}
-                        USING hnsw (embedding vector_cosine_ops)
-                        WITH (m = 16, ef_construction = 64)
-                    """)
-                )
-
-        conn.commit()
-        logger.info(f"Successfully migrated vector indexes to {target_ext}")
-
-
-def ensure_text_search_extension(
-    database_url: str,
-    text_search_extension: str = "native",
-    schema: str | None = None,
-) -> None:
-    """
-    Ensure the text search columns and indexes match the configured extension.
-
-    This function checks the current search_vector column type and index type
-    in the database and adjusts them if necessary:
-    - If they match configured extension: no action needed
-    - If they differ and tables are empty: drop old column/index, recreate with new type
-    - If they differ and tables have data: raise error with migration guidance
-
-    Args:
-        database_url: SQLAlchemy database URL
-        text_search_extension: Configured text search extension ("native" or "vchord")
-        schema: Target PostgreSQL schema name (None for public)
-
-    Raises:
-        RuntimeError: If extension mismatch with existing data
-    """
-    schema_name = schema or "public"
-
-    engine = create_engine(database_url)
-    with engine.connect() as conn:
-        # Tables with search_vector columns to check
-        tables_to_check = [
-            "memory_units",
-            "reflections",  # Renamed from pinned_reflections in p1k2l3m4n5o6 migration
-        ]
-
-        # Determine target column type and index type
-        if text_search_extension == "vchord":
-            target_column_type = "bm25vector"
-            target_index_type = "bm25"
-        elif text_search_extension == "pg_textsearch":
-            target_column_type = "text"
-            target_index_type = "bm25"
-        else:  # native
-            target_column_type = "tsvector"
-            target_index_type = "gin"
-
-        mismatched_tables = []
-        tables_with_data = []
-
-        for table_name in tables_to_check:
-            # Check if table exists
-            table_exists = conn.execute(
-                text("""
-                    SELECT EXISTS (
-                        SELECT 1 FROM information_schema.tables
-                        WHERE table_schema = :schema AND table_name = :table_name
-                    )
-                """),
-                {"schema": schema_name, "table_name": table_name},
-            ).scalar()
-
-            if not table_exists:
-                logger.debug(f"Table {table_name} does not exist in schema '{schema_name}', skipping")
-                continue
-
-            # Get current column type from information_schema
-            current_column_info = conn.execute(
-                text("""
-                    SELECT data_type, udt_name
-                    FROM information_schema.columns
-                    WHERE table_schema = :schema
-                      AND table_name = :table_name
-                      AND column_name = 'search_vector'
-                """),
-                {"schema": schema_name, "table_name": table_name},
-            ).fetchone()
-
-            if not current_column_info:
-                logger.warning(f"No search_vector column found for {table_name}, will create it")
-                mismatched_tables.append((table_name, None, None))
-                continue
-
-            # Check column type (udt_name contains the actual type: tsvector, bm25vector, etc.)
-            current_column_type = current_column_info[1]  # udt_name
-
-            # Get current index type
-            current_index_info = conn.execute(
-                text("""
-                    SELECT am.amname
-                    FROM pg_indexes pi
-                    JOIN pg_class c ON c.relname = pi.indexname
-                    JOIN pg_am am ON am.oid = c.relam
-                    WHERE pi.schemaname = :schema
-                      AND pi.tablename = :table_name
-                      AND pi.indexname LIKE '%text_search%'
-                """),
-                {"schema": schema_name, "table_name": table_name},
-            ).fetchone()
-
-            current_index_type = current_index_info[0] if current_index_info else None
-
-            # Check if column and index types match target
-            column_matches = current_column_type == target_column_type
-            index_matches = current_index_type == target_index_type if current_index_type else False
-
-            if not (column_matches and index_matches):
-                logger.info(
-                    f"Text search mismatch on {table_name}: "
-                    f"column={current_column_type} (want {target_column_type}), "
-                    f"index={current_index_type} (want {target_index_type})"
-                )
-                mismatched_tables.append((table_name, current_column_type, current_index_type))
-
-                # Check if table has data
-                row_count = conn.execute(text(f"SELECT COUNT(*) FROM {schema_name}.{table_name}")).scalar()
-
-                if row_count > 0:
-                    tables_with_data.append((table_name, row_count))
-            else:
-                logger.debug(f"Text search OK for {table_name}: {current_column_type}/{current_index_type}")
-
-        # If no mismatches, we're done
-        if not mismatched_tables:
-            logger.debug(f"All text search columns/indexes match configured extension: {text_search_extension}")
-            return
-
-        # If there's data in any mismatched table, raise error
-        if tables_with_data:
-            table_list = ", ".join([f"{table}({count} rows)" for table, count in tables_with_data])
-            # Detect current extension from column type
-            current_col_type = mismatched_tables[0][1]
-            if current_col_type == "tsvector":
-                current_ext = "native"
-            elif current_col_type == "bm25vector":
-                current_ext = "vchord"
-            elif current_col_type == "text":
-                current_ext = "pg_textsearch"
-            else:
-                current_ext = "unknown"
-            raise RuntimeError(
-                f"Cannot change text search extension from {current_ext} to {text_search_extension}: "
-                f"the following tables contain data: {table_list}. "
-                f"To change text search extension, you must either:\n"
-                f"  1. Clear all data: DELETE FROM {schema_name}.memory_units; "
-                f"DELETE FROM {schema_name}.reflections; then restart\n"
-                f"  2. Use the current text search extension (set HINDSIGHT_API_TEXT_SEARCH_EXTENSION='{current_ext}')"
-            )
-
-        # Tables are empty, safe to recreate columns/indexes
-        logger.info(f"Recreating text search columns/indexes for {text_search_extension}")
-
-        for table_name, current_col_type, current_idx_type in mismatched_tables:
-            # Drop existing index if it exists
-            if current_idx_type:
-                logger.info(f"Dropping {current_idx_type} index on {table_name}")
-                conn.execute(
-                    text(f"""
-                        DROP INDEX IF EXISTS {schema_name}.idx_{table_name.replace(".", "_")}_text_search
-                    """)
-                )
-
-            # Drop existing column if it exists
-            if current_col_type:
-                logger.info(f"Dropping {current_col_type} column on {table_name}")
-                conn.execute(text(f"ALTER TABLE {schema_name}.{table_name} DROP COLUMN IF EXISTS search_vector"))
-
-            # Create new column with appropriate type
-            if text_search_extension == "vchord":
-                logger.info(f"Creating bm25vector column on {table_name}")
-                # Note: vchord_bm25 extension creates types in bm25_catalog schema
-                conn.execute(
-                    text(f"ALTER TABLE {schema_name}.{table_name} ADD COLUMN search_vector bm25_catalog.bm25vector")
-                )
-
-                # Create BM25 index
-                logger.info(f"Creating BM25 index on {table_name}")
-                conn.execute(
-                    text(f"""
-                        CREATE INDEX idx_{table_name.replace(".", "_")}_text_search
-                        ON {schema_name}.{table_name}
-                        USING bm25 (search_vector bm25_catalog.bm25_ops)
-                    """)
-                )
-            elif text_search_extension == "pg_textsearch":
-                logger.info(f"Creating TEXT column on {table_name}")
-                # Dummy TEXT column for consistency (indexes operate on base columns)
-                conn.execute(text(f"ALTER TABLE {schema_name}.{table_name} ADD COLUMN search_vector TEXT"))
-
-                # Create BM25 index on expression
-                logger.info(f"Creating BM25 index on {table_name}")
-                # Different expression for each table
-                if table_name == "memory_units":
-                    index_expr = "(COALESCE(text, '') || ' ' || COALESCE(context, ''))"
-                else:  # reflections
-                    index_expr = "(COALESCE(name, '') || ' ' || content)"
-
-                conn.execute(
-                    text(f"""
-                        CREATE INDEX idx_{table_name.replace(".", "_")}_text_search
-                        ON {schema_name}.{table_name}
-                        USING bm25({index_expr})
-                        WITH (text_config='english')
-                    """)
-                )
-            else:  # native
-                logger.info(f"Creating tsvector column on {table_name}")
-                # Different GENERATED expression for each table
-                if table_name == "memory_units":
-                    generated_expr = "to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))"
-                else:  # reflections
-                    generated_expr = "to_tsvector('english', COALESCE(name, '') || ' ' || content)"
-
-                conn.execute(
-                    text(f"""
-                        ALTER TABLE {schema_name}.{table_name}
-                        ADD COLUMN search_vector tsvector
-                        GENERATED ALWAYS AS ({generated_expr}) STORED
-                    """)
-                )
-
-                # Create GIN index
-                logger.info(f"Creating GIN index on {table_name}")
-                conn.execute(
-                    text(f"""
-                        CREATE INDEX idx_{table_name.replace(".", "_")}_text_search
-                        ON {schema_name}.{table_name}
-                        USING gin(search_vector)
-                    """)
-                )
-
-        conn.commit()
-        logger.info(f"Successfully migrated text search to {text_search_extension}")
@@ -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
@@ -376,12 +376,7 @@ class WorkerPoller:
                        del self._in_flight_by_type[operation_type]

    async def _execute_task_inner(self, task: ClaimedTask):
-        """Inner task execution with error handling.
-
-        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.
-        """
+        """Inner task execution with error handling."""
        task_type = task.task_dict.get("type", "unknown")
        bank_id = task.task_dict.get("bank_id", "unknown")

@@ -391,12 +386,12 @@ class WorkerPoller:
            if task.schema:
                task.task_dict["_schema"] = task.schema
            await self._executor(task.task_dict)
-            logger.debug(f"Task {task.operation_id} execution finished")
+            await self._mark_completed(task.operation_id, task.schema)
+            logger.debug(f"Task {task.operation_id} completed successfully")
        except Exception as 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()
+            error_msg = f"{type(e).__name__}: {e}\n{traceback.format_exc()}"
+            logger.error(f"Task {task.operation_id} failed: {e}")
+            await self._retry_or_fail(task.operation_id, error_msg, task.schema)

    async def recover_own_tasks(self) -> int:
        """
@@ -406,8 +401,6 @@ class WorkerPoller:
        On startup, we reset any tasks stuck in 'processing' for this worker_id
        back to 'pending' so they can be picked up again.

-        Also recovers batch API operations that were in-flight.
-
        If tenant_extension is configured, recovers across all tenant schemas.

        Returns:
@@ -420,16 +413,11 @@ class WorkerPoller:
            try:
                table = fq_table("async_operations", schema)

-                # First, recover batch API operations (before resetting worker tasks)
-                batch_count = await self._recover_batch_operations(schema)
-                total_count += batch_count
-
-                # Then reset normal worker tasks
                result = await self._pool.execute(
                    f"""
                    UPDATE {table}
                    SET status = 'pending', worker_id = NULL, claimed_at = NULL, updated_at = now()
-                    WHERE status = 'processing' AND worker_id = $1 AND result_metadata->>'batch_id' IS NULL
+                    WHERE status = 'processing' AND worker_id = $1
                    """,
                    self._worker_id,
                )
@@ -446,80 +434,6 @@ class WorkerPoller:
            logger.info(f"Worker {self._worker_id} recovered {total_count} stale tasks from previous run")
        return total_count

-    async def _recover_batch_operations(self, schema: str | None) -> int:
-        """
-        Recover batch API operations that were in-flight when worker crashed.
-
-        Finds operations with batch_id in metadata and re-submits them as tasks
-        so polling can resume.
-
-        Args:
-            schema: Database schema to recover from
-
-        Returns:
-            Number of batch operations recovered
-        """
-        table = fq_table("async_operations", schema)
-
-        try:
-            # Find operations with batch_id in metadata (batch API operations)
-            rows = await self._pool.fetch(
-                f"""
-                SELECT operation_id, task_payload, result_metadata
-                FROM {table}
-                WHERE status = 'processing'
-                  AND result_metadata ? 'batch_id'
-                  AND task_payload IS NOT NULL
-                """
-            )
-
-            if not rows:
-                return 0
-
-            recovered = 0
-            for row in rows:
-                operation_id = str(row["operation_id"])
-                task_payload = row["task_payload"]
-                result_metadata = row["result_metadata"]
-
-                # Parse metadata
-                if isinstance(result_metadata, str):
-                    result_metadata = json.loads(result_metadata)
-
-                batch_id = result_metadata.get("batch_id")
-                batch_provider = result_metadata.get("batch_provider", "openai")
-
-                logger.info(
-                    f"Recovering batch operation: operation_id={operation_id}, batch_id={batch_id}, provider={batch_provider}"
-                )
-
-                # Parse task_payload
-                if isinstance(task_payload, str):
-                    task_dict = json.loads(task_payload)
-                else:
-                    task_dict = task_payload
-
-                # Mark operation as ready for re-processing
-                # Reset to pending with task_payload intact so worker picks it up again
-                await self._pool.execute(
-                    f"""
-                    UPDATE {table}
-                    SET status = 'pending', worker_id = NULL, claimed_at = NULL, updated_at = now()
-                    WHERE operation_id = $1
-                    """,
-                    operation_id,
-                )
-
-                recovered += 1
-                logger.info(f"Batch operation {operation_id} reset to pending for re-processing")
-
-            return recovered
-
-        except Exception as e:
-            schema_display = f'"{schema}"' if schema else str(schema)
-            logger.error(f"Failed to recover batch operations for schema {schema_display}: {e}")
-            return 0
-
    async def run(self):
        """
        Main polling loop with fire-and-forget task execution.
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"

 [project]
 name = "hindsight-api"
-version = "0.4.14"
+version = "0.4.10"
 description = "Hindsight: Agent Memory That Works Like Human Memory"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -42,9 +42,6 @@ dependencies = [
    "typer>=0.9.0",
    "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)
    # 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
@@ -67,7 +64,6 @@ test = [
    "pytest-timeout>=2.4.0",
    "pytest-xdist>=3.0.0",
    "filelock>=3.20.1",  # TOCTOU race condition fix
-    "testcontainers>=4.0.0",
 ]

 [project.scripts]
@@ -98,7 +94,7 @@ 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
@@ -117,7 +113,6 @@ dev = [
    "filelock>=3.20.1",  # TOCTOU race condition fix
    "ruff>=0.8.0",
    "ty>=0.0.1",
-    "testcontainers>=4.0.0",
 ]

 [tool.ruff]
@@ -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:
@@ -1,423 +0,0 @@
-"""Test async batch retain with smart batching and parent-child operations."""
-
-import asyncio
-import json
-import uuid
-
-import pytest
-
-from hindsight_api.extensions import RequestContext
-
-
-@pytest.mark.asyncio
-async def test_duplicate_document_ids_rejected_async(memory, request_context):
-    """Test that async retain rejects batches with duplicate document_ids."""
-    bank_id = "test_duplicate_async"
-    contents = [
-        {"content": "First item", "document_id": "doc1"},
-        {"content": "Second item", "document_id": "doc2"},
-        {"content": "Third item", "document_id": "doc1"},  # Duplicate!
-    ]
-
-    # Should raise ValueError due to duplicate document_ids
-    with pytest.raises(ValueError, match="duplicate document_ids.*doc1"):
-        await memory.submit_async_retain(
-            bank_id=bank_id,
-            contents=contents,
-            request_context=request_context,
-        )
-
-
-@pytest.mark.asyncio
-async def test_duplicate_document_ids_rejected_sync(memory, request_context):
-    """Test that sync retain also rejects batches with duplicate document_ids."""
-    bank_id = "test_duplicate_sync"
-    contents = [
-        {"content": "First item", "document_id": "doc1"},
-        {"content": "Second item", "document_id": "doc1"},  # Duplicate!
-    ]
-
-    # Should raise ValueError due to duplicate document_ids
-    with pytest.raises(ValueError, match="duplicate document_ids.*doc1"):
-        await memory.retain_batch_async(
-            bank_id=bank_id,
-            contents=contents,
-            request_context=request_context,
-        )
-
-
-@pytest.mark.asyncio
-async def test_small_async_batch_no_splitting(memory, request_context):
-    """Test that small async batches create parent with single child (simplified code path)."""
-    bank_id = "test_small_async"
-    contents = [{"content": "Alice works at Google", "document_id": f"doc{i}"} for i in range(5)]
-
-    # Calculate total chars (should be well under threshold)
-    total_chars = sum(len(item["content"]) for item in contents)
-    assert total_chars < 10_000, "Test batch should be small"
-
-    # Submit async retain
-    result = await memory.submit_async_retain(
-        bank_id=bank_id,
-        contents=contents,
-        request_context=request_context,
-    )
-
-    # Verify we got an operation_id back
-    assert "operation_id" in result
-    assert "items_count" in result
-    assert result["items_count"] == 5
-
-    operation_id = result["operation_id"]
-
-    # Wait for task to complete (SyncTaskBackend executes immediately)
-    await asyncio.sleep(0.1)
-
-    # Check operation status
-    status = await memory.get_operation_status(
-        bank_id=bank_id,
-        operation_id=operation_id,
-        request_context=request_context,
-    )
-
-    # Should be a parent operation with single child (simplified code path)
-    assert status["status"] == "completed"
-    assert status["operation_type"] == "batch_retain"
-    assert "child_operations" in status
-    assert status["result_metadata"]["num_sub_batches"] == 1  # Single sub-batch
-    assert len(status["child_operations"]) == 1
-    assert status["child_operations"][0]["status"] == "completed"
-
-
-@pytest.mark.asyncio
-async def test_large_async_batch_auto_splits(memory, request_context):
-    """Test that large async batches automatically split into sub-batches with parent operation."""
-    from hindsight_api.engine.memory_engine import count_tokens
-
-    bank_id = "test_large_async"
-
-    # Create a large batch that exceeds the threshold (10k tokens default)
-    # Repeating "A"s gets heavily compressed by tokenizer, use varied content
-    # Use ~22k chars per item = ~5.5k tokens per item, 2 items = ~11k tokens total (exceeds 10k)
-    large_content = "The quick brown fox jumps over the lazy dog. " * 500  # ~22k chars = ~5.5k tokens
-    contents = [{"content": large_content + f" item {i}", "document_id": f"doc{i}"} for i in range(2)]
-
-    # Calculate total tokens (should exceed threshold)
-    total_tokens = sum(count_tokens(item["content"]) for item in contents)
-    assert total_tokens > 10_000, "Test batch should exceed threshold"
-
-    # Submit async retain
-    result = await memory.submit_async_retain(
-        bank_id=bank_id,
-        contents=contents,
-        request_context=request_context,
-    )
-
-    # Verify we got an operation_id back
-    assert "operation_id" in result
-    assert "items_count" in result
-    assert result["items_count"] == 2
-
-    parent_operation_id = result["operation_id"]
-
-    # Wait for tasks to complete
-    await asyncio.sleep(0.5)
-
-    # Check parent operation status
-    parent_status = await memory.get_operation_status(
-        bank_id=bank_id,
-        operation_id=parent_operation_id,
-        request_context=request_context,
-    )
-
-    # Should be a parent operation with children
-    assert parent_status["operation_type"] == "batch_retain"
-    assert "child_operations" in parent_status
-    assert "num_sub_batches" in parent_status["result_metadata"]
-    assert parent_status["result_metadata"]["num_sub_batches"] >= 2  # Should split into at least 2 batches
-    assert parent_status["result_metadata"]["items_count"] == 2
-
-    # Verify child operations
-    child_ops = parent_status["child_operations"]
-    assert len(child_ops) >= 2, "Should have at least 2 child operations"
-
-    # All children should be completed (SyncTaskBackend executes immediately)
-    for child in child_ops:
-        assert child["status"] == "completed"
-        assert child["sub_batch_index"] is not None
-        assert child["items_count"] > 0
-
-    # Parent status should be aggregated as "completed"
-    assert parent_status["status"] == "completed"
-
-
-@pytest.mark.asyncio
-async def test_parent_operation_status_aggregation_pending(memory, request_context):
-    """Test that parent operation shows 'pending' when children are pending."""
-    bank_id = "test_parent_pending"
-    pool = await memory._get_pool()
-
-    # Manually create a parent operation
-    parent_id = uuid.uuid4()
-    async with pool.acquire() as conn:
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            parent_id,
-            bank_id,
-            "batch_retain",
-            json.dumps({"items_count": 20, "num_sub_batches": 2, "is_parent": True}),
-            "pending",
-        )
-
-        # Create 2 child operations - one completed, one pending
-        child1_id = uuid.uuid4()
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            child1_id,
-            bank_id,
-            "retain",
-            json.dumps(
-                {
-                    "items_count": 10,
-                    "parent_operation_id": str(parent_id),
-                    "sub_batch_index": 1,
-                    "total_sub_batches": 2,
-                }
-            ),
-            "completed",
-        )
-
-        child2_id = uuid.uuid4()
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            child2_id,
-            bank_id,
-            "retain",
-            json.dumps(
-                {
-                    "items_count": 10,
-                    "parent_operation_id": str(parent_id),
-                    "sub_batch_index": 2,
-                    "total_sub_batches": 2,
-                }
-            ),
-            "pending",
-        )
-
-    # Check parent status
-    parent_status = await memory.get_operation_status(
-        bank_id=bank_id,
-        operation_id=str(parent_id),
-        request_context=request_context,
-    )
-
-    # Parent should aggregate as "pending" since one child is still pending
-    assert parent_status["status"] == "pending"
-    assert len(parent_status["child_operations"]) == 2
-
-
-@pytest.mark.asyncio
-async def test_parent_operation_status_aggregation_failed(memory, request_context):
-    """Test that parent operation shows 'failed' when any child fails."""
-    bank_id = "test_parent_failed"
-    pool = await memory._get_pool()
-
-    # Manually create a parent operation
-    parent_id = uuid.uuid4()
-    async with pool.acquire() as conn:
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            parent_id,
-            bank_id,
-            "batch_retain",
-            json.dumps({"items_count": 20, "num_sub_batches": 2, "is_parent": True}),
-            "pending",
-        )
-
-        # Create 2 child operations - one completed, one failed
-        child1_id = uuid.uuid4()
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            child1_id,
-            bank_id,
-            "retain",
-            json.dumps(
-                {
-                    "items_count": 10,
-                    "parent_operation_id": str(parent_id),
-                    "sub_batch_index": 1,
-                    "total_sub_batches": 2,
-                }
-            ),
-            "completed",
-        )
-
-        child2_id = uuid.uuid4()
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status, error_message)
-            VALUES ($1, $2, $3, $4, $5, $6)
-            """,
-            child2_id,
-            bank_id,
-            "retain",
-            json.dumps(
-                {
-                    "items_count": 10,
-                    "parent_operation_id": str(parent_id),
-                    "sub_batch_index": 2,
-                    "total_sub_batches": 2,
-                }
-            ),
-            "failed",
-            "Test error",
-        )
-
-    # Check parent status
-    parent_status = await memory.get_operation_status(
-        bank_id=bank_id,
-        operation_id=str(parent_id),
-        request_context=request_context,
-    )
-
-    # Parent should aggregate as "failed" since one child failed
-    assert parent_status["status"] == "failed"
-    assert len(parent_status["child_operations"]) == 2
-
-    # Verify child with error is included
-    failed_child = [c for c in parent_status["child_operations"] if c["status"] == "failed"][0]
-    assert failed_child["error_message"] == "Test error"
-
-
-@pytest.mark.asyncio
-async def test_parent_operation_status_aggregation_completed(memory, request_context):
-    """Test that parent operation shows 'completed' when all children are completed."""
-    bank_id = "test_parent_completed"
-    pool = await memory._get_pool()
-
-    # Manually create a parent operation
-    parent_id = uuid.uuid4()
-    async with pool.acquire() as conn:
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            parent_id,
-            bank_id,
-            "batch_retain",
-            json.dumps({"items_count": 20, "num_sub_batches": 2, "is_parent": True}),
-            "pending",
-        )
-
-        # Create 2 child operations - both completed
-        child1_id = uuid.uuid4()
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            child1_id,
-            bank_id,
-            "retain",
-            json.dumps(
-                {
-                    "items_count": 10,
-                    "parent_operation_id": str(parent_id),
-                    "sub_batch_index": 1,
-                    "total_sub_batches": 2,
-                }
-            ),
-            "completed",
-        )
-
-        child2_id = uuid.uuid4()
-        await conn.execute(
-            """
-            INSERT INTO async_operations (operation_id, bank_id, operation_type, result_metadata, status)
-            VALUES ($1, $2, $3, $4, $5)
-            """,
-            child2_id,
-            bank_id,
-            "retain",
-            json.dumps(
-                {
-                    "items_count": 10,
-                    "parent_operation_id": str(parent_id),
-                    "sub_batch_index": 2,
-                    "total_sub_batches": 2,
-                }
-            ),
-            "completed",
-        )
-
-    # Check parent status
-    parent_status = await memory.get_operation_status(
-        bank_id=bank_id,
-        operation_id=str(parent_id),
-        request_context=request_context,
-    )
-
-    # Parent should aggregate as "completed" since all children are completed
-    assert parent_status["status"] == "completed"
-    assert len(parent_status["child_operations"]) == 2
-    assert all(c["status"] == "completed" for c in parent_status["child_operations"])
-
-
-@pytest.mark.asyncio
-async def test_config_retain_batch_tokens_respected(memory, request_context):
-    """Test that the retain_batch_tokens config setting is respected."""
-    from hindsight_api.config import get_config
-    from hindsight_api.engine.memory_engine import count_tokens
-
-    bank_id = "test_config_batch_tokens"
-    config = get_config()
-
-    # Check that config has the retain_batch_tokens setting
-    assert hasattr(config, "retain_batch_tokens")
-    assert config.retain_batch_tokens > 0
-
-    # Create a batch that's just under the threshold
-    # Use content that produces roughly half the token limit per item
-    content_size = config.retain_batch_tokens * 2  # chars (rough estimate: 1 token ~= 4 chars)
-    contents = [{"content": "A" * content_size, "document_id": f"doc{i}"} for i in range(2)]
-
-    total_tokens = sum(count_tokens(item["content"]) for item in contents)
-    # Should be equal to threshold (boundary case, no splitting since we use > not >=)
-    assert total_tokens <= config.retain_batch_tokens
-
-    # Submit - should NOT split
-    result = await memory.submit_async_retain(
-        bank_id=bank_id,
-        contents=contents,
-        request_context=request_context,
-    )
-
-    # Wait for completion
-    await asyncio.sleep(0.1)
-
-    # Check status - should be a parent with single child (even for small batches)
-    status = await memory.get_operation_status(
-        bank_id=bank_id,
-        operation_id=result["operation_id"],
-        request_context=request_context,
-    )
-
-    # Even small batches use parent-child pattern now (simpler code path)
-    assert "child_operations" in status
-    assert status["result_metadata"]["num_sub_batches"] == 1
@@ -1,94 +0,0 @@
-"""Unit tests for async retain tag propagation."""
-
-from unittest.mock import AsyncMock, MagicMock
-
-import pytest
-
-from hindsight_api.engine.memory_engine import MemoryEngine
-from hindsight_api.models import RequestContext
-
-
-@pytest.mark.asyncio
-async def test_submit_async_retain_includes_document_tags_in_task_payload():
-    """submit_async_retain should include document_tags in queued 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
-    mock_conn = AsyncMock()
-    mock_conn.execute = AsyncMock()
-    mock_conn.transaction = MagicMock()
-    mock_conn.transaction.return_value.__aenter__ = AsyncMock()
-    mock_conn.transaction.return_value.__aexit__ = AsyncMock()
-
-    mock_pool = AsyncMock()
-    mock_pool.acquire = AsyncMock(return_value=mock_conn)
-    mock_pool.release = AsyncMock()
-
-    engine._get_pool = AsyncMock(return_value=mock_pool)
-
-    request_context = RequestContext(tenant_id="tenant-a", api_key_id="key-a")
-    contents = [{"content": "Async retain payload test."}]
-    document_tags = ["scope:tools", "user:alice"]
-
-    result = await MemoryEngine.submit_async_retain(
-        engine,
-        bank_id="bank-1",
-        contents=contents,
-        document_tags=document_tags,
-        request_context=request_context,
-    )
-
-    # Check result structure
-    assert "operation_id" in result
-    assert "items_count" in result
-    assert result["items_count"] == 1
-
-    # Verify authentication was called
-    engine._authenticate_tenant.assert_awaited_once_with(request_context)
-
-    # Verify child operation was submitted
-    engine._submit_async_operation.assert_awaited_once()
-
-    # Verify child operation payload contains document_tags
-    kwargs = engine._submit_async_operation.await_args.kwargs
-    assert kwargs["bank_id"] == "bank-1"
-    assert kwargs["operation_type"] == "retain"
-    assert kwargs["task_type"] == "batch_retain"
-    assert kwargs["task_payload"]["contents"] == contents
-    assert kwargs["task_payload"]["document_tags"] == document_tags
-    assert kwargs["task_payload"]["_tenant_id"] == "tenant-a"
-    assert kwargs["task_payload"]["_api_key_id"] == "key-a"
-
-
-@pytest.mark.asyncio
-async def test_handle_batch_retain_forwards_document_tags_to_retain_batch_async():
-    """Worker handler should forward document_tags from task payload."""
-    engine = MemoryEngine.__new__(MemoryEngine)
-    engine._initialized = True
-    engine.retain_batch_async = AsyncMock(return_value={"items_count": 1})
-
-    task_dict = {
-        "bank_id": "bank-1",
-        "contents": [{"content": "Forward tags test."}],
-        "document_tags": ["scope:client"],
-        "_tenant_id": "tenant-a",
-        "_api_key_id": "key-a",
-    }
-
-    await MemoryEngine._handle_batch_retain(engine, task_dict)
-
-    engine.retain_batch_async.assert_awaited_once()
-    kwargs = engine.retain_batch_async.await_args.kwargs
-    assert kwargs["bank_id"] == "bank-1"
-    assert kwargs["contents"] == task_dict["contents"]
-    assert kwargs["document_tags"] == ["scope:client"]
-
-    request_context = kwargs["request_context"]
-    assert request_context.internal is True
-    assert request_context.user_initiated is True
-    assert request_context.tenant_id == "tenant-a"
-    assert request_context.api_key_id == "key-a"
@@ -1,189 +0,0 @@
-"""
-Integration test for API base path support.
-
-Tests that the API works correctly when deployed with a base path (e.g., /hindsight)
-for reverse proxy deployments.
-"""
-import os
-import pytest
-import pytest_asyncio
-import httpx
-from hindsight_api.api import create_app
-from hindsight_api.config import clear_config_cache
-
-
-@pytest_asyncio.fixture
-async def api_client_with_base_path(memory):
-    """Create an async test client for the FastAPI app with a base path."""
-    # Set base path in environment
-    base_path = "/hindsight"
-    os.environ["HINDSIGHT_API_BASE_PATH"] = base_path
-
-    # Clear config cache to force reload with new base_path
-    clear_config_cache()
-
-    # Memory is already initialized by the conftest fixture (with migrations)
-    app = create_app(memory, initialize_memory=False)
-
-    # Use base_url with base path
-    transport = httpx.ASGITransport(app=app)
-    async with httpx.AsyncClient(
-        transport=transport,
-        base_url=f"http://test{base_path}"
-    ) as client:
-        yield client
-
-    # Cleanup: unset base path
-    os.environ.pop("HINDSIGHT_API_BASE_PATH", None)
-    clear_config_cache()
-
-
-@pytest_asyncio.fixture
-async def api_client_without_base_path(memory):
-    """Create an async test client for the FastAPI app without a base path (root)."""
-    # Ensure no base path is set
-    os.environ.pop("HINDSIGHT_API_BASE_PATH", None)
-    clear_config_cache()
-
-    app = create_app(memory, initialize_memory=False)
-    transport = httpx.ASGITransport(app=app)
-    async with httpx.AsyncClient(transport=transport, base_url="http://test") as client:
-        yield client
-
-
-@pytest.mark.asyncio
-async def test_base_path_health_endpoint(api_client_with_base_path):
-    """Test that health endpoint works with base path."""
-    # With base path set to /hindsight, health should be at /hindsight/health
-    # But since our client base_url is already http://test/hindsight, we request /health
-    response = await api_client_with_base_path.get("/health")
-    assert response.status_code == 200
-    data = response.json()
-    assert "status" in data
-    assert data["status"] in ["ok", "healthy"]  # Accept both formats
-
-
-@pytest.mark.asyncio
-async def test_base_path_banks_endpoint(api_client_with_base_path):
-    """Test that banks endpoint works with base path."""
-    response = await api_client_with_base_path.get("/v1/default/banks")
-    assert response.status_code == 200
-    data = response.json()
-    assert "banks" in data
-
-
-@pytest.mark.asyncio
-async def test_base_path_openapi_schema(api_client_with_base_path):
-    """Test that OpenAPI schema includes correct base path in servers."""
-    response = await api_client_with_base_path.get("/openapi.json")
-    assert response.status_code == 200
-    openapi_schema = response.json()
-
-    # Check that servers array includes base path
-    assert "servers" in openapi_schema
-    servers = openapi_schema["servers"]
-    assert len(servers) > 0
-    # FastAPI should set server URL to the root_path
-    assert servers[0]["url"] == "/hindsight"
-
-
-@pytest.mark.asyncio
-async def test_base_path_docs_redirect(api_client_with_base_path):
-    """Test that /docs redirects correctly with base path."""
-    # FastAPI docs endpoint should work
-    response = await api_client_with_base_path.get("/docs", follow_redirects=False)
-    # Should either return 200 (direct) or 307 (redirect to trailing slash)
-    assert response.status_code in [200, 307]
-
-
-@pytest.mark.asyncio
-async def test_base_path_metrics(api_client_with_base_path):
-    """Test that metrics endpoint works with base path."""
-    response = await api_client_with_base_path.get("/metrics")
-    assert response.status_code == 200
-    # Metrics should be in Prometheus format
-    assert "# HELP" in response.text or "# TYPE" in response.text
-
-
-@pytest.mark.asyncio
-async def test_base_path_full_workflow(api_client_with_base_path):
-    """
-    Test a full retain/recall workflow with base path.
-
-    This ensures that all memory operations work correctly when the API
-    is deployed with a base path.
-    """
-    bank_id = "test_base_path_bank"
-
-    # 1. Create/get bank
-    response = await api_client_with_base_path.get(f"/v1/default/banks/{bank_id}/profile")
-    assert response.status_code == 200
-
-    # 2. Store a memory
-    response = await api_client_with_base_path.post(
-        f"/v1/default/banks/{bank_id}/memories",
-        json={
-            "items": [
-                {
-                    "content": "The API supports base path deployment for reverse proxy use cases.",
-                    "context": "testing base path feature"
-                }
-            ]
-        }
-    )
-    assert response.status_code == 200
-    result = response.json()
-    assert result["success"] is True
-
-    # 3. Recall the memory
-    response = await api_client_with_base_path.post(
-        f"/v1/default/banks/{bank_id}/memories/recall",
-        json={
-            "query": "base path support"
-        }
-    )
-    assert response.status_code == 200
-    recall_result = response.json()
-    # API returns "results" not "memories"
-    assert "results" in recall_result
-    assert len(recall_result["results"]) > 0
-
-
-@pytest.mark.asyncio
-async def test_without_base_path_still_works(api_client_without_base_path):
-    """
-    Regression test: ensure default behavior (no base path) still works.
-
-    This test verifies that when HINDSIGHT_API_BASE_PATH is not set,
-    the API works at the root path as before.
-    """
-    # Health check at root
-    response = await api_client_without_base_path.get("/health")
-    assert response.status_code == 200
-
-    # Banks endpoint at root
-    response = await api_client_without_base_path.get("/v1/default/banks")
-    assert response.status_code == 200
-
-    # OpenAPI schema should have empty or "/" server path
-    response = await api_client_without_base_path.get("/openapi.json")
-    assert response.status_code == 200
-    openapi_schema = response.json()
-    servers = openapi_schema.get("servers", [])
-    if servers:
-        # Server URL should be empty string (root) or "/"
-        assert servers[0]["url"] in ["", "/"]
-
-
-@pytest.mark.skip(reason="MCP endpoint routing with base path needs investigation")
-@pytest.mark.asyncio
-async def test_base_path_mcp_endpoint(api_client_with_base_path):
-    """Test that MCP endpoint is accessible with base path."""
-    bank_id = "test_mcp_bank"
-
-    # MCP endpoint should be mounted at /mcp/{bank_id}/
-    # The MCP server uses a different protocol, so just check the root exists
-    response = await api_client_with_base_path.get(f"/mcp/{bank_id}/")
-    # MCP may return various status codes, but should not be 404 (not found)
-    # Accept 405 (method not allowed), 400 (bad request), etc.
-    assert response.status_code != 404, "MCP endpoint should exist"
@@ -1,508 +0,0 @@
-"""
-Test OpenAI Batch API integration for retain fact extraction.
-
-Tests cover:
- Normal batch API flow (submit, poll, complete)
- Crash recovery (resume from existing batch_id)
- Provider fallback (when batch API not supported)
- Worker recovery on restart
-"""
-import pytest
-import asyncio
-import logging
-import json
-import uuid
-from datetime import datetime, timezone
-from unittest.mock import AsyncMock, MagicMock, patch
-from hindsight_api import RequestContext
-from hindsight_api.engine.retain.fact_extraction import (
-    extract_facts_from_contents_batch_api,
-    extract_facts_from_contents,
-    RetainContent,
-)
-from hindsight_api.config import HindsightConfig
-from hindsight_api.engine.llm_wrapper import create_llm_provider
-from hindsight_api.worker.poller import WorkerPoller
-
-logger = logging.getLogger(__name__)
-
-
-@pytest.fixture
-def mock_llm_config():
-    """Create a mock LLM config with batch API support."""
-    mock = MagicMock()
-    mock.provider = "openai"
-    mock.model = "gpt-4o-mini"
-    mock._provider_impl = AsyncMock()
-    return mock
-
-
-@pytest.fixture
-def test_contents():
-    """Create test content for fact extraction."""
-    return [
-        RetainContent(
-            content="Alice is a senior software engineer at TechCorp. She specializes in distributed systems.",
-            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
-            context="team overview",
-        ),
-        RetainContent(
-            content="Bob joined the team last month as a junior developer. He is learning React.",
-            event_date=datetime(2024, 1, 15, tzinfo=timezone.utc),
-            context="team overview",
-        ),
-    ]
-
-
-@pytest.fixture
-def hindsight_config():
-    """Create test config with batch API enabled."""
-    config = HindsightConfig.from_env()
-    config.retain_batch_enabled = True
-    config.retain_batch_poll_interval_seconds = 1  # Fast polling for tests
-    config.retain_chunk_size = 4000
-    config.retain_extraction_mode = "concise"
-    config.retain_extract_causal_links = False
-    return config
-
-
-@pytest.mark.asyncio
-async def test_batch_api_normal_flow(mock_llm_config, test_contents, hindsight_config, memory, request_context):
-    """Test normal batch API flow: submit, poll, complete."""
-    bank_id = f"test_batch_{datetime.now(timezone.utc).timestamp()}"
-
-    try:
-        # Mock batch API responses
-        batch_id = "batch_test123"
-
-        # Mock supports_batch_api
-        mock_llm_config._provider_impl.supports_batch_api = AsyncMock(return_value=True)
-
-        # Mock submit_batch - returns batch metadata
-        mock_llm_config._provider_impl.submit_batch = AsyncMock(
-            return_value={
-                "batch_id": batch_id,
-                "status": "validating",
-                "request_counts": {"total": 2, "completed": 0, "failed": 0},
-            }
-        )
-
-        # Mock get_batch_status - simulate polling sequence
-        status_sequence = [
-            {"status": "in_progress", "request_counts": {"total": 2, "completed": 1, "failed": 0}},
-            {"status": "completed", "request_counts": {"total": 2, "completed": 2, "failed": 0}},
-        ]
-        mock_llm_config._provider_impl.get_batch_status = AsyncMock(side_effect=status_sequence)
-
-        # Mock retrieve_batch_results - returns fact extraction results
-        mock_results = [
-            {
-                "custom_id": "chunk_0",
-                "response": {
-                    "body": {
-                        "choices": [
-                            {
-                                "message": {
-                                    "content": json.dumps({
-                                        "facts": [
-                                            {
-                                                "what": "Alice is a senior software engineer at TechCorp",
-                                                "when": "present",
-                                                "where": "TechCorp",
-                                                "who": "Alice",
-                                                "why": "Professional background information",
-                                                "fact_type": "world",
-                                                "fact_kind": "conversation",
-                                            }
-                                        ]
-                                    })
-                                }
-                            }
-                        ],
-                        "usage": {"prompt_tokens": 100, "completion_tokens": 50, "total_tokens": 150},
-                    }
-                },
-            },
-            {
-                "custom_id": "chunk_1",
-                "response": {
-                    "body": {
-                        "choices": [
-                            {
-                                "message": {
-                                    "content": json.dumps({
-                                        "facts": [
-                                            {
-                                                "what": "Bob joined the team last month as a junior developer",
-                                                "when": "last month",
-                                                "where": "team",
-                                                "who": "Bob",
-                                                "why": "New team member information",
-                                                "fact_type": "world",
-                                                "fact_kind": "conversation",
-                                            }
-                                        ]
-                                    })
-                                }
-                            }
-                        ],
-                        "usage": {"prompt_tokens": 100, "completion_tokens": 50, "total_tokens": 150},
-                    }
-                },
-            },
-        ]
-        mock_llm_config._provider_impl.retrieve_batch_results = AsyncMock(return_value=mock_results)
-
-        # Call batch API extraction
-        facts, chunks, usage = await extract_facts_from_contents_batch_api(
-            contents=test_contents,
-            llm_config=mock_llm_config,
-            agent_name="test_agent",
-            config=hindsight_config,
-            pool=None,  # No DB pool for this test
-            operation_id=None,
-            schema=None,
-        )
-
-        # Verify results
-        assert len(facts) == 2, "Should extract 2 facts (one per chunk)"
-        # Facts are ExtractedFact objects with .fact_text field
-        assert "Alice" in facts[0].fact_text and "senior software engineer" in facts[0].fact_text
-        assert "Bob" in facts[1].fact_text and "junior developer" in facts[1].fact_text
-
-        # Verify chunks metadata
-        assert len(chunks) == 2, "Should have 2 chunks metadata"
-        assert chunks[0].fact_count == 1
-        assert chunks[1].fact_count == 1
-
-        # Verify token usage
-        assert usage.input_tokens == 200  # 100 per chunk
-        assert usage.output_tokens == 100  # 50 per chunk
-        assert usage.total_tokens == 300
-
-        # Verify API calls
-        mock_llm_config._provider_impl.submit_batch.assert_called_once()
-        assert mock_llm_config._provider_impl.get_batch_status.call_count == 2
-        mock_llm_config._provider_impl.retrieve_batch_results.assert_called_once_with(batch_id)
-
-        logger.info("✅ Normal batch API flow test passed")
-
-    finally:
-        # Cleanup
-        try:
-            await memory.delete_bank(bank_id, request_context=request_context)
-        except Exception:
-            pass
-
-
-@pytest.mark.asyncio
-async def test_batch_api_crash_recovery(mock_llm_config, test_contents, hindsight_config, memory, request_context):
-    """Test crash recovery: resume polling from existing batch_id."""
-    bank_id = f"test_crash_{datetime.now(timezone.utc).timestamp()}"
-    operation_id = str(uuid.uuid4())  # Must be UUID for async_operations table
-
-    try:
-        # Ensure bank exists
-        await memory.get_bank_profile(bank_id, request_context=request_context)
-
-        # Setup: Store batch_id in async_operations table (simulates partial execution)
-        batch_id = "batch_recovered_456"
-        pool = memory._pool
-        schema = request_context.tenant_id
-
-        from hindsight_api.engine.task_backend import fq_table
-        table = fq_table("async_operations", schema)
-
-        # Create operation with batch_id already stored
-        await pool.execute(
-            f"""
-            INSERT INTO {table} (operation_id, operation_type, bank_id, status, result_metadata)
-            VALUES ($1, 'retain', $2, 'processing', $3::jsonb)
-            """,
-            operation_id,
-            bank_id,
-            json.dumps({
-                "batch_id": batch_id,
-                "batch_provider": "openai",
-                "chunk_count": 2,
-            }),
-        )
-
-        # Mock batch API responses for resume scenario
-        mock_llm_config._provider_impl.supports_batch_api = AsyncMock(return_value=True)
-
-        # Mock get_batch_status - batch already in progress
-        mock_llm_config._provider_impl.get_batch_status = AsyncMock(
-            return_value={
-                "status": "completed",
-                "request_counts": {"total": 2, "completed": 2, "failed": 0},
-            }
-        )
-
-        # Mock retrieve_batch_results
-        mock_results = [
-            {
-                "custom_id": "chunk_0",
-                "response": {
-                    "body": {
-                        "choices": [
-                            {
-                                "message": {
-                                    "content": json.dumps({
-                                        "facts": [
-                                            {
-                                                "what": "Alice is a senior software engineer",
-                                                "when": "present",
-                                                "where": "TechCorp",
-                                                "who": "Alice",
-                                                "why": "Background",
-                                                "fact_type": "world",
-                                                "fact_kind": "conversation",
-                                            }
-                                        ]
-                                    })
-                                }
-                            }
-                        ],
-                        "usage": {"prompt_tokens": 100, "completion_tokens": 50, "total_tokens": 150},
-                    }
-                },
-            },
-            {
-                "custom_id": "chunk_1",
-                "response": {
-                    "body": {
-                        "choices": [
-                            {
-                                "message": {
-                                    "content": json.dumps({
-                                        "facts": [
-                                            {
-                                                "what": "Bob is a junior developer",
-                                                "when": "last month",
-                                                "where": "team",
-                                                "who": "Bob",
-                                                "why": "New member",
-                                                "fact_type": "world",
-                                                "fact_kind": "conversation",
-                                            }
-                                        ]
-                                    })
-                                }
-                            }
-                        ],
-                        "usage": {"prompt_tokens": 100, "completion_tokens": 50, "total_tokens": 150},
-                    }
-                },
-            },
-        ]
-        mock_llm_config._provider_impl.retrieve_batch_results = AsyncMock(return_value=mock_results)
-
-        # Call batch API extraction with operation_id (crash recovery scenario)
-        facts, chunks, usage = await extract_facts_from_contents_batch_api(
-            contents=test_contents,
-            llm_config=mock_llm_config,
-            agent_name="test_agent",
-            config=hindsight_config,
-            pool=pool,
-            operation_id=operation_id,  # Provides crash recovery context
-            schema=schema,
-        )
-
-        # Verify results
-        assert len(facts) == 2, "Should extract 2 facts after recovery"
-
-        # CRITICAL: Verify submit_batch was NOT called (because batch_id already exists)
-        mock_llm_config._provider_impl.submit_batch.assert_not_called()
-
-        # Verify get_batch_status WAS called (polling resumed)
-        mock_llm_config._provider_impl.get_batch_status.assert_called()
-
-        # Verify retrieve_batch_results was called with the recovered batch_id
-        mock_llm_config._provider_impl.retrieve_batch_results.assert_called_once_with(batch_id)
-
-        logger.info("✅ Crash recovery test passed - resumed polling without re-submission")
-
-    finally:
-        # Cleanup
-        try:
-            await memory.delete_bank(bank_id, request_context=request_context)
-        except Exception:
-            pass
-
-
-@pytest.mark.asyncio
-async def test_batch_api_fallback_unsupported_provider(mock_llm_config, test_contents, hindsight_config):
-    """Test fallback to sync mode when provider doesn't support batch API."""
-
-    # Mock provider that doesn't support batch API
-    mock_llm_config._provider_impl.supports_batch_api = AsyncMock(return_value=False)
-    mock_llm_config.provider = "groq"  # Example of provider
-
-    # Patch the sync mode function to verify it's called
-    with patch(
-        "hindsight_api.engine.retain.fact_extraction.extract_facts_from_contents"
-    ) as mock_sync_extract:
-        mock_sync_extract.return_value = ([], [], MagicMock())
-
-        # Call batch API extraction (should fallback to sync)
-        await extract_facts_from_contents_batch_api(
-            contents=test_contents,
-            llm_config=mock_llm_config,
-            agent_name="test_agent",
-            config=hindsight_config,
-            pool=None,
-            operation_id=None,
-            schema=None,
-        )
-
-        # Verify fallback occurred
-        mock_sync_extract.assert_called_once()
-
-        # Verify batch API methods were NOT called
-        mock_llm_config._provider_impl.submit_batch.assert_not_called()
-
-        logger.info("✅ Fallback to sync mode test passed")
-
-
-@pytest.mark.asyncio
-async def test_worker_batch_recovery(memory, request_context):
-    """Test that WorkerPoller._recover_batch_operations finds and resets orphaned batches."""
-    bank_id = f"test_worker_recovery_{datetime.now(timezone.utc).timestamp()}"
-    operation_id = str(uuid.uuid4())  # Must be UUID for async_operations table
-
-    try:
-        # Ensure bank exists
-        await memory.get_bank_profile(bank_id, request_context=request_context)
-
-        pool = memory._pool
-        schema = request_context.tenant_id
-
-        from hindsight_api.engine.task_backend import fq_table
-        table = fq_table("async_operations", schema)
-
-        # Create orphaned batch operation (simulates worker crash during polling)
-        batch_id = "batch_orphaned_999"
-        task_payload = {
-            "operation_type": "retain",
-            "bank_id": bank_id,
-            "contents": [{"content": "test", "event_date": "2024-01-15T00:00:00Z"}],
-        }
-
-        await pool.execute(
-            f"""
-            INSERT INTO {table} (operation_id, operation_type, bank_id, status, worker_id, result_metadata, task_payload)
-            VALUES ($1, 'retain', $2, 'processing', 'worker_crashed', $3::jsonb, $4::jsonb)
-            """,
-            operation_id,
-            bank_id,
-            json.dumps({
-                "batch_id": batch_id,
-                "batch_provider": "openai",
-                "chunk_count": 1,
-            }),
-            json.dumps(task_payload),
-        )
-
-        # Create WorkerPoller
-        from hindsight_api.extensions.builtin.tenant import DefaultTenantExtension
-        tenant_extension = DefaultTenantExtension(config={"schema": schema} if schema else {})
-
-        poller = WorkerPoller(
-            pool=pool,
-            worker_id="test_worker_recovery",
-            executor=memory,
-            poll_interval_ms=100,
-            max_retries=3,
-            schema=schema,
-            tenant_extension=tenant_extension,
-            max_slots=5,
-            consolidation_max_slots=2,
-        )
-
-        # Run recovery
-        recovered_count = await poller._recover_batch_operations(schema)
-
-        # Verify recovery
-        assert recovered_count == 1, "Should recover 1 batch operation"
-
-        # Verify operation was reset to pending
-        row = await pool.fetchrow(
-            f"SELECT status, worker_id FROM {table} WHERE operation_id = $1",
-            operation_id,
-        )
-
-        assert row["status"] == "pending", "Operation should be reset to pending"
-        assert row["worker_id"] is None, "Worker ID should be cleared"
-
-        logger.info("✅ Worker batch recovery test passed")
-
-    finally:
-        # Cleanup
-        try:
-            await memory.delete_bank(bank_id, request_context=request_context)
-        except Exception:
-            pass
-
-
-@pytest.mark.asyncio
-async def test_batch_api_via_extract_facts_from_contents(
-    mock_llm_config, test_contents, hindsight_config, memory, request_context
-):
-    """Test that extract_facts_from_contents routes to batch API when enabled."""
-    bank_id = f"test_routing_{datetime.now(timezone.utc).timestamp()}"
-
-    try:
-        # Enable batch API in config
-        hindsight_config.retain_batch_enabled = True
-
-        # Mock batch API support
-        mock_llm_config._provider_impl.supports_batch_api = AsyncMock(return_value=True)
-        mock_llm_config._provider_impl.submit_batch = AsyncMock(
-            return_value={"batch_id": "batch_123", "status": "validating", "request_counts": {}}
-        )
-        mock_llm_config._provider_impl.get_batch_status = AsyncMock(
-            return_value={"status": "completed", "request_counts": {"total": 1, "completed": 1, "failed": 0}}
-        )
-        mock_llm_config._provider_impl.retrieve_batch_results = AsyncMock(
-            return_value=[
-                {
-                    "custom_id": "chunk_0",
-                    "response": {
-                        "body": {
-                            "choices": [
-                                {
-                                    "message": {
-                                        "content": json.dumps({"facts": []})
-                                    }
-                                }
-                            ],
-                            "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
-                        }
-                    },
-                }
-            ]
-        )
-
-        # Call main extract_facts_from_contents (should route to batch API)
-        facts, chunks, usage = await extract_facts_from_contents(
-            contents=test_contents,
-            llm_config=mock_llm_config,
-            agent_name="test_agent",
-            config=hindsight_config,
-            pool=None,
-            operation_id=None,
-            schema=None,
-        )
-
-        # Verify batch API was called
-        mock_llm_config._provider_impl.submit_batch.assert_called_once()
-
-        logger.info("✅ Routing to batch API test passed")
-
-    finally:
-        # Cleanup
-        try:
-            await memory.delete_bank(bank_id, request_context=request_context)
-        except Exception:
-            pass
@@ -1,263 +0,0 @@
-"""
-Real integration test for OpenAI Batch API.
-
-This test makes REAL API calls to OpenAI and measures actual timing.
-It will be slow (minutes to hours) depending on OpenAI's queue.
-
-To run:
-    pytest tests/test_batch_api_integration.py -v -s
-
-To skip in CI:
-    Add @pytest.mark.skip at the test level
-"""
-import pytest
-import os
-import asyncio
-import logging
-import time
-from datetime import datetime, timezone
-from dotenv import load_dotenv
-from hindsight_api import RequestContext
-from hindsight_api.engine.retain.fact_extraction import (
-    extract_facts_from_contents_batch_api,
-    RetainContent,
-)
-from hindsight_api.config import HindsightConfig
-from hindsight_api.engine.llm_wrapper import LLMProvider
-
-logger = logging.getLogger(__name__)
-
-# Load .env file for API keys
-load_dotenv()
-
-
-@pytest.fixture
-def openai_api_key():
-    """Get OpenAI API key from environment."""
-    # Try both current and commented keys from .env
-    api_key = os.getenv("HINDSIGHT_API_LLM_API_KEY")
-
-    # Check if it's an OpenAI key (starts with sk-proj- or sk-)
-    if not api_key or not api_key.startswith("sk-"):
-        # Try the OpenAI-specific env var (if set separately)
-        api_key = os.getenv("OPENAI_API_KEY")
-
-    if not api_key or not api_key.startswith("sk-"):
-        pytest.skip("OpenAI API key not found in environment. Set OPENAI_API_KEY or uncomment OpenAI config in .env")
-
-    return api_key
-
-
-@pytest.fixture
-def real_llm_config(openai_api_key):
-    """Create real LLM config for OpenAI."""
-    # Create config with OpenAI settings
-    config = HindsightConfig.from_env()
-
-    # Use LLMProvider wrapper (which creates _provider_impl internally)
-    llm_config = LLMProvider(
-        provider="openai",
-        api_key=openai_api_key,
-        base_url="https://api.openai.com/v1",
-        model="gpt-4o-mini",  # Fast, cheap model for testing
-        reasoning_effort="medium",  # Required parameter
-    )
-
-    return llm_config
-
-
-@pytest.fixture
-def test_contents_real():
-    """Create realistic test content for fact extraction."""
-    return [
-        RetainContent(
-            content="""
-            Alice is a senior software engineer at TechCorp, where she has been working for 5 years.
-            She specializes in distributed systems and microservices architecture. Alice graduated
-            from MIT with a degree in Computer Science in 2015. She is known for writing clean,
-            well-documented code and mentoring junior developers.
-            """,
-            event_date=datetime(2024, 1, 15, 10, 30, tzinfo=timezone.utc),
-            context="team member profile",
-        ),
-        RetainContent(
-            content="""
-            Bob joined TechCorp last month as a junior developer. He is learning React and Node.js
-            and recently completed his first feature, which was a user authentication flow. Bob
-            graduated from Berkeley with a degree in Computer Science in 2023. He is enthusiastic
-            and asks great questions during code reviews.
-            """,
-            event_date=datetime(2024, 1, 15, 10, 30, tzinfo=timezone.utc),
-            context="team member profile",
-        ),
-        RetainContent(
-            content="""
-            The team uses Kubernetes for container orchestration and deploys to AWS. They follow
-            agile methodologies with two-week sprints. Code reviews are mandatory before merging
-            any pull request. The team meets every morning for a 15-minute standup to discuss
-            progress and blockers.
-            """,
-            event_date=datetime(2024, 1, 15, 10, 30, tzinfo=timezone.utc),
-            context="team processes",
-        ),
-    ]
-
-
-@pytest.fixture
-def integration_config():
-    """Create config for integration test."""
-    config = HindsightConfig.from_env()
-    config.retain_batch_enabled = True
-    config.retain_batch_poll_interval_seconds = 30  # Poll every 30 seconds (reasonable for real API)
-    config.retain_chunk_size = 4000
-    config.retain_extraction_mode = "concise"
-    config.retain_extract_causal_links = False
-    return config
-
-
-@pytest.mark.skip(reason="Real API test - takes minutes and costs money. Run manually with: pytest tests/test_batch_api_integration.py::test_real_openai_batch_api -v -s")
-@pytest.mark.integration  # Mark as integration test
-@pytest.mark.slow  # Mark as slow test
-@pytest.mark.asyncio
-async def test_real_openai_batch_api(real_llm_config, test_contents_real, integration_config, memory, request_context):
-    """
-    REAL integration test: Submit actual batch to OpenAI and measure timing.
-
-    WARNING: This test:
-    - Makes real API calls to OpenAI
-    - Will take minutes to hours to complete
-    - Costs money (though very little with gpt-4o-mini)
-    - Requires valid OpenAI API key
-
-    To skip this test:
-        pytest tests/test_batch_api_integration.py --skip-integration
-    """
-    bank_id = f"test_real_batch_{datetime.now(timezone.utc).timestamp()}"
-
-    logger.info("=" * 80)
-    logger.info("STARTING REAL OPENAI BATCH API INTEGRATION TEST")
-    logger.info("=" * 80)
-    logger.info(f"Test contents: {len(test_contents_real)} items")
-    logger.info(f"Poll interval: {integration_config.retain_batch_poll_interval_seconds}s")
-    logger.info(f"Model: {real_llm_config.model}")
-    logger.info("This may take several minutes to hours depending on OpenAI's queue...")
-    logger.info("=" * 80)
-
-    try:
-        # Ensure bank exists
-        await memory.get_bank_profile(bank_id, request_context=request_context)
-
-        # Get database pool and schema for crash recovery testing
-        pool = memory._pool
-        schema = request_context.tenant_id
-
-        # Track overall timing
-        test_start_time = time.time()
-
-        # Call REAL batch API extraction
-        logger.info("\n📤 Submitting batch to OpenAI...")
-
-        facts, chunks, usage = await extract_facts_from_contents_batch_api(
-            contents=test_contents_real,
-            llm_config=real_llm_config,
-            agent_name="test_agent",
-            config=integration_config,
-            pool=pool,
-            operation_id=None,  # No crash recovery for this test
-            schema=schema,
-        )
-
-        test_end_time = time.time()
-        total_duration = test_end_time - test_start_time
-
-        # Log results
-        logger.info("\n" + "=" * 80)
-        logger.info("✅ BATCH COMPLETED SUCCESSFULLY")
-        logger.info("=" * 80)
-        logger.info(f"Total duration: {total_duration:.1f} seconds ({total_duration/60:.1f} minutes)")
-        logger.info(f"Facts extracted: {len(facts)}")
-        logger.info(f"Chunks processed: {len(chunks)}")
-        logger.info(f"Token usage: {usage.input_tokens} input + {usage.output_tokens} output = {usage.total_tokens} total")
-        logger.info(f"Estimated cost: ${(usage.input_tokens * 0.00015 / 1000 + usage.output_tokens * 0.0006 / 1000):.4f}")
-        logger.info("=" * 80)
-
-        # Log sample facts
-        logger.info("\n📋 Sample extracted facts:")
-        for i, fact in enumerate(facts[:5]):  # Show first 5 facts
-            logger.info(f"\nFact {i+1}:")
-            logger.info(f"  Type: {fact.fact_type}")
-            logger.info(f"  Text: {fact.fact_text[:100]}...")
-            logger.info(f"  Entities: {fact.entities}")
-
-        # Verify results
-        assert len(facts) > 0, "Should extract at least some facts"
-        assert len(chunks) == len(test_contents_real), f"Should have {len(test_contents_real)} chunks"
-        assert usage.total_tokens > 0, "Should have token usage"
-
-        # Verify fact structure
-        for fact in facts:
-            assert hasattr(fact, "fact_text"), "Fact should have fact_text"
-            assert hasattr(fact, "fact_type"), "Fact should have fact_type"
-            assert fact.fact_type in ["world", "experience", "opinion"], f"Invalid fact_type: {fact.fact_type}"
-
-        logger.info("\n✅ All assertions passed!")
-
-        # Write timing report to file for later analysis
-        report_path = "/tmp/openai_batch_api_timing_report.txt"
-        with open(report_path, "w") as f:
-            f.write(f"OpenAI Batch API Integration Test Report\n")
-            f.write(f"={'=' * 60}\n\n")
-            f.write(f"Test Date: {datetime.now(timezone.utc).isoformat()}\n")
-            f.write(f"Model: {real_llm_config.model}\n")
-            f.write(f"Contents: {len(test_contents_real)} items\n")
-            f.write(f"Poll Interval: {integration_config.retain_batch_poll_interval_seconds}s\n\n")
-            f.write(f"Results:\n")
-            f.write(f"  Total Duration: {total_duration:.1f}s ({total_duration/60:.1f} min)\n")
-            f.write(f"  Facts Extracted: {len(facts)}\n")
-            f.write(f"  Chunks Processed: {len(chunks)}\n")
-            f.write(f"  Token Usage: {usage.total_tokens} ({usage.input_tokens} in + {usage.output_tokens} out)\n")
-            f.write(f"  Estimated Cost: ${(usage.input_tokens * 0.00015 / 1000 + usage.output_tokens * 0.0006 / 1000):.4f}\n")
-
-        logger.info(f"\n📄 Timing report written to: {report_path}")
-
-    finally:
-        # Cleanup
-        try:
-            await memory.delete_bank(bank_id, request_context=request_context)
-            logger.info(f"\n🧹 Cleaned up test bank: {bank_id}")
-        except Exception as e:
-            logger.error(f"Failed to cleanup bank: {e}")
-
-
-@pytest.mark.skip(reason="Real API test - requires Groq API key. Run manually if needed.")
-@pytest.mark.integration
-@pytest.mark.slow
-@pytest.mark.asyncio
-async def test_real_batch_supports_groq(integration_config):
-    """
-    Test that Groq also supports batch API (if configured).
-
-    Groq has the same batch API interface as OpenAI.
-    """
-    groq_api_key = os.getenv("HINDSIGHT_API_LLM_API_KEY")
-
-    if not groq_api_key or not groq_api_key.startswith("gsk_"):
-        pytest.skip("Groq API key not found in environment")
-
-    llm_config = LLMProvider(
-        provider="groq",
-        api_key=groq_api_key,
-        base_url="https://api.groq.com/openai/v1",
-        model="llama-3.1-8b-instant",
-        reasoning_effort="medium",
-    )
-
-    # Check if Groq supports batch API
-    supports_batch = await llm_config._provider_impl.supports_batch_api()
-
-    logger.info(f"Groq batch API support: {supports_batch}")
-
-    # Groq should support batch API (same interface as OpenAI)
-    assert supports_batch, "Groq should support batch API"
-
-    logger.info("✅ Groq batch API support confirmed")
@@ -1,38 +0,0 @@
-"""
-Test validation for batch API + synchronous retain.
-
-When HINDSIGHT_API_RETAIN_BATCH_ENABLED=true, synchronous retain operations
-should be rejected with a 400 error since they will timeout.
-"""
-
-import os
-import pytest
-from hindsight_api.engine.memory_engine import MemoryEngine
-from hindsight_api.config import HindsightConfig
-from hindsight_api import RequestContext
-
-
-@pytest.mark.asyncio
-async def test_batch_api_validation(memory, request_context):
-    """
-    Test that attempting synchronous retain with batch API enabled
-    raises an error at the HTTP layer.
-
-    This test verifies the validation logic exists - actual HTTP testing
-    would require full FastAPI app setup.
-    """
-    # Create config with batch API enabled
-    config = HindsightConfig.from_env()
-    config.retain_batch_enabled = True
-    config.retain_batch_poll_interval_seconds = 1
-
-    # Verify the validation exists in memory engine
-    # The actual HTTP validation happens in http.py api_retain()
-    # This test documents the expected behavior
-
-    assert config.retain_batch_enabled is True
-    assert config.retain_batch_poll_interval_seconds == 1
-
-    # When batch API is enabled and async=false, the HTTP endpoint
-    # should return 400 with message:
-    # "Batch API is enabled (HINDSIGHT_API_RETAIN_BATCH_ENABLED=true) but async=false"
@@ -12,7 +12,6 @@ from datetime import datetime
 import pytest

 from hindsight_api import LLMConfig
-from hindsight_api.config import _get_raw_config
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text


@@ -45,7 +44,6 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
-            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -90,7 +88,6 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
-            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -127,7 +124,6 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
-            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract facts about the causal chain"
@@ -177,7 +173,6 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
-            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract facts"
@@ -214,7 +209,6 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
-            config=_get_raw_config(),
        )

        # Verify relation types are all backward-looking
@@ -10,7 +10,6 @@ from datetime import datetime
 import pytest

 from hindsight_api import LLMConfig
-from hindsight_api.config import _get_raw_config
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text


@@ -38,8 +37,7 @@ After searching for weeks, I finally found a cheaper apartment in Brooklyn.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 3, 15), context=context, llm_config=llm_config, agent_name="TestUser",
-            config=_get_raw_config(),
+            text=text, event_date=datetime(2024, 3, 15), context=context, llm_config=llm_config, agent_name="TestUser"
        )

        assert len(facts) >= 3, f"Should extract at least 3 facts from the causal chain. Got {len(facts)}"
@@ -108,8 +106,7 @@ The renovation took three months and cost $15,000.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 6, 1), context=context, llm_config=llm_config, agent_name="TestUser",
-            config=_get_raw_config(),
+            text=text, event_date=datetime(2024, 6, 1), context=context, llm_config=llm_config, agent_name="TestUser"
        )

        assert len(facts) >= 4, f"Should extract at least 4 facts. Got {len(facts)}"
@@ -139,8 +136,7 @@ Machine learning fascinated me so much that I changed my career to data science.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 1, 1), context=context, llm_config=llm_config, agent_name="TestUser",
-            config=_get_raw_config(),
+            text=text, event_date=datetime(2024, 1, 1), context=context, llm_config=llm_config, agent_name="TestUser"
        )

        # Check no fact references itself
@@ -167,8 +163,7 @@ The new role enabled me to lead a team of engineers.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 2, 15), context=context, llm_config=llm_config, agent_name="TestUser",
-            config=_get_raw_config(),
+            text=text, event_date=datetime(2024, 2, 15), context=context, llm_config=llm_config, agent_name="TestUser"
        )

        # Validate all indices (must reference PREVIOUS facts only)
@@ -195,8 +190,7 @@ Reduced spending somewhat affected local businesses.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 4, 1), context=context, llm_config=llm_config, agent_name="TestUser",
-            config=_get_raw_config(),
+            text=text, event_date=datetime(2024, 4, 1), context=context, llm_config=llm_config, agent_name="TestUser"
        )

        for i, fact in enumerate(facts):
@@ -21,9 +21,9 @@ from hindsight_api.engine.reflect.tools import (
@pytest.fixture(autouse=True)
 def enable_observations():
    """Enable observations for all tests in this module."""
-    from hindsight_api.config import _get_raw_config
+    from hindsight_api.config import get_config

-    config = _get_raw_config()
+    config = get_config()
    original_value = config.enable_observations
    config.enable_observations = True
    yield
@@ -500,7 +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:
@@ -519,7 +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:
@@ -565,26 +563,25 @@ class TestConsolidationDisabled:
        self, memory: MemoryEngine, request_context
    ):
        """Test that consolidation returns disabled status when enable_observations is False."""
+        from unittest.mock import patch
+
        bank_id = f"test-consolidation-disabled-{uuid.uuid4().hex[:8]}"

        # Create the bank
        await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)

-        # Disable observations for this bank via bank config
-        await memory._config_resolver.update_bank_config(
-            bank_id=bank_id,
-            updates={"enable_observations": False},
-            context=request_context,
-        )
+        # Disable observations via config
+        with patch("hindsight_api.config.get_config") as mock_config:
+            mock_config.return_value.enable_observations = False

-        result = await run_consolidation_job(
-            memory_engine=memory,
-            bank_id=bank_id,
-            request_context=request_context,
-        )
+            result = await run_consolidation_job(
+                memory_engine=memory,
+                bank_id=bank_id,
+                request_context=request_context,
+            )

-        assert result["status"] == "disabled"
-        assert result["bank_id"] == bank_id
+            assert result["status"] == "disabled"
+            assert result["bank_id"] == bank_id

        # Cleanup
        await memory.delete_bank(bank_id, request_context=request_context)
@@ -830,7 +827,6 @@ 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:
@@ -852,7 +848,6 @@ 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:
@@ -905,7 +900,6 @@ 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:
@@ -924,7 +918,6 @@ 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:
@@ -937,19 +930,22 @@ 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)
@@ -1026,7 +1022,6 @@ 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(
@@ -1034,7 +1029,6 @@ 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:
@@ -1047,10 +1041,11 @@ 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"

@@ -1435,20 +1430,22 @@ 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",
                )

@@ -1715,10 +1712,11 @@ 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"]])
@@ -1931,7 +1929,9 @@ 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(
@@ -1949,26 +1949,12 @@ 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 = [
@@ -1985,335 +1971,19 @@ 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)
@@ -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
@@ -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):
@@ -608,59 +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
@@ -139,228 +135,3 @@ async def test_memory_without_document(memory, request_context):

    finally:
        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-@pytest.mark.asyncio
-async def test_document_persisted_with_zero_facts(memory, request_context):
-    """
-    Test that documents are persisted even when zero facts are extracted.
-
-    This is a regression test for issue #324 where documents with no extractable
-    facts were reported as disappearing from the system.
-    """
-    bank_id = f"test_zero_facts_{datetime.now(timezone.utc).timestamp()}"
-
-    try:
-        document_id = "doc-zero-facts"
-
-        # Retain content that produces zero facts (gibberish/random characters)
-        units = await memory.retain_async(
-            bank_id=bank_id,
-            content="xyzabc123 !!!### @@@ $$$",  # Random characters unlikely to produce facts
-            context="Test zero facts",
-            document_id=document_id,
-            request_context=request_context,
-        )
-
-        # Should return empty unit list (no facts extracted)
-        assert len(units) == 0, "Should extract zero facts from gibberish content"
-
-        # But document should still be persisted and retrievable
-        doc = await memory.get_document(document_id, bank_id, request_context=request_context)
-        assert doc is not None, "Document should be persisted even with zero facts"
-        assert doc["id"] == document_id
-        assert doc["bank_id"] == bank_id
-        assert doc["memory_unit_count"] == 0, "Should have zero memory units"
-        assert len(doc["original_text"]) > 0, "Should have non-zero text length"
-        assert "xyzabc123" in doc["original_text"], "Should contain original content"
-
-        # Document should also appear in list
-        docs_list = await memory.list_documents(
-            bank_id=bank_id,
-            search_query=None,
-            limit=100,
-            offset=0,
-            request_context=request_context,
-        )
-        assert docs_list["total"] == 1, "Document should appear in list"
-        assert any(d["id"] == document_id for d in docs_list["items"]), "Document should be in items"
-
-        listed_doc = next(d for d in docs_list["items"] if d["id"] == document_id)
-        assert listed_doc["memory_unit_count"] == 0, "Listed document should show zero memory units"
-
-    finally:
-        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-@pytest.mark.asyncio
-async def test_document_persisted_with_zero_facts_batch(memory, request_context):
-    """
-    Test that documents are persisted with zero facts in batch retain operations.
-
-    This tests the async batch code path to ensure it also handles zero facts correctly.
-    """
-    bank_id = f"test_zero_facts_batch_{datetime.now(timezone.utc).timestamp()}"
-
-    try:
-        # Mix of content: some produces facts, some produces zero facts
-        contents = [
-            {
-                "content": "Alice works at Google",
-                "document_id": "doc-with-facts",
-            },
-            {
-                "content": "!@# $$$ %%% ^^^ &&& ***",  # Gibberish - zero facts expected
-                "document_id": "doc-zero-facts",
-            },
-        ]
-
-        unit_ids = await memory.retain_batch_async(
-            bank_id=bank_id,
-            contents=contents,
-            request_context=request_context,
-        )
-
-        # First content should produce facts, second should not
-        assert len(unit_ids[0]) > 0, "First content should produce facts"
-        assert len(unit_ids[1]) == 0, "Second content should produce zero facts"
-
-        # Both documents should be persisted
-        doc_with_facts = await memory.get_document("doc-with-facts", bank_id, request_context=request_context)
-        assert doc_with_facts is not None
-        assert doc_with_facts["memory_unit_count"] > 0
-
-        doc_zero_facts = await memory.get_document("doc-zero-facts", bank_id, request_context=request_context)
-        assert doc_zero_facts is not None, "Document with zero facts should be persisted"
-        assert doc_zero_facts["memory_unit_count"] == 0, "Should have zero memory units"
-        assert "!@#" in doc_zero_facts["original_text"]
-
-        # Both should appear in list
-        docs_list = await memory.list_documents(
-            bank_id=bank_id,
-            search_query=None,
-            limit=100,
-            offset=0,
-            request_context=request_context,
-        )
-        assert docs_list["total"] == 2, "Both documents should appear in list"
-
-    finally:
-        await memory.delete_bank(bank_id, request_context=request_context)
-
-
-@pytest.mark.asyncio
-async def test_document_persisted_with_zero_facts_async_submit(memory, request_context):
-    """
-    Test that documents are persisted with zero facts in fire-and-forget async retain.
-
-    This tests the submit_async_retain (background task) code path to ensure it also
-    handles zero facts correctly.
-    """
-    import asyncio
-
-    bank_id = f"test_zero_facts_async_{datetime.now(timezone.utc).timestamp()}"
-
-    try:
-        # Submit async retain with gibberish content
-        result = await memory.submit_async_retain(
-            bank_id=bank_id,
-            contents=[
-                {
-                    "content": "!@# $$$ %%% ^^^ &&& ***",  # Gibberish - zero facts expected
-                    "document_id": "doc-async-zero-facts",
-                }
-            ],
-            request_context=request_context,
-        )
-
-        operation_id = result["operation_id"]
-        assert operation_id is not None, "Should return operation_id"
-
-        # Wait for background task to complete
-        max_wait = 60  # 60 seconds max
-        wait_interval = 0.5
-        elapsed = 0
-
-        while elapsed < max_wait:
-            await asyncio.sleep(wait_interval)
-            elapsed += wait_interval
-
-            # Check if document exists
-            doc = await memory.get_document(
-                "doc-async-zero-facts", bank_id, request_context=request_context
-            )
-            if doc is not None:
-                break
-
-        # Document should be persisted even with zero facts
-        assert doc is not None, "Document should be persisted after async task completes"
-        assert doc["id"] == "doc-async-zero-facts"
-        assert doc["memory_unit_count"] == 0, "Should have zero memory units"
-        assert "!@#" in doc["original_text"]
-
-        # Document should appear in list
-        docs_list = await memory.list_documents(
-            bank_id=bank_id,
-            search_query=None,
-            limit=100,
-            offset=0,
-            request_context=request_context,
-        )
-        assert docs_list["total"] == 1, "Document should appear in list"
-        assert any(d["id"] == "doc-async-zero-facts" for d in docs_list["items"])
-
-        listed_doc = next(d for d in docs_list["items"] if d["id"] == "doc-async-zero-facts")
-        assert listed_doc["memory_unit_count"] == 0, "Listed document should show zero memory units"
-
-    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:
@@ -8,7 +8,7 @@ from datetime import datetime

 import pytest

-from hindsight_api.config import get_config, clear_config_cache, _get_raw_config
+from hindsight_api.config import get_config, clear_config_cache
 from hindsight_api.engine.llm_wrapper import LLMConfig
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text

@@ -58,7 +58,6 @@ async def test_fact_extraction_basic_analysis(llm_config):
        llm_config=llm_config,
        agent_name="test-agent",
        context="Friday Standup meeting",
-        config=_get_raw_config(),
    )

    duration = time.time() - start_time
--- a/Show More
+++ b/Show More