refactor deamin

2025-12-22 22:00:54 +01:00 · 2025-12-22 21:50:19 +01:00 · 2025-12-22 21:43:35 +01:00 · 2025-12-22 21:30:54 +01:00 · 2025-12-22 18:20:00 +01:00 · 2025-12-22 18:19:55 +01:00
676 changed files with 34561 additions and 110924 deletions
@@ -2,30 +2,11 @@
 # Copy this file to .env and fill in your values

 # LLM Configuration (Required)
-# 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=o3-mini
 HINDSIGHT_API_LLM_BASE_URL=https://api.openai.com/v1

-# Example: Anthropic Claude configuration
-# HINDSIGHT_API_LLM_PROVIDER=anthropic
-# HINDSIGHT_API_LLM_API_KEY=your-anthropic-api-key
-# HINDSIGHT_API_LLM_MODEL=claude-sonnet-4-20250514
-
-# Example: Google Vertex AI configuration
-# HINDSIGHT_API_LLM_PROVIDER=vertexai
-# HINDSIGHT_API_LLM_MODEL=google/gemini-2.0-flash-001
-# HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=your-gcp-project-id
-# HINDSIGHT_API_LLM_VERTEXAI_REGION=us-central1
-# HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/path/to/service-account-key.json  # Optional, uses ADC if not set
-
-# Example: LM Studio local configuration (Qwen 2.5 32B recommended)
-# HINDSIGHT_API_LLM_PROVIDER=lmstudio
-# HINDSIGHT_API_LLM_API_KEY=lmstudio
-# HINDSIGHT_API_LLM_BASE_URL=http://localhost:1234/v1
-# HINDSIGHT_API_LLM_MODEL=qwen2.5-32b-instruct
-
 # API Configuration (Optional)
 HINDSIGHT_API_HOST=0.0.0.0
 HINDSIGHT_API_PORT=8888
@@ -33,7 +14,6 @@ HINDSIGHT_API_LOG_LEVEL=info

 # 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)

 # Embeddings Configuration (Optional - uses local by default)
 # Provider: "local" (default) or "tei" (HuggingFace Text Embeddings Inference)
@@ -1,71 +0,0 @@
-name: Bug Report
-description: Report a bug or unexpected behavior
-labels: ["bug", "triage"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to report a bug! Please fill out the sections below.
-
-  - type: textarea
-    id: description
-    attributes:
-      label: Bug Description
-      description: A clear and concise description of the bug
-      placeholder: What happened?
-    validations:
-      required: true
-
-  - type: textarea
-    id: reproduction
-    attributes:
-      label: Steps to Reproduce
-      description: Steps to reproduce the behavior
-      placeholder: |
-        1. Configure '...'
-        2. Call '...'
-        3. See error
-    validations:
-      required: true
-
-  - type: textarea
-    id: expected
-    attributes:
-      label: Expected Behavior
-      description: What did you expect to happen?
-    validations:
-      required: true
-
-  - type: textarea
-    id: actual
-    attributes:
-      label: Actual Behavior
-      description: What actually happened?
-    validations:
-      required: true
-
-  - type: input
-    id: version
-    attributes:
-      label: Version
-      description: What version are you using?
-      placeholder: e.g., 0.1.0 or commit hash
-    validations:
-      required: false
-
-  - type: dropdown
-    id: llm-provider
-    attributes:
-      label: LLM Provider
-      description: Which LLM provider are you using?
-      options:
-        - OpenAI
-        - Anthropic
-        - Gemini
-        - Groq
-        - Ollama
-        - LM Studio
-        - Other
-    validations:
-      required: false
-
@@ -1,8 +0,0 @@
-blank_issues_enabled: false
-contact_links:
-  - name: Questions & Help
-    url: https://github.com/vectorize-io/hindsight/discussions/categories/q-a
-    about: Please ask questions and get help in Discussions instead of opening an issue.
-  - name: Ideas & Feedback
-    url: https://github.com/vectorize-io/hindsight/discussions/categories/ideas
-    about: Share ideas or give feedback in Discussions.
@@ -1,82 +0,0 @@
-name: Feature Request
-description: Suggest a new feature or enhancement
-labels: ["enhancement", "triage"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for suggesting a feature! Please describe what you'd like to see added.
-
-  - type: textarea
-    id: use-case
-    attributes:
-      label: Use Case
-      description: Describe your specific use case. What are you building? What's your goal?
-      placeholder: |
-        I'm building an AI agent that needs to...
-        My application handles...
-    validations:
-      required: true
-
-  - type: textarea
-    id: problem
-    attributes:
-      label: Problem Statement
-      description: What problem are you facing? What's missing or difficult today?
-      placeholder: Currently I have to... which causes...
-    validations:
-      required: true
-
-  - type: textarea
-    id: benefit
-    attributes:
-      label: How This Feature Would Help
-      description: Explain how this feature would improve your workflow or solve your problem
-      placeholder: With this feature, I would be able to...
-    validations:
-      required: true
-
-  - type: textarea
-    id: solution
-    attributes:
-      label: Proposed Solution
-      description: Describe your ideal solution (optional - we may have ideas too!)
-      placeholder: It would be great if Hindsight could...
-    validations:
-      required: false
-
-  - type: textarea
-    id: alternatives
-    attributes:
-      label: Alternatives Considered
-      description: Have you considered any alternative solutions or workarounds?
-    validations:
-      required: false
-
-  - type: dropdown
-    id: priority
-    attributes:
-      label: Priority
-      description: How important is this feature to you?
-      options:
-        - Nice to have
-        - Important - affects my workflow
-        - Critical - blocking my use case
-    validations:
-      required: true
-
-  - type: textarea
-    id: additional
-    attributes:
-      label: Additional Context
-      description: Any other context, mockups, or examples?
-    validations:
-      required: false
-
-  - type: checkboxes
-    id: checklist
-    attributes:
-      label: Checklist
-      options:
-        - label: I would be willing to contribute this feature
-          required: false
@@ -139,55 +139,6 @@ jobs:
        path: hindsight-clients/typescript/*.tgz
        retention-days: 1

-  release-openclawd-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/openclawd
-      run: npm ci
-
-    - name: Build
-      working-directory: ./hindsight-integrations/openclawd
-      run: npm run build
-
-    - name: Publish to npm
-      working-directory: ./hindsight-integrations/openclawd
-      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/openclawd
-      run: npm pack
-
-    - name: Upload artifacts
-      uses: actions/upload-artifact@v4
-      with:
-        name: openclawd-integration
-        path: hindsight-integrations/openclawd/*.tgz
-        retention-days: 1
-
  release-control-plane:
    runs-on: ubuntu-latest
    environment: npm
@@ -415,7 +366,7 @@ jobs:

  create-github-release:
    runs-on: ubuntu-latest
-    needs: [release-python-packages, release-typescript-client, release-openclawd-integration, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
+    needs: [release-python-packages, release-typescript-client, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
    permissions:
      contents: write

@@ -438,12 +389,6 @@ jobs:
        name: typescript-client
        path: ./artifacts/typescript-client

-    - name: Download OpenClawd Integration
-      uses: actions/download-artifact@v4
-      with:
-        name: openclawd-integration
-        path: ./artifacts/openclawd-integration
-
    - name: Download Control Plane
      uses: actions/download-artifact@v4
      with:
@@ -485,8 +430,6 @@ jobs:
        cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
        # TypeScript client
        cp artifacts/typescript-client/*.tgz release-assets/ || true
-        # OpenClawd Integration
-        cp artifacts/openclawd-integration/*.tgz release-assets/ || true
        # Control Plane
        cp artifacts/control-plane/*.tgz release-assets/ || true
        # Rust CLI binaries
@@ -82,29 +82,6 @@ jobs:
    - name: Build TypeScript client
      run: npm run build --workspace=hindsight-clients/typescript

-  build-openclawd-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/openclawd
-      run: npm ci
-
-    - name: Run tests
-      working-directory: ./hindsight-integrations/openclawd
-      run: npm test
-
-    - name: Build
-      working-directory: ./hindsight-integrations/openclawd
-      run: npm run build
-
  build-control-plane:
    runs-on: ubuntu-latest

@@ -176,15 +153,8 @@ jobs:
    - name: Build docs
      run: npm run build --workspace=hindsight-docs

-  test-rust-cli:
+  build-rust-cli:
    runs-on: ubuntu-latest
-    env:
-      HINDSIGHT_API_LLM_PROVIDER: groq
-      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
-      HINDSIGHT_API_URL: http://localhost:8888
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu

    steps:
    - uses: actions/checkout@v4
@@ -201,10 +171,6 @@ jobs:
          hindsight-cli/target
        key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}

-    - name: Run unit tests
-      working-directory: hindsight-cli
-      run: cargo test
-
    - name: Build CLI
      working-directory: hindsight-cli
      run: cargo build --release
@@ -216,60 +182,6 @@ jobs:
        path: hindsight-cli/target/release/hindsight
        retention-days: 1

-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-        prune-cache: false
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Build API
-      working-directory: ./hindsight-api
-      run: uv build
-
-    - name: Install API dependencies
-      working-directory: ./hindsight-api
-      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
-
-    - name: Create .env file
-      run: |
-        cat > .env << EOF
-        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
-        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
-        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        EOF
-
-    - name: Start API server
-      run: |
-        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
-        echo "Waiting for API server to be ready..."
-        for i in {1..60}; do
-          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
-            echo "API server is ready after ${i}s"
-            break
-          fi
-          if [ $i -eq 60 ]; then
-            echo "API server failed to start after 60s"
-            cat /tmp/api-server.log
-            exit 1
-          fi
-          sleep 1
-        done
-
-    - name: Run CLI smoke test
-      run: |
-        HINDSIGHT_CLI=hindsight-cli/target/release/hindsight ./hindsight-cli/smoke-test.sh
-
-    - name: Show API server logs
-      if: always()
-      run: |
-        echo "=== API Server Logs ==="
-        cat /tmp/api-server.log || echo "No API server log found"
-
  lint-helm-chart:
    runs-on: ubuntu-latest

@@ -336,8 +248,6 @@ jobs:
      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)
@@ -363,7 +273,7 @@ jobs:

    - name: Install dependencies
      working-directory: ./hindsight-api
-      run: uv sync --frozen --extra test --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --extra test --no-install-project --index-strategy unsafe-best-match

    - name: Cache HuggingFace models
      uses: actions/cache@v4
@@ -424,11 +334,11 @@ jobs:

    - name: Install client test dependencies
      working-directory: ./hindsight-clients/python
-      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
+      run: uv sync --extra test --index-strategy unsafe-best-match

    - name: Install API dependencies
      working-directory: ./hindsight-api
-      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --no-install-project --index-strategy unsafe-best-match

    - name: Create .env file
      run: |
@@ -501,7 +411,7 @@ jobs:

    - name: Install API dependencies
      working-directory: ./hindsight-api
-      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --no-install-project --index-strategy unsafe-best-match

    - name: Install TypeScript client dependencies
      working-directory: ./hindsight-clients/typescript
@@ -589,7 +499,7 @@ jobs:

    - name: Install API dependencies
      working-directory: ./hindsight-api
-      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --no-install-project --index-strategy unsafe-best-match

    - name: Create .env file
      run: |
@@ -626,97 +536,6 @@ jobs:
        echo "=== API Server Logs ==="
        cat /tmp/api-server.log || echo "No API server log found"

-  test-integration:
-    runs-on: ubuntu-latest
-    env:
-      HINDSIGHT_API_LLM_PROVIDER: groq
-      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
-      HINDSIGHT_API_URL: http://localhost:8888
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-        prune-cache: false
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Build API
-      working-directory: ./hindsight-api
-      run: uv build
-
-    - name: Install API dependencies
-      working-directory: ./hindsight-api
-      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
-
-    - name: Install integration test dependencies
-      working-directory: ./hindsight-integration-tests
-      run: uv sync --frozen
-
-    - name: Cache HuggingFace models
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/huggingface
-        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
-        restore-keys: |
-          ${{ runner.os }}-huggingface-
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
-    - name: Create .env file
-      run: |
-        cat > .env << EOF
-        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
-        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
-        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
-        EOF
-
-    - name: Start API server
-      run: |
-        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
-        echo "Waiting for API server to be ready..."
-        for i in {1..60}; do
-          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
-            echo "API server is ready after ${i}s"
-            break
-          fi
-          if [ $i -eq 60 ]; then
-            echo "API server failed to start after 60s"
-            cat /tmp/api-server.log
-            exit 1
-          fi
-          sleep 1
-        done
-
-    - name: Run integration tests
-      working-directory: ./hindsight-integration-tests
-      run: uv run pytest tests/ -v
-
-    - name: Show API server logs
-      if: always()
-      run: |
-        echo "=== API Server Logs ==="
-        cat /tmp/api-server.log || echo "No API server log found"
-
  test-litellm-integration:
    runs-on: ubuntu-latest

@@ -740,7 +559,7 @@ jobs:

    - name: Install dependencies
      working-directory: ./hindsight-integrations/litellm
-      run: uv sync --frozen --extra dev
+      run: uv sync --extra dev

    - name: Run tests
      working-directory: ./hindsight-integrations/litellm
@@ -771,7 +590,7 @@ jobs:

    - name: Install dependencies
      working-directory: ./hindsight-embed
-      run: uv sync --frozen --index-strategy unsafe-best-match
+      run: uv sync --index-strategy unsafe-best-match

    - name: Cache HuggingFace models
      uses: actions/cache@v4
@@ -788,7 +607,7 @@ jobs:

  test-doc-examples:
    runs-on: ubuntu-latest
-    needs: test-rust-cli
+    needs: build-rust-cli
    env:
      HINDSIGHT_API_LLM_PROVIDER: groq
      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
@@ -831,11 +650,11 @@ jobs:
      working-directory: ./hindsight-api
      run: |
        uv build
-        uv sync --frozen --no-install-project --index-strategy unsafe-best-match
+        uv sync --no-install-project --index-strategy unsafe-best-match

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

    - name: Install TypeScript client
      run: |
@@ -896,136 +715,4 @@ jobs:
      if: always()
      run: |
        echo "=== API Server Logs ==="
-        cat /tmp/api-server.log || echo "No API server log found"
-
-  test-upgrade:
-    runs-on: ubuntu-latest
-    env:
-      HINDSIGHT_API_LLM_PROVIDER: groq
-      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
-
-    steps:
-    - uses: actions/checkout@v4
-      with:
-        fetch-depth: 0  # Full history needed for git clone of tags
-
-    - name: Fetch tags
-      run: git fetch --tags
-
-    - 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: 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: Install hindsight-dev dependencies
-      working-directory: ./hindsight-dev
-      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
-
-    - name: Install current hindsight-api
-      working-directory: ./hindsight-api
-      run: uv sync --frozen --index-strategy unsafe-best-match
-
-    - name: Pre-download models
-      working-directory: ./hindsight-api
-      run: |
-        uv run python -c "
-        from sentence_transformers import SentenceTransformer, CrossEncoder
-        print('Downloading embedding model...')
-        SentenceTransformer('BAAI/bge-small-en-v1.5')
-        print('Downloading cross-encoder model...')
-        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
-        print('Models downloaded successfully')
-        "
-
-    - name: Run upgrade tests
-      working-directory: ./hindsight-dev
-      run: uv run pytest upgrade_tests/ -v --tb=short
-
-  verify-generated-files:
-    runs-on: ubuntu-latest
-    env:
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Set up Node.js
-      uses: actions/setup-node@v4
-      with:
-        node-version: '20'
-        cache: 'npm'
-        cache-dependency-path: package-lock.json
-
-    - name: Install Rust
-      uses: dtolnay/rust-toolchain@stable
-
-    - name: Cache cargo
-      uses: actions/cache@v4
-      with:
-        path: |
-          ~/.cargo/registry
-          ~/.cargo/git
-        key: ${{ runner.os }}-cargo-gen-${{ hashFiles('**/Cargo.lock') }}
-
-    - name: Install Node dependencies
-      run: npm ci
-
-    - name: Install Python dependencies
-      run: |
-        cd hindsight-dev && uv sync --frozen --index-strategy unsafe-best-match
-        cd ../hindsight-api && uv sync --frozen --index-strategy unsafe-best-match
-        cd ../hindsight-embed && uv sync --frozen --index-strategy unsafe-best-match
-
-    - name: Run generate-openapi
-      run: ./scripts/generate-openapi.sh
-
-    - name: Run generate-clients
-      run: ./scripts/generate-clients.sh
-
-    - name: Run lint
-      run: ./scripts/hooks/lint.sh
-
-    - name: Verify no uncommitted changes
-      run: |
-        if [ -n "$(git status --porcelain)" ]; then
-          echo "❌ Error: Generated files are out of sync with committed files."
-          echo ""
-          echo "The following files have changed after running generation scripts:"
-          git status --porcelain
-          echo ""
-          echo "Please run the following commands locally and commit the changes:"
-          echo "  ./scripts/generate-openapi.sh"
-          echo "  ./scripts/generate-clients.sh"
-          echo "  ./scripts/hooks/lint.sh"
-          echo ""
-          git diff --stat
-          exit 1
-        fi
-        echo "✓ All generated files are up to date"
+        cat /tmp/api-server.log || echo "No API server log found"
@@ -5,18 +5,15 @@ build/
 dist/
 wheels/
 *.egg-info
-.mcp.json
-.osgrep
+
 # Virtual environments
 .venv

 # Node
 node_modules/

-# Environment variables and local config
+# Environment variables
 .env
-docker-compose.yml
-docker-compose.override.yml

 # IDE
 .idea/
@@ -27,10 +24,6 @@ docker-compose.override.yml
 # NLTK data (will be downloaded automatically)
 nltk_data/

-# Monitoring stack (Prometheus/Grafana binaries and data)
-.monitoring/
-.pgbouncer/
-
 # Large benchmark datasets (will be downloaded automatically)
 **/longmemeval_s_cleaned.json

@@ -45,12 +38,5 @@ hindsight-docs/static/llms-full.txt

 hindsight-dev/benchmarks/locomo/results/
 hindsight-dev/benchmarks/longmemeval/results/
-hindsight-dev/benchmarks/consolidation/results/
-benchmarks/results/
 hindsight-cli/target
-hindsight-clients/rust/target
-.claude
-whats-next.md
-TASK.md
-# Changelog is now tracked in hindsight-docs/src/pages/changelog.md
-# CHANGELOG.md
+hindsight-clients/rust/target
@@ -1,3 +1,153 @@
 # AGENTS.md

-See [CLAUDE.md](./CLAUDE.md) for project documentation and coding conventions.
+This document captures architectural decisions and coding conventions for the Hindsight project.
+
+## Documentation
+
+- **Main documentation**: [hindsight-docs/docs/developer/](./hindsight-docs/docs/developer/)
+- **Use case patterns**: [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/)
+- **API reference**: Auto-generated from OpenAPI spec
+
+## Project Structure
+
+```
+hindsight/              # Python package for embedded usage
+hindsight-api/          # FastAPI server (core memory engine)
+hindsight-cli/          # Rust CLI client
+hindsight-embed/        # Embedded CLI (no server needed)
+hindsight-control-plane/ # Next.js admin UI
+hindsight-docs/         # Docusaurus documentation site
+hindsight-dev/          # Development tools and benchmarks
+hindsight-integrations/ # Framework integrations (LangChain, etc.)
+hindsight-clients/      # Generated API clients (Python, TypeScript, Rust)
+```
+
+## Core Concepts
+
+### Memory Banks
+- Each bank is an isolated memory store (like a "brain" for one user/agent)
+- Banks contain: memory units (facts), entities, documents, entity links
+- Banks have a **disposition** (personality traits) and **background** (context)
+- Bank isolation is strict - no cross-bank data leakage
+
+### Memory Types
+- **World facts**: General knowledge ("The sky is blue")
+- **Experience facts**: Personal experiences ("I visited Paris in 2023")
+- **Opinion facts**: Beliefs with confidence scores ("Paris is beautiful" - 0.9 confidence)
+
+### Operations
+- **Retain**: Store new memories (extracts facts, entities, relationships)
+- **Recall**: Retrieve memories (semantic, BM25, graph, temporal search)
+- **Reflect**: Deep analysis to form new insights/opinions
+
+## API Design Decisions
+
+### Single Bank Per Request
+- All API endpoints (`recall`, `reflect`, `retain`) operate on a single bank
+- Multi-bank queries are the **client/agent's responsibility** to orchestrate
+- This keeps the API simple and the isolation model clear
+
+### Disposition Traits (3-trait system)
+- **Skepticism** (1-5): How skeptical vs trusting when forming opinions
+- **Literalism** (1-5): How literally to interpret information
+- **Empathy** (1-5): How much to consider emotional context
+- These influence the `reflect` operation, not `recall`
+- Background info also only affects `reflect` (opinion formation)
+
+## Multi-Bank Architecture Patterns
+
+See [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/) for detailed guides:
+
+- **Per-User Memory**: One bank per user, simplest pattern
+- **Support Agent + Shared Knowledge**: User bank + shared docs bank, client orchestrates
+
+## Developer Guide
+
+### Running the API Server
+
+```bash
+# From project root
+./scripts/dev/start-api.sh
+
+# With options
+./scripts/dev/start-api.sh --reload --port 8888 --log-level debug
+```
+
+### Running Tests
+
+```bash
+# API tests
+cd hindsight-api
+uv run pytest tests/
+
+# Specific test
+uv run pytest tests/test_http_api_integration.py -v
+```
+
+### Generating OpenAPI Spec
+
+After changing API endpoints, regenerate the OpenAPI spec and docs:
+
+```bash
+./scripts/generate-openapi.sh
+```
+
+This will:
+1. Generate `openapi.json` at project root
+2. Copy to `hindsight-docs/openapi.json`
+3. Regenerate API reference documentation
+
+### Generating API Clients
+
+After updating the OpenAPI spec, regenerate all clients:
+
+```bash
+./scripts/generate-clients.sh
+```
+
+This generates:
+- **Rust client**: `hindsight-clients/rust/` (via progenitor in build.rs)
+- **Python client**: `hindsight-clients/python/` (via openapi-generator Docker)
+- **TypeScript client**: `hindsight-clients/typescript/` (via @hey-api/openapi-ts)
+
+Note: The maintained wrapper `hindsight_client.py` and `README.md` are preserved during regeneration.
+
+### Running the Documentation Site
+
+```bash
+./scripts/dev/start-docs.sh
+```
+
+### Running the Control Plane
+
+```bash
+./scripts/dev/start-control-plane.sh
+```
+
+## Code Style
+
+### Python (hindsight-api)
+- Use `uv` for package management
+- Async throughout (asyncpg, async FastAPI endpoints)
+- Pydantic models for request/response validation
+- No py files at project root - maintain clean directory structure
+
+### TypeScript (control-plane, clients)
+- Next.js with App Router for control plane
+- Tailwind CSS with shadcn/ui components
+
+### Rust (CLI)
+- Async with tokio
+- reqwest for HTTP client
+- progenitor for API client generation
+
+## Database
+
+- PostgreSQL with pgvector extension
+- Schema managed via Alembic migrations in `hindsight-api/alembic/`, db migrations happen during api startup, no manual commands
+- Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
+
+# Branding
+## Colors
+- Primary: gradient from #0074d9 to #009296
+
@@ -1,282 +0,0 @@
-# CLAUDE.md
-
-This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
-
-## Project Overview
-
-Hindsight is an agent memory system that provides long-term memory for AI agents using biomimetic data structures. Memories are organized as:
- **World facts**: General knowledge ("The sky is blue")
- **Experience facts**: Personal experiences ("I visited Paris in 2023")
- **Mental models**: Consolidated knowledge synthesized from facts ("User prefers functional programming patterns")
-
-## Development Commands
-
-### API Server (Python/FastAPI)
-```bash
-# Start API server (loads .env automatically)
-./scripts/dev/start-api.sh
-
-# Run all tests (parallelized with pytest-xdist)
-cd hindsight-api && uv run pytest tests/
-
-# Run specific test file
-cd hindsight-api && uv run pytest tests/test_http_api_integration.py -v
-
-# Run single test function
-cd hindsight-api && uv run pytest tests/test_retain.py::test_retain_simple -v
-
-# Lint and format
-cd hindsight-api && uv run ruff check .
-cd hindsight-api && uv run ruff format .
-
-# Type checking (uses ty - extremely fast type checker from Astral)
-cd hindsight-api && uv run ty check hindsight_api/
-```
-
-### Control Plane (Next.js)
-```bash
-./scripts/dev/start-control-plane.sh
-# Or manually:
-cd hindsight-control-plane && npm run dev
-```
-
-### Documentation Site (Docusaurus)
-```bash
-./scripts/dev/start-docs.sh
-```
-
-### Generating Clients/OpenAPI
-```bash
-# Regenerate OpenAPI spec after API changes (REQUIRED after changing endpoints)
-./scripts/generate-openapi.sh
-
-# Regenerate all client SDKs (Python, TypeScript, Rust)
-./scripts/generate-clients.sh
-```
-
-### Benchmarks
-```bash
-./scripts/benchmarks/run-longmemeval.sh
-./scripts/benchmarks/run-locomo.sh
-./scripts/benchmarks/start-visualizer.sh  # View results at localhost:8001
-```
-
-## Architecture
-
-### Monorepo Structure
- **hindsight-api/**: Core FastAPI server with memory engine (Python, uv)
- **hindsight/**: Embedded Python bundle (hindsight-all package)
- **hindsight-control-plane/**: Admin UI (Next.js, npm)
- **hindsight-cli/**: CLI tool (Rust, cargo, uses progenitor for API client)
- **hindsight-clients/**: Generated SDK clients (Python, TypeScript, Rust)
- **hindsight-docs/**: Docusaurus documentation site
- **hindsight-integrations/**: Framework integrations (LiteLLM, OpenAI)
- **hindsight-dev/**: Development tools and benchmarks
-
-### Core Engine (hindsight-api/hindsight_api/engine/)
- `memory_engine.py`: Main orchestrator (~170KB) for retain/recall/reflect operations
- `llm_wrapper.py`: LLM abstraction supporting OpenAI, Anthropic, Gemini, Groq, Ollama, LM Studio
- `embeddings.py`: Embedding generation (local sentence-transformers or TEI)
- `cross_encoder.py`: Reranking (local or TEI)
- `entity_resolver.py`: Entity extraction and normalization
- `query_analyzer.py`: Query intent analysis
-
-**retain/**: Memory ingestion pipeline
- `orchestrator.py`: Coordinates the retain flow
- `fact_extraction.py`: LLM-based fact extraction from content
- `link_utils.py`: Entity link creation and management
-
-**search/**: Multi-strategy retrieval
- `retrieval.py`: Main retrieval orchestrator
- `graph_retrieval.py`: Entity/relationship graph traversal
- `mpfp_retrieval.py`: Multi-Path Fact Propagation retrieval
- `fusion.py`: Reciprocal rank fusion for combining results
- `reranking.py`: Cross-encoder reranking
-
-### API Layer (hindsight-api/hindsight_api/api/)
- `http.py`: FastAPI HTTP routers (~80KB) for all REST endpoints
- `mcp.py`: Model Context Protocol server implementation
-
-Main operations:
- **Retain**: Store memories, extracts facts/entities/relationships
- **Recall**: Retrieve memories via 4 parallel strategies (semantic, BM25, graph, temporal) + reranking
- **Reflect**: Disposition-aware reasoning using memories and mental models.
-
-### Database
-PostgreSQL with pgvector. Schema managed via Alembic migrations in `hindsight-api/hindsight_api/alembic/`. Migrations run automatically on API startup.
-
-Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
-
-### Adding Database Migrations
-
-1. **Create a new migration file** in `hindsight-api/hindsight_api/alembic/versions/`:
-   - File name format: `<revision_id>_<description>.py` (e.g., `f1a2b3c4d5e6_add_new_index.py`)
-   - Use a unique hex revision ID (12 chars)
-   - Set `down_revision` to the previous migration's revision ID
-
-2. **Migration template**:
-   ```python
-   """Description of the migration
-
-   Revision ID: f1a2b3c4d5e6
-   Revises: <previous_revision_id>
-   Create Date: YYYY-MM-DD
-   """
-   from collections.abc import Sequence
-   from alembic import context, op
-
-   revision: str = "f1a2b3c4d5e6"
-   down_revision: str | Sequence[str] | None = "<previous_revision_id>"
-   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"CREATE INDEX ... ON {schema}table_name(...)")
-
-   def downgrade() -> None:
-       schema = _get_schema_prefix()
-       op.execute(f"DROP INDEX IF EXISTS {schema}index_name")
-   ```
-
-3. **Run migrations locally**:
-   ```bash
-   # Set database URL and run migrations
-   uv run hindsight-admin run-db-migration
-
-   # Run on a specific tenant schema
-   uv run hindsight-admin run-db-migration --schema tenant_xyz
-   ```
-
-## Key Conventions
-
-### Code Quality
-**Always run the lint script after making Python or TypeScript/Node changes:**
-```bash
-./scripts/hooks/lint.sh
-```
-This runs the same checks as the pre-commit hook (Ruff for Python, ESLint/Prettier for TypeScript).
-
-### Memory Banks
- Each bank is an isolated memory store (like a "brain" for one user/agent)
- Banks have dispositions (skepticism, literalism, empathy traits 1-5) affecting reflect
- Banks can have background context
- Bank isolation is strict - no cross-bank data leakage
-
-### API Design
- All endpoints operate on a single bank per request
- Multi-bank queries are client responsibility to orchestrate
- Disposition traits only affect reflect, not recall
-
-### Control Plane API Routes
-
-When adding or modifying parameters in the dataplane API (hindsight-api), you must also update the control plane routes that proxy to it:
-
-1. **API Routes** (`hindsight-control-plane/src/app/api/`):
-   - `recall/route.ts` - proxies to `/v1/default/banks/{bank_id}/memories/recall`
-   - `reflect/route.ts` - proxies to `/v1/default/banks/{bank_id}/reflect`
-   - `memories/retain/route.ts` - proxies to `/v1/default/banks/{bank_id}/memories/retain`
-   - Other routes follow the same pattern
-
-2. **Client types** (`hindsight-control-plane/src/lib/api.ts`):
-   - Update the TypeScript type definitions for `recall()`, `reflect()`, `retain()` etc.
-
-3. **Checklist when adding new API parameters**:
-   - Add parameter extraction in the route handler (destructure from `body`)
-   - Pass the parameter to the SDK call
-   - Update the client type definition in `lib/api.ts`
-   - Update any UI components that need to use the new parameter
-
-### Python Style
- Python 3.11+, type hints required
- Async throughout (asyncpg, async FastAPI)
- Pydantic models for request/response
- Ruff for linting (line-length 120)
- No Python files at project root - maintain clean directory structure
- **Never use multi-item tuple return values** - prefer dataclass or Pydantic model for structured returns
-
-### Type Safety with Pydantic Models
-**NEVER use raw `dict` types for structured data.** Always use Pydantic models:
- Use Pydantic `BaseModel` for all data structures passed between functions
- Add `@field_validator` for type coercion (e.g., ensuring datetimes are timezone-aware)
- Avoid `dict.get()` patterns - use typed model attributes instead
- Parse external data (JSON, API responses) into Pydantic models at the boundary
- This catches type errors at parse time, not deep in business logic
-
-```python
-# BAD - error-prone dict access
-def process(data: dict) -> str:
-    return data.get("name", "")  # No validation, silent failures
-
-# GOOD - typed and validated
-class UserData(BaseModel):
-    name: str
-    created_at: datetime
-
-    @field_validator("created_at", mode="before")
-    @classmethod
-    def ensure_tz_aware(cls, v):
-        if isinstance(v, str):
-            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
-        if v.tzinfo is None:
-            return v.replace(tzinfo=timezone.utc)
-        return v
-
-def process(data: UserData) -> str:
-    return data.name  # Type-safe, validated at construction
-```
-
-### TypeScript Style
- Next.js App Router for control plane
- Tailwind CSS with shadcn/ui components
-
-### Adding New API Configuration Flags
-
-When adding a new environment variable configuration:
-
-1. **config.py** (`hindsight-api/hindsight_api/config.py`):
-   - Add `ENV_*` constant for the environment variable name
-   - Add `DEFAULT_*` constant for the default value
-   - Add field to `HindsightConfig` dataclass
-   - Add initialization in `from_env()` method
-
-2. **main.py** (`hindsight-api/hindsight_api/main.py`):
-   - Add field to the manual `HindsightConfig()` constructor call (search for "CLI override")
-
-3. **Use the config** in code:
-   ```python
-   from ...config import get_config
-   config = get_config()
-   value = config.your_new_field
-   ```
-
-4. **Documentation** (`hindsight-docs/docs/developer/configuration.md`):
-   - Add to appropriate section table with Variable, Description, Default
-
-## Environment Setup
-
-```bash
-cp .env.example .env
-# Edit .env with LLM API key
-
-# Python deps
-uv sync --directory hindsight-api/
-
-# Node deps (uses npm workspaces)
-npm install
-```
-
-Required env vars:
- `HINDSIGHT_API_LLM_PROVIDER`: openai, anthropic, gemini, groq, ollama, lmstudio
- `HINDSIGHT_API_LLM_API_KEY`: Your API key
- `HINDSIGHT_API_LLM_MODEL`: Model name (e.g., o3-mini, claude-sonnet-4-20250514)
-
-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)
@@ -51,36 +51,7 @@ cd hindsight-api
 uv run pytest tests/
 ```

-### Code Style
-
-We use [Ruff](https://docs.astral.sh/ruff/) for Python linting and formatting, and ESLint/Prettier for TypeScript.
-
-#### Setting up git hooks (recommended)
-
-Set up git hooks to automatically lint and format code before each commit:
-
-```bash
-./scripts/setup-hooks.sh
-```
-
-This configures git to use the hooks in `.githooks/`, which run all scripts in `scripts/hooks/` on commit. The lint hook runs in parallel:
- **Python**: `ruff check --fix`, `ruff format`, `ty check`
- **TypeScript**: `eslint --fix`, `prettier`
-
-#### Manual linting and formatting
-
-```bash
-# Run all lints (same as pre-commit)
-./scripts/hooks/lint.sh
-
-# Or run individually for Python:
-cd hindsight-api
-uv run ruff check --fix .   # Lint and auto-fix
-uv run ruff format .        # Format code
-uv run ty check hindsight_api  # Type check
-```
-
-#### Style guidelines
+### Code style

 - Use Python type hints
 - Follow existing code patterns
@@ -1,11 +1,11 @@
 <div align="center">

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

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

 [![CI](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml)
-[![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
+[![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 ![PyPI - Downloads](https://img.shields.io/pypi/dm/hindsight-api?label=PyPI)
 ![NPM Downloads](https://img.shields.io/npm/dm/%40vectorize-io%2Fhindsight-client?logoColor=orange&label=NPM&color=blue&link=https%3A%2F%2Fwww.npmjs.com%2Fpackage%2F%40vectorize-io%2Fhindsight-client)
@@ -17,31 +17,55 @@

 ## What is Hindsight?

-Hindsight™ is an agent memory system built to create smarter agents that learn over time. Most agent memory systems focus on recalling conversation history. Hindsight is focused on making agents that learn, not just remember.
+Hindsight™ is an agent memory system built to create smarter agents that learn over time. It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.

+Hindsight addresses common challenges that have frustrated AI engineers building agents to automate tasks and assist users with conversational interfaces. Many of these challenges stem directly from a lack of memory.

-<video src="https://github.com/user-attachments/assets/923b798d-3581-4897-bb62-9cfa5a931682" controls></video>
+- **Inconsistency:** Agents complete tasks successfully one time, then fail when asked to complete the same task again. Memory gives the agent a mechanism to remember what worked and what didn't and to use that information to reduce errors and improve consistency.
+- **Hallucinations:** Long term memory can be seeded with external knowledge to ground agent behavior in reliable sources to augment training data.
+- **Cognitive Overload:** As workflows get complex, retrievals, tool calls, user messages and agent responses can grow to fill the context window leading to context rot. Short term memory optimization allows agents to reduce tokens and focus context by removing irrelevant details.

-It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.
+## How is Hindsight Different From Other Memory Systems?
+
+![Overview](./hindsight-docs/static/img/hindsight-overview.webp)
+
+Most agent memory implementation rely on basic vector search or sometimes use a knowledge graph. Hindsight uses biomimetic data structures to organize agent memories in a way that is more like how human memory works:
+
+- **World:** Facts about the world ("The stove gets hot")
+- **Experiences:** Agent's own experiences ("I touched the stove and it really hurt")
+- **Opinion:** Beliefs with confidence scores ("I shouldn't touch the stove again" - .99 confidence)
+- **Observation:** Complex mental models derived by reflecting on facts and experiences ("Curling irons, ovens, and fire are also hot. I shouldn't touch those either.")
+
+Memories in Hindsight are stored in banks (i.e. memory banks). When memories are added to Hindsight, they are pushed into either the world facts or experiences memory pathway. They are then represented as a combination of entities, relationships, and time series with sparse/dense vector representations to aid in later recall.
+
+Hindsight provides three simple methods to interact with the system:
+
+- **Retain:** Provide information to Hindsight that you want it to remember
+- **Recall:** Retrieve memories from Hindsight
+- **Reflect:** Reflect on memories and experiences to generate new observations and insights from existing memories.
+
+### Agent Memory That Learns
+
+A key goal of Hindsight is to build agent memory that enables agents to learn and improve over time. This is the role of the `reflect` operation which provides the agent to form broader opinions and observations over time.
+
+For example, imagine a product support agent that is helping a user troubleshoot a problem. It uses a `search-documentation` tool it found on an MCP server. Later in the conversation, the agent discovers that the documentation returned from the tool wasn't for the product the user was asking about. The agent now has an experience in its memory bank. And just like humans, we want that agent to learn from its experience.
+
+As the agent gains more experiences, `reflect` allows the agent to form observations about what worked, what didn't, and what to do differently the next time it encounters a similar task.
+
+---

 ## Memory Performance & Accuracy

-Hindsight is the most accurate agent memory system ever tested according to benchmark performance. It has achieved state-of-the-art performance on the LongMemEval benchmark, widely used to assess memory system performance across a variety of conversational AI scenarios. The current reported performance of Hindsight and other agent memory solutions as of January 2026 is shown here:
+Hindsight has achieved state-of-the-art performance on the LongMemEval benchmark, widely used to assess memory system performance across a variety of conversational
+AI scenarios. The current reported performance of Hindsight and other agent memory solutions as of December 2025 is shown here:

 ![Overview](./hindsight-docs/static/img/hindsight-bench.jpg)

-The benchmark performance data for Hindsight has been independently reproduced by research collaborators at the Virginia Tech [Sanghani Center for Artificial Intelligence and Data Analytics](https://sanghani.cs.vt.edu/) and The Washington Post. Other scores are self-reported by software vendors.
+The benchmark performance data for Hindsight and GPT-4o (full context) have been reproduced by research collaborators at the Virginia Tech [Sanghani Center for Artificial Intelligence and Data Analytics](https://sanghani.cs.vt.edu/) and The Washington Post. Other scores are self-reported by software vendors.

-Hindsight is being used in production at Fortune 500 enterprises and by a growing number of AI startups. 
-
-## Adding Hindsight to Your AI Agents
-
-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.
-
-![Hindsight Banner](./hindsight-docs/static/img/migration-code.png)
+A thorough examination of the techniques implemented in Hindsight and detailed breakdowns of benchmark performance are [available on arXiv](https://arxiv.org/abs/2512.12818). This research is currently being prepared for conference submission and the wider peer review process.

+The benchmark results from this research can be inspected in our [visual benchmark explorer](https://hindsight-benchmarks.vercel.app). As additional improvements are made to Hindsight, new benchmark data will be available for review using this same tool.

 ## Quick Start

@@ -57,8 +81,6 @@ docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
  ghcr.io/vectorize-io/hindsight:latest
 ```

-You can modify the LLM provider by setting `HINDSIGHT_API_LLM_PROVIDER`. Valid options are `openai`, `anthropic`, `gemini`, `groq`, `ollama`, and `lmstudio`. The documentation provides more details on [supported models](https://hindsight.vectorize.io/developer/models).
-
 API: http://localhost:8888  
 UI: http://localhost:9999

@@ -124,45 +146,8 @@ await client.recall('my-bank', 'What does Alice like?');

 ---

-## Use Cases
-
-
-Hindsight is built to support conversational AI agents as well as agents that are intended to perform tasks autonomously. The ideal use case for Hindsight are agents that require a blend of these features such as AI employees that need to handle open-ended tasks, change behavior based on user feedback, and learn to perform complex tasks to automate work at a level that approximates a human work. Hindsight can be used with simple AI workflows like those built with n8n and other similar tools, but may be overkill for such applications.
-
-### Per-User Memories and Chat History
-
-One of the simpler use cases you can use Hindsight for is to personalize AI chatbots and other conversational agents by storing and recalling memories associated with individual users.
-
-The requirements for this use case usually look something like this:
-
-![Per-User Memories](./hindsight-docs/static/img/per-user-memory-requirements.png)
-
-<video src="https://github.com/user-attachments/assets/4805e8e1-e7d1-47c6-a4f8-2344a5ec8906" controls></video>
-
-Satisfying these requirements in Hindsight is straightforward. When new user inputs and tool calls are ingested into Hindsight using the retain operation, custom metadata can be used to enrich the new memories. Metadata provides a convenient way to isolate memories that need to be restricted to a given user. Once these are fed into the retain operation, any raw memories and mental models that get created can be filtered when retrieving relevant memories. 
-
-![Per-User Memories](./hindsight-docs/static/img/per-user-memory-howto.png)
-
---
-
 ## Architecture & Operations

-![Overview](./hindsight-docs/static/img/hindsight-overview.webp)
-
-Most agent memory implementation rely on basic vector search or sometimes use a knowledge graph. Hindsight uses biomimetic data structures to organize agent memories in a way that is more like how human memory works:
-
- **World:** Facts about the world ("The stove gets hot")
- **Experiences:** Agent's own experiences ("I touched the stove and it really hurt")
- **Mental Models:** Learned understanding of the agent's world formed by reflecting on raw memories and experiences.
-
-Memories in Hindsight are stored in banks (i.e. memory banks). When memories are added to Hindsight, they are pushed into either the world facts or experiences memory pathway. They are then represented as a combination of entities, relationships, and time series with sparse/dense vector representations to aid in later recall.
-
-Hindsight provides three simple methods to interact with the system:
-
- **Retain:** Provide information to Hindsight that you want it to remember
- **Recall:** Retrieve memories from Hindsight
- **Reflect:** Reflect on memories and experiences to generate new observations and insights from existing memories.
-
 ### Retain

 The `retain` operation is used to push new memories into Hindsight. It tells Hindsight to _retain_ the information you pass in as an input.
@@ -221,7 +206,7 @@ The final output is trimmed as needed to fit within the token limit.

 ### Reflect

-The reflect operation is used to perform a more thorough analysis of existing memories. This allows the agent to form new connections between memories and build a more thorough understanding of its world.
+The reflect operation is used to perform a more thorough analysis of existing memories. This allows the agent to form new connections between memories which are then persisted as opinions and/or observations. When building agents, the reflect operation is a key capability to enable the agent to learn from its experiences. 

 For example, the `reflect` operation can be used to support use cases such as:

@@ -255,7 +240,7 @@ client.reflect(bank_id="my-bank", query="What should I know about Alice?")
 - [CLI](https://hindsight.vectorize.io/sdks/cli)

 **Community:**
- [Slack](https://join.slack.com/t/hindsight-space/shared_invite/zt-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
+- [Slack](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
 - [GitHub Issues](https://github.com/vectorize-io/hindsight/issues)

 ---
@@ -2,24 +2,19 @@
 # Supports building API-only, Control Plane-only, or both
 #
 # Build args:
-#   INCLUDE_API=true/false         - Include API (default: true)
-#   INCLUDE_CP=true/false          - Include Control Plane (default: true)
-#   INCLUDE_LOCAL_MODELS=true/false - Include local ML models for embeddings/reranking (default: true)
-#                                     Set to false when using external providers (TEI, OpenAI, Cohere)
-#   PRELOAD_ML_MODELS=true/false   - Pre-download ML models during build (default: true)
-#                                     Only effective when INCLUDE_LOCAL_MODELS=true
+#   INCLUDE_API=true/false       - Include API (default: true)
+#   INCLUDE_CP=true/false        - Include Control Plane (default: true)
+#   PRELOAD_ML_MODELS=true/false - Pre-download ML models during build (default: true)
 #
 # Examples:
 #   docker build -t hindsight .                                          # Both (standalone)
 #   docker build -t hindsight-api --build-arg INCLUDE_CP=false .         # API only
 #   docker build -t hindsight-cp --build-arg INCLUDE_API=false .         # Control Plane only
 #   docker build -t hindsight --build-arg PRELOAD_ML_MODELS=false .      # Skip ML model preload
-#   docker build -t hindsight --build-arg INCLUDE_LOCAL_MODELS=false .   # Skip local ML deps (for external providers)

 ARG INCLUDE_API=true
 ARG INCLUDE_CP=true
 ARG PRELOAD_ML_MODELS=true
-ARG INCLUDE_LOCAL_MODELS=true

 # =============================================================================
 # Stage: API Builder
@@ -27,7 +22,6 @@ ARG INCLUDE_LOCAL_MODELS=true
 FROM python:3.11-slim AS api-builder

 ARG INCLUDE_API
-ARG INCLUDE_LOCAL_MODELS
 RUN if [ "$INCLUDE_API" != "true" ]; then echo "Skipping API build" && exit 0; fi

 WORKDIR /app
@@ -46,15 +40,6 @@ COPY hindsight-api/README.md ./api/

 WORKDIR /app/api

-# Remove local ML model dependencies if INCLUDE_LOCAL_MODELS=false
-# This creates a smaller image when using external providers (TEI, OpenAI, Cohere)
-RUN if [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then \
-        echo "Removing local-models dependencies (sentence-transformers, torch, transformers)..." && \
-        sed -i '/"sentence-transformers/d' pyproject.toml && \
-        sed -i '/"transformers/d' pyproject.toml && \
-        sed -i '/"torch/d' pyproject.toml; \
-    fi
-
 # Sync dependencies (will create lock file if needed)
 RUN uv sync

@@ -140,6 +125,7 @@ FROM python:3.11-slim AS api-only

 WORKDIR /app

+# Install pg0 dependencies (procps provides 'kill' command needed by pg0)
 # Note: libicu version varies by Debian version - try common versions in order
 RUN apt-get update && apt-get install -y \
    curl \
@@ -152,6 +138,7 @@ RUN apt-get update && apt-get install -y \
    && rm -rf /var/lib/apt/lists/* \
    && pip install --no-cache-dir uv

+# Create non-root user (PostgreSQL cannot run as root)
 RUN useradd -m -s /bin/bash hindsight

 # Copy API with virtual environment from builder
@@ -161,17 +148,23 @@ COPY --from=api-builder /app/api /app/api
 COPY docker/standalone/start-all.sh /app/start-all.sh
 RUN chmod +x /app/start-all.sh

-RUN chown -R hindsight:hindsight /app
+# Create data directory for pg0 and set ownership
+RUN mkdir -p /app/data && chown -R hindsight:hindsight /app

+# Switch to non-root user
 USER hindsight

+# Set PATH for hindsight user
 ENV PATH="/app/api/.venv/bin:${PATH}"

+# Pre-cache PostgreSQL binaries by starting/stopping pg0-embedded
+ENV PG0_HOME=/home/hindsight/.pg0-cache
+
+ENV PG0_HOME=/home/hindsight/.pg0
+
 # Pre-download ML models to avoid runtime download (conditional)
-# Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
 ARG PRELOAD_ML_MODELS
-ARG INCLUDE_LOCAL_MODELS
-RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
+RUN if [ "$PRELOAD_ML_MODELS" = "true" ]; then \
    /app/api/.venv/bin/python -c "\
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
@@ -179,7 +172,6 @@ SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
 print('Models cached successfully')"; \
-    elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
    else echo "Skipping ML model preload"; fi

 EXPOSE 8888
@@ -234,7 +226,7 @@ FROM python:3.11-slim AS standalone

 WORKDIR /app

-# Install Node.js, curl, uv, and system dependencies
+# Install Node.js, curl, uv, and pg0 dependencies (procps provides 'kill' command needed by pg0)
 # Note: libicu version varies by Debian version - try common versions in order
 RUN apt-get update && apt-get install -y \
    curl \
@@ -249,6 +241,7 @@ RUN apt-get update && apt-get install -y \
    && rm -rf /var/lib/apt/lists/* \
    && pip install --no-cache-dir uv

+# Create non-root user (PostgreSQL cannot run as root)
 RUN useradd -m -s /bin/bash hindsight

 # Copy API with virtual environment from builder
@@ -269,17 +262,30 @@ WORKDIR /app
 COPY docker/standalone/start-all.sh /app/start-all.sh
 RUN chmod +x /app/start-all.sh

-RUN chown -R hindsight:hindsight /app
+# Create data directory for pg0 and set ownership
+RUN mkdir -p /app/data && chown -R hindsight:hindsight /app

+# Switch to non-root user
 USER hindsight

+# Set PATH for hindsight user
 ENV PATH="/app/api/.venv/bin:${PATH}"

+# Pre-cache PostgreSQL binaries by starting/stopping pg0-embedded
+ENV PG0_HOME=/home/hindsight/.pg0-cache
+RUN /app/api/.venv/bin/python -c "\
+from pg0 import Pg0; \
+print('Pre-caching PostgreSQL binaries...'); \
+pg = Pg0(name='hindsight', port=5555, username='hindsight', password='hindsight', database='hindsight'); \
+pg.start(); \
+pg.stop(); \
+print('PostgreSQL pre-cached to PG0_HOME')" || echo "Pre-download skipped"
+
+ENV PG0_HOME=/home/hindsight/.pg0
+
 # Pre-download ML models to avoid runtime download (conditional)
-# Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
 ARG PRELOAD_ML_MODELS
-ARG INCLUDE_LOCAL_MODELS
-RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
+RUN if [ "$PRELOAD_ML_MODELS" = "true" ]; then \
    /app/api/.venv/bin/python -c "\
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
@@ -287,7 +293,6 @@ SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
 print('Models cached successfully')"; \
-    elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
    else echo "Skipping ML model preload"; fi

 EXPOSE 8888 9999
@@ -5,70 +5,16 @@ set -e
 ENABLE_API="${HINDSIGHT_ENABLE_API:-true}"
 ENABLE_CP="${HINDSIGHT_ENABLE_CP:-true}"

-# =============================================================================
-# Dependency waiting (opt-in via HINDSIGHT_WAIT_FOR_DEPS=true)
-#
-# Problem: When running with LM Studio, the LLM may take time to load models.
-# If Hindsight starts before LM Studio is ready, it fails on LLM verification.
-# This wait loop ensures dependencies are ready before starting.
-# =============================================================================
-if [ "${HINDSIGHT_WAIT_FOR_DEPS:-false}" = "true" ]; then
-    LLM_BASE_URL="${HINDSIGHT_API_LLM_BASE_URL:-http://host.docker.internal:1234/v1}"
-    MAX_RETRIES="${HINDSIGHT_RETRY_MAX:-0}"  # 0 = infinite
-    RETRY_INTERVAL="${HINDSIGHT_RETRY_INTERVAL:-10}"
-
-    # Check if external database is configured (skip check for embedded pg0)
-    SKIP_DB_CHECK=false
-    if [ -z "${HINDSIGHT_API_DATABASE_URL}" ]; then
-        SKIP_DB_CHECK=true
-    else
-        DB_CHECK_HOST=$(echo "$HINDSIGHT_API_DATABASE_URL" | sed -E 's|.*@([^:/]+):([0-9]+)/.*|\1 \2|')
+# Copy pre-cached PostgreSQL data if runtime directory is empty (first run with volume)
+if [ "$ENABLE_API" = "true" ]; then
+    PG0_CACHE="/home/hindsight/.pg0-cache"
+    PG0_HOME="/home/hindsight/.pg0"
+    if [ -d "$PG0_CACHE" ] && [ "$(ls -A $PG0_CACHE 2>/dev/null)" ]; then
+        if [ ! "$(ls -A $PG0_HOME 2>/dev/null)" ]; then
+            echo "📦 Copying pre-cached PostgreSQL data..."
+            cp -r "$PG0_CACHE"/* "$PG0_HOME"/ 2>/dev/null || true
+        fi
    fi
-
-    check_db() {
-        if $SKIP_DB_CHECK; then
-            return 0
-        fi
-        if command -v pg_isready &> /dev/null; then
-            pg_isready -h $(echo $DB_CHECK_HOST | cut -d' ' -f1) -p $(echo $DB_CHECK_HOST | cut -d' ' -f2) &>/dev/null
-        else
-            python3 -c "import socket; s=socket.socket(); s.settimeout(5); exit(0 if s.connect_ex(('$(echo $DB_CHECK_HOST | cut -d' ' -f1)', $(echo $DB_CHECK_HOST | cut -d' ' -f2))) == 0 else 1)" 2>/dev/null
-        fi
-    }
-
-    check_llm() {
-        curl -sf "${LLM_BASE_URL}/models" --connect-timeout 5 &>/dev/null
-    }
-
-    echo "⏳ Waiting for dependencies to be ready..."
-    attempt=1
-
-    while true; do
-        db_ok=false
-        llm_ok=false
-
-        if check_db; then
-            db_ok=true
-        fi
-
-        if check_llm; then
-            llm_ok=true
-        fi
-
-        if $db_ok && $llm_ok; then
-            echo "✅ Dependencies ready!"
-            break
-        fi
-
-        if [ "$MAX_RETRIES" -ne 0 ] && [ "$attempt" -ge "$MAX_RETRIES" ]; then
-            echo "❌ Max retries ($MAX_RETRIES) reached. Dependencies not available."
-            exit 1
-        fi
-
-        echo "   Attempt $attempt: DB=$( $db_ok && echo 'ok' || echo 'waiting' ), LLM=$( $llm_ok && echo 'ok' || echo 'waiting' )"
-        sleep "$RETRY_INTERVAL"
-        ((attempt++))
-    done
 fi

 # Track PIDs for wait
@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.4.2
-appVersion: "0.4.2"
+version: 0.1.12
+appVersion: "0.1.12"
 keywords:
  - ai
  - memory
@@ -80,22 +80,6 @@ Control plane selector labels
 app.kubernetes.io/component: control-plane
 {{- end }}

-{{/*
-Worker labels
-*/}}
-{{- define "hindsight.worker.labels" -}}
-{{ include "hindsight.labels" . }}
-app.kubernetes.io/component: worker
-{{- end }}
-
-{{/*
-Worker selector labels
-*/}}
-{{- define "hindsight.worker.selectorLabels" -}}
-{{ include "hindsight.selectorLabels" . }}
-app.kubernetes.io/component: worker
-{{- end }}
-
 {{/*
 Create the name of the service account to use
 */}}
@@ -126,14 +110,3 @@ API URL for control plane
 {{- define "hindsight.apiUrl" -}}
 {{- printf "http://%s-api:%d" (include "hindsight.fullname" .) (.Values.api.service.port | int) }}
 {{- end }}
-
-{{/*
-Get the name of the secret to use
-*/}}
-{{- define "hindsight.secretName" -}}
-{{- if .Values.existingSecret }}
-{{- .Values.existingSecret }}
-{{- else }}
-{{- printf "%s-secret" (include "hindsight.fullname" .) }}
-{{- end }}
-{{- end }}
@@ -15,9 +15,7 @@ spec:
  template:
    metadata:
      annotations:
-        {{- if not .Values.existingSecret }}
        checksum/secret: {{ include (print $.Template.BasePath "/secret.yaml") . | sha256sum }}
-        {{- end }}
        {{- with .Values.podAnnotations }}
        {{- toYaml . | nindent 8 }}
        {{- end }}
@@ -39,41 +37,27 @@ spec:
        - name: http
          containerPort: {{ .Values.api.service.targetPort }}
          protocol: TCP
-        {{- if .Values.existingSecret }}
-        envFrom:
-        - secretRef:
-            name: {{ .Values.existingSecret }}
-        {{- end }}
        env:
-        {{- /* POSTGRES_PASSWORD must be defined before DATABASE_URL for $(VAR) interpolation */}}
+        - name: HINDSIGHT_API_DATABASE_URL
+          value: {{ include "hindsight.databaseUrl" . | quote }}
        {{- if not .Values.postgresql.enabled }}
        - name: POSTGRES_PASSWORD
          valueFrom:
            secretKeyRef:
-              name: {{ include "hindsight.secretName" . }}
+              name: {{ include "hindsight.fullname" . }}-secret
              key: postgres-password
        {{- end }}
-        - name: HINDSIGHT_API_DATABASE_URL
-          value: {{ include "hindsight.databaseUrl" . | quote }}
-        {{- /* Disable internal worker when dedicated workers are enabled */}}
-        {{- if .Values.worker.enabled }}
-        - name: HINDSIGHT_API_WORKER_ENABLED
-          value: "false"
-        {{- end }}
        {{- range $key, $value := .Values.api.env }}
        - name: {{ $key }}
          value: {{ $value | quote }}
        {{- end }}
-        {{- /* Only use api.secrets when not using existingSecret (for chart-managed secrets) */}}
-        {{- if not .Values.existingSecret }}
        {{- range $key, $value := .Values.api.secrets }}
        - name: {{ $key }}
          valueFrom:
            secretKeyRef:
-              name: {{ include "hindsight.secretName" $ }}
+              name: {{ include "hindsight.fullname" $ }}-secret
              key: {{ $key }}
        {{- end }}
-        {{- end }}
        livenessProbe:
          {{- toYaml .Values.api.livenessProbe | nindent 10 }}
        readinessProbe:
@@ -15,9 +15,7 @@ spec:
  template:
    metadata:
      annotations:
-        {{- if not .Values.existingSecret }}
        checksum/secret: {{ include (print $.Template.BasePath "/secret.yaml") . | sha256sum }}
-        {{- end }}
        {{- with .Values.podAnnotations }}
        {{- toYaml . | nindent 8 }}
        {{- end }}
@@ -39,11 +37,6 @@ spec:
        - name: http
          containerPort: {{ .Values.controlPlane.service.targetPort }}
          protocol: TCP
-        {{- if .Values.existingSecret }}
-        envFrom:
-        - secretRef:
-            name: {{ .Values.existingSecret }}
-        {{- end }}
        env:
        - name: HINDSIGHT_CP_DATAPLANE_API_URL
          value: {{ include "hindsight.apiUrl" . | quote }}
@@ -51,16 +44,13 @@ spec:
        - name: {{ $key }}
          value: {{ $value | quote }}
        {{- end }}
-        {{- /* Only use controlPlane.secrets when not using existingSecret (for chart-managed secrets) */}}
-        {{- if not .Values.existingSecret }}
        {{- range $key, $value := .Values.controlPlane.secrets }}
        - name: {{ $key }}
          valueFrom:
            secretKeyRef:
-              name: {{ include "hindsight.secretName" $ }}
+              name: {{ include "hindsight.fullname" $ }}-secret
              key: {{ $key }}
        {{- end }}
-        {{- end }}
        livenessProbe:
          {{- toYaml .Values.controlPlane.livenessProbe | nindent 10 }}
        readinessProbe:
@@ -1,8 +1,7 @@
-{{- if not .Values.existingSecret }}
 apiVersion: v1
 kind: Secret
 metadata:
-  name: {{ include "hindsight.secretName" . }}
+  name: {{ include "hindsight.fullname" . }}-secret
  labels:
    {{- include "hindsight.labels" . | nindent 4 }}
 type: Opaque
@@ -16,4 +15,3 @@ data:
  {{- if and (not .Values.postgresql.enabled) .Values.postgresql.external.password }}
  postgres-password: {{ .Values.postgresql.external.password | b64enc | quote }}
  {{- end }}
-{{- end }}
@@ -1,25 +0,0 @@
-{{- if .Values.worker.enabled }}
-apiVersion: v1
-kind: Service
-metadata:
-  name: {{ include "hindsight.fullname" . }}-worker
-  labels:
-    {{- include "hindsight.worker.labels" . | nindent 4 }}
-  {{- if .Values.podAnnotations }}
-  annotations:
-    {{- /* Common Prometheus annotations for metrics scraping */}}
-    prometheus.io/scrape: "true"
-    prometheus.io/port: {{ .Values.worker.service.port | quote }}
-    prometheus.io/path: "/metrics"
-  {{- end }}
-spec:
-  # Headless service for StatefulSet (enables stable DNS names like worker-0.worker.namespace)
-  clusterIP: None
-  ports:
-  - port: {{ .Values.worker.service.port }}
-    targetPort: {{ .Values.worker.service.targetPort }}
-    protocol: TCP
-    name: http
-  selector:
-    {{- include "hindsight.worker.selectorLabels" . | nindent 4 }}
-{{- end }}
@@ -1,110 +0,0 @@
-{{- if .Values.worker.enabled }}
-apiVersion: apps/v1
-kind: StatefulSet
-metadata:
-  name: {{ include "hindsight.fullname" . }}-worker
-  labels:
-    {{- include "hindsight.worker.labels" . | nindent 4 }}
-spec:
-  serviceName: {{ include "hindsight.fullname" . }}-worker
-  replicas: {{ .Values.worker.replicaCount }}
-  selector:
-    matchLabels:
-      {{- include "hindsight.worker.selectorLabels" . | nindent 6 }}
-  template:
-    metadata:
-      annotations:
-        {{- if not .Values.existingSecret }}
-        checksum/secret: {{ include (print $.Template.BasePath "/secret.yaml") . | sha256sum }}
-        {{- end }}
-        {{- with .Values.podAnnotations }}
-        {{- toYaml . | nindent 8 }}
-        {{- end }}
-      labels:
-        {{- include "hindsight.worker.selectorLabels" . | nindent 8 }}
-    spec:
-      {{- if .Values.serviceAccount.create }}
-      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
-      {{- end }}
-      securityContext:
-        {{- toYaml .Values.podSecurityContext | nindent 8 }}
-      containers:
-      - name: worker
-        securityContext:
-          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version }}"
-        imagePullPolicy: {{ .Values.worker.image.pullPolicy }}
-        command: ["hindsight-worker"]
-        ports:
-        - name: http
-          containerPort: {{ .Values.worker.service.targetPort }}
-          protocol: TCP
-        {{- if .Values.existingSecret }}
-        envFrom:
-        - secretRef:
-            name: {{ .Values.existingSecret }}
-        {{- end }}
-        env:
-        {{- /* POSTGRES_PASSWORD must be defined before DATABASE_URL for $(VAR) interpolation */}}
-        {{- if not .Values.postgresql.enabled }}
-        - name: POSTGRES_PASSWORD
-          valueFrom:
-            secretKeyRef:
-              name: {{ include "hindsight.secretName" . }}
-              key: postgres-password
-        {{- end }}
-        - name: HINDSIGHT_API_DATABASE_URL
-          value: {{ include "hindsight.databaseUrl" . | quote }}
-        {{- /* Worker ID uses pod name (StatefulSet provides stable names like worker-0, worker-1) */}}
-        - name: HINDSIGHT_API_WORKER_ID
-          valueFrom:
-            fieldRef:
-              fieldPath: metadata.name
-        {{- /* Inherit LLM config from api.env */}}
-        {{- range $key, $value := .Values.api.env }}
-        - name: {{ $key }}
-          value: {{ $value | quote }}
-        {{- end }}
-        {{- /* Worker-specific env vars */}}
-        {{- range $key, $value := .Values.worker.env }}
-        - name: {{ $key }}
-          value: {{ $value | quote }}
-        {{- end }}
-        {{- /* Only use secrets when not using existingSecret */}}
-        {{- if not .Values.existingSecret }}
-        {{- /* Inherit secrets from api.secrets */}}
-        {{- range $key, $value := .Values.api.secrets }}
-        - name: {{ $key }}
-          valueFrom:
-            secretKeyRef:
-              name: {{ include "hindsight.secretName" $ }}
-              key: {{ $key }}
-        {{- end }}
-        {{- /* Worker-specific secrets (can override api.secrets) */}}
-        {{- range $key, $value := .Values.worker.secrets }}
-        - name: {{ $key }}
-          valueFrom:
-            secretKeyRef:
-              name: {{ include "hindsight.secretName" $ }}
-              key: {{ $key }}
-        {{- end }}
-        {{- end }}
-        livenessProbe:
-          {{- toYaml .Values.worker.livenessProbe | nindent 10 }}
-        readinessProbe:
-          {{- toYaml .Values.worker.readinessProbe | nindent 10 }}
-        resources:
-          {{- toYaml .Values.worker.resources | nindent 10 }}
-      {{- with .Values.nodeSelector }}
-      nodeSelector:
-        {{- toYaml . | nindent 8 }}
-      {{- end }}
-      {{- with .Values.affinity }}
-      affinity:
-        {{- toYaml . | nindent 8 }}
-      {{- end }}
-      {{- with .Values.tolerations }}
-      tolerations:
-        {{- toYaml . | nindent 8 }}
-      {{- end }}
-{{- end }}
@@ -3,15 +3,6 @@
 # Chart version - use this to set a consistent image tag across all components
 version: "0.1.1"

-# Use an existing secret instead of creating one from values
-# When set, all keys from this secret are injected as environment variables via envFrom
-# Required keys:
-#   - postgres-password: PostgreSQL password (when postgresql.enabled=false)
-# Optional keys (any key becomes an env var):
-#   - HINDSIGHT_API_LLM_API_KEY: API key for LLM provider
-#   - Any other env vars you want to inject
-# existingSecret: "my-hindsight-secret"
-
 # Global settings
 replicaCount: 1

@@ -67,63 +58,6 @@ api:
    # HINDSIGHT_API_LLM_API_KEY: "your-api-key"
    # HINDSIGHT_API_LLM_BASE_URL: "https://api.groq.com/openai/v1"

-# Worker settings (distributed task processing)
-# When enabled, dedicated worker pods process tasks and the API's internal worker is disabled
-worker:
-  enabled: false
-  replicaCount: 2
-  image:
-    repository: ghcr.io/vectorize-io/hindsight-api
-    pullPolicy: IfNotPresent
-    # tag defaults to .Values.version if not specified
-
-  service:
-    # Service for metrics scraping (headless for StatefulSet)
-    port: 8889
-    targetPort: 8889
-
-  # Resource limits and requests
-  resources:
-    limits:
-      cpu: 2000m
-      memory: 4Gi
-    requests:
-      cpu: 500m
-      memory: 1Gi
-
-  # Liveness and readiness probes
-  livenessProbe:
-    httpGet:
-      path: /health
-      port: 8889
-    initialDelaySeconds: 30
-    periodSeconds: 10
-    timeoutSeconds: 5
-    failureThreshold: 3
-
-  readinessProbe:
-    httpGet:
-      path: /health
-      port: 8889
-    initialDelaySeconds: 10
-    periodSeconds: 5
-    timeoutSeconds: 3
-    failureThreshold: 3
-
-  # Worker-specific environment variables
-  env:
-    # Poll interval in milliseconds (how often to check for new tasks)
-    HINDSIGHT_API_WORKER_POLL_INTERVAL_MS: "500"
-    # Number of tasks to claim per poll cycle
-    HINDSIGHT_API_WORKER_BATCH_SIZE: "10"
-    # Max retries before marking a task as failed
-    HINDSIGHT_API_WORKER_MAX_RETRIES: "3"
-    # HTTP port for metrics/health (matches service.targetPort)
-    HINDSIGHT_API_WORKER_HTTP_PORT: "8889"
-
-  # Secret environment variables (inherited from api.secrets if not specified)
-  secrets: {}
-
 # Image settings for control plane
 controlPlane:
  enabled: true
@@ -80,7 +80,7 @@ Configure via environment variables:
 | Variable | Description | Default |
 |----------|-------------|---------|
 | `HINDSIGHT_API_DATABASE_URL` | PostgreSQL connection string | `pg0` (embedded) |
-| `HINDSIGHT_API_LLM_PROVIDER` | `openai`, `anthropic`, `gemini`, `groq`, `ollama`, `lmstudio` | `openai` |
+| `HINDSIGHT_API_LLM_PROVIDER` | `openai`, `groq`, `gemini`, `ollama` | `openai` |
 | `HINDSIGHT_API_LLM_API_KEY` | API key for LLM provider | - |
 | `HINDSIGHT_API_LLM_MODEL` | Model name | `gpt-4o-mini` |
 | `HINDSIGHT_API_HOST` | Server bind address | `0.0.0.0` |
@@ -46,4 +46,4 @@ __all__ = [
    "RemoteTEICrossEncoder",
    "LLMConfig",
 ]
-__version__ = "0.4.2"
+__version__ = "0.1.0"
@@ -1 +0,0 @@
-# Admin CLI for Hindsight
@@ -1,311 +0,0 @@
-"""
-Hindsight Admin CLI - backup and restore operations.
-"""
-
-import asyncio
-import io
-import json
-import logging
-import zipfile
-from datetime import datetime, timezone
-from pathlib import Path
-from typing import Any
-
-import asyncpg
-import typer
-
-from ..config import HindsightConfig
-from ..pg0 import parse_pg0_url, resolve_database_url
-
-
-def _fq_table(table: str, schema: str) -> str:
-    """Get fully-qualified table name with schema prefix."""
-    return f"{schema}.{table}"
-
-
-# Setup logging
-logging.basicConfig(
-    level=logging.INFO,
-    format="%(message)s",
-)
-logger = logging.getLogger(__name__)
-
-app = typer.Typer(name="hindsight-admin", help="Hindsight administrative commands")
-
-# Tables to backup/restore in dependency order
-# Import must happen in this order due to foreign key constraints
-BACKUP_TABLES = [
-    "banks",
-    "documents",
-    "entities",
-    "chunks",
-    "memory_units",
-    "unit_entities",
-    "entity_cooccurrences",
-    "memory_links",
-]
-
-MANIFEST_VERSION = "1"
-
-
-async def _backup(database_url: str, output_path: Path, schema: str = "public") -> dict[str, Any]:
-    """Backup all tables to a zip file using binary COPY protocol."""
-    conn = await asyncpg.connect(database_url)
-    try:
-        tables: dict[str, Any] = {}
-        manifest: dict[str, Any] = {
-            "version": MANIFEST_VERSION,
-            "created_at": datetime.now(timezone.utc).isoformat(),
-            "schema": schema,
-            "tables": tables,
-        }
-
-        # Use a transaction with REPEATABLE READ isolation to get a consistent
-        # snapshot across all tables. This prevents race conditions where
-        # entity_cooccurrences could reference entities created after the
-        # entities table was backed up.
-        async with conn.transaction(isolation="repeatable_read"):
-            with zipfile.ZipFile(output_path, "w", zipfile.ZIP_DEFLATED) as zf:
-                for i, table in enumerate(BACKUP_TABLES, 1):
-                    typer.echo(f"  [{i}/{len(BACKUP_TABLES)}] Backing up {table}...", nl=False)
-
-                    buffer = io.BytesIO()
-
-                    # Use binary COPY for exact type preservation
-                    # asyncpg requires schema_name as separate parameter
-                    await conn.copy_from_table(table, schema_name=schema, output=buffer, format="binary")
-
-                    data = buffer.getvalue()
-                    zf.writestr(f"{table}.bin", data)
-
-                    # Get row count for manifest
-                    qualified_table = _fq_table(table, schema)
-                    row_count = await conn.fetchval(f"SELECT COUNT(*) FROM {qualified_table}")
-                    tables[table] = {
-                        "rows": row_count,
-                        "size_bytes": len(data),
-                    }
-
-                    typer.echo(f" {row_count} rows")
-
-                zf.writestr("manifest.json", json.dumps(manifest, indent=2))
-
-        return manifest
-    finally:
-        await conn.close()
-
-
-async def _restore(database_url: str, input_path: Path, schema: str = "public") -> dict[str, Any]:
-    """Restore all tables from a zip file using binary COPY protocol."""
-    conn = await asyncpg.connect(database_url)
-    try:
-        with zipfile.ZipFile(input_path, "r") as zf:
-            # Read and validate manifest
-            manifest: dict[str, Any] = json.loads(zf.read("manifest.json"))
-            if manifest.get("version") != MANIFEST_VERSION:
-                raise ValueError(f"Unsupported backup version: {manifest.get('version')}")
-
-            # Use a transaction for atomic restore - either all tables are
-            # restored or none are, preventing partial/inconsistent state.
-            async with conn.transaction():
-                typer.echo("  Clearing existing data...")
-                # Truncate tables in reverse order (respects FK constraints)
-                for table in reversed(BACKUP_TABLES):
-                    qualified_table = _fq_table(table, schema)
-                    await conn.execute(f"TRUNCATE TABLE {qualified_table} CASCADE")
-
-                # Restore tables in forward order
-                for i, table in enumerate(BACKUP_TABLES, 1):
-                    filename = f"{table}.bin"
-                    if filename not in zf.namelist():
-                        typer.echo(f"  [{i}/{len(BACKUP_TABLES)}] {table}: skipped (not in backup)")
-                        continue
-
-                    expected_rows = manifest["tables"].get(table, {}).get("rows", "?")
-                    typer.echo(f"  [{i}/{len(BACKUP_TABLES)}] Restoring {table}... {expected_rows} rows")
-
-                    data = zf.read(filename)
-                    buffer = io.BytesIO(data)
-                    # asyncpg requires schema_name as separate parameter
-                    await conn.copy_to_table(table, schema_name=schema, source=buffer, format="binary")
-
-                # Refresh materialized view
-                typer.echo("  Refreshing materialized views...")
-                await conn.execute(f"REFRESH MATERIALIZED VIEW {_fq_table('memory_units_bm25', schema)}")
-
-        return manifest
-    finally:
-        await conn.close()
-
-
-async def _run_backup(db_url: str, output: Path, schema: str = "public") -> dict[str, Any]:
-    """Resolve database URL and run backup."""
-    is_pg0, instance_name, _ = parse_pg0_url(db_url)
-    if is_pg0:
-        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
-    resolved_url = await resolve_database_url(db_url)
-    return await _backup(resolved_url, output, schema)
-
-
-async def _run_restore(db_url: str, input_file: Path, schema: str = "public") -> dict[str, Any]:
-    """Resolve database URL and run restore."""
-    is_pg0, instance_name, _ = parse_pg0_url(db_url)
-    if is_pg0:
-        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
-    resolved_url = await resolve_database_url(db_url)
-    return await _restore(resolved_url, input_file, schema)
-
-
-@app.command()
-def backup(
-    output: Path = typer.Argument(..., help="Output file path (.zip)"),
-    schema: str = typer.Option("public", "--schema", "-s", help="Database schema to backup"),
-):
-    """Backup the Hindsight database to a zip file."""
-    config = HindsightConfig.from_env()
-
-    if not config.database_url:
-        typer.echo("Error: Database URL not configured.", err=True)
-        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
-        raise typer.Exit(1)
-
-    if output.suffix != ".zip":
-        output = output.with_suffix(".zip")
-
-    typer.echo(f"Backing up database (schema: {schema}) to {output}...")
-
-    manifest = asyncio.run(_run_backup(config.database_url, output, schema))
-
-    total_rows = sum(t["rows"] for t in manifest["tables"].values())
-    typer.echo(f"Backed up {total_rows} rows across {len(BACKUP_TABLES)} tables")
-    typer.echo(f"Backup saved to {output}")
-
-
-@app.command()
-def restore(
-    input_file: Path = typer.Argument(..., help="Input backup file (.zip)"),
-    schema: str = typer.Option("public", "--schema", "-s", help="Database schema to restore to"),
-    yes: bool = typer.Option(False, "--yes", "-y", help="Skip confirmation prompt"),
-):
-    """Restore the database from a backup file. WARNING: This deletes all existing data."""
-    config = HindsightConfig.from_env()
-
-    if not config.database_url:
-        typer.echo("Error: Database URL not configured.", err=True)
-        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
-        raise typer.Exit(1)
-
-    if not input_file.exists():
-        typer.echo(f"Error: File not found: {input_file}", err=True)
-        raise typer.Exit(1)
-
-    if not yes:
-        typer.confirm(
-            "This will DELETE all existing data and replace it with the backup. Continue?",
-            abort=True,
-        )
-
-    typer.echo(f"Restoring database (schema: {schema}) from {input_file}...")
-
-    manifest = asyncio.run(_run_restore(config.database_url, input_file, schema))
-
-    total_rows = sum(t["rows"] for t in manifest["tables"].values())
-    typer.echo(f"Restored {total_rows} rows across {len(BACKUP_TABLES)} tables")
-    typer.echo("Restore complete")
-
-
-async def _run_migration(db_url: str, schema: str = "public") -> None:
-    """Resolve database URL and run migrations."""
-    from ..migrations import run_migrations
-
-    is_pg0, instance_name, _ = parse_pg0_url(db_url)
-    if is_pg0:
-        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
-    resolved_url = await resolve_database_url(db_url)
-    run_migrations(resolved_url, schema=schema)
-
-
-@app.command(name="run-db-migration")
-def run_db_migration(
-    schema: str = typer.Option("public", "--schema", "-s", help="Database schema to run migrations on"),
-):
-    """Run database migrations to the latest version."""
-    config = HindsightConfig.from_env()
-
-    if not config.database_url:
-        typer.echo("Error: Database URL not configured.", err=True)
-        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
-        raise typer.Exit(1)
-
-    typer.echo(f"Running database migrations (schema: {schema})...")
-
-    asyncio.run(_run_migration(config.database_url, schema))
-
-    typer.echo("Database migrations completed successfully")
-
-
-async def _decommission_worker(db_url: str, worker_id: str, schema: str = "public") -> int:
-    """Release all tasks owned by a worker, setting them back to pending status."""
-    is_pg0, instance_name, _ = parse_pg0_url(db_url)
-    if is_pg0:
-        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
-    resolved_url = await resolve_database_url(db_url)
-
-    conn = await asyncpg.connect(resolved_url)
-    try:
-        table = _fq_table("async_operations", schema)
-        result = await conn.fetch(
-            f"""
-            UPDATE {table}
-            SET status = 'pending', worker_id = NULL, claimed_at = NULL, updated_at = now()
-            WHERE worker_id = $1 AND status = 'processing'
-            RETURNING operation_id
-            """,
-            worker_id,
-        )
-        return len(result)
-    finally:
-        await conn.close()
-
-
-@app.command(name="decommission-worker")
-def decommission_worker(
-    worker_id: str = typer.Argument(..., help="Worker ID to decommission"),
-    schema: str = typer.Option("public", "--schema", "-s", help="Database schema"),
-    yes: bool = typer.Option(False, "--yes", "-y", help="Skip confirmation prompt"),
-):
-    """Release all tasks owned by a worker (sets status back to pending).
-
-    Use this command when a worker has crashed or been removed without graceful shutdown.
-    All tasks that were being processed by the worker will be released back to the queue
-    so other workers can pick them up.
-    """
-    config = HindsightConfig.from_env()
-
-    if not config.database_url:
-        typer.echo("Error: Database URL not configured.", err=True)
-        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
-        raise typer.Exit(1)
-
-    if not yes:
-        typer.confirm(
-            f"This will release all tasks owned by worker '{worker_id}' back to pending. Continue?",
-            abort=True,
-        )
-
-    typer.echo(f"Decommissioning worker '{worker_id}' (schema: {schema})...")
-
-    count = asyncio.run(_decommission_worker(config.database_url, worker_id, schema))
-
-    if count > 0:
-        typer.echo(f"Released {count} task(s) from worker '{worker_id}'")
-    else:
-        typer.echo(f"No tasks found for worker '{worker_id}'")
-
-
-def main():
-    app()
-
-
-if __name__ == "__main__":
-    main()
@@ -1,44 +0,0 @@
-"""add_memory_links_from_type_weight_index
-
-Revision ID: f1a2b3c4d5e6
-Revises: e0a1b2c3d4e5
-Create Date: 2025-01-12
-
-Add composite index on memory_links (from_unit_id, link_type, weight DESC)
-to optimize MPFP graph traversal queries that need top-k edges per type.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "f1a2b3c4d5e6"
-down_revision: str | Sequence[str] | None = "e0a1b2c3d4e5"
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-
-def _get_schema_prefix() -> str:
-    """Get schema prefix for table names (e.g., 'tenant_x.' or '' for public)."""
-    schema = context.config.get_main_option("target_schema")
-    return f'"{schema}".' if schema else ""
-
-
-def upgrade() -> None:
-    """Add composite index for efficient MPFP edge loading."""
-    schema = _get_schema_prefix()
-    # Create composite index for efficient top-k per (from_node, link_type) queries
-    # This enables LATERAL joins to use index-only scans with early termination
-    # Note: Not using CONCURRENTLY here as it requires running outside a transaction
-    # For production with large tables, consider running this manually with CONCURRENTLY
-    op.execute(
-        f"CREATE INDEX IF NOT EXISTS idx_memory_links_from_type_weight "
-        f"ON {schema}memory_links(from_unit_id, link_type, weight DESC)"
-    )
-
-
-def downgrade() -> None:
-    """Remove the composite index."""
-    schema = _get_schema_prefix()
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_links_from_type_weight")
@@ -1,48 +0,0 @@
-"""add_tags_column
-
-Revision ID: g2a3b4c5d6e7
-Revises: f1a2b3c4d5e6
-Create Date: 2025-01-13
-
-Add tags column to memory_units and documents tables for visibility scoping.
-Tags enable filtering memories by scope (e.g., user IDs, session IDs) during recall/reflect.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "g2a3b4c5d6e7"
-down_revision: str | Sequence[str] | None = "f1a2b3c4d5e6"
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-
-def _get_schema_prefix() -> str:
-    """Get schema prefix for table names (e.g., 'tenant_x.' or '' for public)."""
-    schema = context.config.get_main_option("target_schema")
-    return f'"{schema}".' if schema else ""
-
-
-def upgrade() -> None:
-    """Add tags column to memory_units and documents tables."""
-    schema = _get_schema_prefix()
-
-    # Add tags column to memory_units table
-    op.execute(f"ALTER TABLE {schema}memory_units ADD COLUMN IF NOT EXISTS tags VARCHAR[] NOT NULL DEFAULT '{{}}'")
-
-    # Create GIN index for efficient array containment queries (tags && ARRAY['x'])
-    op.execute(f"CREATE INDEX IF NOT EXISTS idx_memory_units_tags ON {schema}memory_units USING GIN (tags)")
-
-    # Add tags column to documents table for document-level tags
-    op.execute(f"ALTER TABLE {schema}documents ADD COLUMN IF NOT EXISTS tags VARCHAR[] NOT NULL DEFAULT '{{}}'")
-
-
-def downgrade() -> None:
-    """Remove tags columns and index."""
-    schema = _get_schema_prefix()
-
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_tags")
-    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS tags")
-    op.execute(f"ALTER TABLE {schema}documents DROP COLUMN IF EXISTS tags")
@@ -1,112 +0,0 @@
-"""mental_models_v4
-
-Revision ID: h3c4d5e6f7g8
-Revises: g2a3b4c5d6e7
-Create Date: 2026-01-08 00:00:00.000000
-
-This migration implements the v4 mental models system:
-1. Deletes existing observation memory_units (observations now in mental models)
-2. Adds mission column to banks (replacing background)
-3. Creates mental_models table with final schema
-
-Mental models can reference entities when an entity is "promoted" to a mental model.
-Summary content is stored as JSONB observations with per-observation fact attribution.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "h3c4d5e6f7g8"
-down_revision: str | Sequence[str] | None = "g2a3b4c5d6e7"
-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:
-    """Apply mental models v4 changes."""
-    schema = _get_schema_prefix()
-
-    # Step 1: Delete observation memory_units (cascades to unit_entities links)
-    # Observations are now handled through mental models, not memory_units
-    op.execute(f"DELETE FROM {schema}memory_units WHERE fact_type = 'observation'")
-
-    # Step 2: Drop observation-specific index (if it exists)
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_observation_date")
-
-    # Step 3: Add mission column to banks (replacing background)
-    op.execute(f"ALTER TABLE {schema}banks ADD COLUMN IF NOT EXISTS mission TEXT")
-
-    # Migrate: copy background to mission if background column exists
-    # Use DO block to check column existence first (idempotent for re-runs)
-    schema_name = context.config.get_main_option("target_schema") or "public"
-    op.execute(f"""
-        DO $$
-        BEGIN
-            IF EXISTS (
-                SELECT 1 FROM information_schema.columns
-                WHERE table_schema = '{schema_name}' AND table_name = 'banks' AND column_name = 'background'
-            ) THEN
-                UPDATE {schema}banks
-                SET mission = background
-                WHERE mission IS NULL;
-            END IF;
-        END $$;
-    """)
-
-    # Remove background column (replaced by mission)
-    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS background")
-
-    # Step 4: Create mental_models table with final v4 schema (if not exists)
-    op.execute(f"""
-        CREATE TABLE IF NOT EXISTS {schema}mental_models (
-            id VARCHAR(64) NOT NULL,
-            bank_id VARCHAR(64) NOT NULL,
-            subtype VARCHAR(32) NOT NULL,
-            name VARCHAR(256) NOT NULL,
-            description TEXT NOT NULL,
-            entity_id UUID,
-            observations JSONB DEFAULT '{{"observations": []}}'::jsonb,
-            links VARCHAR[],
-            tags VARCHAR[] DEFAULT '{{}}',
-            last_updated TIMESTAMP WITH TIME ZONE,
-            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
-            PRIMARY KEY (id, bank_id),
-            FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE,
-            FOREIGN KEY (entity_id) REFERENCES {schema}entities(id) ON DELETE SET NULL,
-            CONSTRAINT ck_mental_models_subtype CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned'))
-        )
-    """)
-
-    # Step 5: Create indexes for efficient queries (if not exist)
-    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_bank_id ON {schema}mental_models(bank_id)")
-    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_subtype ON {schema}mental_models(bank_id, subtype)")
-    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_entity_id ON {schema}mental_models(entity_id)")
-    # GIN index for efficient tags array filtering
-    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_tags ON {schema}mental_models USING GIN(tags)")
-
-
-def downgrade() -> None:
-    """Revert mental models v4 changes."""
-    schema = _get_schema_prefix()
-
-    # Drop mental_models table (cascades to indexes)
-    op.execute(f"DROP TABLE IF EXISTS {schema}mental_models CASCADE")
-
-    # Add back background column to banks
-    op.execute(f"ALTER TABLE {schema}banks ADD COLUMN IF NOT EXISTS background TEXT")
-
-    # Migrate mission back to background
-    op.execute(f"UPDATE {schema}banks SET background = mission WHERE background IS NULL")
-
-    # Remove mission column
-    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS mission")
-
-    # Note: Cannot restore deleted observations - they are lost on downgrade
@@ -1,41 +0,0 @@
-"""delete_opinions
-
-Revision ID: i4d5e6f7g8h9
-Revises: h3c4d5e6f7g8
-Create Date: 2026-01-15 00:00:00.000000
-
-This migration removes opinion facts from memory_units.
-Opinions are no longer a separate fact type - they are now represented
-through mental model observations with confidence scores.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "i4d5e6f7g8h9"
-down_revision: str | Sequence[str] | None = "h3c4d5e6f7g8"
-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:
-    """Delete opinion memory_units."""
-    schema = _get_schema_prefix()
-
-    # Delete opinion memory_units (cascades to unit_entities links)
-    # Opinions are now handled through mental model observations
-    op.execute(f"DELETE FROM {schema}memory_units WHERE fact_type = 'opinion'")
-
-
-def downgrade() -> None:
-    """Cannot restore deleted opinions."""
-    # Note: Cannot restore deleted opinions - they are lost on downgrade
-    pass
@@ -1,95 +0,0 @@
-"""mental_model_versions
-
-Revision ID: j5e6f7g8h9i0
-Revises: i4d5e6f7g8h9
-Create Date: 2026-01-16 00:00:00.000000
-
-This migration adds versioning support for mental models:
-1. Creates mental_model_versions table to store observation snapshots
-2. Adds version column to mental_models for tracking current version
-
-This enables changelog/diff functionality for mental model observations.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "j5e6f7g8h9i0"
-down_revision: str | Sequence[str] | None = "i4d5e6f7g8h9"
-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 mental_model_versions table and add version tracking."""
-    schema = _get_schema_prefix()
-
-    # Create mental_model_versions table for storing observation snapshots
-    op.execute(f"""
-        CREATE TABLE {schema}mental_model_versions (
-            id SERIAL PRIMARY KEY,
-            mental_model_id VARCHAR(64) NOT NULL,
-            bank_id VARCHAR(64) NOT NULL,
-            version INT NOT NULL,
-            observations JSONB NOT NULL DEFAULT '{{"observations": []}}'::jsonb,
-            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
-            FOREIGN KEY (mental_model_id, bank_id)
-                REFERENCES {schema}mental_models(id, bank_id) ON DELETE CASCADE,
-            UNIQUE (mental_model_id, bank_id, version)
-        )
-    """)
-
-    # Index for efficient version queries (get latest, list versions)
-    op.execute(f"""
-        CREATE INDEX idx_mental_model_versions_lookup
-        ON {schema}mental_model_versions(mental_model_id, bank_id, version DESC)
-    """)
-
-    # Add version column to mental_models to track current version
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        ADD COLUMN IF NOT EXISTS version INT NOT NULL DEFAULT 0
-    """)
-
-    # Migrate existing mental models: create version 1 for any that have observations
-    op.execute(f"""
-        INSERT INTO {schema}mental_model_versions (mental_model_id, bank_id, version, observations, created_at)
-        SELECT id, bank_id, 1, observations, COALESCE(last_updated, created_at)
-        FROM {schema}mental_models
-        WHERE observations IS NOT NULL
-          AND observations != '{{"observations": []}}'::jsonb
-          AND (observations->'observations') IS NOT NULL
-          AND jsonb_array_length(observations->'observations') > 0
-    """)
-
-    # Update version to 1 for migrated mental models
-    op.execute(f"""
-        UPDATE {schema}mental_models
-        SET version = 1
-        WHERE observations IS NOT NULL
-          AND observations != '{{"observations": []}}'::jsonb
-          AND (observations->'observations') IS NOT NULL
-          AND jsonb_array_length(observations->'observations') > 0
-    """)
-
-
-def downgrade() -> None:
-    """Remove mental_model_versions table and version column."""
-    schema = _get_schema_prefix()
-
-    # Drop index
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_mental_model_versions_lookup")
-
-    # Drop versions table
-    op.execute(f"DROP TABLE IF EXISTS {schema}mental_model_versions")
-
-    # Remove version column from mental_models
-    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS version")
@@ -1,58 +0,0 @@
-"""add_directive_subtype
-
-Revision ID: k6f7g8h9i0j1
-Revises: j5e6f7g8h9i0
-Create Date: 2026-01-16 00:00:00.000000
-
-This migration adds 'directive' to the mental_models subtype constraint.
-Directives are hard rules with user-provided observations that the reflect agent must follow.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "k6f7g8h9i0j1"
-down_revision: str | Sequence[str] | None = "j5e6f7g8h9i0"
-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 'directive' to mental_models subtype constraint."""
-    schema = _get_schema_prefix()
-
-    # Drop existing constraint
-    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
-
-    # Create new constraint with 'directive' added
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        ADD CONSTRAINT ck_mental_models_subtype
-        CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned', 'directive'))
-    """)
-
-
-def downgrade() -> None:
-    """Remove 'directive' from mental_models subtype constraint."""
-    schema = _get_schema_prefix()
-
-    # First delete any directives (cannot downgrade if they exist)
-    op.execute(f"DELETE FROM {schema}mental_models WHERE subtype = 'directive'")
-
-    # Drop constraint with directive
-    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
-
-    # Recreate original constraint without directive
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        ADD CONSTRAINT ck_mental_models_subtype
-        CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned'))
-    """)
@@ -1,109 +0,0 @@
-"""add_worker_columns
-
-Revision ID: l7g8h9i0j1k2
-Revises: k6f7g8h9i0j1
-Create Date: 2026-01-19 00:00:00.000000
-
-This migration adds columns to async_operations for distributed worker support:
- worker_id: ID of the worker that claimed the task
- claimed_at: When the task was claimed
- retry_count: Number of retry attempts
- task_payload: The serialized task dictionary
-"""
-
-from collections.abc import Sequence
-
-import sqlalchemy as sa
-from alembic import context, op
-from sqlalchemy.dialects import postgresql
-
-# revision identifiers, used by Alembic.
-revision: str = "l7g8h9i0j1k2"
-down_revision: str | Sequence[str] | None = "k6f7g8h9i0j1"
-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 worker columns to async_operations."""
-    schema = _get_schema_prefix()
-
-    # Add worker_id column (ID of worker that claimed the task)
-    op.add_column(
-        "async_operations",
-        sa.Column("worker_id", sa.Text(), nullable=True),
-        schema=context.config.get_main_option("target_schema") or None,
-    )
-
-    # Add claimed_at column (when task was claimed by worker)
-    op.add_column(
-        "async_operations",
-        sa.Column("claimed_at", postgresql.TIMESTAMP(timezone=True), nullable=True),
-        schema=context.config.get_main_option("target_schema") or None,
-    )
-
-    # Add retry_count column (number of retry attempts)
-    op.add_column(
-        "async_operations",
-        sa.Column("retry_count", sa.Integer(), server_default="0", nullable=False),
-        schema=context.config.get_main_option("target_schema") or None,
-    )
-
-    # Add task_payload column (serialized task dictionary)
-    op.add_column(
-        "async_operations",
-        sa.Column(
-            "task_payload",
-            postgresql.JSONB(astext_type=sa.Text()),
-            nullable=True,
-        ),
-        schema=context.config.get_main_option("target_schema") or None,
-    )
-
-    # Add index for efficient worker polling (pending tasks ordered by creation time)
-    op.execute(
-        f"CREATE INDEX idx_async_operations_pending_claim ON {schema}async_operations (status, created_at) "
-        f"WHERE status = 'pending' AND task_payload IS NOT NULL"
-    )
-
-    # Add index for finding tasks by worker_id (for decommissioning)
-    op.execute(
-        f"CREATE INDEX idx_async_operations_worker_id ON {schema}async_operations (worker_id) WHERE worker_id IS NOT NULL"
-    )
-
-
-def downgrade() -> None:
-    """Remove worker columns from async_operations."""
-    schema = _get_schema_prefix()
-
-    # Drop indexes
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_pending_claim")
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_worker_id")
-
-    # Drop columns
-    op.drop_column(
-        "async_operations",
-        "task_payload",
-        schema=context.config.get_main_option("target_schema") or None,
-    )
-    op.drop_column(
-        "async_operations",
-        "retry_count",
-        schema=context.config.get_main_option("target_schema") or None,
-    )
-    op.drop_column(
-        "async_operations",
-        "claimed_at",
-        schema=context.config.get_main_option("target_schema") or None,
-    )
-    op.drop_column(
-        "async_operations",
-        "worker_id",
-        schema=context.config.get_main_option("target_schema") or None,
-    )
@@ -1,41 +0,0 @@
-"""mental_model_id_to_text
-
-Revision ID: m8h9i0j1k2l3
-Revises: l7g8h9i0j1k2
-Create Date: 2026-01-19 00:00:00.000000
-
-This migration changes the mental_models.id column from VARCHAR(64) to TEXT
-to support longer model IDs (e.g., entity names that exceed 64 characters).
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "m8h9i0j1k2l3"
-down_revision: str | Sequence[str] | None = "l7g8h9i0j1k2"
-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:
-    """Change mental_models.id from VARCHAR(64) to TEXT."""
-    schema = _get_schema_prefix()
-
-    # Alter the id column type from VARCHAR(64) to TEXT
-    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE TEXT")
-
-
-def downgrade() -> None:
-    """Revert mental_models.id from TEXT to VARCHAR(64)."""
-    schema = _get_schema_prefix()
-
-    # Note: This may fail if any id values exceed 64 characters
-    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE VARCHAR(64)")
@@ -1,134 +0,0 @@
-"""learnings_and_pinned_reflections
-
-Revision ID: n9i0j1k2l3m4
-Revises: m8h9i0j1k2l3
-Create Date: 2026-01-21 00:00:00.000000
-
-This migration:
-1. Creates the 'learnings' table for automatic bottom-up consolidation
-2. Creates the 'pinned_reflections' table for user-curated living documents
-3. Adds consolidation tracking columns to the 'banks' table
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "n9i0j1k2l3m4"
-down_revision: str | Sequence[str] | None = "m8h9i0j1k2l3"
-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 learnings and pinned_reflections tables."""
-    schema = _get_schema_prefix()
-
-    # 1. Create learnings table
-    op.execute(f"""
-        CREATE TABLE {schema}learnings (
-            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            bank_id VARCHAR(64) NOT NULL,
-            text TEXT NOT NULL,
-            proof_count INT NOT NULL DEFAULT 1,
-            history JSONB DEFAULT '[]'::jsonb,
-            mission_context VARCHAR(64),
-            pre_mission_change BOOLEAN DEFAULT FALSE,
-            embedding vector(384),
-            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
-            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
-            updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
-        )
-    """)
-
-    # Add foreign key constraint
-    op.execute(f"""
-        ALTER TABLE {schema}learnings
-        ADD CONSTRAINT fk_learnings_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-
-    # Indexes for learnings
-    op.execute(f"CREATE INDEX idx_learnings_bank_id ON {schema}learnings(bank_id)")
-    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
-    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"""
-        CREATE TABLE {schema}pinned_reflections (
-            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            bank_id VARCHAR(64) NOT NULL,
-            name VARCHAR(256) NOT NULL,
-            source_query TEXT NOT NULL,
-            content TEXT NOT NULL,
-            embedding vector(384),
-            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
-            last_refreshed_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
-            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
-        )
-    """)
-
-    # Add foreign key constraint
-    op.execute(f"""
-        ALTER TABLE {schema}pinned_reflections
-        ADD CONSTRAINT fk_pinned_reflections_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-
-    # Indexes for pinned_reflections
-    op.execute(f"CREATE INDEX idx_pinned_reflections_bank_id ON {schema}pinned_reflections(bank_id)")
-    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
-    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"""
-        ALTER TABLE {schema}banks
-        ADD COLUMN IF NOT EXISTS last_consolidated_at TIMESTAMP WITH TIME ZONE
-    """)
-    op.execute(f"""
-        ALTER TABLE {schema}banks
-        ADD COLUMN IF NOT EXISTS mission_changed_at TIMESTAMP WITH TIME ZONE
-    """)
-
-
-def downgrade() -> None:
-    """Drop learnings and pinned_reflections tables."""
-    schema = _get_schema_prefix()
-
-    # Drop tables
-    op.execute(f"DROP TABLE IF EXISTS {schema}learnings CASCADE")
-    op.execute(f"DROP TABLE IF EXISTS {schema}pinned_reflections CASCADE")
-
-    # Remove columns from banks
-    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS last_consolidated_at")
-    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS mission_changed_at")
@@ -1,113 +0,0 @@
-"""migrate_mental_models_data
-
-Revision ID: o0j1k2l3m4n5
-Revises: n9i0j1k2l3m4
-Create Date: 2026-01-21 00:00:00.000000
-
-This migration:
-1. Migrates existing 'pinned' mental models to the new 'pinned_reflections' table
-2. Migrates existing 'learned' mental models to the new 'learnings' table
-3. Deletes non-directive mental models (structural, emergent, pinned, learned)
-4. Drops the mental_model_versions table (no longer used)
-5. Adds a CHECK constraint that only 'directive' subtype is allowed
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "o0j1k2l3m4n5"
-down_revision: str | Sequence[str] | None = "n9i0j1k2l3m4"
-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:
-    """Migrate data and clean up old mental models."""
-    schema = _get_schema_prefix()
-
-    # 1. Migrate 'pinned' mental models to pinned_reflections
-    # For pinned models, the first observation's content becomes the pinned reflection content
-    op.execute(f"""
-        INSERT INTO {schema}pinned_reflections (bank_id, name, source_query, content, tags, created_at)
-        SELECT
-            bank_id,
-            name,
-            description AS source_query,
-            COALESCE(
-                observations->'observations'->0->>'content',
-                description,
-                ''
-            ) AS content,
-            tags,
-            created_at
-        FROM {schema}mental_models
-        WHERE subtype = 'pinned'
-        ON CONFLICT DO NOTHING
-    """)
-
-    # 2. Migrate 'learned' mental models to learnings
-    # Each observation in a learned model becomes a separate learning
-    op.execute(f"""
-        INSERT INTO {schema}learnings (bank_id, text, proof_count, tags, created_at)
-        SELECT
-            mm.bank_id,
-            obs->>'content' AS text,
-            GREATEST(1, COALESCE(jsonb_array_length(obs->'evidence'), 1)) AS proof_count,
-            mm.tags,
-            mm.created_at
-        FROM {schema}mental_models mm,
-        LATERAL jsonb_array_elements(mm.observations->'observations') AS obs
-        WHERE mm.subtype = 'learned'
-          AND obs->>'content' IS NOT NULL
-          AND obs->>'content' != ''
-        ON CONFLICT DO NOTHING
-    """)
-
-    # 3. Delete all non-directive mental models (they've been migrated or are obsolete)
-    op.execute(f"""
-        DELETE FROM {schema}mental_models
-        WHERE subtype != 'directive'
-    """)
-
-    # 4. Drop the mental_model_versions table (no longer used)
-    op.execute(f"DROP TABLE IF EXISTS {schema}mental_model_versions CASCADE")
-
-    # 5. Drop old constraints and add new one that only allows 'directive'
-    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        ADD CONSTRAINT ck_mental_models_subtype CHECK (subtype = 'directive')
-    """)
-
-
-def downgrade() -> None:
-    """Reverse the migration (data migration is one-way, so this just removes constraints)."""
-    schema = _get_schema_prefix()
-
-    # Remove the directive-only constraint
-    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
-
-    # Re-create mental_model_versions table
-    op.execute(f"""
-        CREATE TABLE IF NOT EXISTS {schema}mental_model_versions (
-            id SERIAL PRIMARY KEY,
-            bank_id VARCHAR(64) NOT NULL,
-            model_id VARCHAR(128) NOT NULL,
-            version INT NOT NULL,
-            observations JSONB NOT NULL,
-            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
-        )
-    """)
-    op.execute(
-        f"CREATE INDEX IF NOT EXISTS idx_mm_versions_lookup ON {schema}mental_model_versions(bank_id, model_id, version DESC)"
-    )
-
-    # Note: Data migration cannot be reversed - pinned_reflections and learnings data remains
@@ -1,194 +0,0 @@
-"""new_knowledge_architecture
-
-Revision ID: p1k2l3m4n5o6
-Revises: o0j1k2l3m4n5
-Create Date: 2026-01-21 00:00:00.000000
-
-This migration implements the new knowledge architecture:
-1. Drops the 'learnings' table (mental models are now in memory_units)
-2. Renames 'pinned_reflections' to 'reflections'
-3. Drops the 'mental_models' table completely
-4. Creates 'directives' table for hard rules
-5. Adds mental model support columns to 'memory_units' (proof_count, source_memory_ids, history)
-
-The new architecture:
- Directives: Hard rules in their own table
- Mental Models: Stored in memory_units with fact_type='mental_model'
- Reflections: User-curated documents (renamed from pinned_reflections)
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "p1k2l3m4n5o6"
-down_revision: str | Sequence[str] | None = "o0j1k2l3m4n5"
-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:
-    """Implement new knowledge architecture."""
-    schema = _get_schema_prefix()
-
-    # 1. Drop the learnings table (mental models will be in memory_units)
-    op.execute(f"DROP TABLE IF EXISTS {schema}learnings CASCADE")
-
-    # 2. Rename pinned_reflections to reflections
-    op.execute(f"ALTER TABLE IF EXISTS {schema}pinned_reflections RENAME TO reflections")
-
-    # Rename indexes for reflections
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_bank_id RENAME TO idx_reflections_bank_id")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_embedding RENAME TO idx_reflections_embedding")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_tags RENAME TO idx_reflections_tags")
-    op.execute(
-        f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_text_search RENAME TO idx_reflections_text_search"
-    )
-
-    # Rename foreign key constraint
-    op.execute(f"""
-        ALTER TABLE {schema}reflections
-        DROP CONSTRAINT IF EXISTS fk_pinned_reflections_bank_id
-    """)
-    op.execute(f"""
-        ALTER TABLE {schema}reflections
-        ADD CONSTRAINT fk_reflections_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-
-    # 3. Drop the mental_models table completely
-    op.execute(f"DROP TABLE IF EXISTS {schema}mental_models CASCADE")
-
-    # 4. Create directives table
-    op.execute(f"""
-        CREATE TABLE {schema}directives (
-            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            bank_id VARCHAR(64) NOT NULL,
-            name VARCHAR(256) NOT NULL,
-            content TEXT NOT NULL,
-            priority INT NOT NULL DEFAULT 0,
-            is_active BOOLEAN NOT NULL DEFAULT TRUE,
-            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
-            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
-            updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
-        )
-    """)
-
-    # Add foreign key and indexes for directives
-    op.execute(f"""
-        ALTER TABLE {schema}directives
-        ADD CONSTRAINT fk_directives_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-    op.execute(f"CREATE INDEX idx_directives_bank_id ON {schema}directives(bank_id)")
-    op.execute(f"CREATE INDEX idx_directives_bank_active ON {schema}directives(bank_id, is_active)")
-    op.execute(f"CREATE INDEX idx_directives_tags ON {schema}directives USING GIN(tags)")
-
-    # 5. Add mental model support columns to memory_units
-    # proof_count: Number of memories that support this mental model
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD COLUMN IF NOT EXISTS proof_count INT DEFAULT 1
-    """)
-
-    # source_memory_ids: Array of memory IDs that consolidated into this mental model
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD COLUMN IF NOT EXISTS source_memory_ids UUID[] DEFAULT ARRAY[]::UUID[]
-    """)
-
-    # history: JSONB array tracking changes to mental models
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD COLUMN IF NOT EXISTS history JSONB DEFAULT '[]'::jsonb
-    """)
-
-    # Add index for finding mental models
-    op.execute(f"""
-        CREATE INDEX IF NOT EXISTS idx_memory_units_mental_models
-        ON {schema}memory_units(bank_id, fact_type)
-        WHERE fact_type = 'mental_model'
-    """)
-
-    # 6. Update fact_type check constraint to include 'mental_model'
-    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD CONSTRAINT memory_units_fact_type_check
-        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation', 'mental_model'))
-    """)
-
-
-def downgrade() -> None:
-    """Reverse the migration."""
-    schema = _get_schema_prefix()
-
-    # Restore original fact_type check constraint (without 'mental_model')
-    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD CONSTRAINT memory_units_fact_type_check
-        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation'))
-    """)
-
-    # Drop mental model columns from memory_units
-    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS proof_count")
-    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS source_memory_ids")
-    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS history")
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_mental_models")
-
-    # Drop directives table
-    op.execute(f"DROP TABLE IF EXISTS {schema}directives CASCADE")
-
-    # Rename reflections back to pinned_reflections
-    op.execute(f"ALTER TABLE IF EXISTS {schema}reflections RENAME TO pinned_reflections")
-
-    # Restore indexes
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_bank_id RENAME TO idx_pinned_reflections_bank_id")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_embedding RENAME TO idx_pinned_reflections_embedding")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_tags RENAME TO idx_pinned_reflections_tags")
-    op.execute(
-        f"ALTER INDEX IF EXISTS {schema}idx_reflections_text_search RENAME TO idx_pinned_reflections_text_search"
-    )
-
-    # Restore foreign key
-    op.execute(f"""
-        ALTER TABLE {schema}pinned_reflections
-        DROP CONSTRAINT IF EXISTS fk_reflections_bank_id
-    """)
-    op.execute(f"""
-        ALTER TABLE {schema}pinned_reflections
-        ADD CONSTRAINT fk_pinned_reflections_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-
-    # Re-create learnings table
-    op.execute(f"""
-        CREATE TABLE IF NOT EXISTS {schema}learnings (
-            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            bank_id VARCHAR(64) NOT NULL,
-            text TEXT NOT NULL,
-            proof_count INT NOT NULL DEFAULT 1,
-            history JSONB DEFAULT '[]'::jsonb,
-            mission_context VARCHAR(64),
-            pre_mission_change BOOLEAN DEFAULT FALSE,
-            embedding vector(384),
-            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
-            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
-            updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
-        )
-    """)
-    op.execute(f"""
-        ALTER TABLE {schema}learnings
-        ADD CONSTRAINT fk_learnings_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-
-    # Note: mental_models table recreation is complex and would need separate handling
@@ -1,50 +0,0 @@
-"""fix_mental_model_fact_type
-
-Revision ID: q2l3m4n5o6p7
-Revises: p1k2l3m4n5o6
-Create Date: 2026-01-21 13:30:00.000000
-
-Fix the fact_type check constraint to include 'mental_model'.
-This is a fix for p1k2l3m4n5o6 which should have included this change.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-# revision identifiers, used by Alembic.
-revision: str = "q2l3m4n5o6p7"
-down_revision: str | Sequence[str] | None = "p1k2l3m4n5o6"
-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 'mental_model' to the fact_type check constraint."""
-    schema = _get_schema_prefix()
-
-    # Drop the old constraint and add the new one with mental_model included
-    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD CONSTRAINT memory_units_fact_type_check
-        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation', 'mental_model'))
-    """)
-
-
-def downgrade() -> None:
-    """Remove 'mental_model' from the fact_type check constraint."""
-    schema = _get_schema_prefix()
-
-    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD CONSTRAINT memory_units_fact_type_check
-        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation'))
-    """)
@@ -1,47 +0,0 @@
-"""Add reflect_response JSONB column to reflections
-
-Revision ID: r3m4n5o6p7q8
-Revises: q2l3m4n5o6p7
-Create Date: 2026-01-21
-
-This migration adds a reflect_response JSONB column to store the full
-reflect API response payload, including based_on facts and trace data.
-
-Note: Table was renamed from pinned_reflections to reflections in p1k2l3m4n5o6.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "r3m4n5o6p7q8"
-down_revision: str | Sequence[str] | None = "q2l3m4n5o6p7"
-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 reflect_response JSONB column to reflections."""
-    schema = _get_schema_prefix()
-
-    # Add reflect_response column to store the full reflect API response
-    op.execute(f"""
-        ALTER TABLE {schema}reflections
-        ADD COLUMN IF NOT EXISTS reflect_response JSONB
-    """)
-
-
-def downgrade() -> None:
-    """Remove reflect_response column from reflections."""
-    schema = _get_schema_prefix()
-
-    op.execute(f"""
-        ALTER TABLE {schema}reflections
-        DROP COLUMN IF EXISTS reflect_response
-    """)
@@ -1,53 +0,0 @@
-"""Add consolidated_at column to memory_units for incremental consolidation tracking.
-
-This allows consolidation to track progress at the memory level rather than
-using a bank-level watermark. If consolidation crashes, already-processed
-memories won't be reprocessed.
-
-Revision ID: s4n5o6p7q8r9
-Revises: r3m4n5o6p7q8
-Create Date: 2025-01-22
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "s4n5o6p7q8r9"
-down_revision: str | Sequence[str] | None = "r3m4n5o6p7q8"
-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()
-
-    # Add consolidated_at column to memory_units
-    op.execute(
-        f"""
-        ALTER TABLE {schema}memory_units
-        ADD COLUMN IF NOT EXISTS consolidated_at TIMESTAMPTZ DEFAULT NULL
-        """
-    )
-
-    # Create index for efficient querying of unconsolidated memories
-    op.execute(
-        f"""
-        CREATE INDEX IF NOT EXISTS idx_memory_units_unconsolidated
-        ON {schema}memory_units (bank_id, created_at)
-        WHERE consolidated_at IS NULL AND fact_type IN ('experience', 'world')
-        """
-    )
-
-
-def downgrade() -> None:
-    schema = _get_schema_prefix()
-
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_unconsolidated")
-    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS consolidated_at")
@@ -1,134 +0,0 @@
-"""Rename mental_model fact_type to observation and reflections table to mental_models
-
-Revision ID: t5o6p7q8r9s0
-Revises: s4n5o6p7q8r9
-Create Date: 2026-01-26
-
-This migration implements the terminology rename:
-1. mental_model (fact_type in memory_units) -> observation
-2. reflections table -> mental_models table
-
-The new terminology:
- Observations: Consolidated knowledge synthesized from facts (was mental_model)
- Mental Models: Stored reflect responses (was reflections)
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "t5o6p7q8r9s0"
-down_revision: str | Sequence[str] | None = "s4n5o6p7q8r9"
-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:
-    """Rename mental_model -> observation and reflections -> mental_models."""
-    schema = _get_schema_prefix()
-
-    # 1. Update fact_type values: mental_model -> observation
-    op.execute(f"""
-        UPDATE {schema}memory_units
-        SET fact_type = 'observation'
-        WHERE fact_type = 'mental_model'
-    """)
-
-    # 2. Update the CHECK constraint - remove mental_model, keep observation
-    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD CONSTRAINT memory_units_fact_type_check
-        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation'))
-    """)
-
-    # 3. Rename the index for observations (was for mental_models)
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_mental_models")
-    op.execute(f"""
-        CREATE INDEX IF NOT EXISTS idx_memory_units_observations
-        ON {schema}memory_units(bank_id, fact_type)
-        WHERE fact_type = 'observation'
-    """)
-
-    # 4. Update the unconsolidated index to not filter by fact_type since observations
-    # are now the consolidated type
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_unconsolidated")
-    op.execute(f"""
-        CREATE INDEX IF NOT EXISTS idx_memory_units_unconsolidated
-        ON {schema}memory_units (bank_id, created_at)
-        WHERE consolidated_at IS NULL AND fact_type IN ('experience', 'world')
-    """)
-
-    # 5. Rename reflections table to mental_models
-    op.execute(f"ALTER TABLE IF EXISTS {schema}reflections RENAME TO mental_models")
-
-    # 6. Rename indexes for mental_models (was reflections)
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_bank_id RENAME TO idx_mental_models_bank_id")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_embedding RENAME TO idx_mental_models_embedding")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_tags RENAME TO idx_mental_models_tags")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_text_search RENAME TO idx_mental_models_text_search")
-
-    # 7. Rename foreign key constraint
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        DROP CONSTRAINT IF EXISTS fk_reflections_bank_id
-    """)
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        ADD CONSTRAINT fk_mental_models_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-
-
-def downgrade() -> None:
-    """Reverse: observation -> mental_model and mental_models -> reflections."""
-    schema = _get_schema_prefix()
-
-    # 1. Rename mental_models table back to reflections
-    op.execute(f"ALTER TABLE IF EXISTS {schema}mental_models RENAME TO reflections")
-
-    # 2. Rename indexes back
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_bank_id RENAME TO idx_reflections_bank_id")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_embedding RENAME TO idx_reflections_embedding")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_tags RENAME TO idx_reflections_tags")
-    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_text_search RENAME TO idx_reflections_text_search")
-
-    # 3. Rename foreign key back
-    op.execute(f"""
-        ALTER TABLE {schema}reflections
-        DROP CONSTRAINT IF EXISTS fk_mental_models_bank_id
-    """)
-    op.execute(f"""
-        ALTER TABLE {schema}reflections
-        ADD CONSTRAINT fk_reflections_bank_id
-        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
-    """)
-
-    # 4. Update fact_type values: observation -> mental_model
-    op.execute(f"""
-        UPDATE {schema}memory_units
-        SET fact_type = 'mental_model'
-        WHERE fact_type = 'observation'
-    """)
-
-    # 5. Update the CHECK constraint back
-    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
-    op.execute(f"""
-        ALTER TABLE {schema}memory_units
-        ADD CONSTRAINT memory_units_fact_type_check
-        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation', 'mental_model'))
-    """)
-
-    # 6. Rename index back
-    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_observations")
-    op.execute(f"""
-        CREATE INDEX IF NOT EXISTS idx_memory_units_mental_models
-        ON {schema}memory_units(bank_id, fact_type)
-        WHERE fact_type = 'mental_model'
-    """)
@@ -1,41 +0,0 @@
-"""Change mental_models.id from UUID to TEXT
-
-Revision ID: u6p7q8r9s0t1
-Revises: t5o6p7q8r9s0
-Create Date: 2026-01-27
-
-This migration changes the mental_models.id column from UUID to TEXT
-to support user-defined text identifiers like 'team-communication' instead of UUIDs.
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "u6p7q8r9s0t1"
-down_revision: str | Sequence[str] | None = "t5o6p7q8r9s0"
-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:
-    """Change mental_models.id from UUID to TEXT."""
-    schema = _get_schema_prefix()
-
-    # Change the id column type from UUID to TEXT
-    # Existing UUIDs will be converted to their string representation
-    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE TEXT USING id::TEXT")
-
-
-def downgrade() -> None:
-    """Revert mental_models.id from TEXT to UUID."""
-    schema = _get_schema_prefix()
-
-    # Note: This will fail if any id values are not valid UUIDs
-    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE UUID USING id::UUID")
@@ -1,50 +0,0 @@
-"""Add max_tokens and trigger columns to mental_models
-
-Revision ID: v7q8r9s0t1u2
-Revises: u6p7q8r9s0t1
-Create Date: 2026-01-27
-
-This migration adds:
- max_tokens column: token limit for content generation during refresh
- trigger column: JSONB for trigger settings (e.g., refresh_after_consolidation)
-"""
-
-from collections.abc import Sequence
-
-from alembic import context, op
-
-revision: str = "v7q8r9s0t1u2"
-down_revision: str | Sequence[str] | None = "u6p7q8r9s0t1"
-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 max_tokens and trigger columns to mental_models."""
-    schema = _get_schema_prefix()
-
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        ADD COLUMN IF NOT EXISTS max_tokens INT NOT NULL DEFAULT 2048
-    """)
-
-    # trigger column stores trigger settings as JSONB
-    # Default: refresh_after_consolidation = false (not "real time")
-    op.execute(f"""
-        ALTER TABLE {schema}mental_models
-        ADD COLUMN IF NOT EXISTS trigger JSONB NOT NULL DEFAULT '{{"refresh_after_consolidation": false}}'::jsonb
-    """)
-
-
-def downgrade() -> None:
-    """Remove max_tokens and trigger columns from mental_models."""
-    schema = _get_schema_prefix()
-
-    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS max_tokens")
-    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS trigger")
@@ -5,7 +5,6 @@ Provides both HTTP REST API and MCP (Model Context Protocol) server.
 """

 import logging
-from contextlib import asynccontextmanager
 from typing import Optional

 from fastapi import FastAPI
@@ -46,18 +45,6 @@ def create_app(
        # Both HTTP and MCP
        app = create_app(memory, mcp_api_enabled=True)
    """
-    mcp_app = None
-
-    # Create MCP app first if enabled (we need its lifespan for chaining)
-    if mcp_api_enabled:
-        try:
-            from .mcp import create_mcp_app
-
-            mcp_app = create_mcp_app(memory=memory)
-        except ImportError as e:
-            logger.error(f"MCP server requested but dependencies not available: {e}")
-            logger.error("Install with: pip install hindsight-api[mcp]")
-            raise

    # Import and create HTTP API if enabled
    if http_api_enabled:
@@ -70,31 +57,20 @@ def create_app(
        app = FastAPI(title="Hindsight API", version="0.0.7")
        logger.info("HTTP REST API disabled")

-    # Mount MCP server and chain its lifespan if enabled
-    if mcp_app is not None:
-        # Get the MCP app's underlying Starlette app for lifespan access
-        mcp_starlette_app = mcp_app.mcp_app
+    # Mount MCP server if enabled
+    if mcp_api_enabled:
+        try:
+            from .mcp import create_mcp_app

-        # Store the original lifespan
-        original_lifespan = app.router.lifespan_context
-
-        @asynccontextmanager
-        async def chained_lifespan(app_instance: FastAPI):
-            """Chain the MCP lifespan with the main app lifespan."""
-            # Start MCP lifespan first
-            async with mcp_starlette_app.router.lifespan_context(mcp_starlette_app):
-                logger.info("MCP lifespan started")
-                # Then start the original app lifespan
-                async with original_lifespan(app_instance):
-                    yield
-            logger.info("MCP lifespan stopped")
-
-        # Replace the app's lifespan with the chained version
-        app.router.lifespan_context = chained_lifespan
-
-        # Mount the MCP middleware
-        app.mount(mcp_mount_path, mcp_app)
-        logger.info(f"MCP server enabled at {mcp_mount_path}/")
+            # Create MCP app with dynamic bank_id support
+            # Supports: /mcp/{bank_id}/sse (bank-specific SSE endpoint)
+            mcp_app = create_mcp_app(memory=memory)
+            app.mount(mcp_mount_path, mcp_app)
+            logger.info(f"MCP server enabled at {mcp_mount_path}/{{bank_id}}/sse")
+        except ImportError as e:
+            logger.error(f"MCP server requested but dependencies not available: {e}")
+            logger.error("Install with: pip install hindsight-api[mcp]")
+            raise

    return app

@@ -1,4 +1,4 @@
-"""Hindsight MCP Server implementation using FastMCP (HTTP transport)."""
+"""Hindsight MCP Server implementation using FastMCP."""

 import json
 import logging
@@ -8,7 +8,8 @@ from contextvars import ContextVar
 from fastmcp import FastMCP

 from hindsight_api import MemoryEngine
-from hindsight_api.mcp_tools import MCPToolsConfig, register_mcp_tools
+from hindsight_api.engine.response_models import VALID_RECALL_FACT_TYPES
+from hindsight_api.models import RequestContext

 # Configure logging from HINDSIGHT_API_LOG_LEVEL environment variable
 _log_level_str = os.environ.get("HINDSIGHT_API_LOG_LEVEL", "info").lower()
@@ -26,29 +27,15 @@ logging.basicConfig(
 )
 logger = logging.getLogger(__name__)

-# Default bank_id from environment variable
-DEFAULT_BANK_ID = os.environ.get("HINDSIGHT_MCP_BANK_ID", "default")
-
-# MCP authentication token (optional - if set, Bearer token auth is required)
-MCP_AUTH_TOKEN = os.environ.get("HINDSIGHT_API_MCP_AUTH_TOKEN")
-
-# Context variable to hold the current bank_id
+# Context variable to hold the current bank_id from the URL path
 _current_bank_id: ContextVar[str | None] = ContextVar("current_bank_id", default=None)

-# Context variable to hold the current API key (for tenant auth propagation)
-_current_api_key: ContextVar[str | None] = ContextVar("current_api_key", default=None)
-

 def get_current_bank_id() -> str | None:
-    """Get the current bank_id from context."""
+    """Get the current bank_id from context (set from URL path)."""
    return _current_bank_id.get()


-def get_current_api_key() -> str | None:
-    """Get the current API key from context."""
-    return _current_api_key.get()
-
-
 def create_mcp_server(memory: MemoryEngine) -> FastMCP:
    """
    Create and configure the Hindsight MCP server.
@@ -57,79 +44,102 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
        memory: MemoryEngine instance (required)

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

-    # 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
-        include_bank_id_param=True,  # HTTP MCP supports multi-bank via parameter
-        tools=None,  # All tools
-        retain_fire_and_forget=False,  # HTTP MCP supports sync/async modes
-    )
+    @mcp.tool()
+    async def retain(content: str, context: str = "general") -> str:
+        """
+        Store important information to long-term memory.

-    register_mcp_tools(mcp, memory, config)
+        Use this tool PROACTIVELY whenever the user shares:
+        - Personal facts, preferences, or interests
+        - Important events or milestones
+        - User history, experiences, or background
+        - Decisions, opinions, or stated preferences
+        - Goals, plans, or future intentions
+        - Relationships or people mentioned
+        - Work context, projects, or responsibilities
+
+        Args:
+            content: The fact/memory to store (be specific and include relevant details)
+            context: Category for the memory (e.g., 'preferences', 'work', 'hobbies', 'family'). Default: 'general'
+        """
+        try:
+            bank_id = get_current_bank_id()
+            if bank_id is None:
+                return "Error: No bank_id configured"
+            await memory.retain_batch_async(
+                bank_id=bank_id, contents=[{"content": content, "context": context}], request_context=RequestContext()
+            )
+            return "Memory stored successfully"
+        except Exception as e:
+            logger.error(f"Error storing memory: {e}", exc_info=True)
+            return f"Error: {str(e)}"
+
+    @mcp.tool()
+    async def recall(query: str, max_results: int = 10) -> str:
+        """
+        Search memories to provide personalized, context-aware responses.
+
+        Use this tool PROACTIVELY to:
+        - Check user's preferences before making suggestions
+        - Recall user's history to provide continuity
+        - Remember user's goals and context
+        - Personalize responses based on past interactions
+
+        Args:
+            query: Natural language search query (e.g., "user's food preferences", "what projects is user working on")
+            max_results: Maximum number of results to return (default: 10)
+        """
+        try:
+            bank_id = get_current_bank_id()
+            if bank_id is None:
+                return "Error: No bank_id configured"
+            from hindsight_api.engine.memory_engine import Budget
+
+            search_result = await memory.recall_async(
+                bank_id=bank_id,
+                query=query,
+                fact_type=list(VALID_RECALL_FACT_TYPES),
+                budget=Budget.LOW,
+                request_context=RequestContext(),
+            )
+
+            results = [
+                {
+                    "id": fact.id,
+                    "text": fact.text,
+                    "type": fact.fact_type,
+                    "context": fact.context,
+                    "occurred_start": fact.occurred_start,
+                }
+                for fact in search_result.results[:max_results]
+            ]
+
+            return json.dumps({"results": results}, indent=2)
+        except Exception as e:
+            logger.error(f"Error searching: {e}", exc_info=True)
+            return json.dumps({"error": str(e), "results": []})

    return mcp


 class MCPMiddleware:
-    """ASGI middleware that handles authentication and extracts bank_id from header or path.
-
-    Authentication:
-        If HINDSIGHT_API_MCP_AUTH_TOKEN is set, all requests must include a valid
-        Authorization header with Bearer token or direct token matching the configured value.
-
-    Bank ID can be provided via:
-    1. X-Bank-Id header (recommended for Claude Code)
-    2. URL path: /mcp/{bank_id}/
-    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback default)
-
-    For Claude Code, configure with:
-        claude mcp add --transport http hindsight http://localhost:8888/mcp \\
-            --header "X-Bank-Id: my-bank" --header "Authorization: Bearer <token>"
-    """
+    """ASGI middleware that extracts bank_id from path and sets context."""

    def __init__(self, app, memory: MemoryEngine):
        self.app = app
        self.memory = memory
        self.mcp_server = create_mcp_server(memory)
-        self.mcp_app = self.mcp_server.http_app(path="/")
-        # Expose the lifespan for the parent app to chain
-        self.lifespan = self.mcp_app.lifespan_handler if hasattr(self.mcp_app, "lifespan_handler") else None
-
-    def _get_header(self, scope: dict, name: str) -> str | None:
-        """Extract a header value from ASGI scope."""
-        name_lower = name.lower().encode()
-        for header_name, header_value in scope.get("headers", []):
-            if header_name.lower() == name_lower:
-                return header_value.decode()
-        return None
+        self.mcp_app = self.mcp_server.http_app()

    async def __call__(self, scope, receive, send):
        if scope["type"] != "http":
            await self.mcp_app(scope, receive, send)
            return

-        # Extract auth token from header (for tenant auth propagation)
-        auth_header = self._get_header(scope, "Authorization")
-        auth_token: str | None = None
-        if auth_header:
-            # Support both "Bearer <token>" and direct token
-            auth_token = auth_header[7:].strip() if auth_header.startswith("Bearer ") else auth_header.strip()
-
-        # Authenticate if MCP_AUTH_TOKEN is configured
-        if MCP_AUTH_TOKEN:
-            if not auth_token:
-                await self._send_error(send, 401, "Authorization header required")
-                return
-            if auth_token != MCP_AUTH_TOKEN:
-                await self._send_error(send, 401, "Invalid authentication token")
-                return
-
        path = scope.get("path", "")

        # Strip any mount prefix (e.g., /mcp) that FastAPI might not have stripped
@@ -140,41 +150,32 @@ class MCPMiddleware:
        # Also handle case where mount path wasn't stripped (e.g., /mcp/...)
        if path.startswith("/mcp/"):
            path = path[4:]  # Remove /mcp prefix
-        elif path == "/mcp":
-            path = "/"

-        # Try to get bank_id from header first (for Claude Code compatibility)
-        bank_id = self._get_header(scope, "X-Bank-Id")
+        # Extract bank_id from path: /{bank_id}/ or /{bank_id}
+        # http_app expects requests at /
+        if not path.startswith("/") or len(path) <= 1:
+            # No bank_id in path - return error
+            await self._send_error(send, 400, "bank_id required in path: /mcp/{bank_id}/")
+            return

-        # MCP endpoint paths that should not be treated as bank_ids
-        MCP_ENDPOINTS = {"sse", "messages"}
+        # Extract bank_id from first path segment
+        parts = path[1:].split("/", 1)
+        if not parts[0]:
+            await self._send_error(send, 400, "bank_id required in path: /mcp/{bank_id}/")
+            return

-        # If no header, try to extract from path: /{bank_id}/...
-        new_path = path
-        if not bank_id and path.startswith("/") and len(path) > 1:
-            parts = path[1:].split("/", 1)
-            # Don't treat MCP endpoints as bank_ids
-            if parts[0] and parts[0] not in MCP_ENDPOINTS:
-                # First segment looks like a bank_id
-                bank_id = parts[0]
-                new_path = "/" + parts[1] if len(parts) > 1 else "/"
+        bank_id = parts[0]
+        new_path = "/" + parts[1] if len(parts) > 1 else "/"

-        # Fall back to default bank_id
-        if not bank_id:
-            bank_id = DEFAULT_BANK_ID
-            logger.debug(f"Using default bank_id: {bank_id}")
-
-        # Set bank_id and api_key context
-        bank_id_token = _current_bank_id.set(bank_id)
-        # Store the auth token for tenant extension to validate
-        api_key_token = _current_api_key.set(auth_token) if auth_token else None
+        # Set bank_id context
+        token = _current_bank_id.set(bank_id)
        try:
            new_scope = scope.copy()
            new_scope["path"] = new_path
-            # Clear root_path since we're passing directly to the app
-            new_scope["root_path"] = ""

-            # Wrap send to rewrite the SSE endpoint URL to include bank_id if using path-based routing
+            # Wrap send to rewrite the SSE endpoint URL to include bank_id
+            # The SSE app sends "event: endpoint\ndata: /messages\n" but we need
+            # the client to POST to /{bank_id}/messages instead
            async def send_wrapper(message):
                if message["type"] == "http.response.body":
                    body = message.get("body", b"")
@@ -186,9 +187,7 @@ class MCPMiddleware:

            await self.mcp_app(new_scope, receive, send_wrapper)
        finally:
-            _current_bank_id.reset(bank_id_token)
-            if api_key_token is not None:
-                _current_api_key.reset(api_key_token)
+            _current_bank_id.reset(token)

    async def _send_error(self, send, status: int, message: str):
        """Send an error response."""
@@ -212,14 +211,9 @@ def create_mcp_app(memory: MemoryEngine):
    """
    Create an ASGI app that handles MCP requests.

-    Authentication:
-        Set HINDSIGHT_API_MCP_AUTH_TOKEN to require Bearer token authentication.
-        If not set, MCP endpoint is open (for local development).
+    URL pattern: /mcp/{bank_id}/

-    Bank ID can be provided via:
-    1. X-Bank-Id header: claude mcp add --transport http hindsight http://localhost:8888/mcp --header "X-Bank-Id: my-bank"
-    2. URL path: /mcp/{bank_id}/
-    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback, default: "default")
+    The bank_id is extracted from the URL path and made available to tools.

    Args:
        memory: MemoryEngine instance
@@ -4,236 +4,56 @@ Centralized configuration for Hindsight API.
 All environment variables and their defaults are defined here.
 """

-import json
 import logging
 import os
-import sys
 from dataclasses import dataclass
-from datetime import datetime, timezone
-
-from dotenv import find_dotenv, load_dotenv
-
-# Load .env file, searching current and parent directories (overrides existing env vars)
-load_dotenv(find_dotenv(usecwd=True), override=True)

 logger = logging.getLogger(__name__)

 # Environment variable names
 ENV_DATABASE_URL = "HINDSIGHT_API_DATABASE_URL"
-ENV_DATABASE_SCHEMA = "HINDSIGHT_API_DATABASE_SCHEMA"
 ENV_LLM_PROVIDER = "HINDSIGHT_API_LLM_PROVIDER"
 ENV_LLM_API_KEY = "HINDSIGHT_API_LLM_API_KEY"
 ENV_LLM_MODEL = "HINDSIGHT_API_LLM_MODEL"
 ENV_LLM_BASE_URL = "HINDSIGHT_API_LLM_BASE_URL"
-ENV_LLM_MAX_CONCURRENT = "HINDSIGHT_API_LLM_MAX_CONCURRENT"
-ENV_LLM_MAX_RETRIES = "HINDSIGHT_API_LLM_MAX_RETRIES"
-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"
-
-# Per-operation LLM configuration (optional, falls back to global LLM config)
-ENV_RETAIN_LLM_PROVIDER = "HINDSIGHT_API_RETAIN_LLM_PROVIDER"
-ENV_RETAIN_LLM_API_KEY = "HINDSIGHT_API_RETAIN_LLM_API_KEY"
-ENV_RETAIN_LLM_MODEL = "HINDSIGHT_API_RETAIN_LLM_MODEL"
-ENV_RETAIN_LLM_BASE_URL = "HINDSIGHT_API_RETAIN_LLM_BASE_URL"
-ENV_RETAIN_LLM_MAX_CONCURRENT = "HINDSIGHT_API_RETAIN_LLM_MAX_CONCURRENT"
-ENV_RETAIN_LLM_MAX_RETRIES = "HINDSIGHT_API_RETAIN_LLM_MAX_RETRIES"
-ENV_RETAIN_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_RETAIN_LLM_INITIAL_BACKOFF"
-ENV_RETAIN_LLM_MAX_BACKOFF = "HINDSIGHT_API_RETAIN_LLM_MAX_BACKOFF"
-ENV_RETAIN_LLM_TIMEOUT = "HINDSIGHT_API_RETAIN_LLM_TIMEOUT"
-
-ENV_REFLECT_LLM_PROVIDER = "HINDSIGHT_API_REFLECT_LLM_PROVIDER"
-ENV_REFLECT_LLM_API_KEY = "HINDSIGHT_API_REFLECT_LLM_API_KEY"
-ENV_REFLECT_LLM_MODEL = "HINDSIGHT_API_REFLECT_LLM_MODEL"
-ENV_REFLECT_LLM_BASE_URL = "HINDSIGHT_API_REFLECT_LLM_BASE_URL"
-ENV_REFLECT_LLM_MAX_CONCURRENT = "HINDSIGHT_API_REFLECT_LLM_MAX_CONCURRENT"
-ENV_REFLECT_LLM_MAX_RETRIES = "HINDSIGHT_API_REFLECT_LLM_MAX_RETRIES"
-ENV_REFLECT_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_REFLECT_LLM_INITIAL_BACKOFF"
-ENV_REFLECT_LLM_MAX_BACKOFF = "HINDSIGHT_API_REFLECT_LLM_MAX_BACKOFF"
-ENV_REFLECT_LLM_TIMEOUT = "HINDSIGHT_API_REFLECT_LLM_TIMEOUT"
-
-ENV_CONSOLIDATION_LLM_PROVIDER = "HINDSIGHT_API_CONSOLIDATION_LLM_PROVIDER"
-ENV_CONSOLIDATION_LLM_API_KEY = "HINDSIGHT_API_CONSOLIDATION_LLM_API_KEY"
-ENV_CONSOLIDATION_LLM_MODEL = "HINDSIGHT_API_CONSOLIDATION_LLM_MODEL"
-ENV_CONSOLIDATION_LLM_BASE_URL = "HINDSIGHT_API_CONSOLIDATION_LLM_BASE_URL"
-ENV_CONSOLIDATION_LLM_MAX_CONCURRENT = "HINDSIGHT_API_CONSOLIDATION_LLM_MAX_CONCURRENT"
-ENV_CONSOLIDATION_LLM_MAX_RETRIES = "HINDSIGHT_API_CONSOLIDATION_LLM_MAX_RETRIES"
-ENV_CONSOLIDATION_LLM_INITIAL_BACKOFF = "HINDSIGHT_API_CONSOLIDATION_LLM_INITIAL_BACKOFF"
-ENV_CONSOLIDATION_LLM_MAX_BACKOFF = "HINDSIGHT_API_CONSOLIDATION_LLM_MAX_BACKOFF"
-ENV_CONSOLIDATION_LLM_TIMEOUT = "HINDSIGHT_API_CONSOLIDATION_LLM_TIMEOUT"

 ENV_EMBEDDINGS_PROVIDER = "HINDSIGHT_API_EMBEDDINGS_PROVIDER"
 ENV_EMBEDDINGS_LOCAL_MODEL = "HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL"
-ENV_EMBEDDINGS_LOCAL_FORCE_CPU = "HINDSIGHT_API_EMBEDDINGS_LOCAL_FORCE_CPU"
 ENV_EMBEDDINGS_TEI_URL = "HINDSIGHT_API_EMBEDDINGS_TEI_URL"
-ENV_EMBEDDINGS_OPENAI_API_KEY = "HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY"
-ENV_EMBEDDINGS_OPENAI_MODEL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_MODEL"
-ENV_EMBEDDINGS_OPENAI_BASE_URL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_BASE_URL"
-
-ENV_COHERE_API_KEY = "HINDSIGHT_API_COHERE_API_KEY"
-ENV_EMBEDDINGS_COHERE_MODEL = "HINDSIGHT_API_EMBEDDINGS_COHERE_MODEL"
-ENV_EMBEDDINGS_COHERE_BASE_URL = "HINDSIGHT_API_EMBEDDINGS_COHERE_BASE_URL"
-ENV_RERANKER_COHERE_MODEL = "HINDSIGHT_API_RERANKER_COHERE_MODEL"
-ENV_RERANKER_COHERE_BASE_URL = "HINDSIGHT_API_RERANKER_COHERE_BASE_URL"
-
-# LiteLLM gateway configuration (for embeddings and reranker via LiteLLM proxy)
-ENV_LITELLM_API_BASE = "HINDSIGHT_API_LITELLM_API_BASE"
-ENV_LITELLM_API_KEY = "HINDSIGHT_API_LITELLM_API_KEY"
-ENV_EMBEDDINGS_LITELLM_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_MODEL"
-ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"

 ENV_RERANKER_PROVIDER = "HINDSIGHT_API_RERANKER_PROVIDER"
 ENV_RERANKER_LOCAL_MODEL = "HINDSIGHT_API_RERANKER_LOCAL_MODEL"
-ENV_RERANKER_LOCAL_FORCE_CPU = "HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU"
-ENV_RERANKER_LOCAL_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT"
 ENV_RERANKER_TEI_URL = "HINDSIGHT_API_RERANKER_TEI_URL"
-ENV_RERANKER_TEI_BATCH_SIZE = "HINDSIGHT_API_RERANKER_TEI_BATCH_SIZE"
-ENV_RERANKER_TEI_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_TEI_MAX_CONCURRENT"
-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"

 ENV_HOST = "HINDSIGHT_API_HOST"
 ENV_PORT = "HINDSIGHT_API_PORT"
 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_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"
-
-# Vertex AI configuration
-ENV_LLM_VERTEXAI_PROJECT_ID = "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID"
-ENV_LLM_VERTEXAI_REGION = "HINDSIGHT_API_LLM_VERTEXAI_REGION"
-ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY = "HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY"
-
-# Retain settings
-ENV_RETAIN_MAX_COMPLETION_TOKENS = "HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS"
-ENV_RETAIN_CHUNK_SIZE = "HINDSIGHT_API_RETAIN_CHUNK_SIZE"
-ENV_RETAIN_EXTRACT_CAUSAL_LINKS = "HINDSIGHT_API_RETAIN_EXTRACT_CAUSAL_LINKS"
-ENV_RETAIN_EXTRACTION_MODE = "HINDSIGHT_API_RETAIN_EXTRACTION_MODE"
-ENV_RETAIN_CUSTOM_INSTRUCTIONS = "HINDSIGHT_API_RETAIN_CUSTOM_INSTRUCTIONS"
-
-# Observations settings (consolidated knowledge from facts)
-ENV_ENABLE_OBSERVATIONS = "HINDSIGHT_API_ENABLE_OBSERVATIONS"
-ENV_CONSOLIDATION_BATCH_SIZE = "HINDSIGHT_API_CONSOLIDATION_BATCH_SIZE"
-ENV_CONSOLIDATION_MAX_TOKENS = "HINDSIGHT_API_CONSOLIDATION_MAX_TOKENS"

 # Optimization flags
 ENV_SKIP_LLM_VERIFICATION = "HINDSIGHT_API_SKIP_LLM_VERIFICATION"
 ENV_LAZY_RERANKER = "HINDSIGHT_API_LAZY_RERANKER"

-# Database migrations
-ENV_RUN_MIGRATIONS_ON_STARTUP = "HINDSIGHT_API_RUN_MIGRATIONS_ON_STARTUP"
-
-# Database connection pool
-ENV_DB_POOL_MIN_SIZE = "HINDSIGHT_API_DB_POOL_MIN_SIZE"
-ENV_DB_POOL_MAX_SIZE = "HINDSIGHT_API_DB_POOL_MAX_SIZE"
-ENV_DB_COMMAND_TIMEOUT = "HINDSIGHT_API_DB_COMMAND_TIMEOUT"
-ENV_DB_ACQUIRE_TIMEOUT = "HINDSIGHT_API_DB_ACQUIRE_TIMEOUT"
-
-# Worker configuration (distributed task processing)
-ENV_WORKER_ENABLED = "HINDSIGHT_API_WORKER_ENABLED"
-ENV_WORKER_ID = "HINDSIGHT_API_WORKER_ID"
-ENV_WORKER_POLL_INTERVAL_MS = "HINDSIGHT_API_WORKER_POLL_INTERVAL_MS"
-ENV_WORKER_MAX_RETRIES = "HINDSIGHT_API_WORKER_MAX_RETRIES"
-ENV_WORKER_BATCH_SIZE = "HINDSIGHT_API_WORKER_BATCH_SIZE"
-ENV_WORKER_HTTP_PORT = "HINDSIGHT_API_WORKER_HTTP_PORT"
-
-# Reflect agent settings
-ENV_REFLECT_MAX_ITERATIONS = "HINDSIGHT_API_REFLECT_MAX_ITERATIONS"
-
 # Default values
 DEFAULT_DATABASE_URL = "pg0"
-DEFAULT_DATABASE_SCHEMA = "public"
 DEFAULT_LLM_PROVIDER = "openai"
 DEFAULT_LLM_MODEL = "gpt-5-mini"
-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
-DEFAULT_LLM_MAX_BACKOFF = 60.0  # Max backoff cap in seconds for retry exponential backoff
-DEFAULT_LLM_TIMEOUT = 120.0  # seconds
-
-# Vertex AI defaults
-DEFAULT_LLM_VERTEXAI_PROJECT_ID = None  # Required for Vertex AI
-DEFAULT_LLM_VERTEXAI_REGION = "us-central1"
-DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY = None  # Optional, uses ADC if not set

 DEFAULT_EMBEDDINGS_PROVIDER = "local"
 DEFAULT_EMBEDDINGS_LOCAL_MODEL = "BAAI/bge-small-en-v1.5"
-DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU = False  # Force CPU mode for local embeddings (avoids MPS/XPC issues on macOS)
-DEFAULT_EMBEDDINGS_OPENAI_MODEL = "text-embedding-3-small"
-DEFAULT_EMBEDDING_DIMENSION = 384

 DEFAULT_RERANKER_PROVIDER = "local"
 DEFAULT_RERANKER_LOCAL_MODEL = "cross-encoder/ms-marco-MiniLM-L-6-v2"
-DEFAULT_RERANKER_LOCAL_FORCE_CPU = False  # Force CPU mode for local reranker (avoids MPS/XPC issues on macOS)
-DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4  # Limit concurrent CPU-bound reranking to prevent thrashing
-DEFAULT_RERANKER_TEI_BATCH_SIZE = 128
-DEFAULT_RERANKER_TEI_MAX_CONCURRENT = 8
-DEFAULT_RERANKER_MAX_CANDIDATES = 300
-DEFAULT_RERANKER_FLASHRANK_MODEL = "ms-marco-MiniLM-L-12-v2"  # Best balance of speed and quality
-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"
-
-# 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"

 DEFAULT_HOST = "0.0.0.0"
 DEFAULT_PORT = 8888
 DEFAULT_LOG_LEVEL = "info"
-DEFAULT_LOG_FORMAT = "text"  # Options: "text", "json"
-DEFAULT_WORKERS = 1
 DEFAULT_MCP_ENABLED = 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_GRAPH_RETRIEVER = "bfs"  # Options: "bfs", "mpfp"
 DEFAULT_MCP_LOCAL_BANK_ID = "mcp"
-DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY = 8  # Max concurrent mental model refreshes
-
-# Retain settings
-DEFAULT_RETAIN_MAX_COMPLETION_TOKENS = 64000  # Max tokens for fact extraction LLM call
-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_CUSTOM_INSTRUCTIONS = None  # Custom extraction guidelines (only used when mode="custom")
-
-# 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_MAX_TOKENS = 1024  # Max tokens for recall when finding related observations
-
-# Database migrations
-DEFAULT_RUN_MIGRATIONS_ON_STARTUP = True
-
-# Database connection pool
-DEFAULT_DB_POOL_MIN_SIZE = 5
-DEFAULT_DB_POOL_MAX_SIZE = 100
-DEFAULT_DB_COMMAND_TIMEOUT = 60  # seconds
-DEFAULT_DB_ACQUIRE_TIMEOUT = 30  # seconds
-
-# Worker configuration (distributed task processing)
-DEFAULT_WORKER_ENABLED = True  # API runs worker by default (standalone mode)
-DEFAULT_WORKER_ID = None  # Will use hostname if not specified
-DEFAULT_WORKER_POLL_INTERVAL_MS = 500  # Poll database every 500ms
-DEFAULT_WORKER_MAX_RETRIES = 3  # Max retries before marking task failed
-DEFAULT_WORKER_BATCH_SIZE = 10  # Tasks to claim per poll cycle
-DEFAULT_WORKER_HTTP_PORT = 8889  # HTTP port for worker metrics/health
-
-# Reflect agent settings
-DEFAULT_REFLECT_MAX_ITERATIONS = 10  # Max tool call iterations before forcing response

 # Default MCP tool descriptions (can be customized via env vars)
 DEFAULT_MCP_RETAIN_DESCRIPTION = """Store important information to long-term memory.
@@ -255,50 +75,8 @@ Use this tool PROACTIVELY to:
 - Remember user's goals and context
 - Personalize responses based on past interactions"""

-# Default embedding dimension (used by initial migration, adjusted at runtime)
-EMBEDDING_DIMENSION = DEFAULT_EMBEDDING_DIMENSION
-
-
-class JsonFormatter(logging.Formatter):
-    """JSON formatter for structured logging.
-
-    Outputs logs in JSON format with a 'severity' field that cloud logging
-    systems (GCP, AWS CloudWatch, etc.) can parse to correctly categorize log levels.
-    """
-
-    SEVERITY_MAP = {
-        logging.DEBUG: "DEBUG",
-        logging.INFO: "INFO",
-        logging.WARNING: "WARNING",
-        logging.ERROR: "ERROR",
-        logging.CRITICAL: "CRITICAL",
-    }
-
-    def format(self, record: logging.LogRecord) -> str:
-        log_entry = {
-            "severity": self.SEVERITY_MAP.get(record.levelno, "DEFAULT"),
-            "message": record.getMessage(),
-            "timestamp": datetime.now(timezone.utc).isoformat(),
-            "logger": record.name,
-        }
-
-        # Add exception info if present
-        if record.exc_info:
-            log_entry["exception"] = self.formatException(record.exc_info)
-
-        return json.dumps(log_entry)
-
-
-def _validate_extraction_mode(mode: str) -> str:
-    """Validate and normalize extraction mode."""
-    mode_lower = mode.lower()
-    if mode_lower not in RETAIN_EXTRACTION_MODES:
-        logger.warning(
-            f"Invalid extraction mode '{mode}', must be one of {RETAIN_EXTRACTION_MODES}. "
-            f"Defaulting to '{DEFAULT_RETAIN_EXTRACTION_MODE}'."
-        )
-        return DEFAULT_RETAIN_EXTRACTION_MODE
-    return mode_lower
+# Required embedding dimension for database schema
+EMBEDDING_DIMENSION = 384


@dataclass
@@ -307,285 +85,65 @@ class HindsightConfig:

    # Database
    database_url: str
-    database_schema: str

-    # LLM (default, used as fallback for per-operation config)
+    # LLM
    llm_provider: str
    llm_api_key: str | None
    llm_model: str
    llm_base_url: str | None
-    llm_max_concurrent: int
-    llm_max_retries: int
-    llm_initial_backoff: float
-    llm_max_backoff: float
-    llm_timeout: float
-
-    # Vertex AI configuration
-    llm_vertexai_project_id: str | None
-    llm_vertexai_region: str
-    llm_vertexai_service_account_key: str | None
-
-    # Per-operation LLM configuration (None = use default LLM config)
-    retain_llm_provider: str | None
-    retain_llm_api_key: str | None
-    retain_llm_model: str | None
-    retain_llm_base_url: str | None
-    retain_llm_max_concurrent: int | None
-    retain_llm_max_retries: int | None
-    retain_llm_initial_backoff: float | None
-    retain_llm_max_backoff: float | None
-    retain_llm_timeout: float | None
-
-    reflect_llm_provider: str | None
-    reflect_llm_api_key: str | None
-    reflect_llm_model: str | None
-    reflect_llm_base_url: str | None
-    reflect_llm_max_concurrent: int | None
-    reflect_llm_max_retries: int | None
-    reflect_llm_initial_backoff: float | None
-    reflect_llm_max_backoff: float | None
-    reflect_llm_timeout: float | None
-
-    consolidation_llm_provider: str | None
-    consolidation_llm_api_key: str | None
-    consolidation_llm_model: str | None
-    consolidation_llm_base_url: str | None
-    consolidation_llm_max_concurrent: int | None
-    consolidation_llm_max_retries: int | None
-    consolidation_llm_initial_backoff: float | None
-    consolidation_llm_max_backoff: float | None
-    consolidation_llm_timeout: float | None

    # Embeddings
    embeddings_provider: str
    embeddings_local_model: str
-    embeddings_local_force_cpu: bool
    embeddings_tei_url: str | None
-    embeddings_openai_base_url: str | None
-    embeddings_cohere_base_url: str | None

    # Reranker
    reranker_provider: str
    reranker_local_model: str
-    reranker_local_force_cpu: bool
-    reranker_local_max_concurrent: int
    reranker_tei_url: str | None
-    reranker_tei_batch_size: int
-    reranker_tei_max_concurrent: int
-    reranker_max_candidates: int
-    reranker_cohere_base_url: str | None

    # Server
    host: str
    port: int
    log_level: str
-    log_format: str
    mcp_enabled: bool

    # Recall
    graph_retriever: str
-    mpfp_top_k_neighbors: int
-    recall_max_concurrent: int
-    recall_connection_budget: int
-    mental_model_refresh_concurrency: int
-
-    # Retain settings
-    retain_max_completion_tokens: int
-    retain_chunk_size: int
-    retain_extract_causal_links: bool
-    retain_extraction_mode: str
-    retain_custom_instructions: str | None
-
-    # Observations settings (consolidated knowledge from facts)
-    enable_observations: bool
-    consolidation_batch_size: int
-    consolidation_max_tokens: int

    # Optimization flags
    skip_llm_verification: bool
    lazy_reranker: bool

-    # Database migrations
-    run_migrations_on_startup: bool
-
-    # Database connection pool
-    db_pool_min_size: int
-    db_pool_max_size: int
-    db_command_timeout: int
-    db_acquire_timeout: int
-
-    # Worker configuration (distributed task processing)
-    worker_enabled: bool
-    worker_id: str | None
-    worker_poll_interval_ms: int
-    worker_max_retries: int
-    worker_batch_size: int
-    worker_http_port: int
-
-    # Reflect agent settings
-    reflect_max_iterations: int
-
    @classmethod
    def from_env(cls) -> "HindsightConfig":
        """Create configuration from environment variables."""
        return cls(
            # Database
            database_url=os.getenv(ENV_DATABASE_URL, DEFAULT_DATABASE_URL),
-            database_schema=os.getenv(ENV_DATABASE_SCHEMA, DEFAULT_DATABASE_SCHEMA),
            # LLM
            llm_provider=os.getenv(ENV_LLM_PROVIDER, DEFAULT_LLM_PROVIDER),
            llm_api_key=os.getenv(ENV_LLM_API_KEY),
            llm_model=os.getenv(ENV_LLM_MODEL, DEFAULT_LLM_MODEL),
            llm_base_url=os.getenv(ENV_LLM_BASE_URL) or None,
-            llm_max_concurrent=int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_CONCURRENT))),
-            llm_max_retries=int(os.getenv(ENV_LLM_MAX_RETRIES, str(DEFAULT_LLM_MAX_RETRIES))),
-            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))),
-            # 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),
-            llm_vertexai_service_account_key=os.getenv(ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY)
-            or DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY,
-            # Per-operation LLM config (None = use default)
-            retain_llm_provider=os.getenv(ENV_RETAIN_LLM_PROVIDER) or None,
-            retain_llm_api_key=os.getenv(ENV_RETAIN_LLM_API_KEY) or None,
-            retain_llm_model=os.getenv(ENV_RETAIN_LLM_MODEL) or None,
-            retain_llm_base_url=os.getenv(ENV_RETAIN_LLM_BASE_URL) or None,
-            retain_llm_max_concurrent=int(os.getenv(ENV_RETAIN_LLM_MAX_CONCURRENT))
-            if os.getenv(ENV_RETAIN_LLM_MAX_CONCURRENT)
-            else None,
-            retain_llm_max_retries=int(os.getenv(ENV_RETAIN_LLM_MAX_RETRIES))
-            if os.getenv(ENV_RETAIN_LLM_MAX_RETRIES)
-            else None,
-            retain_llm_initial_backoff=float(os.getenv(ENV_RETAIN_LLM_INITIAL_BACKOFF))
-            if os.getenv(ENV_RETAIN_LLM_INITIAL_BACKOFF)
-            else None,
-            retain_llm_max_backoff=float(os.getenv(ENV_RETAIN_LLM_MAX_BACKOFF))
-            if os.getenv(ENV_RETAIN_LLM_MAX_BACKOFF)
-            else None,
-            retain_llm_timeout=float(os.getenv(ENV_RETAIN_LLM_TIMEOUT)) if os.getenv(ENV_RETAIN_LLM_TIMEOUT) else None,
-            reflect_llm_provider=os.getenv(ENV_REFLECT_LLM_PROVIDER) or None,
-            reflect_llm_api_key=os.getenv(ENV_REFLECT_LLM_API_KEY) or None,
-            reflect_llm_model=os.getenv(ENV_REFLECT_LLM_MODEL) or None,
-            reflect_llm_base_url=os.getenv(ENV_REFLECT_LLM_BASE_URL) or None,
-            reflect_llm_max_concurrent=int(os.getenv(ENV_REFLECT_LLM_MAX_CONCURRENT))
-            if os.getenv(ENV_REFLECT_LLM_MAX_CONCURRENT)
-            else None,
-            reflect_llm_max_retries=int(os.getenv(ENV_REFLECT_LLM_MAX_RETRIES))
-            if os.getenv(ENV_REFLECT_LLM_MAX_RETRIES)
-            else None,
-            reflect_llm_initial_backoff=float(os.getenv(ENV_REFLECT_LLM_INITIAL_BACKOFF))
-            if os.getenv(ENV_REFLECT_LLM_INITIAL_BACKOFF)
-            else None,
-            reflect_llm_max_backoff=float(os.getenv(ENV_REFLECT_LLM_MAX_BACKOFF))
-            if os.getenv(ENV_REFLECT_LLM_MAX_BACKOFF)
-            else None,
-            reflect_llm_timeout=float(os.getenv(ENV_REFLECT_LLM_TIMEOUT))
-            if os.getenv(ENV_REFLECT_LLM_TIMEOUT)
-            else None,
-            consolidation_llm_provider=os.getenv(ENV_CONSOLIDATION_LLM_PROVIDER) or None,
-            consolidation_llm_api_key=os.getenv(ENV_CONSOLIDATION_LLM_API_KEY) or None,
-            consolidation_llm_model=os.getenv(ENV_CONSOLIDATION_LLM_MODEL) or None,
-            consolidation_llm_base_url=os.getenv(ENV_CONSOLIDATION_LLM_BASE_URL) or None,
-            consolidation_llm_max_concurrent=int(os.getenv(ENV_CONSOLIDATION_LLM_MAX_CONCURRENT))
-            if os.getenv(ENV_CONSOLIDATION_LLM_MAX_CONCURRENT)
-            else None,
-            consolidation_llm_max_retries=int(os.getenv(ENV_CONSOLIDATION_LLM_MAX_RETRIES))
-            if os.getenv(ENV_CONSOLIDATION_LLM_MAX_RETRIES)
-            else None,
-            consolidation_llm_initial_backoff=float(os.getenv(ENV_CONSOLIDATION_LLM_INITIAL_BACKOFF))
-            if os.getenv(ENV_CONSOLIDATION_LLM_INITIAL_BACKOFF)
-            else None,
-            consolidation_llm_max_backoff=float(os.getenv(ENV_CONSOLIDATION_LLM_MAX_BACKOFF))
-            if os.getenv(ENV_CONSOLIDATION_LLM_MAX_BACKOFF)
-            else None,
-            consolidation_llm_timeout=float(os.getenv(ENV_CONSOLIDATION_LLM_TIMEOUT))
-            if os.getenv(ENV_CONSOLIDATION_LLM_TIMEOUT)
-            else None,
            # Embeddings
            embeddings_provider=os.getenv(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER),
            embeddings_local_model=os.getenv(ENV_EMBEDDINGS_LOCAL_MODEL, DEFAULT_EMBEDDINGS_LOCAL_MODEL),
-            embeddings_local_force_cpu=os.getenv(
-                ENV_EMBEDDINGS_LOCAL_FORCE_CPU, str(DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU)
-            ).lower()
-            in ("true", "1"),
            embeddings_tei_url=os.getenv(ENV_EMBEDDINGS_TEI_URL),
-            embeddings_openai_base_url=os.getenv(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None,
-            embeddings_cohere_base_url=os.getenv(ENV_EMBEDDINGS_COHERE_BASE_URL) 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),
-            reranker_local_force_cpu=os.getenv(
-                ENV_RERANKER_LOCAL_FORCE_CPU, str(DEFAULT_RERANKER_LOCAL_FORCE_CPU)
-            ).lower()
-            in ("true", "1"),
-            reranker_local_max_concurrent=int(
-                os.getenv(ENV_RERANKER_LOCAL_MAX_CONCURRENT, str(DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT))
-            ),
            reranker_tei_url=os.getenv(ENV_RERANKER_TEI_URL),
-            reranker_tei_batch_size=int(os.getenv(ENV_RERANKER_TEI_BATCH_SIZE, str(DEFAULT_RERANKER_TEI_BATCH_SIZE))),
-            reranker_tei_max_concurrent=int(
-                os.getenv(ENV_RERANKER_TEI_MAX_CONCURRENT, str(DEFAULT_RERANKER_TEI_MAX_CONCURRENT))
-            ),
-            reranker_max_candidates=int(os.getenv(ENV_RERANKER_MAX_CANDIDATES, str(DEFAULT_RERANKER_MAX_CANDIDATES))),
-            reranker_cohere_base_url=os.getenv(ENV_RERANKER_COHERE_BASE_URL) or None,
            # Server
            host=os.getenv(ENV_HOST, DEFAULT_HOST),
            port=int(os.getenv(ENV_PORT, DEFAULT_PORT)),
            log_level=os.getenv(ENV_LOG_LEVEL, DEFAULT_LOG_LEVEL),
-            log_format=os.getenv(ENV_LOG_FORMAT, DEFAULT_LOG_FORMAT).lower(),
            mcp_enabled=os.getenv(ENV_MCP_ENABLED, str(DEFAULT_MCP_ENABLED)).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))),
-            recall_max_concurrent=int(os.getenv(ENV_RECALL_MAX_CONCURRENT, str(DEFAULT_RECALL_MAX_CONCURRENT))),
-            recall_connection_budget=int(
-                os.getenv(ENV_RECALL_CONNECTION_BUDGET, str(DEFAULT_RECALL_CONNECTION_BUDGET))
-            ),
-            mental_model_refresh_concurrency=int(
-                os.getenv(ENV_MENTAL_MODEL_REFRESH_CONCURRENCY, str(DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY))
-            ),
            # Optimization flags
            skip_llm_verification=os.getenv(ENV_SKIP_LLM_VERIFICATION, "false").lower() == "true",
            lazy_reranker=os.getenv(ENV_LAZY_RERANKER, "false").lower() == "true",
-            # Retain settings
-            retain_max_completion_tokens=int(
-                os.getenv(ENV_RETAIN_MAX_COMPLETION_TOKENS, str(DEFAULT_RETAIN_MAX_COMPLETION_TOKENS))
-            ),
-            retain_chunk_size=int(os.getenv(ENV_RETAIN_CHUNK_SIZE, str(DEFAULT_RETAIN_CHUNK_SIZE))),
-            retain_extract_causal_links=os.getenv(
-                ENV_RETAIN_EXTRACT_CAUSAL_LINKS, str(DEFAULT_RETAIN_EXTRACT_CAUSAL_LINKS)
-            ).lower()
-            == "true",
-            retain_extraction_mode=_validate_extraction_mode(
-                os.getenv(ENV_RETAIN_EXTRACTION_MODE, DEFAULT_RETAIN_EXTRACTION_MODE)
-            ),
-            retain_custom_instructions=os.getenv(ENV_RETAIN_CUSTOM_INSTRUCTIONS) or DEFAULT_RETAIN_CUSTOM_INSTRUCTIONS,
-            # 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_max_tokens=int(
-                os.getenv(ENV_CONSOLIDATION_MAX_TOKENS, str(DEFAULT_CONSOLIDATION_MAX_TOKENS))
-            ),
-            # Database migrations
-            run_migrations_on_startup=os.getenv(ENV_RUN_MIGRATIONS_ON_STARTUP, "true").lower() == "true",
-            # Database connection pool
-            db_pool_min_size=int(os.getenv(ENV_DB_POOL_MIN_SIZE, str(DEFAULT_DB_POOL_MIN_SIZE))),
-            db_pool_max_size=int(os.getenv(ENV_DB_POOL_MAX_SIZE, str(DEFAULT_DB_POOL_MAX_SIZE))),
-            db_command_timeout=int(os.getenv(ENV_DB_COMMAND_TIMEOUT, str(DEFAULT_DB_COMMAND_TIMEOUT))),
-            db_acquire_timeout=int(os.getenv(ENV_DB_ACQUIRE_TIMEOUT, str(DEFAULT_DB_ACQUIRE_TIMEOUT))),
-            # Worker configuration
-            worker_enabled=os.getenv(ENV_WORKER_ENABLED, str(DEFAULT_WORKER_ENABLED)).lower() == "true",
-            worker_id=os.getenv(ENV_WORKER_ID) or DEFAULT_WORKER_ID,
-            worker_poll_interval_ms=int(os.getenv(ENV_WORKER_POLL_INTERVAL_MS, str(DEFAULT_WORKER_POLL_INTERVAL_MS))),
-            worker_max_retries=int(os.getenv(ENV_WORKER_MAX_RETRIES, str(DEFAULT_WORKER_MAX_RETRIES))),
-            worker_batch_size=int(os.getenv(ENV_WORKER_BATCH_SIZE, str(DEFAULT_WORKER_BATCH_SIZE))),
-            worker_http_port=int(os.getenv(ENV_WORKER_HTTP_PORT, str(DEFAULT_WORKER_HTTP_PORT))),
-            # Reflect agent settings
-            reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
        )

    def get_llm_base_url(self) -> str:
@@ -598,8 +156,6 @@ class HindsightConfig:
            return "https://api.groq.com/openai/v1"
        elif provider == "ollama":
            return "http://localhost:11434/v1"
-        elif provider == "lmstudio":
-            return "http://localhost:1234/v1"
        else:
            return ""

@@ -616,63 +172,22 @@ class HindsightConfig:
        return log_level_map.get(self.log_level.lower(), logging.INFO)

    def configure_logging(self) -> None:
-        """Configure Python logging based on the log level and format.
-
-        When log_format is "json", outputs structured JSON logs with a severity
-        field that GCP Cloud Logging can parse for proper log level categorization.
-        """
-        root_logger = logging.getLogger()
-        root_logger.setLevel(self.get_python_log_level())
-
-        # Remove existing handlers
-        for handler in root_logger.handlers[:]:
-            root_logger.removeHandler(handler)
-
-        # Create handler writing to stdout (GCP treats stderr as ERROR)
-        handler = logging.StreamHandler(sys.stdout)
-        handler.setLevel(self.get_python_log_level())
-
-        if self.log_format == "json":
-            handler.setFormatter(JsonFormatter())
-        else:
-            handler.setFormatter(logging.Formatter("%(asctime)s - %(levelname)s - %(name)s - %(message)s"))
-
-        root_logger.addHandler(handler)
+        """Configure Python logging based on the log level."""
+        logging.basicConfig(
+            level=self.get_python_log_level(),
+            format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+            force=True,  # Override any existing configuration
+        )

    def log_config(self) -> None:
        """Log the current configuration (without sensitive values)."""
-        logger.info(f"Database: {self.database_url} (schema: {self.database_schema})")
+        logger.info(f"Database: {self.database_url}")
        logger.info(f"LLM: provider={self.llm_provider}, model={self.llm_model}")
-        if self.retain_llm_provider or self.retain_llm_model:
-            retain_provider = self.retain_llm_provider or self.llm_provider
-            retain_model = self.retain_llm_model or self.llm_model
-            logger.info(f"LLM (retain): provider={retain_provider}, model={retain_model}")
-        if self.reflect_llm_provider or self.reflect_llm_model:
-            reflect_provider = self.reflect_llm_provider or self.llm_provider
-            reflect_model = self.reflect_llm_model or self.llm_model
-            logger.info(f"LLM (reflect): provider={reflect_provider}, model={reflect_model}")
-        if self.consolidation_llm_provider or self.consolidation_llm_model:
-            consolidation_provider = self.consolidation_llm_provider or self.llm_provider
-            consolidation_model = self.consolidation_llm_model or self.llm_model
-            logger.info(f"LLM (consolidation): provider={consolidation_provider}, model={consolidation_model}")
        logger.info(f"Embeddings: provider={self.embeddings_provider}")
        logger.info(f"Reranker: provider={self.reranker_provider}")
        logger.info(f"Graph retriever: {self.graph_retriever}")


-# Cached config instance
-_config_cache: HindsightConfig | None = None
-
-
 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()
-    return _config_cache
-
-
-def clear_config_cache() -> None:
-    """Clear the config cache. Useful for testing or reloading config."""
-    global _config_cache
-    _config_cache = None
+    """Get the current configuration from environment variables."""
+    return HindsightConfig.from_env()
@@ -52,10 +52,7 @@ class IdleTimeoutMiddleware:
                logger.info(f"Idle timeout reached ({self.idle_timeout}s), shutting down daemon")
                # Give a moment for any in-flight requests
                await asyncio.sleep(1)
-                # Send SIGTERM to ourselves to trigger graceful shutdown
-                import signal
-
-                os.kill(os.getpid(), signal.SIGTERM)
+                os._exit(0)


 class DaemonLock:
@@ -1,5 +0,0 @@
-"""Consolidation engine for automatic learning creation from memories."""
-
-from .consolidator import run_consolidation_job
-
-__all__ = ["run_consolidation_job"]
@@ -1,962 +0,0 @@
-"""Consolidation engine for automatic observation creation from memories.
-
-The consolidation engine runs as a background job after retain operations complete.
-It processes new memories and either:
- Creates new observations from novel facts
- Updates existing observations when new evidence supports/contradicts/refines them
-
-Observations are stored in memory_units with fact_type='observation' and include:
- proof_count: Number of supporting memories
- source_memory_ids: Array of memory UUIDs that contribute to this observation
- history: JSONB tracking changes over time
-"""
-
-import json
-import logging
-import time
-import uuid
-from datetime import datetime, timezone
-from typing import TYPE_CHECKING, Any
-
-from ..memory_engine import fq_table
-from ..retain import embedding_utils
-from .prompts import (
-    CONSOLIDATION_SYSTEM_PROMPT,
-    CONSOLIDATION_USER_PROMPT,
-)
-
-if TYPE_CHECKING:
-    from asyncpg import Connection
-
-    from ...api.http import RequestContext
-    from ..memory_engine import MemoryEngine
-
-logger = logging.getLogger(__name__)
-
-
-class ConsolidationPerfLog:
-    """Performance logging for consolidation operations."""
-
-    def __init__(self, bank_id: str):
-        self.bank_id = bank_id
-        self.start_time = time.time()
-        self.lines: list[str] = []
-        self.timings: dict[str, float] = {}
-
-    def log(self, message: str) -> None:
-        """Add a log line."""
-        self.lines.append(message)
-
-    def record_timing(self, key: str, duration: float) -> None:
-        """Record a timing measurement."""
-        if key in self.timings:
-            self.timings[key] += duration
-        else:
-            self.timings[key] = duration
-
-    def flush(self) -> None:
-        """Flush all log lines to the logger."""
-        total_time = time.time() - self.start_time
-        header = f"\n{'=' * 60}\nCONSOLIDATION for bank {self.bank_id}"
-        footer = f"{'=' * 60}\nCONSOLIDATION COMPLETE: {total_time:.3f}s total\n{'=' * 60}"
-
-        log_output = header + "\n" + "\n".join(self.lines) + "\n" + footer
-        logger.info(log_output)
-
-
-async def run_consolidation_job(
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    request_context: "RequestContext",
-) -> dict[str, Any]:
-    """
-    Run consolidation job for a bank.
-
-    This is called after retain operations to consolidate new memories into mental models.
-
-    Args:
-        memory_engine: MemoryEngine instance
-        bank_id: Bank identifier
-        request_context: Request context for authentication
-
-    Returns:
-        Dict with consolidation results
-    """
-    from ...config import get_config
-
-    config = get_config()
-    perf = ConsolidationPerfLog(bank_id)
-    max_memories_per_batch = config.consolidation_batch_size
-
-    # Check if consolidation is enabled
-    if not config.enable_observations:
-        logger.debug(f"Consolidation disabled for bank {bank_id}")
-        return {"status": "disabled", "bank_id": bank_id}
-
-    pool = memory_engine._pool
-
-    # Get bank profile
-    async with pool.acquire() as conn:
-        t0 = time.time()
-        bank_row = await conn.fetchrow(
-            f"""
-            SELECT bank_id, name, mission
-            FROM {fq_table("banks")}
-            WHERE bank_id = $1
-            """,
-            bank_id,
-        )
-
-        if not bank_row:
-            logger.warning(f"Bank {bank_id} not found for consolidation")
-            return {"status": "bank_not_found", "bank_id": bank_id}
-
-        mission = bank_row["mission"] or "General memory consolidation"
-        perf.record_timing("fetch_bank", time.time() - t0)
-
-        # Count total unconsolidated memories for progress logging
-        total_count = await conn.fetchval(
-            f"""
-            SELECT COUNT(*)
-            FROM {fq_table("memory_units")}
-            WHERE bank_id = $1
-              AND consolidated_at IS NULL
-              AND fact_type IN ('experience', 'world')
-            """,
-            bank_id,
-        )
-
-    if total_count == 0:
-        logger.debug(f"No new memories to consolidate for bank {bank_id}")
-        return {"status": "no_new_memories", "bank_id": bank_id, "memories_processed": 0}
-
-    logger.info(f"[CONSOLIDATION] bank={bank_id} total_unconsolidated={total_count}")
-    perf.log(f"[1] Found {total_count} pending memories to consolidate")
-
-    # Process each memory with individual commits for crash recovery
-    stats = {
-        "memories_processed": 0,
-        "observations_created": 0,
-        "observations_updated": 0,
-        "observations_merged": 0,
-        "actions_executed": 0,
-        "skipped": 0,
-    }
-
-    batch_num = 0
-    last_progress_timings = {}  # Track timings at last progress log
-    while True:
-        batch_num += 1
-        batch_start = time.time()
-
-        # Snapshot timings at batch start for per-batch calculation
-        batch_start_timings = perf.timings.copy()
-
-        # Fetch next batch of unconsolidated memories
-        async with pool.acquire() as conn:
-            t0 = time.time()
-            memories = await conn.fetch(
-                f"""
-                SELECT id, text, fact_type, occurred_start, occurred_end, event_date, tags, mentioned_at
-                FROM {fq_table("memory_units")}
-                WHERE bank_id = $1
-                  AND consolidated_at IS NULL
-                  AND fact_type IN ('experience', 'world')
-                ORDER BY created_at ASC
-                LIMIT $2
-                """,
-                bank_id,
-                max_memories_per_batch,
-            )
-            perf.record_timing("fetch_memories", time.time() - t0)
-
-        if not memories:
-            break  # No more unconsolidated memories
-
-        for memory in memories:
-            mem_start = time.time()
-
-            # Process the memory (uses its own connection internally)
-            async with pool.acquire() as conn:
-                result = await _process_memory(
-                    conn=conn,
-                    memory_engine=memory_engine,
-                    bank_id=bank_id,
-                    memory=dict(memory),
-                    mission=mission,
-                    request_context=request_context,
-                    perf=perf,
-                )
-
-                # Mark memory as consolidated (committed immediately)
-                await conn.execute(
-                    f"""
-                    UPDATE {fq_table("memory_units")}
-                    SET consolidated_at = NOW()
-                    WHERE id = $1
-                    """,
-                    memory["id"],
-                )
-
-            mem_time = time.time() - mem_start
-            perf.record_timing("process_memory_total", mem_time)
-
-            stats["memories_processed"] += 1
-
-            action = result.get("action")
-            if action == "created":
-                stats["observations_created"] += 1
-                stats["actions_executed"] += 1
-            elif action == "updated":
-                stats["observations_updated"] += 1
-                stats["actions_executed"] += 1
-            elif action == "merged":
-                stats["observations_merged"] += 1
-                stats["actions_executed"] += 1
-            elif action == "multiple":
-                stats["observations_created"] += result.get("created", 0)
-                stats["observations_updated"] += result.get("updated", 0)
-                stats["observations_merged"] += result.get("merged", 0)
-                stats["actions_executed"] += result.get("total_actions", 0)
-            elif action == "skipped":
-                stats["skipped"] += 1
-
-            # Log progress periodically with timing breakdown
-            if stats["memories_processed"] % 10 == 0:
-                # Calculate timing deltas since last progress log
-                timing_parts = []
-                for key in ["recall", "llm", "embedding", "db_write"]:
-                    if key in perf.timings:
-                        delta = perf.timings[key] - last_progress_timings.get(key, 0)
-                        timing_parts.append(f"{key}={delta:.2f}s")
-
-                timing_str = f" | {', '.join(timing_parts)}" if timing_parts else ""
-                logger.info(
-                    f"[CONSOLIDATION] bank={bank_id} progress: "
-                    f"{stats['memories_processed']}/{total_count} memories processed{timing_str}"
-                )
-
-                # Update last progress snapshot
-                last_progress_timings = perf.timings.copy()
-
-        batch_time = time.time() - batch_start
-        perf.log(
-            f"[2] Batch {batch_num}: {len(memories)} memories in {batch_time:.3f}s "
-            f"(avg {batch_time / len(memories):.3f}s/memory)"
-        )
-
-        # Log timing breakdown after each batch (delta from batch start)
-        timing_parts = []
-        for key in ["recall", "llm", "embedding", "db_write"]:
-            if key in perf.timings:
-                delta = perf.timings[key] - batch_start_timings.get(key, 0)
-                timing_parts.append(f"{key}={delta:.3f}s")
-
-        if timing_parts:
-            avg_per_memory = batch_time / len(memories) if memories else 0
-            logger.info(
-                f"[CONSOLIDATION] bank={bank_id} batch {batch_num}/{len(memories)} memories: "
-                f"{', '.join(timing_parts)} | avg={avg_per_memory:.3f}s/memory"
-            )
-
-    # Build summary
-    perf.log(
-        f"[3] Results: {stats['memories_processed']} memories -> "
-        f"{stats['actions_executed']} actions "
-        f"({stats['observations_created']} created, "
-        f"{stats['observations_updated']} updated, "
-        f"{stats['observations_merged']} merged, "
-        f"{stats['skipped']} skipped)"
-    )
-
-    # Add timing breakdown
-    timing_parts = []
-    if "recall" in perf.timings:
-        timing_parts.append(f"recall={perf.timings['recall']:.3f}s")
-    if "llm" in perf.timings:
-        timing_parts.append(f"llm={perf.timings['llm']:.3f}s")
-    if "embedding" in perf.timings:
-        timing_parts.append(f"embedding={perf.timings['embedding']:.3f}s")
-    if "db_write" in perf.timings:
-        timing_parts.append(f"db_write={perf.timings['db_write']:.3f}s")
-
-    if timing_parts:
-        perf.log(f"[4] Timing breakdown: {', '.join(timing_parts)}")
-
-    # Trigger mental model refreshes for models with refresh_after_consolidation=true
-    mental_models_refreshed = await _trigger_mental_model_refreshes(
-        memory_engine=memory_engine,
-        bank_id=bank_id,
-        request_context=request_context,
-        perf=perf,
-    )
-    stats["mental_models_refreshed"] = mental_models_refreshed
-
-    perf.flush()
-
-    return {"status": "completed", "bank_id": bank_id, **stats}
-
-
-async def _trigger_mental_model_refreshes(
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    request_context: "RequestContext",
-    perf: ConsolidationPerfLog | None = None,
-) -> int:
-    """
-    Trigger refreshes for mental models with refresh_after_consolidation=true.
-
-    Args:
-        memory_engine: MemoryEngine instance
-        bank_id: Bank identifier
-        request_context: Request context for authentication
-        perf: Performance logging
-
-    Returns:
-        Number of mental models scheduled for refresh
-    """
-    pool = memory_engine._pool
-
-    # Find mental models with refresh_after_consolidation=true
-    async with pool.acquire() as conn:
-        rows = await conn.fetch(
-            f"""
-            SELECT id, name
-            FROM {fq_table("mental_models")}
-            WHERE bank_id = $1
-              AND (trigger->>'refresh_after_consolidation')::boolean = true
-            """,
-            bank_id,
-        )
-
-    if not rows:
-        return 0
-
-    if perf:
-        perf.log(f"[5] Triggering refresh for {len(rows)} mental models with refresh_after_consolidation=true")
-
-    # Submit refresh tasks for each mental model
-    refreshed_count = 0
-    for row in rows:
-        mental_model_id = row["id"]
-        try:
-            await memory_engine.submit_async_refresh_mental_model(
-                bank_id=bank_id,
-                mental_model_id=mental_model_id,
-                request_context=request_context,
-            )
-            refreshed_count += 1
-            logger.info(
-                f"[CONSOLIDATION] Triggered refresh for mental model {mental_model_id} "
-                f"(name: {row['name']}) in bank {bank_id}"
-            )
-        except Exception as e:
-            logger.warning(f"[CONSOLIDATION] Failed to trigger refresh for mental model {mental_model_id}: {e}")
-
-    return refreshed_count
-
-
-async def _process_memory(
-    conn: "Connection",
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    memory: dict[str, Any],
-    mission: str,
-    request_context: "RequestContext",
-    perf: ConsolidationPerfLog | None = None,
-) -> dict[str, Any]:
-    """
-    Process a single memory for consolidation using a SINGLE LLM call.
-
-    This function:
-    1. Finds related observations (can be empty)
-    2. Uses ONE LLM call to extract durable knowledge AND decide on actions
-    3. Executes array of actions (can be multiple creates/updates)
-
-    The LLM handles all cases:
-    - No related observations: returns create action(s) with extracted durable knowledge
-    - Related observations exist: returns update/create actions based on tag routing
-    - Purely ephemeral fact: returns empty array (skip)
-
-    Returns:
-        Dict with action summary: created/updated/merged counts
-    """
-    fact_text = memory["text"]
-    memory_id = memory["id"]
-    fact_tags = memory.get("tags") or []
-
-    # Find related observations using the full recall system (NO tag filtering)
-    t0 = time.time()
-    related_observations = await _find_related_observations(
-        conn=conn,
-        memory_engine=memory_engine,
-        bank_id=bank_id,
-        query=fact_text,
-        request_context=request_context,
-    )
-    if perf:
-        perf.record_timing("recall", time.time() - t0)
-
-    # Single LLM call handles ALL cases (with or without existing observations)
-    # Note: Tags are NOT passed to LLM - they are handled algorithmically
-    t0 = time.time()
-    actions = await _consolidate_with_llm(
-        memory_engine=memory_engine,
-        fact_text=fact_text,
-        observations=related_observations,  # Can be empty list
-        mission=mission,
-    )
-    if perf:
-        perf.record_timing("llm", time.time() - t0)
-
-    if not actions:
-        # LLM returned empty array - fact is purely ephemeral, skip
-        return {"action": "skipped", "reason": "no_durable_knowledge"}
-
-    # Execute all actions and collect results
-    results = []
-    for action in actions:
-        action_type = action.get("action")
-        if action_type == "update":
-            result = await _execute_update_action(
-                conn=conn,
-                memory_engine=memory_engine,
-                bank_id=bank_id,
-                memory_id=memory_id,
-                action=action,
-                observations=related_observations,
-                source_fact_tags=fact_tags,  # Pass source fact's tags for security
-                source_occurred_start=memory.get("occurred_start"),
-                source_occurred_end=memory.get("occurred_end"),
-                source_mentioned_at=memory.get("mentioned_at"),
-                perf=perf,
-            )
-            results.append(result)
-        elif action_type == "create":
-            result = await _execute_create_action(
-                conn=conn,
-                memory_engine=memory_engine,
-                bank_id=bank_id,
-                memory_id=memory_id,
-                action=action,
-                source_fact_tags=fact_tags,  # Pass source fact's tags for security
-                event_date=memory.get("event_date"),
-                occurred_start=memory.get("occurred_start"),
-                occurred_end=memory.get("occurred_end"),
-                mentioned_at=memory.get("mentioned_at"),
-                perf=perf,
-            )
-            results.append(result)
-
-    if not results:
-        # No valid actions executed
-        return {"action": "skipped", "reason": "no_valid_actions"}
-
-    # Summarize results
-    created = sum(1 for r in results if r.get("action") == "created")
-    updated = sum(1 for r in results if r.get("action") == "updated")
-    merged = sum(1 for r in results if r.get("action") == "merged")
-
-    if len(results) == 1:
-        return results[0]
-
-    return {
-        "action": "multiple",
-        "created": created,
-        "updated": updated,
-        "merged": merged,
-        "total_actions": len(results),
-    }
-
-
-async def _execute_update_action(
-    conn: "Connection",
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    memory_id: uuid.UUID,
-    action: dict[str, Any],
-    observations: list[dict[str, Any]],
-    source_fact_tags: list[str] | None = None,
-    source_occurred_start: datetime | None = None,
-    source_occurred_end: datetime | None = None,
-    source_mentioned_at: datetime | None = None,
-    perf: ConsolidationPerfLog | None = None,
-) -> dict[str, Any]:
-    """
-    Execute an update action on an existing observation.
-
-    Updates the observation text, adds to history, increments proof_count,
-    and updates temporal fields:
-    - occurred_start: uses LEAST to keep the earliest start time
-    - occurred_end: uses GREATEST to keep the most recent end time
-    - mentioned_at: uses GREATEST to keep the most recent mention time
-
-    SECURITY: Merges source fact's tags into the observation's existing tags.
-    This ensures all contributors can see the observation they contributed to.
-    For example, if Lisa's observation (tags=['user_lisa']) is updated with
-    Mike's fact (tags=['user_mike']), the observation will have both tags.
-    """
-    learning_id = action.get("learning_id")
-    new_text = action.get("text")
-    reason = action.get("reason", "Updated with new fact")
-
-    if not learning_id or not new_text:
-        return {"action": "skipped", "reason": "missing_learning_id_or_text"}
-
-    # Find the observation
-    model = next((m for m in observations if str(m["id"]) == learning_id), None)
-    if not model:
-        return {"action": "skipped", "reason": "learning_not_found"}
-
-    # Build history entry
-    history = list(model.get("history", []))
-    history.append(
-        {
-            "previous_text": model["text"],
-            "changed_at": datetime.now(timezone.utc).isoformat(),
-            "reason": reason,
-            "source_memory_id": str(memory_id),
-        }
-    )
-
-    # Update source_memory_ids
-    source_ids = list(model.get("source_memory_ids", []))
-    source_ids.append(memory_id)
-
-    # SECURITY: Merge source fact's tags into existing observation tags
-    # This ensures all contributors can see the observation they contributed to
-    existing_tags = set(model.get("tags", []) or [])
-    source_tags = set(source_fact_tags or [])
-    merged_tags = list(existing_tags | source_tags)  # Union of both tag sets
-    if source_tags and source_tags != existing_tags:
-        logger.debug(
-            f"Security: Merging tags for observation {learning_id}: "
-            f"existing={list(existing_tags)}, source={list(source_tags)}, merged={merged_tags}"
-        )
-
-    # Generate new embedding for updated text
-    t0 = time.time()
-    embeddings = await embedding_utils.generate_embeddings_batch(memory_engine.embeddings, [new_text])
-    embedding_str = str(embeddings[0]) if embeddings else None
-    if perf:
-        perf.record_timing("embedding", time.time() - t0)
-
-    # Update the observation
-    # - occurred_start: LEAST keeps the earliest start time across all source facts
-    # - occurred_end: GREATEST keeps the most recent end time across all source facts
-    # - mentioned_at: GREATEST keeps the most recent mention time
-    # - tags: merged from existing + source fact (for visibility)
-    t0 = time.time()
-    await conn.execute(
-        f"""
-        UPDATE {fq_table("memory_units")}
-        SET text = $1,
-            embedding = $2::vector,
-            history = $3,
-            source_memory_ids = $4,
-            proof_count = $5,
-            tags = $10,
-            updated_at = now(),
-            occurred_start = LEAST(occurred_start, COALESCE($7, occurred_start)),
-            occurred_end = GREATEST(occurred_end, COALESCE($8, occurred_end)),
-            mentioned_at = GREATEST(mentioned_at, COALESCE($9, mentioned_at))
-        WHERE id = $6
-        """,
-        new_text,
-        embedding_str,
-        json.dumps(history),
-        source_ids,
-        len(source_ids),
-        uuid.UUID(learning_id),
-        source_occurred_start,
-        source_occurred_end,
-        source_mentioned_at,
-        merged_tags,
-    )
-
-    # Create links from memory to observation
-    await _create_memory_links(conn, memory_id, uuid.UUID(learning_id))
-    if perf:
-        perf.record_timing("db_write", time.time() - t0)
-
-    logger.debug(f"Updated observation {learning_id} with memory {memory_id}")
-
-    return {"action": "updated", "observation_id": learning_id}
-
-
-async def _execute_create_action(
-    conn: "Connection",
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    memory_id: uuid.UUID,
-    action: dict[str, Any],
-    source_fact_tags: list[str] | None = None,
-    event_date: datetime | None = None,
-    occurred_start: datetime | None = None,
-    occurred_end: datetime | None = None,
-    mentioned_at: datetime | None = None,
-    perf: ConsolidationPerfLog | None = None,
-) -> dict[str, Any]:
-    """
-    Execute a create action for a new observation.
-
-    Creates a new observation with the specified text.
-    The text comes directly from the classify LLM - no second LLM call needed.
-
-    Tags are determined algorithmically (not by LLM):
-    - Observations always inherit their source fact's tags
-    - This ensures visibility scope is maintained (security)
-    """
-    text = action.get("text")
-
-    # Tags are determined algorithmically - always use source fact's tags
-    # This ensures private memories create private observations
-    tags = source_fact_tags or []
-
-    if not text:
-        return {"action": "skipped", "reason": "missing_text"}
-
-    # Use text directly from classify - skip the redundant LLM call
-    result = await _create_observation_directly(
-        conn=conn,
-        memory_engine=memory_engine,
-        bank_id=bank_id,
-        source_memory_id=memory_id,
-        observation_text=text,  # Text already processed by classify LLM
-        tags=tags,
-        event_date=event_date,
-        occurred_start=occurred_start,
-        occurred_end=occurred_end,
-        mentioned_at=mentioned_at,
-        perf=perf,
-    )
-
-    logger.debug(f"Created observation {result.get('observation_id')} from memory {memory_id} (tags: {tags})")
-
-    return result
-
-
-async def _create_memory_links(
-    conn: "Connection",
-    memory_id: uuid.UUID,
-    observation_id: uuid.UUID,
-) -> None:
-    """
-    Placeholder for observation link creation.
-
-    Observations do NOT get any memory_links copied from their source facts.
-    Instead, retrieval uses source_memory_ids to traverse:
-    - Entity connections: observation → source_memory_ids → unit_entities
-    - Semantic similarity: observations have their own embeddings
-    - Temporal proximity: observations have their own temporal fields
-
-    This avoids data duplication and ensures observations are always
-    connected via their source facts' relationships.
-
-    The memory_id and observation_id parameters are kept for interface
-    compatibility but no links are created.
-    """
-    # No links are created - observations rely on source_memory_ids for traversal
-    pass
-
-
-async def _find_related_observations(
-    conn: "Connection",
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    query: str,
-    request_context: "RequestContext",
-) -> list[dict[str, Any]]:
-    """
-    Find observations related to the given query using optimized recall.
-
-    IMPORTANT: We do NOT filter by tags here. Consolidation needs to see ALL
-    potentially related observations regardless of scope, so the LLM can
-    decide on tag routing (same scope update vs cross-scope create).
-
-    Uses max_tokens to naturally limit observations (no artificial count limit).
-    Includes source memories with dates for LLM context.
-
-    Returns:
-        List of related observations with their tags, source memories, and dates
-    """
-    # Use recall to find related observations with token budget
-    # max_tokens naturally limits how many observations are returned
-    from ...config import get_config
-
-    config = get_config()
-    recall_result = await memory_engine.recall_async(
-        bank_id=bank_id,
-        query=query,
-        max_tokens=config.consolidation_max_tokens,  # Token budget for observations (configurable)
-        fact_type=["observation"],  # Only retrieve observations
-        request_context=request_context,
-        _quiet=True,  # Suppress logging
-        # NO tags parameter - intentionally get ALL observations
-    )
-
-    # If no observations returned, return empty list
-    if not recall_result.results:
-        return []
-
-    # Batch fetch all observations in a single query (no artificial limit)
-    observation_ids = [uuid.UUID(obs.id) for obs in recall_result.results]
-
-    rows = await conn.fetch(
-        f"""
-        SELECT id, text, proof_count, history, tags, source_memory_ids, created_at, updated_at,
-               occurred_start, occurred_end, mentioned_at
-        FROM {fq_table("memory_units")}
-        WHERE id = ANY($1) AND bank_id = $2 AND fact_type = 'observation'
-        """,
-        observation_ids,
-        bank_id,
-    )
-
-    # Build results list preserving recall order
-    id_to_row = {row["id"]: row for row in rows}
-    results = []
-
-    for obs in recall_result.results:
-        obs_id = uuid.UUID(obs.id)
-        if obs_id not in id_to_row:
-            continue
-
-        row = id_to_row[obs_id]
-        history = row["history"]
-        if isinstance(history, str):
-            history = json.loads(history)
-        elif history is None:
-            history = []
-
-        # Fetch source memories to include their text and dates
-        source_memory_ids = row["source_memory_ids"] or []
-        source_memories = []
-
-        if source_memory_ids:
-            source_rows = await conn.fetch(
-                f"""
-                SELECT text, occurred_start, occurred_end, mentioned_at, event_date
-                FROM {fq_table("memory_units")}
-                WHERE id = ANY($1) AND bank_id = $2
-                ORDER BY created_at ASC
-                LIMIT 5
-                """,
-                source_memory_ids[:5],  # Limit to first 5 source memories for token efficiency
-                bank_id,
-            )
-
-            for src_row in source_rows:
-                source_memories.append(
-                    {
-                        "text": src_row["text"],
-                        "occurred_start": src_row["occurred_start"],
-                        "occurred_end": src_row["occurred_end"],
-                        "mentioned_at": src_row["mentioned_at"],
-                        "event_date": src_row["event_date"],
-                    }
-                )
-
-        results.append(
-            {
-                "id": row["id"],
-                "text": row["text"],
-                "proof_count": row["proof_count"] or 1,
-                "tags": row["tags"] or [],
-                "source_memories": source_memories,
-                "occurred_start": row["occurred_start"],
-                "occurred_end": row["occurred_end"],
-                "mentioned_at": row["mentioned_at"],
-                "created_at": row["created_at"],
-                "updated_at": row["updated_at"],
-            }
-        )
-
-    return results
-
-
-async def _consolidate_with_llm(
-    memory_engine: "MemoryEngine",
-    fact_text: str,
-    observations: list[dict[str, Any]],
-    mission: str,
-) -> list[dict[str, Any]]:
-    """
-    Single LLM call to extract durable knowledge and decide on consolidation actions.
-
-    This handles ALL cases:
-    - No related observations: extracts durable knowledge, returns create action
-    - Related observations exist: compares and returns update/create actions
-    - Purely ephemeral fact: returns empty array
-
-    Note: Tags are NOT handled by the LLM. They are determined algorithmically:
-    - CREATE: observation inherits source fact's tags
-    - UPDATE: observation merges source fact's tags with existing tags
-
-    Returns:
-        List of actions, each being:
-        - {"action": "update", "learning_id": "uuid", "text": "...", "reason": "..."}
-        - {"action": "create", "text": "...", "reason": "..."}
-        - [] if fact is purely ephemeral (no durable knowledge)
-    """
-    # Format observations as JSON with source memories and dates
-    if observations:
-        obs_list = []
-        for obs in observations:
-            obs_data = {
-                "id": str(obs["id"]),
-                "text": obs["text"],
-                "proof_count": obs["proof_count"],
-                "tags": obs["tags"],
-                "created_at": obs["created_at"].isoformat() if obs.get("created_at") else None,
-                "updated_at": obs["updated_at"].isoformat() if obs.get("updated_at") else None,
-            }
-
-            # Include temporal info if available
-            if obs.get("occurred_start"):
-                obs_data["occurred_start"] = obs["occurred_start"].isoformat()
-            if obs.get("occurred_end"):
-                obs_data["occurred_end"] = obs["occurred_end"].isoformat()
-            if obs.get("mentioned_at"):
-                obs_data["mentioned_at"] = obs["mentioned_at"].isoformat()
-
-            # Include source memories (up to 3 for brevity)
-            if obs.get("source_memories"):
-                obs_data["source_memories"] = [
-                    {
-                        "text": sm["text"],
-                        "event_date": sm["event_date"].isoformat() if sm.get("event_date") else None,
-                        "occurred_start": sm["occurred_start"].isoformat() if sm.get("occurred_start") else None,
-                    }
-                    for sm in obs["source_memories"][:3]  # Limit to 3 for token efficiency
-                ]
-
-            obs_list.append(obs_data)
-
-        observations_text = json.dumps(obs_list, indent=2)
-    else:
-        observations_text = "[]"
-
-    # Only include mission section if mission is set and not the default
-    mission_section = ""
-    if mission and mission != "General memory consolidation":
-        mission_section = f"""
-MISSION CONTEXT: {mission}
-
-Focus on DURABLE knowledge that serves this mission, not ephemeral state.
-"""
-
-    user_prompt = CONSOLIDATION_USER_PROMPT.format(
-        mission_section=mission_section,
-        fact_text=fact_text,
-        observations_text=observations_text,
-    )
-
-    messages = [
-        {"role": "system", "content": CONSOLIDATION_SYSTEM_PROMPT},
-        {"role": "user", "content": user_prompt},
-    ]
-
-    try:
-        result = await memory_engine._consolidation_llm_config.call(
-            messages=messages,
-            skip_validation=True,  # Raw JSON response
-            scope="consolidation",
-        )
-        # Parse JSON response - should be an array
-        if isinstance(result, str):
-            # Strip markdown code fences (some models wrap JSON in ```json ... ```)
-            clean = result.strip()
-            if clean.startswith("```"):
-                clean = clean.split("\n", 1)[1] if "\n" in clean else clean[3:]
-                if clean.endswith("```"):
-                    clean = clean[:-3]
-                clean = clean.strip()
-            result = json.loads(clean)
-        # Ensure result is a list
-        if isinstance(result, list):
-            return result
-        # Handle legacy single-action format for backward compatibility
-        if isinstance(result, dict):
-            if result.get("related_ids") and result.get("consolidated_text"):
-                # Convert old format to new format
-                return [
-                    {
-                        "action": "update",
-                        "learning_id": result["related_ids"][0],
-                        "text": result["consolidated_text"],
-                        "reason": result.get("reason", ""),
-                    }
-                ]
-            return []
-        return []
-    except Exception as e:
-        logger.warning(f"Error in consolidation LLM call: {e}")
-        return []
-
-
-async def _create_observation_directly(
-    conn: "Connection",
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    source_memory_id: uuid.UUID,
-    observation_text: str,
-    tags: list[str] | None = None,
-    event_date: datetime | None = None,
-    occurred_start: datetime | None = None,
-    occurred_end: datetime | None = None,
-    mentioned_at: datetime | None = None,
-    perf: ConsolidationPerfLog | None = None,
-) -> dict[str, Any]:
-    """
-    Create an observation directly with pre-processed text (no LLM call).
-
-    Used when the classify LLM has already provided the learning text.
-    This avoids the redundant second LLM call.
-    """
-    # Generate embedding for the observation (convert to string for pgvector)
-    t0 = time.time()
-    embeddings = await embedding_utils.generate_embeddings_batch(memory_engine.embeddings, [observation_text])
-    embedding_str = str(embeddings[0]) if embeddings else None
-    if perf:
-        perf.record_timing("embedding", time.time() - t0)
-
-    # Create the observation as a memory_unit
-    now = datetime.now(timezone.utc)
-    obs_event_date = event_date or now
-    obs_occurred_start = occurred_start or now
-    obs_occurred_end = occurred_end or now
-    obs_mentioned_at = mentioned_at or now
-    obs_tags = tags or []
-
-    t0 = time.time()
-    observation_id = uuid.uuid4()
-    row = await conn.fetchrow(
-        f"""
-        INSERT INTO {fq_table("memory_units")} (
-            id, bank_id, text, fact_type, embedding, proof_count, source_memory_ids, history,
-            tags, event_date, occurred_start, occurred_end, mentioned_at
-        )
-        VALUES ($1, $2, $3, 'observation', $4::vector, 1, $5, '[]'::jsonb, $6, $7, $8, $9, $10)
-        RETURNING id
-        """,
-        observation_id,
-        bank_id,
-        observation_text,
-        embedding_str,
-        [source_memory_id],
-        obs_tags,
-        obs_event_date,
-        obs_occurred_start,
-        obs_occurred_end,
-        obs_mentioned_at,
-    )
-
-    # Create links between memory and observation (includes entity links, memory_links)
-    await _create_memory_links(conn, source_memory_id, observation_id)
-    if perf:
-        perf.record_timing("db_write", time.time() - t0)
-
-    logger.debug(f"Created observation {observation_id} from memory {source_memory_id} (tags: {obs_tags})")
-
-    return {"action": "created", "observation_id": str(row["id"]), "tags": obs_tags}
@@ -1,77 +0,0 @@
-"""Prompts for the consolidation engine."""
-
-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.
-
-You must output ONLY valid JSON with no markdown formatting, no code blocks, and no additional 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.
-
-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 history (e.g., "used to X, now Y")
-3. UPDATE: New state replacing old state → update with history
-
-## CRITICAL RULES:
- NEVER merge facts about DIFFERENT people
- NEVER merge unrelated topics (food preferences vs work vs hobbies)
- When merging contradictions, capture the CHANGE (before → after)
- 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
-
-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 JSON array of actions:
-[
-  {{"action": "update", "learning_id": "uuid-from-observations", "text": "updated knowledge", "reason": "..."}},
-  {{"action": "create", "text": "new durable knowledge", "reason": "..."}}
-]
-
-Return [] if fact contains no durable knowledge."""
@@ -6,40 +6,17 @@ Provides an interface for reranking with different backends.
 Configuration via environment variables - see hindsight_api.config for all env var names.
 """

-import asyncio
 import logging
 import os
 from abc import ABC, abstractmethod
-from concurrent.futures import ThreadPoolExecutor

 import httpx

 from ..config import (
-    DEFAULT_LITELLM_API_BASE,
-    DEFAULT_RERANKER_COHERE_MODEL,
-    DEFAULT_RERANKER_FLASHRANK_CACHE_DIR,
-    DEFAULT_RERANKER_FLASHRANK_MODEL,
-    DEFAULT_RERANKER_LITELLM_MODEL,
-    DEFAULT_RERANKER_LOCAL_FORCE_CPU,
-    DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT,
    DEFAULT_RERANKER_LOCAL_MODEL,
    DEFAULT_RERANKER_PROVIDER,
-    DEFAULT_RERANKER_TEI_BATCH_SIZE,
-    DEFAULT_RERANKER_TEI_MAX_CONCURRENT,
-    ENV_COHERE_API_KEY,
-    ENV_LITELLM_API_BASE,
-    ENV_LITELLM_API_KEY,
-    ENV_RERANKER_COHERE_BASE_URL,
-    ENV_RERANKER_COHERE_MODEL,
-    ENV_RERANKER_FLASHRANK_CACHE_DIR,
-    ENV_RERANKER_FLASHRANK_MODEL,
-    ENV_RERANKER_LITELLM_MODEL,
-    ENV_RERANKER_LOCAL_FORCE_CPU,
-    ENV_RERANKER_LOCAL_MAX_CONCURRENT,
    ENV_RERANKER_LOCAL_MODEL,
    ENV_RERANKER_PROVIDER,
-    ENV_RERANKER_TEI_BATCH_SIZE,
-    ENV_RERANKER_TEI_MAX_CONCURRENT,
    ENV_RERANKER_TEI_URL,
 )

@@ -70,7 +47,7 @@ class CrossEncoderModel(ABC):
        pass

    @abstractmethod
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+    def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
        """
        Score query-document pairs for relevance.

@@ -93,37 +70,25 @@ class LocalSTCrossEncoder(CrossEncoderModel):
    - Fast inference (~80ms for 100 pairs on CPU)
    - Small model (80MB)
    - Trained for passage re-ranking
-
-    Uses a dedicated thread pool to limit concurrent CPU-bound work.
    """

-    # Shared executor across all instances (one model loaded anyway)
-    _executor: ThreadPoolExecutor | None = None
-    _max_concurrent: int = 4  # Limit concurrent CPU-bound reranking calls
-
-    def __init__(self, model_name: str | None = None, max_concurrent: int = 4, force_cpu: bool = False):
+    def __init__(self, model_name: str | None = None):
        """
        Initialize local SentenceTransformers cross-encoder.

        Args:
            model_name: Name of the CrossEncoder model to use.
                       Default: cross-encoder/ms-marco-MiniLM-L-6-v2
-            max_concurrent: Maximum concurrent reranking calls (default: 2).
-                           Higher values may cause CPU thrashing under load.
-            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
-                      Default: False
        """
        self.model_name = model_name or DEFAULT_RERANKER_LOCAL_MODEL
-        self.force_cpu = force_cpu
        self._model = None
-        LocalSTCrossEncoder._max_concurrent = max_concurrent

    @property
    def provider_name(self) -> str:
        return "local"

    async def initialize(self) -> None:
-        """Load the cross-encoder model and initialize the executor."""
+        """Load the cross-encoder model."""
        if self._model is not None:
            return

@@ -136,59 +101,13 @@ class LocalSTCrossEncoder(CrossEncoderModel):
            )

        logger.info(f"Reranker: initializing local provider with model {self.model_name}")
+        self._model = CrossEncoder(self.model_name)
+        logger.info("Reranker: local provider initialized")

-        # Determine device based on hardware availability.
-        # We always set low_cpu_mem_usage=False to prevent lazy loading (meta tensors)
-        # which can cause issues when accelerate is installed but no GPU is available.
-        # Note: We do NOT use device_map because CrossEncoder internally calls .to(device)
-        # after loading, which conflicts with accelerate's device_map handling.
-        import torch
-
-        # Force CPU mode if configured (used in daemon mode to avoid MPS/XPC issues on macOS)
-        if self.force_cpu:
-            device = "cpu"
-            logger.info("Reranker: forcing CPU mode (HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU=1)")
-        else:
-            # Check for GPU (CUDA) or Apple Silicon (MPS)
-            # Wrap in try-except to gracefully handle any device detection issues
-            # (e.g., in CI environments or when PyTorch is built without GPU support)
-            device = "cpu"  # Default to CPU
-            try:
-                has_gpu = torch.cuda.is_available() or (
-                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
-                )
-                if has_gpu:
-                    device = None  # Let sentence-transformers auto-detect GPU/MPS
-            except Exception as e:
-                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")
-
-        self._model = CrossEncoder(
-            self.model_name,
-            device=device,
-            model_kwargs={"low_cpu_mem_usage": False},
-        )
-
-        # Initialize shared executor (limited workers naturally limits concurrency)
-        if LocalSTCrossEncoder._executor is None:
-            LocalSTCrossEncoder._executor = ThreadPoolExecutor(
-                max_workers=LocalSTCrossEncoder._max_concurrent,
-                thread_name_prefix="reranker",
-            )
-            logger.info(f"Reranker: local provider initialized (max_concurrent={LocalSTCrossEncoder._max_concurrent})")
-        else:
-            logger.info("Reranker: local provider initialized (using existing executor)")
-
-    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """Synchronous prediction wrapper for thread pool execution."""
-        scores = self._model.predict(pairs, show_progress_bar=False)
-        return scores.tolist() if hasattr(scores, "tolist") else list(scores)
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+    def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
        """
        Score query-document pairs for relevance.

-        Uses a dedicated thread pool with limited workers to prevent CPU thrashing.
-
        Args:
            pairs: List of (query, document) tuples to score

@@ -197,14 +116,8 @@ class LocalSTCrossEncoder(CrossEncoderModel):
        """
        if self._model is None:
            raise RuntimeError("Reranker not initialized. Call initialize() first.")
-
-        # Use dedicated executor - limited workers naturally limits concurrency
-        loop = asyncio.get_event_loop()
-        return await loop.run_in_executor(
-            LocalSTCrossEncoder._executor,
-            self._predict_sync,
-            pairs,
-        )
+        scores = self._model.predict(pairs, show_progress_bar=False)
+        return scores.tolist() if hasattr(scores, "tolist") else list(scores)


 class RemoteTEICrossEncoder(CrossEncoderModel):
@@ -215,21 +128,13 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
    See: https://github.com/huggingface/text-embeddings-inference

    Note: The TEI server must be running a cross-encoder/reranker model.
-
-    Requests are made in parallel with configurable batch size and max concurrency (backpressure).
-    Uses a GLOBAL semaphore to limit concurrent requests across ALL recall operations.
    """

-    # Global semaphore shared across all instances and calls to prevent thundering herd
-    _global_semaphore: asyncio.Semaphore | None = None
-    _global_max_concurrent: int = DEFAULT_RERANKER_TEI_MAX_CONCURRENT
-
    def __init__(
        self,
        base_url: str,
        timeout: float = 30.0,
-        batch_size: int = DEFAULT_RERANKER_TEI_BATCH_SIZE,
-        max_concurrent: int = DEFAULT_RERANKER_TEI_MAX_CONCURRENT,
+        batch_size: int = 32,
        max_retries: int = 3,
        retry_delay: float = 0.5,
    ):
@@ -239,246 +144,80 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
        Args:
            base_url: Base URL of the TEI server (e.g., "http://localhost:8080")
            timeout: Request timeout in seconds (default: 30.0)
-            batch_size: Maximum batch size for rerank requests (default: 128)
-            max_concurrent: Maximum concurrent requests for backpressure (default: 8).
-                           This is a GLOBAL limit across all parallel recall operations.
+            batch_size: Maximum batch size for rerank requests (default: 32)
            max_retries: Maximum number of retries for failed requests (default: 3)
            retry_delay: Initial delay between retries in seconds, doubles each retry (default: 0.5)
        """
        self.base_url = base_url.rstrip("/")
        self.timeout = timeout
        self.batch_size = batch_size
-        self.max_concurrent = max_concurrent
        self.max_retries = max_retries
        self.retry_delay = retry_delay
-        self._async_client: httpx.AsyncClient | None = None
+        self._client: httpx.Client | None = None
        self._model_id: str | None = None

-        # Update global semaphore if max_concurrent changed
-        if (
-            RemoteTEICrossEncoder._global_semaphore is None
-            or RemoteTEICrossEncoder._global_max_concurrent != max_concurrent
-        ):
-            RemoteTEICrossEncoder._global_max_concurrent = max_concurrent
-            RemoteTEICrossEncoder._global_semaphore = asyncio.Semaphore(max_concurrent)
-
    @property
    def provider_name(self) -> str:
        return "tei"

-    async def _async_request_with_retry(
-        self,
-        client: httpx.AsyncClient,
-        semaphore: asyncio.Semaphore,
-        method: str,
-        url: str,
-        **kwargs,
-    ) -> httpx.Response:
-        """Make an async HTTP request with automatic retries on transient errors and semaphore for backpressure."""
+    def _request_with_retry(self, method: str, url: str, **kwargs) -> httpx.Response:
+        """Make an HTTP request with automatic retries on transient errors."""
+        import time
+
        last_error = None
        delay = self.retry_delay

-        async with semaphore:
-            for attempt in range(self.max_retries + 1):
-                try:
-                    if method == "GET":
-                        response = await client.get(url, **kwargs)
-                    else:
-                        response = await client.post(url, **kwargs)
-                    response.raise_for_status()
-                    return response
-                except (httpx.ConnectError, httpx.ReadTimeout, httpx.WriteTimeout) as e:
+        for attempt in range(self.max_retries + 1):
+            try:
+                if method == "GET":
+                    response = self._client.get(url, **kwargs)
+                else:
+                    response = self._client.post(url, **kwargs)
+                response.raise_for_status()
+                return response
+            except (httpx.ConnectError, httpx.ReadTimeout, httpx.WriteTimeout) as e:
+                last_error = e
+                if attempt < self.max_retries:
+                    logger.warning(
+                        f"TEI request failed (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s..."
+                    )
+                    time.sleep(delay)
+                    delay *= 2  # Exponential backoff
+            except httpx.HTTPStatusError as e:
+                # Retry on 5xx server errors
+                if e.response.status_code >= 500 and attempt < self.max_retries:
                    last_error = e
-                    if attempt < self.max_retries:
-                        logger.warning(
-                            f"TEI request failed (attempt {attempt + 1}/{self.max_retries + 1}): {e}. "
-                            f"Retrying in {delay}s..."
-                        )
-                        await asyncio.sleep(delay)
-                        delay *= 2  # Exponential backoff
-                except httpx.HTTPStatusError as e:
-                    # Retry on 5xx server errors
-                    if e.response.status_code >= 500 and attempt < self.max_retries:
-                        last_error = e
-                        logger.warning(
-                            f"TEI server error (attempt {attempt + 1}/{self.max_retries + 1}): {e}. "
-                            f"Retrying in {delay}s..."
-                        )
-                        await asyncio.sleep(delay)
-                        delay *= 2
-                    else:
-                        raise
+                    logger.warning(
+                        f"TEI server error (attempt {attempt + 1}/{self.max_retries + 1}): {e}. Retrying in {delay}s..."
+                    )
+                    time.sleep(delay)
+                    delay *= 2
+                else:
+                    raise

        raise last_error

    async def initialize(self) -> None:
        """Initialize the HTTP client and verify server connectivity."""
-        if self._async_client is not None:
+        if self._client is not None:
            return

-        logger.info(
-            f"Reranker: initializing TEI provider at {self.base_url} "
-            f"(batch_size={self.batch_size}, max_concurrent={self.max_concurrent})"
-        )
-        self._async_client = httpx.AsyncClient(timeout=self.timeout)
+        logger.info(f"Reranker: initializing TEI provider at {self.base_url}")
+        self._client = httpx.Client(timeout=self.timeout)

        # Verify server is reachable and get model info
-        # Use a temporary semaphore for initialization
-        init_semaphore = asyncio.Semaphore(1)
        try:
-            response = await self._async_request_with_retry(
-                self._async_client, init_semaphore, "GET", f"{self.base_url}/info"
-            )
+            response = self._request_with_retry("GET", f"{self.base_url}/info")
            info = response.json()
            self._model_id = info.get("model_id", "unknown")
            logger.info(f"Reranker: TEI provider initialized (model: {self._model_id})")
        except httpx.HTTPError as e:
-            self._async_client = None
            raise RuntimeError(f"Failed to connect to TEI server at {self.base_url}: {e}")

-    async def _rerank_query_group(
-        self,
-        client: httpx.AsyncClient,
-        semaphore: asyncio.Semaphore,
-        query: str,
-        texts: list[str],
-    ) -> list[tuple[int, float]]:
-        """Rerank a single query group and return list of (original_index, score) tuples."""
-        try:
-            response = await self._async_request_with_retry(
-                client,
-                semaphore,
-                "POST",
-                f"{self.base_url}/rerank",
-                json={
-                    "query": query,
-                    "texts": texts,
-                    "return_text": False,
-                },
-            )
-            results = response.json()
-            # TEI returns results sorted by score descending, with original index
-            return [(result["index"], result["score"]) for result in results]
-        except httpx.HTTPError as e:
-            raise RuntimeError(f"TEI rerank request failed: {e}")
-
-    async def _predict_async(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """Async implementation of predict that runs requests in parallel with backpressure."""
-        if not pairs:
-            return []
-
-        # Group all pairs by query
-        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))
-
-        # Split each query group into batches
-        tasks_info: list[tuple[str, list[int], list[str]]] = []  # (query, indices, texts)
-        for query, indexed_texts in query_groups.items():
-            indices = [idx for idx, _ in indexed_texts]
-            texts = [text for _, text in indexed_texts]
-
-            # Split into batches
-            for i in range(0, len(texts), self.batch_size):
-                batch_indices = indices[i : i + self.batch_size]
-                batch_texts = texts[i : i + self.batch_size]
-                tasks_info.append((query, batch_indices, batch_texts))
-
-        # Run all requests in parallel with GLOBAL semaphore for backpressure
-        # This ensures max_concurrent is respected across ALL parallel recall operations
-        all_scores = [0.0] * len(pairs)
-        semaphore = RemoteTEICrossEncoder._global_semaphore
-
-        tasks = [
-            self._rerank_query_group(self._async_client, semaphore, query, texts) for query, _, texts in tasks_info
-        ]
-        results = await asyncio.gather(*tasks)
-
-        # Map scores back to original positions
-        for (_, indices, _), result_scores in zip(tasks_info, results):
-            for original_idx_in_batch, score in result_scores:
-                global_idx = indices[original_idx_in_batch]
-                all_scores[global_idx] = score
-
-        return all_scores
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+    def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
        """
        Score query-document pairs using the remote TEI reranker.

-        Requests are made in parallel with configurable backpressure.
-
-        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.")
-
-        return await self._predict_async(pairs)
-
-
-class CohereCrossEncoder(CrossEncoderModel):
-    """
-    Cohere cross-encoder implementation using the Cohere Rerank API.
-
-    Supports rerank-english-v3.0 and rerank-multilingual-v3.0 models.
-    """
-
-    def __init__(
-        self,
-        api_key: str,
-        model: str = DEFAULT_RERANKER_COHERE_MODEL,
-        base_url: str | None = None,
-        timeout: float = 60.0,
-    ):
-        """
-        Initialize Cohere cross-encoder client.
-
-        Args:
-            api_key: Cohere API key
-            model: Cohere rerank model name (default: rerank-english-v3.0)
-            base_url: Custom base URL for Cohere-compatible API (e.g., Azure-hosted endpoint)
-            timeout: Request timeout in seconds (default: 60.0)
-        """
-        self.api_key = api_key
-        self.model = model
-        self.base_url = base_url
-        self.timeout = timeout
-        self._client = None
-
-    @property
-    def provider_name(self) -> str:
-        return "cohere"
-
-    async def initialize(self) -> None:
-        """Initialize the Cohere client."""
-        if self._client is not None:
-            return
-
-        try:
-            import cohere
-        except ImportError:
-            raise ImportError("cohere is required for CohereCrossEncoder. Install it with: pip install cohere")
-
-        base_url_msg = f" at {self.base_url}" if self.base_url else ""
-        logger.info(f"Reranker: initializing Cohere provider with model {self.model}{base_url_msg}")
-
-        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
-        client_kwargs = {"api_key": self.api_key, "timeout": self.timeout}
-        if self.base_url:
-            client_kwargs["base_url"] = self.base_url
-        self._client = cohere.Client(**client_kwargs)
-        logger.info("Reranker: Cohere provider initialized")
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """
-        Score query-document pairs using the Cohere Rerank API.
-
        Args:
            pairs: List of (query, document) tuples to score

@@ -491,364 +230,73 @@ class CohereCrossEncoder(CrossEncoderModel):
        if not pairs:
            return []

-        # Run sync Cohere API calls in thread pool
-        loop = asyncio.get_event_loop()
-        return await loop.run_in_executor(None, self._predict_sync, pairs)
+        all_scores = []

-    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """Synchronous predict implementation for Cohere API."""
-        # Group pairs by query for efficient batching
-        # Cohere 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))
+        # Process in batches
+        for i in range(0, len(pairs), self.batch_size):
+            batch = pairs[i : i + self.batch_size]

-        all_scores = [0.0] * len(pairs)
+            # TEI rerank endpoint expects query and texts separately
+            # All pairs in a batch should have the same query for optimal performance
+            # but we handle mixed queries by making separate requests per unique query
+            query_groups: dict[str, list[tuple[int, str]]] = {}
+            for idx, (query, text) in enumerate(batch):
+                if query not in query_groups:
+                    query_groups[query] = []
+                query_groups[query].append((idx, text))

-        for query, indexed_texts in query_groups.items():
-            texts = [text for _, text in indexed_texts]
-            indices = [idx for idx, _ in indexed_texts]
+            batch_scores = [0.0] * len(batch)

-            response = self._client.rerank(
-                query=query,
-                documents=texts,
-                model=self.model,
-                return_documents=False,
-            )
+            for query, indexed_texts in query_groups.items():
+                texts = [text for _, text in indexed_texts]
+                indices = [idx for idx, _ in indexed_texts]

-            # Map scores back to original positions
-            for result in response.results:
-                original_idx = result.index
-                score = result.relevance_score
-                all_scores[indices[original_idx]] = score
+                try:
+                    response = self._request_with_retry(
+                        "POST",
+                        f"{self.base_url}/rerank",
+                        json={
+                            "query": query,
+                            "texts": texts,
+                            "return_text": False,
+                        },
+                    )
+                    results = response.json()

-        return all_scores
+                    # TEI returns results sorted by score descending, with original index
+                    for result in results:
+                        original_idx = result["index"]
+                        score = result["score"]
+                        # Map back to batch position
+                        batch_scores[indices[original_idx]] = score

+                except httpx.HTTPError as e:
+                    raise RuntimeError(f"TEI rerank request failed: {e}")

-class RRFPassthroughCrossEncoder(CrossEncoderModel):
-    """
-    Passthrough cross-encoder that preserves RRF scores without neural reranking.
-
-    This is useful for:
-    - Testing retrieval quality without reranking overhead
-    - Deployments where reranking latency is unacceptable
-    - Debugging to isolate retrieval vs reranking issues
-    """
-
-    def __init__(self):
-        """Initialize RRF passthrough cross-encoder."""
-        pass
-
-    @property
-    def provider_name(self) -> str:
-        return "rrf"
-
-    async def initialize(self) -> None:
-        """No initialization needed."""
-        logger.info("Reranker: RRF passthrough provider initialized (neural reranking disabled)")
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """
-        Return neutral scores - actual ranking uses RRF scores from retrieval.
-
-        Args:
-            pairs: List of (query, document) tuples (ignored)
-
-        Returns:
-            List of 0.5 scores (neutral, lets RRF scores dominate)
-        """
-        # Return neutral scores so RRF ranking is preserved
-        return [0.5] * len(pairs)
-
-
-class FlashRankCrossEncoder(CrossEncoderModel):
-    """
-    FlashRank cross-encoder implementation.
-
-    FlashRank is an ultra-lite reranking library that runs on CPU without
-    requiring PyTorch or Transformers. It's ideal for serverless deployments
-    with minimal cold-start overhead.
-
-    Available models:
-    - ms-marco-TinyBERT-L-2-v2: Fastest, ~4MB
-    - ms-marco-MiniLM-L-12-v2: Best quality, ~34MB (default)
-    - rank-T5-flan: Best zero-shot, ~110MB
-    - ms-marco-MultiBERT-L-12: Multi-lingual, ~150MB
-    """
-
-    # Shared executor for CPU-bound reranking
-    _executor: ThreadPoolExecutor | None = None
-    _max_concurrent: int = 4
-
-    def __init__(
-        self,
-        model_name: str | None = None,
-        cache_dir: str | None = None,
-        max_length: int = 512,
-        max_concurrent: int = 4,
-    ):
-        """
-        Initialize FlashRank cross-encoder.
-
-        Args:
-            model_name: FlashRank model name. Default: ms-marco-MiniLM-L-12-v2
-            cache_dir: Directory to cache downloaded models. Default: system cache
-            max_length: Maximum sequence length for reranking. Default: 512
-            max_concurrent: Maximum concurrent reranking calls. Default: 4
-        """
-        self.model_name = model_name or DEFAULT_RERANKER_FLASHRANK_MODEL
-        self.cache_dir = cache_dir or DEFAULT_RERANKER_FLASHRANK_CACHE_DIR
-        self.max_length = max_length
-        self._ranker = None
-        FlashRankCrossEncoder._max_concurrent = max_concurrent
-
-    @property
-    def provider_name(self) -> str:
-        return "flashrank"
-
-    async def initialize(self) -> None:
-        """Load the FlashRank model."""
-        if self._ranker is not None:
-            return
-
-        try:
-            from flashrank import Ranker
-        except ImportError:
-            raise ImportError("flashrank is required for FlashRankCrossEncoder. Install it with: pip install flashrank")
-
-        logger.info(f"Reranker: initializing FlashRank provider with model {self.model_name}")
-
-        # Initialize ranker with optional cache directory
-        ranker_kwargs = {"model_name": self.model_name, "max_length": self.max_length}
-        if self.cache_dir:
-            ranker_kwargs["cache_dir"] = self.cache_dir
-
-        self._ranker = Ranker(**ranker_kwargs)
-
-        # Initialize shared executor
-        if FlashRankCrossEncoder._executor is None:
-            FlashRankCrossEncoder._executor = ThreadPoolExecutor(
-                max_workers=FlashRankCrossEncoder._max_concurrent,
-                thread_name_prefix="flashrank",
-            )
-            logger.info(
-                f"Reranker: FlashRank provider initialized (max_concurrent={FlashRankCrossEncoder._max_concurrent})"
-            )
-        else:
-            logger.info("Reranker: FlashRank provider initialized (using existing executor)")
-
-    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """Synchronous predict - processes each query group."""
-        from flashrank import RerankRequest
-
-        if not pairs:
-            return []
-
-        # Group pairs by query
-        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():
-            # Build passages list for FlashRank
-            passages = [{"id": i, "text": text} for i, (_, text) in enumerate(indexed_texts)]
-            global_indices = [idx for idx, _ in indexed_texts]
-
-            # Create rerank request
-            request = RerankRequest(query=query, passages=passages)
-            results = self._ranker.rerank(request)
-
-            # Map scores back to original positions
-            for result in results:
-                local_idx = result["id"]
-                score = result["score"]
-                global_idx = global_indices[local_idx]
-                all_scores[global_idx] = score
-
-        return all_scores
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """
-        Score query-document pairs using FlashRank.
-
-        Args:
-            pairs: List of (query, document) tuples to score
-
-        Returns:
-            List of relevance scores (higher = more relevant)
-        """
-        if self._ranker is None:
-            raise RuntimeError("Reranker not initialized. Call initialize() first.")
-
-        # Run in thread pool to avoid blocking event loop
-        loop = asyncio.get_event_loop()
-        return await loop.run_in_executor(FlashRankCrossEncoder._executor, self._predict_sync, pairs)
-
-
-class LiteLLMCrossEncoder(CrossEncoderModel):
-    """
-    LiteLLM cross-encoder implementation using LiteLLM proxy's /rerank endpoint.
-
-    LiteLLM provides a unified interface for multiple reranking providers via
-    the Cohere-compatible /rerank endpoint.
-    See: https://docs.litellm.ai/docs/rerank
-
-    Supported providers via LiteLLM:
-    - Cohere (rerank-english-v3.0, etc.) - prefix with cohere/
-    - Together AI - prefix with together_ai/
-    - Azure AI - prefix with azure_ai/
-    - Jina AI - prefix with jina_ai/
-    - AWS Bedrock - prefix with bedrock/
-    - Voyage AI - prefix with voyage/
-    """
-
-    def __init__(
-        self,
-        api_base: str = DEFAULT_LITELLM_API_BASE,
-        api_key: str | None = None,
-        model: str = DEFAULT_RERANKER_LITELLM_MODEL,
-        timeout: float = 60.0,
-    ):
-        """
-        Initialize LiteLLM cross-encoder client.
-
-        Args:
-            api_base: Base URL of the LiteLLM proxy (default: http://localhost:4000)
-            api_key: API key for the LiteLLM proxy (optional, depends on proxy config)
-            model: Reranking model name (default: cohere/rerank-english-v3.0)
-                   Use provider prefix (e.g., cohere/, together_ai/, voyage/)
-            timeout: Request timeout in seconds (default: 60.0)
-        """
-        self.api_base = api_base.rstrip("/")
-        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 "litellm"
-
-    async def initialize(self) -> None:
-        """Initialize the async HTTP client."""
-        if self._async_client is not None:
-            return
-
-        logger.info(f"Reranker: initializing LiteLLM provider at {self.api_base} with model {self.model}")
-
-        headers = {"Content-Type": "application/json"}
-        if self.api_key:
-            headers["Authorization"] = f"Bearer {self.api_key}"
-
-        self._async_client = httpx.AsyncClient(timeout=self.timeout, headers=headers)
-        logger.info("Reranker: LiteLLM provider initialized")
-
-    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """
-        Score query-document pairs using the LiteLLM proxy's /rerank endpoint.
-
-        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 (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]
-
-            # LiteLLM /rerank follows Cohere API format
-            response = await self._async_client.post(
-                f"{self.api_base}/rerank",
-                json={
-                    "model": self.model,
-                    "query": query,
-                    "documents": texts,
-                    "top_n": len(texts),  # Return all scores
-                },
-            )
-            response.raise_for_status()
-            result = response.json()
-
-            # Map scores back to original positions
-            # Response format: {"results": [{"index": 0, "relevance_score": 0.9}, ...]}
-            for item in result.get("results", []):
-                original_idx = item["index"]
-                score = item.get("relevance_score", item.get("score", 0.0))
-                all_scores[indices[original_idx]] = score
+            all_scores.extend(batch_scores)

        return all_scores


 def create_cross_encoder_from_env() -> CrossEncoderModel:
    """
-    Create a CrossEncoderModel instance based on configuration.
+    Create a CrossEncoderModel instance based on environment variables.

-    Reads configuration via get_config() to ensure consistency across the codebase.
+    See hindsight_api.config for environment variable names and defaults.

    Returns:
        Configured CrossEncoderModel instance
    """
-    from ..config import get_config
-
-    config = get_config()
-    provider = config.reranker_provider.lower()
+    provider = os.environ.get(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER).lower()

    if provider == "tei":
-        url = config.reranker_tei_url
+        url = os.environ.get(ENV_RERANKER_TEI_URL)
        if not url:
            raise ValueError(f"{ENV_RERANKER_TEI_URL} is required when {ENV_RERANKER_PROVIDER} is 'tei'")
-        return RemoteTEICrossEncoder(
-            base_url=url,
-            batch_size=config.reranker_tei_batch_size,
-            max_concurrent=config.reranker_tei_max_concurrent,
-        )
+        return RemoteTEICrossEncoder(base_url=url)
    elif provider == "local":
-        return LocalSTCrossEncoder(
-            model_name=config.reranker_local_model,
-            max_concurrent=config.reranker_local_max_concurrent,
-            force_cpu=config.reranker_local_force_cpu,
-        )
-    elif provider == "cohere":
-        api_key = os.environ.get(ENV_COHERE_API_KEY)
-        if not api_key:
-            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'cohere'")
-        model = os.environ.get(ENV_RERANKER_COHERE_MODEL, DEFAULT_RERANKER_COHERE_MODEL)
-        base_url = os.environ.get(ENV_RERANKER_COHERE_BASE_URL) or None
-        return CohereCrossEncoder(api_key=api_key, model=model, base_url=base_url)
-    elif provider == "flashrank":
-        model = os.environ.get(ENV_RERANKER_FLASHRANK_MODEL, DEFAULT_RERANKER_FLASHRANK_MODEL)
-        cache_dir = os.environ.get(ENV_RERANKER_FLASHRANK_CACHE_DIR, DEFAULT_RERANKER_FLASHRANK_CACHE_DIR)
-        return FlashRankCrossEncoder(model_name=model, cache_dir=cache_dir)
-    elif provider == "litellm":
-        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
-        api_key = os.environ.get(ENV_LITELLM_API_KEY)
-        model = os.environ.get(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL)
-        return LiteLLMCrossEncoder(api_base=api_base, api_key=api_key, model=model)
-    elif provider == "rrf":
-        return RRFPassthroughCrossEncoder()
+        model = os.environ.get(ENV_RERANKER_LOCAL_MODEL)
+        model_name = model or DEFAULT_RERANKER_LOCAL_MODEL
+        return LocalSTCrossEncoder(model_name=model_name)
    else:
-        raise ValueError(
-            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'flashrank', 'litellm', 'rrf'"
-        )
+        raise ValueError(f"Unknown reranker provider: {provider}. Supported: 'local', 'tei'")
@@ -1,284 +0,0 @@
-"""
-Database connection budget management.
-
-Limits concurrent database connections per operation to prevent
-a single operation (e.g., recall with parallel queries) from
-exhausting the connection pool.
-"""
-
-import asyncio
-import logging
-import uuid
-from contextlib import asynccontextmanager
-from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, AsyncIterator
-
-if TYPE_CHECKING:
-    import asyncpg
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class OperationBudget:
-    """
-    Tracks connection budget for a single operation.
-
-    Each operation gets a semaphore limiting its concurrent connections.
-    """
-
-    operation_id: str
-    max_connections: int
-    semaphore: asyncio.Semaphore = field(init=False)
-    active_count: int = field(default=0, init=False)
-
-    def __post_init__(self):
-        self.semaphore = asyncio.Semaphore(self.max_connections)
-
-
-class ConnectionBudgetManager:
-    """
-    Manages per-operation connection budgets.
-
-    Usage:
-        manager = ConnectionBudgetManager(default_budget=4)
-
-        # Start an operation
-        async with manager.operation(max_connections=2) as op:
-            # Acquire connections within the budget
-            async with op.acquire(pool) as conn:
-                await conn.fetch(...)
-
-            # Multiple connections respect the budget
-            async with op.acquire(pool) as conn1, op.acquire(pool) as conn2:
-                # At most 2 concurrent connections for this operation
-                ...
-    """
-
-    def __init__(self, default_budget: int = 4):
-        """
-        Initialize the budget manager.
-
-        Args:
-            default_budget: Default max connections per operation
-        """
-        self.default_budget = default_budget
-        self._operations: dict[str, OperationBudget] = {}
-        self._lock = asyncio.Lock()
-
-    @asynccontextmanager
-    async def operation(
-        self,
-        max_connections: int | None = None,
-        operation_id: str | None = None,
-    ) -> AsyncIterator["BudgetedOperation"]:
-        """
-        Create a budgeted operation context.
-
-        Args:
-            max_connections: Max concurrent connections for this operation.
-                           Defaults to manager's default_budget.
-            operation_id: Optional custom operation ID. Auto-generated if not provided.
-
-        Yields:
-            BudgetedOperation context for acquiring connections
-        """
-        op_id = operation_id or f"op-{uuid.uuid4().hex[:12]}"
-        budget = max_connections or self.default_budget
-
-        async with self._lock:
-            if op_id in self._operations:
-                raise ValueError(f"Operation {op_id} already exists")
-            self._operations[op_id] = OperationBudget(op_id, budget)
-
-        try:
-            yield BudgetedOperation(self, op_id)
-        finally:
-            async with self._lock:
-                self._operations.pop(op_id, None)
-
-    def _get_budget(self, operation_id: str) -> OperationBudget:
-        """Get budget for an operation (internal use)."""
-        budget = self._operations.get(operation_id)
-        if not budget:
-            raise ValueError(f"Operation {operation_id} not found")
-        return budget
-
-
-class BudgetedOperation:
-    """
-    A single operation with connection budget.
-
-    Provides methods to acquire connections within the budget.
-    """
-
-    def __init__(self, manager: ConnectionBudgetManager, operation_id: str):
-        self._manager = manager
-        self.operation_id = operation_id
-
-    @property
-    def budget(self) -> OperationBudget:
-        """Get the budget for this operation."""
-        return self._manager._get_budget(self.operation_id)
-
-    @asynccontextmanager
-    async def acquire(self, pool: "asyncpg.Pool") -> AsyncIterator["asyncpg.Connection"]:
-        """
-        Acquire a connection within the operation's budget.
-
-        Blocks if the operation has reached its connection limit.
-
-        Args:
-            pool: asyncpg connection pool
-
-        Yields:
-            Database connection
-        """
-        budget = self.budget
-        async with budget.semaphore:
-            budget.active_count += 1
-            conn = await pool.acquire()
-            try:
-                yield conn
-            finally:
-                budget.active_count -= 1
-                await pool.release(conn)
-
-    def wrap_pool(self, pool: "asyncpg.Pool") -> "BudgetedPool":
-        """
-        Wrap a pool with this operation's budget.
-
-        The returned BudgetedPool can be passed to functions expecting a pool,
-        and all acquire() calls will be limited by this operation's budget.
-
-        Args:
-            pool: asyncpg connection pool to wrap
-
-        Returns:
-            BudgetedPool that limits connections to this operation's budget
-        """
-        return BudgetedPool(pool, self)
-
-    async def acquire_many(
-        self,
-        pool: "asyncpg.Pool",
-        count: int,
-    ) -> AsyncIterator[list["asyncpg.Connection"]]:
-        """
-        Acquire multiple connections within the budget.
-
-        Note: This acquires connections sequentially to respect the budget.
-        For parallel acquisition, use multiple acquire() calls with asyncio.gather().
-
-        Args:
-            pool: asyncpg connection pool
-            count: Number of connections to acquire
-
-        Yields:
-            List of database connections
-        """
-        connections = []
-        try:
-            for _ in range(count):
-                conn = await pool.acquire()
-                connections.append(conn)
-            yield connections
-        finally:
-            for conn in connections:
-                await pool.release(conn)
-
-
-# Global default manager instance
-_default_manager: ConnectionBudgetManager | None = None
-
-
-def get_budget_manager(default_budget: int = 4) -> ConnectionBudgetManager:
-    """
-    Get or create the global budget manager.
-
-    Args:
-        default_budget: Default max connections per operation
-
-    Returns:
-        Global ConnectionBudgetManager instance
-    """
-    global _default_manager
-    if _default_manager is None:
-        _default_manager = ConnectionBudgetManager(default_budget=default_budget)
-    return _default_manager
-
-
-@asynccontextmanager
-async def budgeted_operation(
-    max_connections: int | None = None,
-    operation_id: str | None = None,
-    default_budget: int = 4,
-) -> AsyncIterator[BudgetedOperation]:
-    """
-    Convenience function to create a budgeted operation.
-
-    Args:
-        max_connections: Max concurrent connections for this operation
-        operation_id: Optional custom operation ID
-        default_budget: Default budget if manager not yet created
-
-    Yields:
-        BudgetedOperation context
-
-    Example:
-        async with budgeted_operation(max_connections=2) as op:
-            async with op.acquire(pool) as conn:
-                await conn.fetch(...)
-    """
-    manager = get_budget_manager(default_budget)
-    async with manager.operation(max_connections, operation_id) as op:
-        yield op
-
-
-class BudgetedPool:
-    """
-    A pool wrapper that limits concurrent connection acquisitions.
-
-    This can be passed to functions expecting a pool, and acquire()
-    calls will be limited by the budget semaphore.
-
-    Usage:
-        async with budgeted_operation(max_connections=4) as op:
-            budgeted_pool = op.wrap_pool(pool)
-            # Pass budgeted_pool to functions that expect a pool
-            await some_function(budgeted_pool, ...)
-    """
-
-    def __init__(self, pool: "asyncpg.Pool", operation: BudgetedOperation):
-        self._pool = pool
-        self._operation = operation
-
-    async def acquire(self) -> "asyncpg.Connection":
-        """
-        Acquire a connection within the budget.
-
-        Note: Caller must release the connection when done.
-        Prefer using as context manager via acquire_with_retry or op.acquire().
-        """
-        budget = self._operation.budget
-        await budget.semaphore.acquire()
-        budget.active_count += 1
-        try:
-            return await self._pool.acquire()
-        except Exception:
-            budget.active_count -= 1
-            budget.semaphore.release()
-            raise
-
-    async def release(self, conn: "asyncpg.Connection") -> None:
-        """Release a connection back to the pool."""
-        budget = self._operation.budget
-        try:
-            await self._pool.release(conn)
-        finally:
-            budget.active_count -= 1
-            budget.semaphore.release()
-
-    def __getattr__(self, name):
-        """Proxy other attributes to the underlying pool."""
-        return getattr(self._pool, name)
@@ -83,22 +83,11 @@ async def acquire_with_retry(pool: asyncpg.Pool, max_retries: int = DEFAULT_MAX_
    Yields:
        An asyncpg connection
    """
-    import time
-
-    start = time.time()

    async def acquire():
        return await pool.acquire()

    conn = await retry_with_backoff(acquire, max_retries=max_retries)
-    acquire_time = time.time() - start
-
-    # Log slow connection acquisitions (indicates pool contention)
-    if acquire_time > 0.05:  # 50ms threshold
-        pool_size = pool.get_size()
-        pool_free = pool.get_idle_size()
-        logger.warning(f"[DB POOL] Slow acquire: {acquire_time:.3f}s | size={pool_size}, idle={pool_free}")
-
    try:
        yield conn
    finally:
@@ -1,5 +0,0 @@
-"""Directives module for hard rules injected into prompts."""
-
-from .models import Directive
-
-__all__ = ["Directive"]
@@ -1,37 +0,0 @@
-"""Pydantic models for directives."""
-
-from datetime import datetime, timezone
-from uuid import UUID
-
-from pydantic import BaseModel, Field
-
-
-class Directive(BaseModel):
-    """A directive is a hard rule injected into prompts.
-
-    Directives are user-defined rules that guide agent behavior. Unlike mental models
-    which are automatically consolidated from memories, directives are explicit
-    instructions that are always included in relevant prompts.
-
-    Examples:
-    - "Always respond in formal English"
-    - "Never share personal data with third parties"
-    - "Prefer conservative investment recommendations"
-    """
-
-    id: UUID = Field(description="Unique identifier")
-    bank_id: str = Field(description="Bank this directive belongs to")
-    name: str = Field(description="Human-readable name")
-    content: str = Field(description="The directive text to inject into prompts")
-    priority: int = Field(default=0, description="Higher priority directives are injected first")
-    is_active: bool = Field(default=True, description="Whether this directive is currently active")
-    tags: list[str] = Field(default_factory=list, description="Tags for filtering")
-    created_at: datetime = Field(
-        default_factory=lambda: datetime.now(timezone.utc), description="When this directive was created"
-    )
-    updated_at: datetime = Field(
-        default_factory=lambda: datetime.now(timezone.utc), description="When this directive was last updated"
-    )
-
-    class Config:
-        from_attributes = True
@@ -3,8 +3,8 @@ Embeddings abstraction for the memory system.

 Provides an interface for generating embeddings with different backends.

-The embedding dimension is auto-detected from the model at initialization.
-The database schema is automatically adjusted to match the model's dimension.
+IMPORTANT: All embeddings must produce 384-dimensional vectors to match
+the database schema (pgvector column defined as vector(384)).

 Configuration via environment variables - see hindsight_api.config for all env var names.
 """
@@ -16,27 +16,12 @@ from abc import ABC, abstractmethod
 import httpx

 from ..config import (
-    DEFAULT_EMBEDDINGS_COHERE_MODEL,
-    DEFAULT_EMBEDDINGS_LITELLM_MODEL,
-    DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU,
    DEFAULT_EMBEDDINGS_LOCAL_MODEL,
-    DEFAULT_EMBEDDINGS_OPENAI_MODEL,
    DEFAULT_EMBEDDINGS_PROVIDER,
-    DEFAULT_LITELLM_API_BASE,
-    ENV_COHERE_API_KEY,
-    ENV_EMBEDDINGS_COHERE_BASE_URL,
-    ENV_EMBEDDINGS_COHERE_MODEL,
-    ENV_EMBEDDINGS_LITELLM_MODEL,
-    ENV_EMBEDDINGS_LOCAL_FORCE_CPU,
+    EMBEDDING_DIMENSION,
    ENV_EMBEDDINGS_LOCAL_MODEL,
-    ENV_EMBEDDINGS_OPENAI_API_KEY,
-    ENV_EMBEDDINGS_OPENAI_BASE_URL,
-    ENV_EMBEDDINGS_OPENAI_MODEL,
    ENV_EMBEDDINGS_PROVIDER,
    ENV_EMBEDDINGS_TEI_URL,
-    ENV_LITELLM_API_BASE,
-    ENV_LITELLM_API_KEY,
-    ENV_LLM_API_KEY,
 )

 logger = logging.getLogger(__name__)
@@ -46,8 +31,8 @@ class Embeddings(ABC):
    """
    Abstract base class for embedding generation.

-    The embedding dimension is determined by the model and detected at initialization.
-    The database schema is automatically adjusted to match the model's dimension.
+    All implementations MUST generate 384-dimensional embeddings to match
+    the database schema.
    """

    @property
@@ -56,12 +41,6 @@ class Embeddings(ABC):
        """Return a human-readable name for this provider (e.g., 'local', 'tei')."""
        pass

-    @property
-    @abstractmethod
-    def dimension(self) -> int:
-        """Return the embedding dimension produced by this model."""
-        pass
-
    @abstractmethod
    async def initialize(self) -> None:
        """
@@ -75,13 +54,13 @@ class Embeddings(ABC):
    @abstractmethod
    def encode(self, texts: list[str]) -> list[list[float]]:
        """
-        Generate embeddings for a list of texts.
+        Generate 384-dimensional embeddings for a list of texts.

        Args:
            texts: List of text strings to encode

        Returns:
-            List of embedding vectors (each is a list of floats)
+            List of 384-dimensional embedding vectors (each is a list of floats)
        """
        pass

@@ -91,34 +70,27 @@ class LocalSTEmbeddings(Embeddings):
    Local embeddings implementation using SentenceTransformers.

    Call initialize() during startup to load the model and avoid cold starts.
-    The embedding dimension is auto-detected from the model.
+
+    Default model is BAAI/bge-small-en-v1.5 which produces 384-dimensional
+    embeddings matching the database schema.
    """

-    def __init__(self, model_name: str | None = None, force_cpu: bool = False):
+    def __init__(self, model_name: str | None = None):
        """
        Initialize local SentenceTransformers embeddings.

        Args:
            model_name: Name of the SentenceTransformer model to use.
+                       Must produce 384-dimensional embeddings.
                       Default: BAAI/bge-small-en-v1.5
-            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
-                      Default: False
        """
        self.model_name = model_name or DEFAULT_EMBEDDINGS_LOCAL_MODEL
-        self.force_cpu = force_cpu
        self._model = None
-        self._dimension: int | None = None

    @property
    def provider_name(self) -> str:
        return "local"

-    @property
-    def dimension(self) -> int:
-        if self._dimension is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-        return self._dimension
-
    async def initialize(self) -> None:
        """Load the embedding model."""
        if self._model is not None:
@@ -133,52 +105,36 @@ class LocalSTEmbeddings(Embeddings):
            )

        logger.info(f"Embeddings: initializing local provider with model {self.model_name}")
-
-        # Determine device based on hardware availability.
-        # We always set low_cpu_mem_usage=False to prevent lazy loading (meta tensors)
-        # which can cause issues when accelerate is installed but no GPU is available.
-        import torch
-
-        # Force CPU mode if configured (used in daemon mode to avoid MPS/XPC issues on macOS)
-        if self.force_cpu:
-            device = "cpu"
-            logger.info("Embeddings: forcing CPU mode")
-        else:
-            # Check for GPU (CUDA) or Apple Silicon (MPS)
-            # Wrap in try-except to gracefully handle any device detection issues
-            # (e.g., in CI environments or when PyTorch is built without GPU support)
-            device = "cpu"  # Default to CPU
-            try:
-                has_gpu = torch.cuda.is_available() or (
-                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
-                )
-                if has_gpu:
-                    device = None  # Let sentence-transformers auto-detect GPU/MPS
-            except Exception as e:
-                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")
-
+        # Disable lazy loading (meta tensors) which causes issues with newer transformers/accelerate
+        # Setting low_cpu_mem_usage=False and device_map=None ensures tensors are fully materialized
        self._model = SentenceTransformer(
            self.model_name,
-            device=device,
-            model_kwargs={"low_cpu_mem_usage": False},
+            model_kwargs={"low_cpu_mem_usage": False, "device_map": None},
        )

-        self._dimension = self._model.get_sentence_embedding_dimension()
-        logger.info(f"Embeddings: local provider initialized (dim: {self._dimension})")
+        # Validate dimension matches database schema
+        model_dim = self._model.get_sentence_embedding_dimension()
+        if model_dim != EMBEDDING_DIMENSION:
+            raise ValueError(
+                f"Model {self.model_name} produces {model_dim}-dimensional embeddings, "
+                f"but database schema requires {EMBEDDING_DIMENSION} dimensions. "
+                f"Use a model that produces {EMBEDDING_DIMENSION}-dimensional embeddings."
+            )
+
+        logger.info(f"Embeddings: local provider initialized (dim: {model_dim})")

    def encode(self, texts: list[str]) -> list[list[float]]:
        """
-        Generate embeddings for a list of texts.
+        Generate 384-dimensional embeddings for a list of texts.

        Args:
            texts: List of text strings to encode

        Returns:
-            List of embedding vectors
+            List of 384-dimensional embedding vectors
        """
        if self._model is None:
            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-
        embeddings = self._model.encode(texts, convert_to_numpy=True, show_progress_bar=False)
        return [emb.tolist() for emb in embeddings]

@@ -190,7 +146,7 @@ class RemoteTEIEmbeddings(Embeddings):
    TEI provides a high-performance inference server for embedding models.
    See: https://github.com/huggingface/text-embeddings-inference

-    The embedding dimension is auto-detected from the server at initialization.
+    The server should be running a model that produces 384-dimensional embeddings.
    """

    def __init__(
@@ -218,18 +174,11 @@ class RemoteTEIEmbeddings(Embeddings):
        self.retry_delay = retry_delay
        self._client: httpx.Client | None = None
        self._model_id: str | None = None
-        self._dimension: int | None = None

    @property
    def provider_name(self) -> str:
        return "tei"

-    @property
-    def dimension(self) -> int:
-        if self._dimension is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-        return self._dimension
-
    def _request_with_retry(self, method: str, url: str, **kwargs) -> httpx.Response:
        """Make an HTTP request with automatic retries on transient errors."""
        import time
@@ -280,24 +229,7 @@ class RemoteTEIEmbeddings(Embeddings):
            response = self._request_with_retry("GET", f"{self.base_url}/info")
            info = response.json()
            self._model_id = info.get("model_id", "unknown")
-
-            # Get dimension from server info or by doing a test embedding
-            if "max_input_length" in info and "model_dtype" in info:
-                # Try to get dimension from info endpoint (some TEI versions expose it)
-                # If not available, do a test embedding
-                pass
-
-            # Do a test embedding to detect dimension
-            test_response = self._request_with_retry(
-                "POST",
-                f"{self.base_url}/embed",
-                json={"inputs": ["test"]},
-            )
-            test_embeddings = test_response.json()
-            if test_embeddings and len(test_embeddings) > 0:
-                self._dimension = len(test_embeddings[0])
-
-            logger.info(f"Embeddings: TEI provider initialized (model: {self._model_id}, dim: {self._dimension})")
+            logger.info(f"Embeddings: TEI provider initialized (model: {self._model_id})")
        except httpx.HTTPError as e:
            raise RuntimeError(f"Failed to connect to TEI server at {self.base_url}: {e}")

@@ -337,417 +269,25 @@ class RemoteTEIEmbeddings(Embeddings):
        return all_embeddings


-class OpenAIEmbeddings(Embeddings):
-    """
-    OpenAI embeddings implementation using the OpenAI API.
-
-    Supports text-embedding-3-small (1536 dims), text-embedding-3-large (3072 dims),
-    and text-embedding-ada-002 (1536 dims, legacy).
-
-    The embedding dimension is auto-detected from the model at initialization.
-    """
-
-    # Known dimensions for OpenAI embedding models
-    MODEL_DIMENSIONS = {
-        "text-embedding-3-small": 1536,
-        "text-embedding-3-large": 3072,
-        "text-embedding-ada-002": 1536,
-    }
-
-    def __init__(
-        self,
-        api_key: str,
-        model: str = DEFAULT_EMBEDDINGS_OPENAI_MODEL,
-        base_url: str | None = None,
-        batch_size: int = 100,
-        max_retries: int = 3,
-    ):
-        """
-        Initialize OpenAI embeddings client.
-
-        Args:
-            api_key: OpenAI API key
-            model: OpenAI embedding model name (default: text-embedding-3-small)
-            base_url: Custom base URL for OpenAI-compatible API (e.g., Azure OpenAI endpoint)
-            batch_size: Maximum batch size for embedding requests (default: 100)
-            max_retries: Maximum number of retries for failed requests (default: 3)
-        """
-        self.api_key = api_key
-        self.model = model
-        self.base_url = base_url
-        self.batch_size = batch_size
-        self.max_retries = max_retries
-        self._client = None
-        self._dimension: int | None = None
-
-    @property
-    def provider_name(self) -> str:
-        return "openai"
-
-    @property
-    def dimension(self) -> int:
-        if self._dimension is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-        return self._dimension
-
-    async def initialize(self) -> None:
-        """Initialize the OpenAI client and detect dimension."""
-        if self._client is not None:
-            return
-
-        try:
-            from openai import OpenAI
-        except ImportError:
-            raise ImportError("openai is required for OpenAIEmbeddings. Install it with: pip install openai")
-
-        base_url_msg = f" at {self.base_url}" if self.base_url else ""
-        logger.info(f"Embeddings: initializing OpenAI provider with model {self.model}{base_url_msg}")
-
-        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
-        client_kwargs = {"api_key": self.api_key, "max_retries": self.max_retries}
-        if self.base_url:
-            client_kwargs["base_url"] = self.base_url
-        self._client = OpenAI(**client_kwargs)
-
-        # Try to get dimension from known models, otherwise do a test embedding
-        if self.model in self.MODEL_DIMENSIONS:
-            self._dimension = self.MODEL_DIMENSIONS[self.model]
-        else:
-            # Do a test embedding to detect dimension
-            response = self._client.embeddings.create(
-                model=self.model,
-                input=["test"],
-            )
-            if response.data:
-                self._dimension = len(response.data[0].embedding)
-
-        logger.info(f"Embeddings: OpenAI provider initialized (model: {self.model}, dim: {self._dimension})")
-
-    def encode(self, texts: list[str]) -> list[list[float]]:
-        """
-        Generate embeddings using the OpenAI API.
-
-        Args:
-            texts: List of text strings to encode
-
-        Returns:
-            List of embedding vectors
-        """
-        if self._client is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-
-        if not texts:
-            return []
-
-        all_embeddings = []
-
-        # Process in batches
-        for i in range(0, len(texts), self.batch_size):
-            batch = texts[i : i + self.batch_size]
-
-            response = self._client.embeddings.create(
-                model=self.model,
-                input=batch,
-            )
-
-            # Sort by index to ensure correct order
-            batch_embeddings = sorted(response.data, key=lambda x: x.index)
-            all_embeddings.extend([e.embedding for e in batch_embeddings])
-
-        return all_embeddings
-
-
-class CohereEmbeddings(Embeddings):
-    """
-    Cohere embeddings implementation using the Cohere API.
-
-    Supports embed-english-v3.0 (1024 dims) and embed-multilingual-v3.0 (1024 dims).
-
-    The embedding dimension is auto-detected from the model at initialization.
-    """
-
-    # Known dimensions for Cohere embedding models
-    MODEL_DIMENSIONS = {
-        "embed-english-v3.0": 1024,
-        "embed-multilingual-v3.0": 1024,
-        "embed-english-light-v3.0": 384,
-        "embed-multilingual-light-v3.0": 384,
-        "embed-english-v2.0": 4096,
-        "embed-multilingual-v2.0": 768,
-    }
-
-    def __init__(
-        self,
-        api_key: str,
-        model: str = DEFAULT_EMBEDDINGS_COHERE_MODEL,
-        base_url: str | None = None,
-        batch_size: int = 96,
-        timeout: float = 60.0,
-        input_type: str = "search_document",
-    ):
-        """
-        Initialize Cohere embeddings client.
-
-        Args:
-            api_key: Cohere API key
-            model: Cohere embedding model name (default: embed-english-v3.0)
-            base_url: Custom base URL for Cohere-compatible API (e.g., Azure-hosted endpoint)
-            batch_size: Maximum batch size for embedding requests (default: 96, Cohere's limit)
-            timeout: Request timeout in seconds (default: 60.0)
-            input_type: Input type for embeddings (default: search_document).
-                       Options: search_document, search_query, classification, clustering
-        """
-        self.api_key = api_key
-        self.model = model
-        self.base_url = base_url
-        self.batch_size = batch_size
-        self.timeout = timeout
-        self.input_type = input_type
-        self._client = None
-        self._dimension: int | None = None
-
-    @property
-    def provider_name(self) -> str:
-        return "cohere"
-
-    @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 Cohere client and detect dimension."""
-        if self._client is not None:
-            return
-
-        try:
-            import cohere
-        except ImportError:
-            raise ImportError("cohere is required for CohereEmbeddings. Install it with: pip install cohere")
-
-        base_url_msg = f" at {self.base_url}" if self.base_url else ""
-        logger.info(f"Embeddings: initializing Cohere provider with model {self.model}{base_url_msg}")
-
-        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
-        client_kwargs = {"api_key": self.api_key, "timeout": self.timeout}
-        if self.base_url:
-            client_kwargs["base_url"] = self.base_url
-        self._client = cohere.Client(**client_kwargs)
-
-        # Try to get dimension from known models, otherwise do a test embedding
-        if self.model in self.MODEL_DIMENSIONS:
-            self._dimension = self.MODEL_DIMENSIONS[self.model]
-        else:
-            # Do a test embedding to detect dimension
-            response = self._client.embed(
-                texts=["test"],
-                model=self.model,
-                input_type=self.input_type,
-            )
-            if response.embeddings and isinstance(response.embeddings, list):
-                self._dimension = len(response.embeddings[0])
-
-        logger.info(f"Embeddings: Cohere provider initialized (model: {self.model}, dim: {self._dimension})")
-
-    def encode(self, texts: list[str]) -> list[list[float]]:
-        """
-        Generate embeddings using the Cohere API.
-
-        Args:
-            texts: List of text strings to encode
-
-        Returns:
-            List of embedding vectors
-        """
-        if self._client is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-
-        if not texts:
-            return []
-
-        all_embeddings = []
-
-        # Process in batches
-        for i in range(0, len(texts), self.batch_size):
-            batch = texts[i : i + self.batch_size]
-
-            response = self._client.embed(
-                texts=batch,
-                model=self.model,
-                input_type=self.input_type,
-            )
-
-            all_embeddings.extend(response.embeddings)
-
-        return all_embeddings
-
-
-class LiteLLMEmbeddings(Embeddings):
-    """
-    LiteLLM embeddings implementation using LiteLLM proxy's /embeddings endpoint.
-
-    LiteLLM provides a unified interface for multiple embedding providers.
-    The proxy exposes an OpenAI-compatible /embeddings endpoint.
-    See: https://docs.litellm.ai/docs/embedding/supported_embedding
-
-    Supported providers via LiteLLM:
-    - OpenAI (text-embedding-3-small, text-embedding-ada-002, etc.)
-    - Cohere (embed-english-v3.0, etc.) - prefix with cohere/
-    - Vertex AI (textembedding-gecko, etc.) - prefix with vertex_ai/
-    - HuggingFace, Mistral, Voyage AI, etc.
-
-    The embedding dimension is auto-detected from the model at initialization.
-    """
-
-    def __init__(
-        self,
-        api_base: str = DEFAULT_LITELLM_API_BASE,
-        api_key: str | None = None,
-        model: str = DEFAULT_EMBEDDINGS_LITELLM_MODEL,
-        batch_size: int = 100,
-        timeout: float = 60.0,
-    ):
-        """
-        Initialize LiteLLM embeddings client.
-
-        Args:
-            api_base: Base URL of the LiteLLM proxy (default: http://localhost:4000)
-            api_key: API key for the LiteLLM proxy (optional, depends on proxy config)
-            model: Embedding model name (default: text-embedding-3-small)
-                   Use provider prefix for non-OpenAI models (e.g., cohere/embed-english-v3.0)
-            batch_size: Maximum batch size for embedding requests (default: 100)
-            timeout: Request timeout in seconds (default: 60.0)
-        """
-        self.api_base = api_base.rstrip("/")
-        self.api_key = api_key
-        self.model = model
-        self.batch_size = batch_size
-        self.timeout = timeout
-        self._client: httpx.Client | None = None
-        self._dimension: int | None = None
-
-    @property
-    def provider_name(self) -> str:
-        return "litellm"
-
-    @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 HTTP client and detect embedding dimension."""
-        if self._client is not None:
-            return
-
-        logger.info(f"Embeddings: initializing LiteLLM provider at {self.api_base} with model {self.model}")
-
-        headers = {"Content-Type": "application/json"}
-        if self.api_key:
-            headers["Authorization"] = f"Bearer {self.api_key}"
-
-        self._client = httpx.Client(timeout=self.timeout, headers=headers)
-
-        # Do a test embedding to detect dimension
-        try:
-            response = self._client.post(
-                f"{self.api_base}/embeddings",
-                json={"model": self.model, "input": ["test"]},
-            )
-            response.raise_for_status()
-            result = response.json()
-            if result.get("data") and len(result["data"]) > 0:
-                self._dimension = len(result["data"][0]["embedding"])
-            logger.info(f"Embeddings: LiteLLM provider initialized (model: {self.model}, dim: {self._dimension})")
-        except httpx.HTTPError as e:
-            raise RuntimeError(f"Failed to connect to LiteLLM proxy at {self.api_base}: {e}")
-
-    def encode(self, texts: list[str]) -> list[list[float]]:
-        """
-        Generate embeddings using the LiteLLM proxy.
-
-        Args:
-            texts: List of text strings to encode
-
-        Returns:
-            List of embedding vectors
-        """
-        if self._client is None:
-            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-
-        if not texts:
-            return []
-
-        all_embeddings = []
-
-        # Process in batches
-        for i in range(0, len(texts), self.batch_size):
-            batch = texts[i : i + self.batch_size]
-
-            response = self._client.post(
-                f"{self.api_base}/embeddings",
-                json={"model": self.model, "input": batch},
-            )
-            response.raise_for_status()
-            result = response.json()
-
-            # Sort by index to ensure correct order
-            batch_embeddings = sorted(result["data"], key=lambda x: x["index"])
-            all_embeddings.extend([e["embedding"] for e in batch_embeddings])
-
-        return all_embeddings
-
-
 def create_embeddings_from_env() -> Embeddings:
    """
-    Create an Embeddings instance based on configuration.
+    Create an Embeddings instance based on environment variables.

-    Reads configuration via get_config() to ensure consistency across the codebase.
+    See hindsight_api.config for environment variable names and defaults.

    Returns:
        Configured Embeddings instance
    """
-    from ..config import get_config
-
-    config = get_config()
-    provider = config.embeddings_provider.lower()
+    provider = os.environ.get(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER).lower()

    if provider == "tei":
-        url = config.embeddings_tei_url
+        url = os.environ.get(ENV_EMBEDDINGS_TEI_URL)
        if not url:
            raise ValueError(f"{ENV_EMBEDDINGS_TEI_URL} is required when {ENV_EMBEDDINGS_PROVIDER} is 'tei'")
        return RemoteTEIEmbeddings(base_url=url)
    elif provider == "local":
-        return LocalSTEmbeddings(
-            model_name=config.embeddings_local_model,
-            force_cpu=config.embeddings_local_force_cpu,
-        )
-    elif provider == "openai":
-        # Use dedicated embeddings API key, or fall back to LLM API key
-        api_key = os.environ.get(ENV_EMBEDDINGS_OPENAI_API_KEY) or os.environ.get(ENV_LLM_API_KEY)
-        if not api_key:
-            raise ValueError(
-                f"{ENV_EMBEDDINGS_OPENAI_API_KEY} or {ENV_LLM_API_KEY} is required "
-                f"when {ENV_EMBEDDINGS_PROVIDER} is 'openai'"
-            )
-        model = os.environ.get(ENV_EMBEDDINGS_OPENAI_MODEL, DEFAULT_EMBEDDINGS_OPENAI_MODEL)
-        base_url = os.environ.get(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None
-        return OpenAIEmbeddings(api_key=api_key, model=model, base_url=base_url)
-    elif provider == "cohere":
-        api_key = os.environ.get(ENV_COHERE_API_KEY)
-        if not api_key:
-            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'cohere'")
-        model = os.environ.get(ENV_EMBEDDINGS_COHERE_MODEL, DEFAULT_EMBEDDINGS_COHERE_MODEL)
-        base_url = os.environ.get(ENV_EMBEDDINGS_COHERE_BASE_URL) or None
-        return CohereEmbeddings(api_key=api_key, model=model, base_url=base_url)
-    elif provider == "litellm":
-        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
-        api_key = os.environ.get(ENV_LITELLM_API_KEY)
-        model = os.environ.get(ENV_EMBEDDINGS_LITELLM_MODEL, DEFAULT_EMBEDDINGS_LITELLM_MODEL)
-        return LiteLLMEmbeddings(api_base=api_base, api_key=api_key, model=model)
+        model = os.environ.get(ENV_EMBEDDINGS_LOCAL_MODEL)
+        model_name = model or DEFAULT_EMBEDDINGS_LOCAL_MODEL
+        return LocalSTEmbeddings(model_name=model_name)
    else:
-        raise ValueError(
-            f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai', 'cohere', 'litellm'"
-        )
+        raise ValueError(f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei'")
@@ -209,7 +209,7 @@ class EntityResolver:
        # This handles duplicates via ON CONFLICT and returns all IDs
        if entities_to_create:
            # Group entities by canonical name (lowercase) to handle duplicates within batch
-            # For duplicates, we only insert once and reuse the ID, but track the count
+            # For duplicates, we only insert once and reuse the ID
            unique_entities = {}  # lowercase_name -> (entity_data, event_date, [indices])
            for idx, entity_data, event_date in entities_to_create:
                name_lower = entity_data["text"].lower()
@@ -223,32 +223,29 @@ class EntityResolver:
            # Use a single query with unnest for speed
            entity_names = []
            entity_dates = []
-            entity_counts = []  # Track how many times each entity appears in this batch
            indices_map = []  # Maps result index -> list of original indices

            for name_lower, (entity_data, event_date, indices) in unique_entities.items():
                entity_names.append(entity_data["text"])
                entity_dates.append(event_date)
-                entity_counts.append(len(indices))  # Count of occurrences in this batch
                indices_map.append(indices)

            # Batch INSERT ... ON CONFLICT with RETURNING
-            # Uses the batch count for mention_count instead of always 1
+            # This is much faster than individual inserts
            rows = await conn.fetch(
                f"""
                INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
-                SELECT $1, name, event_date, event_date, cnt
-                FROM unnest($2::text[], $3::timestamptz[], $4::int[]) AS t(name, event_date, cnt)
+                SELECT $1, name, event_date, event_date, 1
+                FROM unnest($2::text[], $3::timestamptz[]) AS t(name, event_date)
                ON CONFLICT (bank_id, LOWER(canonical_name))
                DO UPDATE SET
-                    mention_count = {fq_table("entities")}.mention_count + EXCLUDED.mention_count,
+                    mention_count = {fq_table("entities")}.mention_count + 1,
                    last_seen = EXCLUDED.last_seen
                RETURNING id
                """,
                bank_id,
                entity_names,
                entity_dates,
-                entity_counts,
            )

            # Map returned IDs back to original indices
@@ -110,8 +110,6 @@ class MemoryEngineInterface(ABC):
        *,
        budget: "Budget | None" = None,
        context: str | None = None,
-        max_tokens: int = 4096,
-        response_schema: dict | None = None,
        request_context: "RequestContext",
    ) -> "ReflectResult":
        """
@@ -122,8 +120,6 @@ class MemoryEngineInterface(ABC):
            query: The question to reflect on.
            budget: Search budget for retrieving context.
            context: Additional context for the reflection.
-            max_tokens: Maximum tokens for the response.
-            response_schema: Optional JSON Schema for structured output.
            request_context: Request context for authentication.

        Returns:
@@ -160,14 +156,14 @@ class MemoryEngineInterface(ABC):
        request_context: "RequestContext",
    ) -> dict[str, Any]:
        """
-        Get bank profile including disposition and mission.
+        Get bank profile including disposition and background.

        Args:
            bank_id: The memory bank ID.
            request_context: Request context for authentication.

        Returns:
-            Bank profile dict with bank_id, name, disposition, and mission.
+            Bank profile dict.
        """
        ...

@@ -190,44 +186,25 @@ class MemoryEngineInterface(ABC):
        ...

    @abstractmethod
-    async def merge_bank_mission(
+    async def merge_bank_background(
        self,
        bank_id: str,
        new_info: str,
        *,
+        update_disposition: bool = True,
        request_context: "RequestContext",
    ) -> dict[str, Any]:
        """
-        Merge new mission information into bank profile.
+        Merge new background information into bank profile.

        Args:
            bank_id: The memory bank ID.
-            new_info: New mission information to merge.
+            new_info: New background information to merge.
+            update_disposition: Whether to infer disposition from background.
            request_context: Request context for authentication.

        Returns:
-            Updated mission info.
-        """
-        ...
-
-    @abstractmethod
-    async def set_bank_mission(
-        self,
-        bank_id: str,
-        mission: str,
-        *,
-        request_context: "RequestContext",
-    ) -> dict[str, Any]:
-        """
-        Set the bank's mission (replaces existing).
-
-        Args:
-            bank_id: The memory bank ID.
-            mission: The mission text.
-            request_context: Request context for authentication.
-
-        Returns:
-            Dict with bank_id and mission.
+            Updated background info.
        """
        ...

@@ -308,7 +285,6 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        fact_type: str | None = None,
-        limit: int = 1000,
        request_context: "RequestContext",
    ) -> dict[str, Any]:
        """
@@ -317,11 +293,10 @@ class MemoryEngineInterface(ABC):
        Args:
            bank_id: The memory bank ID.
            fact_type: Filter by fact type.
-            limit: Maximum number of items to return (default: 1000).
            request_context: Request context for authentication.

        Returns:
-            Dict with nodes, edges, table_rows, total_units, limit.
+            Dict with nodes, edges, table_rows, total_units.
        """
        ...

@@ -425,20 +400,61 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        limit: int = 100,
-        offset: int = 0,
        request_context: "RequestContext",
-    ) -> dict[str, Any]:
+    ) -> list[dict[str, Any]]:
        """
-        List entities for a bank with pagination.
+        List entities for a bank.

        Args:
            bank_id: The memory bank ID.
            limit: Maximum results.
-            offset: Offset for pagination.
            request_context: Request context for authentication.

        Returns:
-            Dict with items, total, limit, offset.
+            List of entity dicts.
+        """
+        ...
+
+    @abstractmethod
+    async def get_entity_observations(
+        self,
+        bank_id: str,
+        entity_id: str,
+        *,
+        limit: int = 10,
+        request_context: "RequestContext",
+    ) -> list[Any]:
+        """
+        Get observations for an entity.
+
+        Args:
+            bank_id: The memory bank ID.
+            entity_id: The entity ID.
+            limit: Maximum observations.
+            request_context: Request context for authentication.
+
+        Returns:
+            List of EntityObservation objects.
+        """
+        ...
+
+    @abstractmethod
+    async def regenerate_entity_observations(
+        self,
+        bank_id: str,
+        entity_id: str,
+        entity_name: str,
+        *,
+        request_context: "RequestContext",
+    ) -> None:
+        """
+        Regenerate observations for an entity.
+
+        Args:
+            bank_id: The memory bank ID.
+            entity_id: The entity ID.
+            entity_name: The entity's canonical name.
+            request_context: Request context for authentication.
        """
        ...

@@ -494,7 +510,7 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        request_context: "RequestContext",
-    ) -> dict[str, Any]:
+    ) -> list[dict[str, Any]]:
        """
        List async operations for a bank.

@@ -503,7 +519,7 @@ class MemoryEngineInterface(ABC):
            request_context: Request context for authentication.

        Returns:
-            Dict with 'total' (int) and 'operations' (list of operation dicts).
+            List of operation dicts with id, task_type, status, etc.
        """
        ...

@@ -537,16 +553,16 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        name: str | None = None,
-        mission: str | None = None,
+        background: str | None = None,
        request_context: "RequestContext",
    ) -> dict[str, Any]:
        """
-        Update bank name and/or mission.
+        Update bank name and/or background.

        Args:
            bank_id: The memory bank ID.
            name: New bank name (optional).
-            mission: New mission text (optional, replaces existing).
+            background: New background text (optional, replaces existing).
            request_context: Request context for authentication.

        Returns:
@@ -1,14 +0,0 @@
-"""
-Mental models module for Hindsight.
-
-Mental models contain directives - hard rules that are injected into reflect prompts.
-Directives are user-defined and their observations are user-provided (not LLM-generated).
-
-Other types of consolidated knowledge are handled by:
- Learnings: Automatic bottom-up consolidation from facts
- Pinned Reflections: User-curated living documents
-"""
-
-from .models import MentalModel, MentalModelSubtype
-
-__all__ = ["MentalModel", "MentalModelSubtype"]
@@ -1,53 +0,0 @@
-"""
-Pydantic models for mental models.
-"""
-
-from datetime import datetime, timezone
-from enum import Enum
-
-from pydantic import BaseModel, Field
-
-
-class MentalModelSubtype(str, Enum):
-    """Subtype of mental model.
-
-    Currently only DIRECTIVE is supported. Other types of consolidated knowledge
-    are handled by:
-    - Learnings: Automatic bottom-up consolidation from facts
-    - Pinned Reflections: User-curated living documents
-    """
-
-    DIRECTIVE = "directive"  # User-defined hard rules, observations user-provided
-
-
-class MentalModel(BaseModel):
-    """
-    A mental model representing synthesized understanding.
-
-    Mental models are the agent's consolidated knowledge. Unlike raw facts,
-    mental models provide:
-    - A one-liner description for quick scanning/retrieval
-    - A full summary for deep understanding
-    - Links to related mental models
-    """
-
-    id: str = Field(description="Unique identifier within the bank")
-    bank_id: str = Field(description="Bank this mental model belongs to")
-    subtype: MentalModelSubtype = Field(description="How this model was created")
-    name: str = Field(description="Human-readable name")
-    description: str = Field(description="One-liner for quick scanning and retrieval matching")
-    summary: str | None = Field(default=None, description="Full synthesized understanding")
-
-    # References
-    entity_id: str | None = Field(default=None, description="Reference to entities table when type=entity")
-    source_facts: list[str] = Field(default_factory=list, description="Fact IDs used to generate summary")
-    links: list[str] = Field(default_factory=list, description="Related mental model IDs")
-
-    # Tags for scoped visibility (similar to document tags)
-    tags: list[str] = Field(default_factory=list, description="Tags for scoped visibility filtering")
-
-    # Timestamps
-    last_updated: datetime | None = Field(default=None, description="When summary was last regenerated")
-    created_at: datetime = Field(
-        default_factory=lambda: datetime.now(timezone.utc), description="When this model was created"
-    )
@@ -84,7 +84,7 @@ class DateparserQueryAnalyzer(QueryAnalyzer):

    Performance:
    - ~10-50ms per query
-    - No model loading required (lazy import on first use)
+    - No model loading required
    """

    def __init__(self):
@@ -112,6 +112,8 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
        Returns:
            QueryAnalysis with temporal_constraint if found
        """
+        self.load()
+
        if reference_date is None:
            reference_date = datetime.now()

@@ -121,9 +123,6 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
        if period_result is not None:
            return QueryAnalysis(temporal_constraint=period_result)

-        # Lazy load dateparser (only imports on first call, then cached)
-        self.load()
-
        # Use dateparser's search_dates to find temporal expressions
        settings = {
            "RELATIVE_BASE": reference_date,
@@ -1,18 +0,0 @@
-"""
-Reflect agent module for agentic reflection with tools.
-
-The reflect agent uses an iterative loop with tools to:
-1. Lookup mental models (existing knowledge)
-2. Recall facts (semantic + temporal search)
-3. Expand memories (get chunk/document context)
-"""
-
-from .agent import ReflectAgentResult, run_reflect_agent
-from .models import ReflectAction, ReflectActionBatch
-
-__all__ = [
-    "run_reflect_agent",
-    "ReflectAgentResult",
-    "ReflectAction",
-    "ReflectActionBatch",
-]
@@ -1,933 +0,0 @@
-"""
-Reflect agent - agentic loop for reflection with native tool calling.
-
-Uses hierarchical retrieval:
-1. search_mental_models - User-curated summaries (highest quality)
-2. search_observations - Consolidated knowledge with freshness
-3. recall - Raw facts as ground truth
-"""
-
-import asyncio
-import json
-import logging
-import re
-import time
-from typing import TYPE_CHECKING, Any, Awaitable, Callable
-
-from .models import DirectiveInfo, LLMCall, ReflectAgentResult, TokenUsageSummary, ToolCall
-from .prompts import FINAL_SYSTEM_PROMPT, _extract_directive_rules, build_final_prompt, build_system_prompt_for_tools
-from .tools_schema import get_reflect_tools
-
-
-def _build_directives_applied(directives: list[dict[str, Any]] | None) -> list[DirectiveInfo]:
-    """Build list of DirectiveInfo from 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 []
-
-    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:
-    from ..llm_wrapper import LLMProvider
-    from ..response_models import LLMToolCall
-
-logger = logging.getLogger(__name__)
-
-DEFAULT_MAX_ITERATIONS = 10
-
-
-def _normalize_tool_name(name: str) -> str:
-    """Normalize tool name from various LLM output formats.
-
-    Some LLMs output tool names in non-standard formats:
-    - 'functions.done' (OpenAI-style prefix)
-    - 'call=functions.done' (some models)
-    - 'call=done' (some models)
-    - 'done<|channel|>commentary' (malformed special tokens appended)
-
-    Returns the normalized tool name (e.g., 'done', 'recall', etc.)
-    """
-    # Handle 'call=functions.name' or 'call=name' format
-    if name.startswith("call="):
-        name = name[len("call=") :]
-
-    # Handle 'functions.name' format
-    if name.startswith("functions."):
-        name = name[len("functions.") :]
-
-    # Handle malformed special tokens appended to tool name
-    # e.g., 'done<|channel|>commentary' -> 'done'
-    if "<|" in name:
-        name = name.split("<|")[0]
-
-    return name
-
-
-def _is_done_tool(name: str) -> bool:
-    """Check if the tool name represents the 'done' tool."""
-    return _normalize_tool_name(name) == "done"
-
-
-# Pattern to match done() call as text - handles done({...}) with nested JSON
-_DONE_CALL_PATTERN = re.compile(r"done\s*\(\s*\{.*$", re.DOTALL)
-
-# Patterns for leaked structured output in the answer field
-_LEAKED_JSON_SUFFIX = re.compile(
-    r'\s*```(?:json)?\s*\{[^}]*(?:"(?:observation_ids|memory_ids|mental_model_ids)"|\})\s*```\s*$',
-    re.DOTALL | re.IGNORECASE,
-)
-_LEAKED_JSON_OBJECT = re.compile(
-    r'\s*\{[^{]*"(?:observation_ids|memory_ids|mental_model_ids|answer)"[^}]*\}\s*$', re.DOTALL
-)
-_TRAILING_IDS_PATTERN = re.compile(
-    r"\s*(?:observation_ids|memory_ids|mental_model_ids)\s*[=:]\s*\[.*?\]\s*$", re.DOTALL | re.IGNORECASE
-)
-
-
-def _clean_answer_text(text: str) -> str:
-    """Clean up answer text by removing any done() tool call syntax.
-
-    Some LLMs output the done() call as text instead of a proper tool call.
-    This strips out patterns like: done({"answer": "...", ...})
-    """
-    # Remove done() call pattern from the end of the text
-    cleaned = _DONE_CALL_PATTERN.sub("", text).strip()
-    return cleaned if cleaned else text
-
-
-def _clean_done_answer(text: str) -> str:
-    """Clean up the answer field from a done() tool call.
-
-    Some LLMs leak structured output patterns into the answer text, such as:
-    - JSON code blocks with observation_ids/memory_ids at the end
-    - Raw JSON objects with these fields
-    - Plain text like "observation_ids: [...]"
-
-    This cleans those patterns while preserving the actual answer content.
-    """
-    if not text:
-        return text
-
-    cleaned = text
-
-    # Remove leaked JSON in code blocks at the end
-    cleaned = _LEAKED_JSON_SUFFIX.sub("", cleaned).strip()
-
-    # Remove leaked raw JSON objects at the end
-    cleaned = _LEAKED_JSON_OBJECT.sub("", cleaned).strip()
-
-    # Remove trailing ID patterns
-    cleaned = _TRAILING_IDS_PATTERN.sub("", cleaned).strip()
-
-    return cleaned if cleaned else text
-
-
-async def _generate_structured_output(
-    answer: str,
-    response_schema: dict,
-    llm_config: "LLMProvider",
-    reflect_id: str,
-) -> tuple[dict[str, Any] | None, int, int]:
-    """Generate structured output from an answer using the provided JSON schema.
-
-    Args:
-        answer: The text answer to extract structured data from
-        response_schema: JSON Schema for the expected output structure
-        llm_config: LLM provider for making the extraction call
-        reflect_id: Reflect ID for logging
-
-    Returns:
-        Tuple of (structured_output, input_tokens, output_tokens).
-        structured_output is None if generation fails.
-    """
-    try:
-        from typing import Any as TypingAny
-
-        from pydantic import create_model
-
-        def _json_schema_type_to_python(field_schema: dict) -> type:
-            """Map JSON schema type to Python type for better LLM guidance."""
-            json_type = field_schema.get("type", "string")
-            if json_type == "array":
-                return list
-            elif json_type == "object":
-                return dict
-            elif json_type == "integer":
-                return int
-            elif json_type == "number":
-                return float
-            elif json_type == "boolean":
-                return bool
-            else:
-                return str
-
-        # Build fields from JSON schema properties
-        schema_props = response_schema.get("properties", {})
-        required_fields = set(response_schema.get("required", []))
-        fields: dict[str, TypingAny] = {}
-        for field_name, field_schema in schema_props.items():
-            field_type = _json_schema_type_to_python(field_schema)
-            default = ... if field_name in required_fields else None
-            fields[field_name] = (field_type, default)
-
-        if not fields:
-            logger.warning(f"[REFLECT {reflect_id}] No fields found in response_schema, skipping structured output")
-            return None, 0, 0
-
-        DynamicModel = create_model("StructuredResponse", **fields)
-
-        # Include the full schema in the prompt for better LLM guidance
-        schema_str = json.dumps(response_schema, indent=2)
-
-        # Build field descriptions for the prompt
-        field_descriptions = []
-        for field_name, field_schema in schema_props.items():
-            field_type = field_schema.get("type", "string")
-            field_desc = field_schema.get("description", "")
-            is_required = field_name in required_fields
-            req_marker = " (REQUIRED)" if is_required else " (optional)"
-            field_descriptions.append(f"- {field_name} ({field_type}){req_marker}: {field_desc}")
-        fields_text = "\n".join(field_descriptions)
-
-        # Call LLM with the answer to extract structured data
-        structured_prompt = f"""Your task is to extract specific information from the answer below and format it as JSON.
-
-ANSWER TO EXTRACT FROM:
-\"\"\"
-{answer}
-\"\"\"
-
-REQUIRED OUTPUT FORMAT - Extract the following fields from the answer above:
-{fields_text}
-
-JSON Schema:
-```json
-{schema_str}
-```
-
-INSTRUCTIONS:
-1. Read the answer carefully and identify the information that matches each field
-2. Extract the ACTUAL content from the answer - do NOT leave fields empty if information is present
-3. For string fields: use the exact text or a clear summary from the answer
-4. For array fields: return a JSON array (e.g., ["item1", "item2"]), NOT a string
-5. For required fields: you MUST provide a value extracted from the answer
-6. Return ONLY the JSON object, no explanation
-
-OUTPUT:"""
-
-        structured_result, usage = await llm_config.call(
-            messages=[
-                {
-                    "role": "system",
-                    "content": "You are a precise data extraction assistant. Extract information from text and return it as valid JSON matching the provided schema. Always extract actual content - never return empty strings for required fields if information is available.",
-                },
-                {"role": "user", "content": structured_prompt},
-            ],
-            response_format=DynamicModel,
-            scope="reflect_structured",
-            skip_validation=True,  # We'll handle the dict ourselves
-            return_usage=True,
-        )
-
-        # Convert to dict
-        if hasattr(structured_result, "model_dump"):
-            structured_output = structured_result.model_dump()
-        elif isinstance(structured_result, dict):
-            structured_output = structured_result
-        else:
-            # Try to parse as JSON
-            structured_output = json.loads(str(structured_result))
-
-        # Validate that required fields have non-empty values
-        for field_name in required_fields:
-            value = structured_output.get(field_name)
-            if value is None or value == "" or value == []:
-                logger.warning(f"[REFLECT {reflect_id}] Required field '{field_name}' is empty in structured output")
-
-        logger.info(f"[REFLECT {reflect_id}] Generated structured output with {len(structured_output)} fields")
-        return structured_output, usage.input_tokens, usage.output_tokens
-
-    except Exception as e:
-        logger.warning(f"[REFLECT {reflect_id}] Failed to generate structured output: {e}")
-        return None, 0, 0
-
-
-async def run_reflect_agent(
-    llm_config: "LLMProvider",
-    bank_id: str,
-    query: str,
-    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], Awaitable[dict[str, Any]]],
-    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
-    context: str | None = None,
-    max_iterations: int = DEFAULT_MAX_ITERATIONS,
-    max_tokens: int | None = None,
-    response_schema: dict | None = None,
-    directives: list[dict[str, Any]] | None = None,
-    has_mental_models: bool = False,
-    budget: str | None = None,
-) -> ReflectAgentResult:
-    """
-    Execute the reflect agent loop using native tool calling.
-
-    The agent uses hierarchical retrieval:
-    1. search_mental_models - User-curated summaries (try first)
-    2. search_observations - Consolidated knowledge with freshness
-    3. recall - Raw facts as ground truth
-
-    Args:
-        llm_config: LLM provider for agent calls
-        bank_id: Bank identifier
-        query: Question to answer
-        bank_profile: Bank profile with name and mission
-        search_mental_models_fn: Tool callback for searching mental models (query, max_results) -> result
-        search_observations_fn: Tool callback for searching observations (query, max_results) -> result
-        recall_fn: Tool callback for recall (query, max_tokens) -> result
-        expand_fn: Tool callback for expand (memory_ids, depth) -> result
-        context: Optional additional context
-        max_iterations: Maximum number of iterations before forcing response
-        max_tokens: Maximum tokens for the final response
-        response_schema: Optional JSON Schema for structured output in final response
-        directives: Optional list of directive mental models to inject as hard rules
-
-    Returns:
-        ReflectAgentResult with final answer and metadata
-    """
-    reflect_id = f"{bank_id[:8]}-{int(time.time() * 1000) % 100000}"
-    start_time = time.time()
-
-    # Build directives_applied for the trace
-    directives_applied = _build_directives_applied(directives)
-
-    # Extract directive rules for tool schema (if any)
-    directive_rules = _extract_directive_rules(directives) if directives else None
-
-    # Get tools for this agent (with directive compliance field if directives exist)
-    tools = get_reflect_tools(directive_rules=directive_rules)
-
-    # Build initial messages (directives are injected into system prompt at START and END)
-    system_prompt = build_system_prompt_for_tools(
-        bank_profile, context, directives=directives, has_mental_models=has_mental_models, budget=budget
-    )
-    messages: list[dict[str, Any]] = [
-        {"role": "system", "content": system_prompt},
-        {"role": "user", "content": query},
-    ]
-
-    # Tracking
-    total_tools_called = 0
-    tool_trace: list[ToolCall] = []
-    tool_trace_summary: list[dict[str, Any]] = []
-    llm_trace: list[dict[str, Any]] = []
-    context_history: list[dict[str, Any]] = []  # For final prompt fallback
-
-    # Token usage tracking - accumulate across all LLM calls
-    total_input_tokens = 0
-    total_output_tokens = 0
-
-    # Track available IDs for validation (prevents hallucinated citations)
-    available_memory_ids: set[str] = set()
-    available_mental_model_ids: set[str] = set()
-    available_observation_ids: set[str] = set()
-
-    def _get_llm_trace() -> list[LLMCall]:
-        return [
-            LLMCall(
-                scope=c["scope"],
-                duration_ms=c["duration_ms"],
-                input_tokens=c.get("input_tokens", 0),
-                output_tokens=c.get("output_tokens", 0),
-            )
-            for c in llm_trace
-        ]
-
-    def _get_usage() -> TokenUsageSummary:
-        return TokenUsageSummary(
-            input_tokens=total_input_tokens,
-            output_tokens=total_output_tokens,
-            total_tokens=total_input_tokens + total_output_tokens,
-        )
-
-    def _log_completion(answer: str, iterations: int, forced: bool = False):
-        elapsed_ms = int((time.time() - start_time) * 1000)
-        tools_summary = (
-            ", ".join(
-                f"{t['tool']}({t['input_summary']})={t['duration_ms']}ms/{t.get('output_chars', 0)}c"
-                for t in tool_trace_summary
-            )
-            or "none"
-        )
-        llm_summary = ", ".join(f"{c['scope']}={c['duration_ms']}ms" for c in llm_trace) or "none"
-        total_llm_ms = sum(c["duration_ms"] for c in llm_trace)
-        total_tools_ms = sum(t["duration_ms"] for t in tool_trace_summary)
-
-        answer_preview = answer[:100] + "..." if len(answer) > 100 else answer
-        mode = "forced" if forced else "done"
-        logger.info(
-            f"[REFLECT {reflect_id}] {mode} | "
-            f"query='{query[:50]}...' | "
-            f"iterations={iterations} | "
-            f"llm=[{llm_summary}] ({total_llm_ms}ms) | "
-            f"tools=[{tools_summary}] ({total_tools_ms}ms) | "
-            f"answer='{answer_preview}' | "
-            f"total={elapsed_ms}ms"
-        )
-
-    for iteration in range(max_iterations):
-        is_last = iteration == max_iterations - 1
-
-        if is_last:
-            # Force text response on last iteration - no tools
-            prompt = build_final_prompt(query, context_history, bank_profile, context)
-            llm_start = time.time()
-            response, usage = await llm_config.call(
-                messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
-                    {"role": "user", "content": prompt},
-                ],
-                scope="reflect_agent_final",
-                max_completion_tokens=max_tokens,
-                return_usage=True,
-            )
-            llm_duration = int((time.time() - llm_start) * 1000)
-            total_input_tokens += usage.input_tokens
-            total_output_tokens += usage.output_tokens
-            llm_trace.append(
-                {
-                    "scope": "final",
-                    "duration_ms": llm_duration,
-                    "input_tokens": usage.input_tokens,
-                    "output_tokens": usage.output_tokens,
-                }
-            )
-            answer = _clean_answer_text(response.strip())
-
-            # Generate structured output if schema provided
-            structured_output = None
-            if response_schema and answer:
-                structured_output, struct_in, struct_out = await _generate_structured_output(
-                    answer, response_schema, llm_config, reflect_id
-                )
-                total_input_tokens += struct_in
-                total_output_tokens += struct_out
-
-            _log_completion(answer, iteration + 1, forced=True)
-            return ReflectAgentResult(
-                text=answer,
-                structured_output=structured_output,
-                iterations=iteration + 1,
-                tools_called=total_tools_called,
-                tool_trace=tool_trace,
-                llm_trace=_get_llm_trace(),
-                usage=_get_usage(),
-                directives_applied=directives_applied,
-            )
-
-        # Call LLM with tools
-        llm_start = time.time()
-
-        try:
-            result = await llm_config.call_with_tools(
-                messages=messages,
-                tools=tools,
-                scope="reflect_agent",
-                tool_choice="required" if iteration == 0 else "auto",  # Force tool use on first iteration
-            )
-            llm_duration = int((time.time() - llm_start) * 1000)
-            total_input_tokens += result.input_tokens
-            total_output_tokens += result.output_tokens
-            llm_trace.append(
-                {
-                    "scope": f"agent_{iteration + 1}",
-                    "duration_ms": llm_duration,
-                    "input_tokens": result.input_tokens,
-                    "output_tokens": result.output_tokens,
-                }
-            )
-
-        except Exception as e:
-            err_duration = int((time.time() - llm_start) * 1000)
-            logger.warning(f"[REFLECT {reflect_id}] LLM error on iteration {iteration + 1}: {e} ({err_duration}ms)")
-            llm_trace.append({"scope": f"agent_{iteration + 1}_err", "duration_ms": err_duration})
-            # Guardrail: If no evidence gathered yet, retry
-            has_gathered_evidence = (
-                bool(available_memory_ids) or bool(available_mental_model_ids) or bool(available_observation_ids)
-            )
-            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()
-            response, usage = await llm_config.call(
-                messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
-                    {"role": "user", "content": prompt},
-                ],
-                scope="reflect_agent_final",
-                max_completion_tokens=max_tokens,
-                return_usage=True,
-            )
-            llm_duration = int((time.time() - llm_start) * 1000)
-            total_input_tokens += usage.input_tokens
-            total_output_tokens += usage.output_tokens
-            llm_trace.append(
-                {
-                    "scope": "final",
-                    "duration_ms": llm_duration,
-                    "input_tokens": usage.input_tokens,
-                    "output_tokens": usage.output_tokens,
-                }
-            )
-            answer = _clean_answer_text(response.strip())
-
-            # Generate structured output if schema provided
-            structured_output = None
-            if response_schema and answer:
-                structured_output, struct_in, struct_out = await _generate_structured_output(
-                    answer, response_schema, llm_config, reflect_id
-                )
-                total_input_tokens += struct_in
-                total_output_tokens += struct_out
-
-            _log_completion(answer, iteration + 1, forced=True)
-            return ReflectAgentResult(
-                text=answer,
-                structured_output=structured_output,
-                iterations=iteration + 1,
-                tools_called=total_tools_called,
-                tool_trace=tool_trace,
-                llm_trace=_get_llm_trace(),
-                usage=_get_usage(),
-                directives_applied=directives_applied,
-            )
-
-        # No tool calls - LLM wants to respond with text
-        if not result.tool_calls:
-            if result.content:
-                answer = _clean_answer_text(result.content.strip())
-
-                # Generate structured output if schema provided
-                structured_output = None
-                if response_schema and answer:
-                    structured_output, struct_in, struct_out = await _generate_structured_output(
-                        answer, response_schema, llm_config, reflect_id
-                    )
-                    total_input_tokens += struct_in
-                    total_output_tokens += struct_out
-
-                _log_completion(answer, iteration + 1)
-                return ReflectAgentResult(
-                    text=answer,
-                    structured_output=structured_output,
-                    iterations=iteration + 1,
-                    tools_called=total_tools_called,
-                    tool_trace=tool_trace,
-                    llm_trace=_get_llm_trace(),
-                    usage=_get_usage(),
-                    directives_applied=directives_applied,
-                )
-            # Empty response, force final
-            prompt = build_final_prompt(query, context_history, bank_profile, context)
-            llm_start = time.time()
-            response, usage = await llm_config.call(
-                messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
-                    {"role": "user", "content": prompt},
-                ],
-                scope="reflect_agent_final",
-                max_completion_tokens=max_tokens,
-                return_usage=True,
-            )
-            llm_duration = int((time.time() - llm_start) * 1000)
-            total_input_tokens += usage.input_tokens
-            total_output_tokens += usage.output_tokens
-            llm_trace.append(
-                {
-                    "scope": "final",
-                    "duration_ms": llm_duration,
-                    "input_tokens": usage.input_tokens,
-                    "output_tokens": usage.output_tokens,
-                }
-            )
-            answer = _clean_answer_text(response.strip())
-
-            # Generate structured output if schema provided
-            structured_output = None
-            if response_schema and answer:
-                structured_output, struct_in, struct_out = await _generate_structured_output(
-                    answer, response_schema, llm_config, reflect_id
-                )
-                total_input_tokens += struct_in
-                total_output_tokens += struct_out
-
-            _log_completion(answer, iteration + 1, forced=True)
-            return ReflectAgentResult(
-                text=answer,
-                structured_output=structured_output,
-                iterations=iteration + 1,
-                tools_called=total_tools_called,
-                tool_trace=tool_trace,
-                llm_trace=_get_llm_trace(),
-                usage=_get_usage(),
-                directives_applied=directives_applied,
-            )
-
-        # Check for done tool call (handle various LLM output formats)
-        done_call = next((tc for tc in result.tool_calls if _is_done_tool(tc.name)), None)
-        if done_call:
-            # Guardrail: Require evidence before done
-            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:
-                # Add assistant message and fake tool result asking for evidence
-                messages.append(
-                    {
-                        "role": "assistant",
-                        "tool_calls": [_tool_call_to_dict(done_call)],
-                    }
-                )
-                messages.append(
-                    {
-                        "role": "tool",
-                        "tool_call_id": done_call.id,
-                        "name": done_call.name,  # Required by Gemini
-                        "content": json.dumps(
-                            {
-                                "error": "You must search for information first. Use search_mental_models(), search_observations(), or recall() before providing your final answer."
-                            }
-                        ),
-                    }
-                )
-                continue
-
-            # Process done tool
-            return await _process_done_tool(
-                done_call,
-                available_memory_ids,
-                available_mental_model_ids,
-                available_observation_ids,
-                iteration + 1,
-                total_tools_called,
-                tool_trace,
-                _get_llm_trace(),
-                _get_usage(),
-                _log_completion,
-                reflect_id,
-                directives_applied=directives_applied,
-                llm_config=llm_config,
-                response_schema=response_schema,
-            )
-
-        # Execute other tools in parallel (exclude done tool in all its format variants)
-        other_tools = [tc for tc in result.tool_calls if not _is_done_tool(tc.name)]
-        if other_tools:
-            # Add assistant message with tool calls
-            messages.append(
-                {
-                    "role": "assistant",
-                    "tool_calls": [_tool_call_to_dict(tc) for tc in other_tools],
-                }
-            )
-
-            # Execute tools in parallel
-            tool_tasks = [
-                _execute_tool_with_timing(
-                    tc,
-                    search_mental_models_fn,
-                    search_observations_fn,
-                    recall_fn,
-                    expand_fn,
-                )
-                for tc in other_tools
-            ]
-            tool_results = await asyncio.gather(*tool_tasks, return_exceptions=True)
-            total_tools_called += len(other_tools)
-
-            # Process results and add to messages
-            for tc, result_data in zip(other_tools, tool_results):
-                if isinstance(result_data, Exception):
-                    # Tool execution failed - send error back to LLM so it can try again
-                    logger.warning(f"[REFLECT {reflect_id}] Tool {tc.name} failed with exception: {result_data}")
-                    output = {"error": f"Tool execution failed: {result_data}"}
-                    duration_ms = 0
-                else:
-                    output, duration_ms = result_data
-
-                # Normalize tool name for consistent tracking
-                normalized_tool_name = _normalize_tool_name(tc.name)
-
-                # Check if tool returned an error response - log but continue (LLM will see the error)
-                if isinstance(output, dict) and "error" in output:
-                    logger.warning(
-                        f"[REFLECT {reflect_id}] Tool {normalized_tool_name} returned error: {output['error']}"
-                    )
-
-                # Track available IDs from tool results (only for successful responses)
-                if (
-                    normalized_tool_name == "search_mental_models"
-                    and isinstance(output, dict)
-                    and "mental_models" in output
-                ):
-                    for mm in output["mental_models"]:
-                        if "id" in mm:
-                            available_mental_model_ids.add(mm["id"])
-
-                if (
-                    normalized_tool_name == "search_observations"
-                    and isinstance(output, dict)
-                    and "observations" in output
-                ):
-                    for obs in output["observations"]:
-                        if "id" in obs:
-                            available_observation_ids.add(obs["id"])
-
-                if normalized_tool_name == "recall" and isinstance(output, dict) and "memories" in output:
-                    for memory in output["memories"]:
-                        if "id" in memory:
-                            available_memory_ids.add(memory["id"])
-
-                # Add tool result message
-                messages.append(
-                    {
-                        "role": "tool",
-                        "tool_call_id": tc.id,
-                        "name": tc.name,  # Required by Gemini
-                        "content": json.dumps(output, default=str),
-                    }
-                )
-
-                # Track for logging and context history
-                input_dict = {"tool": tc.name, **tc.arguments}
-                input_summary = _summarize_input(tc.name, tc.arguments)
-
-                # Extract reason from tool arguments (if provided)
-                tool_reason = tc.arguments.get("reason")
-
-                tool_trace.append(
-                    ToolCall(
-                        tool=tc.name,
-                        reason=tool_reason,
-                        input=input_dict,
-                        output=output,
-                        duration_ms=duration_ms,
-                        iteration=iteration + 1,
-                    )
-                )
-
-                try:
-                    output_chars = len(json.dumps(output))
-                except (TypeError, ValueError):
-                    output_chars = len(str(output))
-
-                tool_trace_summary.append(
-                    {
-                        "tool": tc.name,
-                        "input_summary": input_summary,
-                        "duration_ms": duration_ms,
-                        "output_chars": output_chars,
-                    }
-                )
-
-                # Keep context history for fallback final prompt
-                context_history.append({"tool": tc.name, "input": input_dict, "output": output})
-
-    # Should not reach here
-    answer = "I was unable to formulate a complete answer within the iteration limit."
-    _log_completion(answer, max_iterations, forced=True)
-    return ReflectAgentResult(
-        text=answer,
-        iterations=max_iterations,
-        tools_called=total_tools_called,
-        tool_trace=tool_trace,
-        llm_trace=_get_llm_trace(),
-        usage=_get_usage(),
-        directives_applied=directives_applied,
-    )
-
-
-def _tool_call_to_dict(tc: "LLMToolCall") -> dict[str, Any]:
-    """Convert LLMToolCall to OpenAI message format."""
-    return {
-        "id": tc.id,
-        "type": "function",
-        "function": {
-            "name": tc.name,
-            "arguments": json.dumps(tc.arguments),
-        },
-    }
-
-
-async def _process_done_tool(
-    done_call: "LLMToolCall",
-    available_memory_ids: set[str],
-    available_mental_model_ids: set[str],
-    available_observation_ids: set[str],
-    iterations: int,
-    total_tools_called: int,
-    tool_trace: list[ToolCall],
-    llm_trace: list[LLMCall],
-    usage: TokenUsageSummary,
-    log_completion: Callable,
-    reflect_id: str,
-    directives_applied: list[DirectiveInfo],
-    llm_config: "LLMProvider | None" = None,
-    response_schema: dict | None = None,
-) -> ReflectAgentResult:
-    """Process the done tool call and return the result."""
-    args = done_call.arguments
-
-    # Extract and clean the answer - some LLMs leak structured output into the answer text
-    raw_answer = args.get("answer", "").strip()
-    answer = _clean_done_answer(raw_answer) if raw_answer else ""
-    if not answer:
-        answer = "No answer provided."
-
-    # Validate IDs (only include IDs that were actually retrieved)
-    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
-    final_usage = usage
-    if response_schema and llm_config and answer:
-        structured_output, struct_in, struct_out = await _generate_structured_output(
-            answer, response_schema, llm_config, reflect_id
-        )
-        # Add structured output tokens to usage
-        final_usage = TokenUsageSummary(
-            input_tokens=usage.input_tokens + struct_in,
-            output_tokens=usage.output_tokens + struct_out,
-            total_tokens=usage.total_tokens + struct_in + struct_out,
-        )
-
-    log_completion(answer, iterations)
-    return ReflectAgentResult(
-        text=answer,
-        structured_output=structured_output,
-        iterations=iterations,
-        tools_called=total_tools_called,
-        tool_trace=tool_trace,
-        llm_trace=llm_trace,
-        usage=final_usage,
-        used_memory_ids=used_memory_ids,
-        used_mental_model_ids=used_mental_model_ids,
-        used_observation_ids=used_observation_ids,
-        directives_applied=directives_applied,
-    )
-
-
-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], 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."""
-    start = time.time()
-    result = await _execute_tool(
-        tc.name,
-        tc.arguments,
-        search_mental_models_fn,
-        search_observations_fn,
-        recall_fn,
-        expand_fn,
-    )
-    duration_ms = int((time.time() - start) * 1000)
-    return result, duration_ms
-
-
-async def _execute_tool(
-    tool_name: str,
-    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], Awaitable[dict[str, Any]]],
-    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
-) -> dict[str, Any]:
-    """Execute a single tool by name."""
-    # Normalize tool name for various LLM output formats
-    tool_name = _normalize_tool_name(tool_name)
-
-    if tool_name == "search_mental_models":
-        query = args.get("query")
-        if not query:
-            return {"error": "search_mental_models requires a query parameter"}
-        max_results = args.get("max_results") or 5
-        return await search_mental_models_fn(query, max_results)
-
-    elif tool_name == "search_observations":
-        query = args.get("query")
-        if not query:
-            return {"error": "search_observations requires a query parameter"}
-        max_tokens = max(args.get("max_tokens") or 5000, 1000)  # Default 5000, min 1000
-        return await search_observations_fn(query, max_tokens)
-
-    elif tool_name == "recall":
-        query = args.get("query")
-        if not query:
-            return {"error": "recall requires a query parameter"}
-        max_tokens = max(args.get("max_tokens") or 2048, 1000)  # Default 2048, min 1000
-        return await recall_fn(query, max_tokens)
-
-    elif tool_name == "expand":
-        memory_ids = args.get("memory_ids", [])
-        if not memory_ids:
-            return {"error": "expand requires memory_ids"}
-        depth = args.get("depth", "chunk")
-        return await expand_fn(memory_ids, depth)
-
-    else:
-        return {"error": f"Unknown tool: {tool_name}"}
-
-
-def _summarize_input(tool_name: str, args: dict[str, Any]) -> str:
-    """Create a summary of tool input for logging, showing all params."""
-    if tool_name == "search_mental_models":
-        query = args.get("query", "")
-        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
-        max_results = args.get("max_results") or 5
-        return f"(query={query_preview}, max_results={max_results})"
-    elif tool_name == "search_observations":
-        query = args.get("query", "")
-        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
-        max_tokens = max(args.get("max_tokens") or 5000, 1000)
-        return f"(query={query_preview}, max_tokens={max_tokens})"
-    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(args.get("max_tokens") or 2048, 1000)
-        return f"(query={query_preview}, max_tokens={max_tokens})"
-    elif tool_name == "expand":
-        memory_ids = args.get("memory_ids", [])
-        depth = args.get("depth", "chunk")
-        return f"(memory_ids=[{len(memory_ids)} ids], depth={depth})"
-    elif tool_name == "done":
-        answer = args.get("answer", "")
-        answer_preview = f"'{answer[:30]}...'" if len(answer) > 30 else f"'{answer}'"
-        memory_ids = args.get("memory_ids", [])
-        mental_model_ids = args.get("mental_model_ids", [])
-        observation_ids = args.get("observation_ids", [])
-        return (
-            f"(answer={answer_preview}, mem={len(memory_ids)}, mm={len(mental_model_ids)}, obs={len(observation_ids)})"
-        )
-    return str(args)
@@ -1,109 +0,0 @@
-"""
-Pydantic models for the reflect agent.
-"""
-
-from typing import Any, Literal
-
-from pydantic import BaseModel, Field
-
-
-class ObservationSection(BaseModel):
-    """A section within an observation with its supporting memories."""
-
-    title: str = Field(description="Section header (can be empty for intro)")
-    text: str = Field(description="Section content - no headers, use lists/tables/bold")
-    memory_ids: list[str] = Field(default_factory=list, description="Memory IDs supporting this section")
-
-
-class ReflectAction(BaseModel):
-    """Single action the reflect agent can take."""
-
-    tool: Literal["list_observations", "get_observation", "recall", "expand", "done"] = Field(
-        description="Tool to invoke: list_observations, get_observation, recall, expand, or done"
-    )
-    # Tool-specific parameters
-    observation_id: str | None = Field(default=None, description="Observation ID for get_observation")
-    query: str | None = Field(default=None, description="Search query for recall")
-    max_tokens: int | None = Field(default=None, description="Max tokens for recall results (default 2048)")
-    memory_ids: list[str] | None = Field(default=None, description="Memory unit IDs for expand (batched)")
-    depth: Literal["chunk", "document"] | None = Field(default=None, description="Expansion depth for expand")
-    observation_sections: list[ObservationSection] | None = Field(
-        default=None, description="Observation sections for done action (when output_mode=observations)"
-    )
-    # Plain text answer fields (for output_mode=answer)
-    answer: str | None = Field(default=None, description="Plain text answer for done action (no markdown)")
-    answer_memory_ids: list[str] | None = Field(
-        default=None, description="Memory IDs supporting the answer", alias="memory_ids"
-    )
-    answer_model_ids: list[str] | None = Field(
-        default=None, description="Mental model IDs supporting the answer", alias="model_ids"
-    )
-    reasoning: str | None = Field(default=None, description="Brief reasoning for this action")
-
-
-class ReflectActionBatch(BaseModel):
-    """Batch of actions for parallel execution."""
-
-    actions: list[ReflectAction] = Field(description="List of actions to execute in parallel")
-
-
-class ToolCall(BaseModel):
-    """A single tool call made during reflect."""
-
-    tool: str = Field(description="Tool name: lookup, recall, expand")
-    reason: str | None = Field(default=None, description="Agent's reasoning for making this tool call")
-    input: dict = Field(description="Tool input parameters")
-    output: dict = Field(description="Tool output/result")
-    duration_ms: int = Field(description="Execution time in milliseconds")
-    iteration: int = Field(default=0, description="Iteration number (1-based) when this tool was called")
-
-
-class LLMCall(BaseModel):
-    """A single LLM call made during reflect."""
-
-    scope: str = Field(description="Call scope: agent_1, agent_2, final, etc.")
-    duration_ms: int = Field(description="Execution time in milliseconds")
-    input_tokens: int = Field(default=0, description="Input tokens used")
-    output_tokens: int = Field(default=0, description="Output tokens used")
-
-
-class DirectiveInfo(BaseModel):
-    """Information about a directive that was applied during reflect."""
-
-    id: str = Field(description="Directive mental model ID")
-    name: str = Field(description="Directive name")
-    content: str = Field(description="Directive content")
-
-
-class TokenUsageSummary(BaseModel):
-    """Total token usage across all LLM calls."""
-
-    input_tokens: int = Field(default=0, description="Total input tokens used")
-    output_tokens: int = Field(default=0, description="Total output tokens used")
-    total_tokens: int = Field(default=0, description="Total tokens (input + output)")
-
-
-class ReflectAgentResult(BaseModel):
-    """Result from the reflect agent."""
-
-    text: str = Field(description="Final answer text")
-    structured_output: dict[str, Any] | None = Field(
-        default=None, description="Structured output parsed according to provided response_schema"
-    )
-    iterations: int = Field(default=0, description="Number of iterations taken")
-    tools_called: int = Field(default=0, description="Total number of tool calls made")
-    tool_trace: list[ToolCall] = Field(default_factory=list, description="Trace of all tool calls made")
-    llm_trace: list[LLMCall] = Field(default_factory=list, description="Trace of all LLM calls made")
-    usage: TokenUsageSummary = Field(
-        default_factory=TokenUsageSummary, description="Total token usage across all LLM calls"
-    )
-    used_memory_ids: list[str] = Field(default_factory=list, description="Validated memory IDs actually used in answer")
-    used_mental_model_ids: list[str] = Field(
-        default_factory=list, description="Validated mental model IDs actually used in answer"
-    )
-    used_observation_ids: list[str] = Field(
-        default_factory=list, description="Validated observation IDs actually used in answer"
-    )
-    directives_applied: list[DirectiveInfo] = Field(
-        default_factory=list, description="Directive mental models that affected this reflection"
-    )
@@ -1,186 +0,0 @@
-"""
-Models and utilities for evidence-grounded observations with computed trends.
-
-Observations are part of mental models and represent patterns/beliefs derived
-from memories. Each observation must be grounded in specific evidence (quotes)
-from memories, and trends are computed algorithmically from evidence timestamps.
-"""
-
-from datetime import datetime, timedelta, timezone
-from enum import Enum
-
-from pydantic import BaseModel, Field, computed_field, field_validator
-
-
-class Trend(str, Enum):
-    """Computed trend for an observation based on evidence timestamps.
-
-    Trends indicate how an observation's evidence is distributed over time:
-    - STABLE: Evidence spread across time, continues to present
-    - STRENGTHENING: More/denser evidence recently than before
-    - WEAKENING: Evidence mostly old, sparse recently
-    - NEW: All evidence within recent window
-    - STALE: No evidence in recent window (may no longer apply)
-    """
-
-    STABLE = "stable"
-    STRENGTHENING = "strengthening"
-    WEAKENING = "weakening"
-    NEW = "new"
-    STALE = "stale"
-
-
-class ObservationEvidence(BaseModel):
-    """A single piece of evidence supporting an observation.
-
-    Each evidence item must include an exact quote from the source memory
-    to ensure observations are grounded and verifiable.
-    """
-
-    memory_id: str = Field(description="ID of the memory unit this evidence comes from")
-    quote: str = Field(description="Exact quote from the memory supporting the observation")
-    relevance: str = Field(default="", description="Brief explanation of how this quote supports the observation")
-    timestamp: datetime = Field(description="When the source memory was created")
-
-    @field_validator("timestamp", mode="before")
-    @classmethod
-    def ensure_timezone_aware(cls, v: datetime | str | None) -> datetime:
-        """Ensure timestamp is always timezone-aware UTC."""
-        if v is None:
-            return datetime.now(timezone.utc)
-        if isinstance(v, str):
-            # Parse ISO format string, handling 'Z' suffix
-            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
-        if isinstance(v, datetime):
-            if v.tzinfo is None:
-                return v.replace(tzinfo=timezone.utc)
-            return v
-        raise ValueError(f"Invalid timestamp type: {type(v)}")
-
-
-class Observation(BaseModel):
-    """A single observation within a mental model.
-
-    Observations represent patterns, preferences, beliefs, or other insights
-    derived from memories. Each observation must be grounded in evidence
-    with exact quotes from source memories.
-    """
-
-    title: str = Field(description="Short summary title for the observation (5-10 words)")
-    content: str = Field(description="The observation content - detailed explanation of what we believe to be true")
-    evidence: list[ObservationEvidence] = Field(default_factory=list, description="Supporting evidence with quotes")
-    created_at: datetime = Field(
-        default_factory=lambda: datetime.now(timezone.utc), description="When this observation was first created"
-    )
-
-    @field_validator("created_at", mode="before")
-    @classmethod
-    def ensure_created_at_timezone_aware(cls, v: datetime | str | None) -> datetime:
-        """Ensure created_at is always timezone-aware UTC."""
-        if v is None:
-            return datetime.now(timezone.utc)
-        if isinstance(v, str):
-            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
-        if isinstance(v, datetime):
-            if v.tzinfo is None:
-                return v.replace(tzinfo=timezone.utc)
-            return v
-        raise ValueError(f"Invalid created_at type: {type(v)}")
-
-    @computed_field
-    @property
-    def trend(self) -> Trend:
-        """Compute trend from evidence timestamps."""
-        return compute_trend(self.evidence)
-
-    @computed_field
-    @property
-    def evidence_span(self) -> dict[str, str | None]:
-        """Get the time span covered by evidence."""
-        if not self.evidence:
-            return {"from": None, "to": None}
-        timestamps = [e.timestamp for e in self.evidence]
-        return {
-            "from": min(timestamps).isoformat(),
-            "to": max(timestamps).isoformat(),
-        }
-
-    @computed_field
-    @property
-    def evidence_count(self) -> int:
-        """Number of evidence items supporting this observation."""
-        return len(self.evidence)
-
-
-def compute_trend(
-    evidence: list[ObservationEvidence],
-    now: datetime | None = None,
-    recent_days: int = 30,
-    old_days: int = 90,
-) -> Trend:
-    """Compute the trend for an observation based on evidence timestamps.
-
-    The trend indicates how the evidence is distributed over time:
-    - STABLE: Evidence spread across time, continues to present
-    - STRENGTHENING: More evidence recently than historically
-    - WEAKENING: Evidence mostly old, sparse recently
-    - NEW: All evidence is recent (within recent_days)
-    - STALE: No evidence in recent window
-
-    Args:
-        evidence: List of evidence items with timestamps
-        now: Reference time for calculations (defaults to current UTC time)
-        recent_days: Number of days to consider "recent" (default 30)
-        old_days: Number of days to consider "old" (default 90)
-
-    Returns:
-        Computed Trend enum value
-    """
-    if now is None:
-        now = datetime.now(timezone.utc)
-
-    # Ensure now is timezone-aware
-    if now.tzinfo is None:
-        now = now.replace(tzinfo=timezone.utc)
-
-    if not evidence:
-        return Trend.STALE
-
-    recent_cutoff = now - timedelta(days=recent_days)
-    old_cutoff = now - timedelta(days=old_days)
-
-    # Normalize timestamps to UTC for comparison
-    def normalize_ts(ts: datetime) -> datetime:
-        if ts.tzinfo is None:
-            return ts.replace(tzinfo=timezone.utc)
-        return ts
-
-    recent = [e for e in evidence if normalize_ts(e.timestamp) > recent_cutoff]
-    old = [e for e in evidence if normalize_ts(e.timestamp) < old_cutoff]
-    middle = [e for e in evidence if old_cutoff <= normalize_ts(e.timestamp) <= recent_cutoff]
-
-    # No recent evidence = stale
-    if not recent:
-        return Trend.STALE
-
-    # All evidence is recent = new
-    if not old and not middle:
-        return Trend.NEW
-
-    # Compare density (evidence per day)
-    recent_density = len(recent) / recent_days if recent_days > 0 else 0
-    older_period = old_days - recent_days
-    older_density = (len(old) + len(middle)) / older_period if older_period > 0 else 0
-
-    # Avoid division by zero
-    if older_density == 0:
-        return Trend.NEW
-
-    ratio = recent_density / older_density
-
-    if ratio > 1.5:
-        return Trend.STRENGTHENING
-    elif ratio < 0.5:
-        return Trend.WEAKENING
-    else:
-        return Trend.STABLE
@@ -1,483 +0,0 @@
-"""
-System prompts for the reflect agent.
-
-The reflect agent uses hierarchical retrieval:
-1. search_mental_models - User-curated summaries (highest quality)
-2. search_observations - Consolidated knowledge with freshness awareness
-3. recall - Raw facts as ground truth fallback
-"""
-
-import json
-from typing import Any
-
-
-def _extract_directive_rules(directives: list[dict[str, Any]]) -> list[str]:
-    """
-    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:
-        directive_name = directive.get("name", "")
-        # New format: directives have direct content field
-        content = directive.get("content", "")
-        if 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.
-
-    Directives are hard rules that MUST be followed in all responses.
-
-    Args:
-        directives: List of directive mental models with observations
-    """
-    if not directives:
-        return ""
-
-    rules = _extract_directive_rules(directives)
-    if not rules:
-        return ""
-
-    parts = [
-        "## DIRECTIVES (MANDATORY)",
-        "These are hard rules you MUST follow in ALL responses:",
-        "",
-    ]
-
-    for rule in rules:
-        parts.append(f"- {rule}")
-
-    parts.extend(
-        [
-            "",
-            "NEVER violate these directives, even if other context suggests otherwise.",
-            "IMPORTANT: Do NOT explain or justify how you handled directives in your answer. Just follow them silently.",
-            "",
-        ]
-    )
-    return "\n".join(parts)
-
-
-def build_directives_reminder(directives: list[dict[str, Any]]) -> str:
-    """
-    Build a reminder section for directives to place at the end of the prompt.
-
-    Args:
-        directives: List of directive mental models with observations
-    """
-    if not directives:
-        return ""
-
-    rules = _extract_directive_rules(directives)
-    if not rules:
-        return ""
-
-    parts = [
-        "",
-        "## REMINDER: MANDATORY DIRECTIVES",
-        "Before responding, ensure your answer complies with ALL of these directives:",
-        "",
-    ]
-
-    for i, rule in enumerate(rules, 1):
-        parts.append(f"{i}. {rule}")
-
-    parts.append("")
-    parts.append("Your response will be REJECTED if it violates any directive above.")
-    parts.append("Do NOT include any commentary about how you handled directives - just follow them.")
-    return "\n".join(parts)
-
-
-def build_system_prompt_for_tools(
-    bank_profile: dict[str, Any],
-    context: str | None = None,
-    directives: list[dict[str, Any]] | None = None,
-    has_mental_models: bool = False,
-    budget: str | None = None,
-) -> str:
-    """
-    Build the system prompt for tool-calling reflect agent.
-
-    The agent uses hierarchical retrieval:
-    1. search_mental_models - User-curated summaries (try first, if available)
-    2. search_observations - Consolidated knowledge with freshness
-    3. recall - Raw facts as ground truth
-
-    Args:
-        bank_profile: Bank profile with name and mission
-        context: Optional additional context
-        directives: Optional list of directive mental models to inject as hard rules
-        has_mental_models: Whether the bank has any mental models (skip if not)
-        budget: Search depth budget - "low", "mid", or "high". Controls exploration thoroughness.
-    """
-    name = bank_profile.get("name", "Assistant")
-    mission = bank_profile.get("mission", "")
-
-    parts = []
-
-    # Inject directives at the VERY START for maximum prominence
-    if directives:
-        parts.append(build_directives_section(directives))
-
-    parts.extend(
-        [
-            "You are a reflection agent that answers questions by reasoning over retrieved memories.",
-            "",
-        ]
-    )
-
-    parts.extend(
-        [
-            "## CRITICAL RULES",
-            "- You must NEVER fabricate information that has no basis in retrieved data",
-            "- You SHOULD synthesize, infer, and reason from the retrieved memories",
-            "- You MUST search before saying you don't have information",
-            "",
-            "## How to Reason",
-            "- If memories mention someone did an activity, you can infer they likely enjoyed it",
-            "- Synthesize a coherent narrative from related memories",
-            "- Be a thoughtful interpreter, not just a literal repeater",
-            "- When the exact answer isn't stated, use what IS stated to give the best answer",
-            "",
-            "## HIERARCHICAL RETRIEVAL STRATEGY",
-            "",
-        ]
-    )
-
-    # Build retrieval levels based on what's available
-    if has_mental_models:
-        parts.extend(
-            [
-                "You have access to THREE levels of knowledge. Use them in this order:",
-                "",
-                "### 1. MENTAL MODELS (search_mental_models) - Try First",
-                "- User-curated summaries about specific topics",
-                "- HIGHEST quality - manually created and maintained",
-                "- If a relevant mental model exists and is FRESH, it may fully answer the question",
-                "- Check `is_stale` field - if stale, also verify with lower levels",
-                "",
-                "### 2. OBSERVATIONS (search_observations) - Second Priority",
-                "- Auto-consolidated knowledge from memories",
-                "- Check `is_stale` field - if stale, ALSO use recall() to verify",
-                "- Good for understanding patterns and summaries",
-                "",
-                "### 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",
-                "- This is the source of truth that other levels are built from",
-                "",
-            ]
-        )
-    else:
-        parts.extend(
-            [
-                "You have access to TWO levels of knowledge. Use them in this order:",
-                "",
-                "### 1. OBSERVATIONS (search_observations) - Try First",
-                "- Auto-consolidated knowledge from memories",
-                "- Check `is_stale` field - if stale, ALSO use recall() to verify",
-                "- Good for understanding patterns and summaries",
-                "",
-                "### 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",
-                "- This is the source of truth that observations are built from",
-                "",
-            ]
-        )
-
-    parts.extend(
-        [
-            "## Query Strategy",
-            "recall() uses semantic search. NEVER just echo the user's question - decompose it into targeted searches:",
-            "",
-            "BAD: User asks 'recurring lesson themes between students' → recall('recurring lesson themes between students')",
-            "GOOD: Break it down into component searches:",
-            "  1. recall('lessons') - find all lesson-related memories",
-            "  2. recall('teaching sessions') - alternative phrasing",
-            "  3. recall('student progress') - find student-related memories",
-            "",
-            "Think: What ENTITIES and CONCEPTS does this question involve? Search for each separately.",
-            "",
-        ]
-    )
-
-    # Add budget guidance
-    if budget:
-        budget_lower = budget.lower()
-        if budget_lower == "low":
-            parts.extend(
-                [
-                    "## RESEARCH DEPTH: SHALLOW (Quick Response)",
-                    "- Prioritize speed over completeness",
-                    "- If mental models or observations provide a reasonable answer, stop there",
-                    "- Only dig deeper if the initial results are clearly insufficient",
-                    "- Prefer a quick overview rather than exhaustive details",
-                    "- Answer promptly with available information",
-                    "",
-                ]
-            )
-        elif budget_lower == "mid":
-            parts.extend(
-                [
-                    "## RESEARCH DEPTH: MODERATE (Balanced)",
-                    "- Balance thoroughness with efficiency",
-                    "- Check multiple sources when the question warrants it",
-                    "- Verify stale data if it's central to the answer",
-                    "- Don't over-explore, but ensure reasonable coverage",
-                    "",
-                ]
-            )
-        elif budget_lower == "high":
-            parts.extend(
-                [
-                    "## RESEARCH DEPTH: DEEP (Thorough Exploration)",
-                    "- Explore comprehensively before answering",
-                    "- Search across all available knowledge levels",
-                    "- Use multiple query variations to ensure coverage",
-                    "- Verify information across different retrieval levels",
-                    "- Use expand() to get full context on important memories",
-                    "- Take time to synthesize a complete, well-researched answer",
-                    "",
-                ]
-            )
-
-    parts.append("## Workflow")
-
-    if has_mental_models:
-        parts.extend(
-            [
-                "1. First, try search_mental_models() - check if a curated summary exists",
-                "2. If no mental model or it's stale, try search_observations() for consolidated knowledge",
-                "3. If observations are stale OR you need specific details, use recall() for raw facts",
-                "4. Use expand() if you need more context on specific memories",
-                "5. When ready, call done() with your answer and supporting IDs",
-            ]
-        )
-    else:
-        parts.extend(
-            [
-                "1. First, try search_observations() - check for consolidated knowledge",
-                "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",
-            ]
-        )
-
-    parts.extend(
-        [
-            "",
-            "## Output Format: Plain Text Answer",
-            "Call done() with a plain text 'answer' field.",
-            "- Do NOT use markdown formatting",
-            "- 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",
-        ]
-    )
-
-    parts.append("")
-    parts.append(f"## Memory Bank: {name}")
-
-    if mission:
-        parts.append(f"Mission: {mission}")
-
-    # Disposition traits
-    disposition = bank_profile.get("disposition", {})
-    if disposition:
-        traits = []
-        if "skepticism" in disposition:
-            traits.append(f"skepticism={disposition['skepticism']}")
-        if "literalism" in disposition:
-            traits.append(f"literalism={disposition['literalism']}")
-        if "empathy" in disposition:
-            traits.append(f"empathy={disposition['empathy']}")
-        if traits:
-            parts.append(f"Disposition: {', '.join(traits)}")
-
-    if context:
-        parts.append(f"\n## Additional Context\n{context}")
-
-    # Add directive reminder at the END for recency effect
-    if directives:
-        parts.append(build_directives_reminder(directives))
-
-    return "\n".join(parts)
-
-
-def build_agent_prompt(
-    query: str,
-    context_history: list[dict],
-    bank_profile: dict,
-    additional_context: str | None = None,
-) -> str:
-    """Build the user prompt for the reflect agent."""
-    parts = []
-
-    # Bank identity
-    name = bank_profile.get("name", "Assistant")
-    mission = bank_profile.get("mission", "")
-
-    parts.append(f"## Memory Bank Context\nName: {name}")
-    if mission:
-        parts.append(f"Mission: {mission}")
-
-    # Disposition traits if present
-    disposition = bank_profile.get("disposition", {})
-    if disposition:
-        traits = []
-        if "skepticism" in disposition:
-            traits.append(f"skepticism={disposition['skepticism']}")
-        if "literalism" in disposition:
-            traits.append(f"literalism={disposition['literalism']}")
-        if "empathy" in disposition:
-            traits.append(f"empathy={disposition['empathy']}")
-        if traits:
-            parts.append(f"Disposition: {', '.join(traits)}")
-
-    # Additional context from caller
-    if additional_context:
-        parts.append(f"\n## Additional Context\n{additional_context}")
-
-    # Tool call history
-    if context_history:
-        parts.append("\n## Tool Results (synthesize and reason from this data)")
-        for i, entry in enumerate(context_history, 1):
-            tool = entry["tool"]
-            output = entry["output"]
-            # Format as proper JSON for LLM readability
-            try:
-                output_str = json.dumps(output, indent=2, default=str)
-            except (TypeError, ValueError):
-                output_str = str(output)
-            parts.append(f"\n### Call {i}: {tool}\n```json\n{output_str}\n```")
-
-    # The question
-    parts.append(f"\n## Question\n{query}")
-
-    # Instructions
-    if context_history:
-        parts.append(
-            "\n## Instructions\n"
-            "Based on the tool results above, either call more tools or provide your final answer. "
-            "Synthesize and reason from the data - make reasonable inferences when helpful. "
-            "If you have related information, use it to give the best possible answer."
-        )
-    else:
-        parts.append(
-            "\n## Instructions\n"
-            "Start by searching for relevant information using the hierarchical retrieval strategy:\n"
-            "1. Try search_mental_models() first for curated summaries\n"
-            "2. Try search_observations() for consolidated knowledge\n"
-            "3. Use recall() for specific details or to verify stale data"
-        )
-
-    return "\n".join(parts)
-
-
-def build_final_prompt(
-    query: str,
-    context_history: list[dict],
-    bank_profile: dict,
-    additional_context: str | None = None,
-) -> str:
-    """Build the final prompt when forcing a text response (no tools)."""
-    parts = []
-
-    # Bank identity
-    name = bank_profile.get("name", "Assistant")
-    mission = bank_profile.get("mission", "")
-
-    parts.append(f"## Memory Bank Context\nName: {name}")
-    if mission:
-        parts.append(f"Mission: {mission}")
-
-    # Disposition traits if present
-    disposition = bank_profile.get("disposition", {})
-    if disposition:
-        traits = []
-        if "skepticism" in disposition:
-            traits.append(f"skepticism={disposition['skepticism']}")
-        if "literalism" in disposition:
-            traits.append(f"literalism={disposition['literalism']}")
-        if "empathy" in disposition:
-            traits.append(f"empathy={disposition['empathy']}")
-        if traits:
-            parts.append(f"Disposition: {', '.join(traits)}")
-
-    # Additional context from caller
-    if additional_context:
-        parts.append(f"\n## Additional Context\n{additional_context}")
-
-    # Tool call history
-    if context_history:
-        parts.append("\n## Retrieved Data (synthesize and reason from this data)")
-        for entry in context_history:
-            tool = entry["tool"]
-            output = entry["output"]
-            # Format as proper JSON for LLM readability
-            try:
-                output_str = json.dumps(output, indent=2, default=str)
-            except (TypeError, ValueError):
-                output_str = str(output)
-            parts.append(f"\n### From {tool}:\n```json\n{output_str}\n```")
-    else:
-        parts.append("\n## Retrieved Data\nNo data was retrieved.")
-
-    # The question
-    parts.append(f"\n## Question\n{query}")
-
-    # Final instructions
-    parts.append(
-        "\n## Instructions\n"
-        "Provide a thoughtful answer by synthesizing and reasoning from the retrieved data above. "
-        "You can make reasonable inferences from the memories, but don't completely fabricate information."
-        "If the exact answer isn't stated, use what IS stated to give the best possible answer. "
-        "Only say 'I don't have information' if the retrieved data is truly unrelated to the question."
-    )
-
-    return "\n".join(parts)
-
-
-FINAL_SYSTEM_PROMPT = """You are a thoughtful assistant that synthesizes answers from retrieved memories.
-
-Your approach:
- Reason over the retrieved memories to answer the question
- Make reasonable inferences when the exact answer isn't explicitly stated
- Connect related memories to form a complete picture
- Be helpful - if you have related information, use it to give the best possible answer
-
-Only say "I don't have information" if the retrieved data is truly unrelated to the question.
-Do NOT fabricate information that has no basis in the retrieved data."""
@@ -1,437 +0,0 @@
-"""
-Tool implementations for the reflect agent.
-
-Implements hierarchical retrieval:
-1. search_mental_models - User-curated stored reflect responses (highest quality)
-2. search_observations - Consolidated knowledge with freshness
-3. recall - Raw facts as ground truth
-"""
-
-import logging
-import uuid
-from datetime import datetime, timedelta, timezone
-from typing import TYPE_CHECKING, Any
-
-if TYPE_CHECKING:
-    from asyncpg import Connection
-
-    from ...api.http import RequestContext
-    from ..memory_engine import MemoryEngine
-
-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",
-    bank_id: str,
-    query: str,
-    query_embedding: list[float],
-    max_results: int = 5,
-    tags: list[str] | None = None,
-    tags_match: str = "any",
-    exclude_ids: list[str] | None = None,
-) -> dict[str, Any]:
-    """
-    Search user-curated mental models by semantic similarity.
-
-    Mental models are high-quality, manually created summaries about specific topics.
-    They should be searched FIRST as they represent the most reliable synthesized knowledge.
-
-    Args:
-        conn: Database connection
-        bank_id: Bank identifier
-        query: Search query (for logging/tracing)
-        query_embedding: Pre-computed embedding for semantic search
-        max_results: Maximum number of mental models to return
-        tags: Optional tags to filter mental models
-        tags_match: How to match tags - "any" (OR), "all" (AND)
-        exclude_ids: Optional list of mental model IDs to exclude (e.g., when refreshing a mental model)
-
-    Returns:
-        Dict with matching mental models including content and freshness info
-    """
-    from ..memory_engine import fq_table
-
-    # Build filters dynamically
-    filters = ""
-    params: list[Any] = [bank_id, str(query_embedding), max_results]
-    next_param = 4
-
-    if tags:
-        if tags_match == "all":
-            filters += f" AND tags @> ${next_param}::varchar[]"
-        else:
-            filters += f" AND (tags && ${next_param}::varchar[] OR tags IS NULL OR tags = '{{}}')"
-        params.append(tags)
-        next_param += 1
-
-    if exclude_ids:
-        filters += f" AND id != ALL(${next_param}::text[])"
-        params.append(exclude_ids)
-        next_param += 1
-
-    # Search mental models by embedding similarity
-    rows = await conn.fetch(
-        f"""
-        SELECT
-            id, name, content,
-            tags, created_at, last_refreshed_at,
-            1 - (embedding <=> $2::vector) as relevance
-        FROM {fq_table("mental_models")}
-        WHERE bank_id = $1 AND embedding IS NOT NULL {filters}
-        ORDER BY embedding <=> $2::vector
-        LIMIT $3
-        """,
-        *params,
-    )
-
-    now = datetime.now(timezone.utc)
-    mental_models = []
-
-    for row in rows:
-        last_refreshed_at = row["last_refreshed_at"]
-        if last_refreshed_at and last_refreshed_at.tzinfo is None:
-            last_refreshed_at = last_refreshed_at.replace(tzinfo=timezone.utc)
-
-        # 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(
-            {
-                "id": str(row["id"]),
-                "name": row["name"],
-                "content": row["content"],
-                "tags": row["tags"] or [],
-                "relevance": round(row["relevance"], 4),
-                "updated_at": last_refreshed_at.isoformat() if last_refreshed_at else None,
-                "is_stale": is_stale,
-            }
-        )
-
-    return {
-        "query": query,
-        "count": len(mental_models),
-        "mental_models": mental_models,
-    }
-
-
-async def tool_search_observations(
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    query: str,
-    request_context: "RequestContext",
-    max_tokens: int = 5000,
-    tags: list[str] | None = None,
-    tags_match: str = "any",
-    last_consolidated_at: datetime | None = None,
-    pending_consolidation: int = 0,
-) -> dict[str, Any]:
-    """
-    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().
-
-    Args:
-        memory_engine: Memory engine instance
-        bank_id: Bank identifier
-        query: Search query
-        request_context: Request context for authentication
-        max_tokens: Maximum tokens for results (default 5000)
-        tags: Optional tags to filter observations
-        tags_match: How to match tags - "any" (OR), "all" (AND)
-        last_consolidated_at: When consolidation last ran (for staleness check)
-        pending_consolidation: Number of memories waiting to be consolidated
-
-    Returns:
-        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"],  # 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,
-        _connection_budget=1,
-        _quiet=True,
-    )
-
-    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:
-        freshness = "slightly_stale"
-    else:
-        freshness = "stale"
-
-    return {
-        "query": query,
-        "count": len(observations),
-        "observations": observations,
-        "freshness": freshness,
-    }
-
-
-async def tool_recall(
-    memory_engine: "MemoryEngine",
-    bank_id: str,
-    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,
-) -> dict[str, Any]:
-    """
-    Search memories using TEMPR retrieval.
-
-    This is the ground truth - raw facts and experiences.
-    Use when mental models/observations don't exist, are stale, or need verification.
-
-    Args:
-        memory_engine: Memory engine instance
-        bank_id: Bank identifier
-        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)
-
-    Returns:
-        Dict with list of matching memories
-    """
-    result = await memory_engine.recall_async(
-        bank_id=bank_id,
-        query=query,
-        fact_type=["experience", "world"],  # Exclude opinions and observations
-        max_tokens=max_tokens,
-        enable_trace=False,
-        request_context=request_context,
-        tags=tags,
-        tags_match=tags_match,
-        _connection_budget=connection_budget,
-        _quiet=True,  # Suppress logging for internal operations
-    )
-
-    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,
-        "count": len(memories),
-        "memories": memories,
-    }
-
-
-async def tool_expand(
-    conn: "Connection",
-    bank_id: str,
-    memory_ids: list[str],
-    depth: str,
-) -> dict[str, Any]:
-    """
-    Expand multiple memories to get chunk or document context.
-
-    Args:
-        conn: Database connection
-        bank_id: Bank identifier
-        memory_ids: List of memory unit IDs
-        depth: "chunk" or "document"
-
-    Returns:
-        Dict with results array, each containing memory, chunk, and optionally document data
-    """
-    from ..memory_engine import fq_table
-
-    if not memory_ids:
-        return {"error": "memory_ids is required and must not be empty"}
-
-    # Validate and convert UUIDs
-    valid_uuids: list[uuid.UUID] = []
-    errors: dict[str, str] = {}
-    for mid in memory_ids:
-        try:
-            valid_uuids.append(uuid.UUID(mid))
-        except ValueError:
-            errors[mid] = f"Invalid memory_id format: {mid}"
-
-    if not valid_uuids:
-        return {"error": "No valid memory IDs provided", "details": errors}
-
-    # Batch fetch all memory units
-    memories = await conn.fetch(
-        f"""
-        SELECT id, text, chunk_id, document_id, fact_type, context
-        FROM {fq_table("memory_units")}
-        WHERE id = ANY($1) AND bank_id = $2
-        """,
-        valid_uuids,
-        bank_id,
-    )
-    memory_map = {row["id"]: row for row in memories}
-
-    # Collect chunk_ids and document_ids for batch fetching
-    chunk_ids = [m["chunk_id"] for m in memories if m["chunk_id"]]
-    doc_ids_from_chunks: set[str] = set()
-    doc_ids_direct: set[str] = set()
-
-    # Batch fetch all chunks
-    chunk_map: dict[str, Any] = {}
-    if chunk_ids:
-        chunks = await conn.fetch(
-            f"""
-            SELECT chunk_id, chunk_text, chunk_index, document_id
-            FROM {fq_table("chunks")}
-            WHERE chunk_id = ANY($1)
-            """,
-            chunk_ids,
-        )
-        chunk_map = {row["chunk_id"]: row for row in chunks}
-        if depth == "document":
-            doc_ids_from_chunks = {c["document_id"] for c in chunks if c["document_id"]}
-
-    # Collect direct document IDs (memories without chunks)
-    if depth == "document":
-        for m in memories:
-            if not m["chunk_id"] and m["document_id"]:
-                doc_ids_direct.add(m["document_id"])
-
-    # Batch fetch all documents
-    doc_map: dict[str, Any] = {}
-    all_doc_ids = list(doc_ids_from_chunks | doc_ids_direct)
-    if all_doc_ids:
-        docs = await conn.fetch(
-            f"""
-            SELECT id, original_text, metadata, retain_params
-            FROM {fq_table("documents")}
-            WHERE id = ANY($1) AND bank_id = $2
-            """,
-            all_doc_ids,
-            bank_id,
-        )
-        doc_map = {row["id"]: row for row in docs}
-
-    # Build results
-    results: list[dict[str, Any]] = []
-    for mid, mem_uuid in zip(memory_ids, valid_uuids):
-        if mid in errors:
-            results.append({"memory_id": mid, "error": errors[mid]})
-            continue
-
-        memory = memory_map.get(mem_uuid)
-        if not memory:
-            results.append({"memory_id": mid, "error": f"Memory not found: {mid}"})
-            continue
-
-        item: dict[str, Any] = {
-            "memory_id": mid,
-            "memory": {
-                "id": str(memory["id"]),
-                "text": memory["text"],
-                "type": memory["fact_type"],
-                "context": memory["context"],
-            },
-        }
-
-        # Add chunk if available
-        if memory["chunk_id"] and memory["chunk_id"] in chunk_map:
-            chunk = chunk_map[memory["chunk_id"]]
-            item["chunk"] = {
-                "id": chunk["chunk_id"],
-                "text": chunk["chunk_text"],
-                "index": chunk["chunk_index"],
-                "document_id": chunk["document_id"],
-            }
-            # Add document if depth=document
-            if depth == "document" and chunk["document_id"] in doc_map:
-                doc = doc_map[chunk["document_id"]]
-                item["document"] = {
-                    "id": doc["id"],
-                    "full_text": doc["original_text"],
-                    "metadata": doc["metadata"],
-                    "retain_params": doc["retain_params"],
-                }
-        elif memory["document_id"] and depth == "document" and memory["document_id"] in doc_map:
-            # No chunk, but has document_id
-            doc = doc_map[memory["document_id"]]
-            item["document"] = {
-                "id": doc["id"],
-                "full_text": doc["original_text"],
-                "metadata": doc["metadata"],
-                "retain_params": doc["retain_params"],
-            }
-
-        results.append(item)
-
-    return {"results": results, "count": len(results)}
@@ -1,250 +0,0 @@
-"""
-Tool schema definitions for the reflect agent.
-
-These are OpenAI-format tool definitions used with native tool calling.
-The reflect agent uses a hierarchical retrieval strategy:
-1. search_mental_models - User-curated stored reflect responses (highest quality, if applicable)
-2. search_observations - Consolidated knowledge with freshness awareness
-3. recall - Raw facts (world/experience) as ground truth fallback
-"""
-
-# Tool definitions in OpenAI format
-
-TOOL_SEARCH_MENTAL_MODELS = {
-    "type": "function",
-    "function": {
-        "name": "search_mental_models",
-        "description": (
-            "Search user-curated mental models (stored reflect responses). These are high-quality, manually created "
-            "summaries about specific topics. Use FIRST when the question might be covered by an "
-            "existing mental model. Returns mental models with their content and last refresh time."
-        ),
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "reason": {
-                    "type": "string",
-                    "description": "Brief explanation of why you're making this search (for debugging)",
-                },
-                "query": {
-                    "type": "string",
-                    "description": "Search query to find relevant mental models",
-                },
-                "max_results": {
-                    "type": "integer",
-                    "description": "Maximum number of mental models to return (default 5)",
-                },
-            },
-            "required": ["reason", "query"],
-        },
-    },
-}
-
-TOOL_SEARCH_OBSERVATIONS = {
-    "type": "function",
-    "function": {
-        "name": "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."
-        ),
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "reason": {
-                    "type": "string",
-                    "description": "Brief explanation of why you're making this search (for debugging)",
-                },
-                "query": {
-                    "type": "string",
-                    "description": "Search query to find relevant observations",
-                },
-                "max_tokens": {
-                    "type": "integer",
-                    "description": "Maximum tokens for results (default 5000). Use higher values for broader searches.",
-                },
-            },
-            "required": ["reason", "query"],
-        },
-    },
-}
-
-TOOL_RECALL = {
-    "type": "function",
-    "function": {
-        "name": "recall",
-        "description": (
-            "Search raw memories (facts and experiences). This is the ground truth data. "
-            "Use when: (1) no reflections/mental models exist, (2) mental models are stale, "
-            "(3) you need specific details not in synthesized knowledge. "
-            "Returns individual memory facts with their timestamps."
-        ),
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "reason": {
-                    "type": "string",
-                    "description": "Brief explanation of why you're making this search (for debugging)",
-                },
-                "query": {
-                    "type": "string",
-                    "description": "Search query string",
-                },
-                "max_tokens": {
-                    "type": "integer",
-                    "description": "Optional limit on result size (default 2048). Use higher values for broader searches.",
-                },
-            },
-            "required": ["reason", "query"],
-        },
-    },
-}
-
-TOOL_EXPAND = {
-    "type": "function",
-    "function": {
-        "name": "expand",
-        "description": "Get more context for one or more memories. Memory hierarchy: memory -> chunk -> document.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "reason": {
-                    "type": "string",
-                    "description": "Brief explanation of why you need more context (for debugging)",
-                },
-                "memory_ids": {
-                    "type": "array",
-                    "items": {"type": "string"},
-                    "description": "Array of memory IDs from recall results (batch multiple for efficiency)",
-                },
-                "depth": {
-                    "type": "string",
-                    "enum": ["chunk", "document"],
-                    "description": "chunk: surrounding text chunk, document: full source document",
-                },
-            },
-            "required": ["reason", "memory_ids", "depth"],
-        },
-    },
-}
-
-TOOL_DONE_ANSWER = {
-    "type": "function",
-    "function": {
-        "name": "done",
-        "description": "Signal completion with your final answer. Use this when you have gathered enough information to answer the question.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "answer": {
-                    "type": "string",
-                    "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
-                },
-                "memory_ids": {
-                    "type": "array",
-                    "items": {"type": "string"},
-                    "description": "Array of memory IDs that support your answer (put IDs here, NOT in answer text)",
-                },
-                "mental_model_ids": {
-                    "type": "array",
-                    "items": {"type": "string"},
-                    "description": "Array of mental model IDs that support your answer",
-                },
-                "observation_ids": {
-                    "type": "array",
-                    "items": {"type": "string"},
-                    "description": "Array of observation IDs that support your answer",
-                },
-            },
-            "required": ["answer"],
-        },
-    },
-}
-
-
-def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
-    """
-    Build the done tool schema with directive compliance field.
-
-    When directives are present, adds a required field that forces the agent
-    to confirm compliance with each directive before submitting.
-
-    Args:
-        directive_rules: List of directive rule strings
-    """
-    # Build rules list for description
-    rules_list = "\n".join(f"  {i + 1}. {rule}" for i, rule in enumerate(directive_rules))
-
-    # Build the tool with directive compliance field
-    return {
-        "type": "function",
-        "function": {
-            "name": "done",
-            "description": (
-                "Signal completion with your final answer. IMPORTANT: You must confirm directive compliance before submitting. "
-                "Your answer will be REJECTED if it violates any directive."
-            ),
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "answer": {
-                        "type": "string",
-                        "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
-                    },
-                    "memory_ids": {
-                        "type": "array",
-                        "items": {"type": "string"},
-                        "description": "Array of memory IDs that support your answer (put IDs here, NOT in answer text)",
-                    },
-                    "mental_model_ids": {
-                        "type": "array",
-                        "items": {"type": "string"},
-                        "description": "Array of mental model IDs that support your answer",
-                    },
-                    "observation_ids": {
-                        "type": "array",
-                        "items": {"type": "string"},
-                        "description": "Array of observation IDs that support your answer",
-                    },
-                    "directive_compliance": {
-                        "type": "string",
-                        "description": f"REQUIRED: Confirm your answer complies with ALL directives. List each directive and how your answer follows it:\n{rules_list}\n\nFormat: 'Directive 1: [how answer complies]. Directive 2: [how answer complies]...'",
-                    },
-                },
-                "required": ["answer", "directive_compliance"],
-            },
-        },
-    }
-
-
-def get_reflect_tools(directive_rules: list[str] | None = None) -> list[dict]:
-    """
-    Get the list of tools for the reflect agent.
-
-    The tools support a hierarchical retrieval strategy:
-    1. search_mental_models - User-curated stored reflect responses (try first)
-    2. search_observations - Consolidated knowledge with freshness
-    3. recall - Raw facts as ground truth
-
-    Args:
-        directive_rules: Optional list of directive rule strings. If provided,
-                        the done() tool will require directive compliance confirmation.
-
-    Returns:
-        List of tool definitions in OpenAI format
-    """
-    tools = [
-        TOOL_SEARCH_MENTAL_MODELS,
-        TOOL_SEARCH_OBSERVATIONS,
-        TOOL_RECALL,
-        TOOL_EXPAND,
-    ]
-
-    # Use directive-aware done tool if directives are present
-    if directive_rules:
-        tools.append(_build_done_tool_with_directives(directive_rules))
-    else:
-        tools.append(TOOL_DONE_ANSWER)
-
-    return tools
@@ -10,94 +10,8 @@ from typing import Any

 from pydantic import BaseModel, ConfigDict, Field

-# Valid fact types for recall operations (excludes 'opinion' which is deprecated)
-VALID_RECALL_FACT_TYPES = frozenset(["world", "experience", "observation"])
-
-
-class LLMToolCall(BaseModel):
-    """A tool call requested by the LLM."""
-
-    id: str = Field(description="Unique identifier for this tool call")
-    name: str = Field(description="Name of the tool to call")
-    arguments: dict[str, Any] = Field(description="Arguments to pass to the tool")
-
-
-class LLMToolCallResult(BaseModel):
-    """Result from an LLM call that may include tool calls."""
-
-    content: str | None = Field(default=None, description="Text content if any")
-    tool_calls: list[LLMToolCall] = Field(default_factory=list, description="Tool calls requested by the LLM")
-    finish_reason: str | None = Field(default=None, description="Reason the LLM stopped: 'stop', 'tool_calls', etc.")
-    input_tokens: int = Field(default=0, description="Input tokens used in this call")
-    output_tokens: int = Field(default=0, description="Output tokens used in this call")
-
-
-class ToolCallTrace(BaseModel):
-    """A single tool call made during reflect."""
-
-    tool: str = Field(description="Tool name: lookup, recall, learn, expand")
-    reason: str | None = Field(default=None, description="Agent's reasoning for making this tool call")
-    input: dict = Field(description="Tool input parameters")
-    output: dict = Field(description="Tool output/result")
-    duration_ms: int = Field(description="Execution time in milliseconds")
-    iteration: int = Field(default=0, description="Iteration number (1-based) when this tool was called")
-
-
-class LLMCallTrace(BaseModel):
-    """A single LLM call made during reflect."""
-
-    scope: str = Field(description="Call scope: agent_1, agent_2, final, etc.")
-    duration_ms: int = Field(description="Execution time in milliseconds")
-
-
-class ObservationRef(BaseModel):
-    """Reference to an observation accessed during reflect."""
-
-    id: str = Field(description="Observation ID")
-    name: str = Field(description="Observation name")
-    type: str = Field(description="Observation type: entity, concept, event")
-    subtype: str = Field(description="Observation subtype: structural, emergent, learned")
-    description: str = Field(description="Brief description")
-    summary: str | None = Field(default=None, description="Full summary (when looked up in detail)")
-
-
-class DirectiveRef(BaseModel):
-    """Reference to a directive that was applied during reflect."""
-
-    id: str = Field(description="Directive mental model ID")
-    name: str = Field(description="Directive name")
-    content: str = Field(description="Directive content")
-
-
-class TokenUsage(BaseModel):
-    """
-    Token usage metrics for LLM calls.
-
-    Tracks input/output tokens for a single request to enable
-    per-request cost tracking and monitoring.
-    """
-
-    model_config = ConfigDict(
-        json_schema_extra={
-            "example": {
-                "input_tokens": 1500,
-                "output_tokens": 500,
-                "total_tokens": 2000,
-            }
-        }
-    )
-
-    input_tokens: int = Field(default=0, description="Number of input/prompt tokens consumed")
-    output_tokens: int = Field(default=0, description="Number of output/completion tokens generated")
-    total_tokens: int = Field(default=0, description="Total tokens (input + output)")
-
-    def __add__(self, other: "TokenUsage") -> "TokenUsage":
-        """Allow aggregating token usage from multiple calls."""
-        return TokenUsage(
-            input_tokens=self.input_tokens + other.input_tokens,
-            output_tokens=self.output_tokens + other.output_tokens,
-            total_tokens=self.total_tokens + other.total_tokens,
-        )
+# Valid fact types for recall operations (excludes 'observation' which is internal)
+VALID_RECALL_FACT_TYPES = frozenset(["world", "experience", "opinion"])


 class DispositionTraits(BaseModel):
@@ -140,7 +54,6 @@ class MemoryFact(BaseModel):
                "metadata": {"source": "slack"},
                "chunk_id": "bank123_session_abc123_0",
                "activation": 0.95,
-                "tags": ["user_a", "session_123"],
            }
        }
    )
@@ -158,7 +71,6 @@ class MemoryFact(BaseModel):
    chunk_id: str | None = Field(
        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")


 class ChunkInfo(BaseModel):
@@ -169,28 +81,6 @@ class ChunkInfo(BaseModel):
    truncated: bool = Field(default=False, description="Whether the chunk was truncated due to token limits")


-class ObservationResult(BaseModel):
-    """An observation result from recall (consolidated knowledge synthesized from facts)."""
-
-    id: str = Field(description="Unique observation ID")
-    text: str = Field(description="The observation text")
-    proof_count: int = Field(description="Number of facts supporting this observation")
-    relevance: float = Field(default=0.0, description="Relevance score to the query")
-    tags: list[str] | None = Field(default=None, description="Tags for visibility scoping")
-    source_memory_ids: list[str] = Field(
-        default_factory=list, description="IDs of facts that contribute to this observation"
-    )
-
-
-class MentalModelResult(BaseModel):
-    """A mental model result from recall (stored reflect response)."""
-
-    id: str = Field(description="Unique mental model ID")
-    name: str = Field(description="Human-readable name")
-    content: str = Field(description="The synthesized content")
-    relevance: float = Field(default=0.0, description="Relevance score to the query")
-
-
 class RecallResult(BaseModel):
    """
    Result from a recall operation.
@@ -233,8 +123,7 @@ class ReflectResult(BaseModel):
    Result from a reflect operation.

    Contains the formulated answer, the facts it was based on (organized by type),
-    any new opinions that were formed during the reflection process, and optionally
-    structured output if a response schema was provided.
+    and any new opinions that were formed during the reflection process.
    """

    model_config = ConfigDict(
@@ -254,46 +143,36 @@ class ReflectResult(BaseModel):
                    ],
                    "experience": [],
                    "opinion": [],
-                    "mental_models": [],
-                    "directives": [
-                        {
-                            "id": "directive-123",
-                            "name": "Response Style",
-                            "rules": ["Always be concise"],
-                        }
-                    ],
                },
-                "structured_output": {"summary": "ML in healthcare", "confidence": 0.9},
-                "usage": {"input_tokens": 1500, "output_tokens": 500, "total_tokens": 2000},
+                "new_opinions": ["Machine learning has great potential in healthcare"],
            }
        }
    )

    text: str = Field(description="The formulated answer text")
-    based_on: dict[str, Any] = Field(
-        description="Facts used to formulate the answer, organized by type (world, experience, mental_models, directives)"
+    based_on: dict[str, list[MemoryFact]] = Field(
+        description="Facts used to formulate the answer, organized by type (world, experience, opinion)"
    )
-    structured_output: dict[str, Any] | None = Field(
-        default=None,
-        description="Structured output parsed according to the provided response schema. Only present when response_schema was provided.",
-    )
-    usage: TokenUsage | None = Field(
-        default=None,
-        description="Token usage metrics for the LLM calls made during this reflect operation.",
-    )
-    tool_trace: list[ToolCallTrace] = Field(
-        default_factory=list,
-        description="Trace of tool calls made during reflection. Only present when include.tool_calls is enabled.",
-    )
-    llm_trace: list[LLMCallTrace] = Field(
-        default_factory=list,
-        description="Trace of LLM calls made during reflection. Only present when include.tool_calls is enabled.",
-    )
-    directives_applied: list[DirectiveRef] = Field(
-        default_factory=list,
-        description="Directive mental models that were applied during this reflection.",
+    new_opinions: list[str] = Field(default_factory=list, description="List of newly formed opinions during reflection")
+
+
+class Opinion(BaseModel):
+    """
+    An opinion with confidence score.
+
+    Opinions represent the bank's formed perspectives on topics,
+    with a confidence level indicating strength of belief.
+    """
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {"text": "Machine learning has great potential in healthcare", "confidence": 0.85}
+        }
    )

+    text: str = Field(description="The opinion text")
+    confidence: float = Field(description="Confidence score between 0.0 and 1.0")
+

 class EntityObservation(BaseModel):
    """
@@ -338,32 +217,3 @@ class EntityState(BaseModel):
    observations: list[EntityObservation] = Field(
        default_factory=list, description="List of observations about this entity"
    )
-
-
-class MentalModel(BaseModel):
-    """
-    A manually configured mental model for tracking specific topics/areas.
-
-    Mental models are user-defined focus areas that the agent should track
-    and maintain summaries for, unlike auto-extracted entities.
-    """
-
-    model_config = ConfigDict(
-        json_schema_extra={
-            "example": {
-                "id": "team-dynamics",
-                "name": "Team Dynamics",
-                "description": "Track how the team collaborates, communication patterns, conflicts, and resolutions",
-                "summary": "The team has strong collaboration...",
-                "summary_updated_at": "2024-01-15T10:30:00Z",
-                "created_at": "2024-01-10T08:00:00Z",
-            }
-        }
-    )
-
-    id: str = Field(description="Unique identifier (alphanumeric lowercase)")
-    name: str = Field(description="Display name for the mental model")
-    description: str = Field(description="Prompt/directions for what to track and summarize")
-    summary: str | None = Field(None, description="Generated summary based on relevant facts")
-    summary_updated_at: str | None = Field(None, description="ISO format date when summary was last updated")
-    created_at: str = Field(description="ISO format date when the mental model was created")
@@ -1,5 +1,5 @@
 """
-bank profile utilities for disposition and mission management.
+bank profile utilities for disposition and background management.
 """

 import json
@@ -27,18 +27,19 @@ class BankProfile(TypedDict):

    name: str
    disposition: DispositionTraits
-    mission: str
+    background: str


-class MissionMergeResponse(BaseModel):
-    """LLM response for mission merge."""
+class BackgroundMergeResponse(BaseModel):
+    """LLM response for background merge with disposition inference."""

-    mission: str = Field(description="Merged mission in first person perspective")
+    background: str = Field(description="Merged background in first person perspective")
+    disposition: DispositionTraits = Field(description="Inferred disposition traits (skepticism, literalism, empathy)")


 async def get_bank_profile(pool, bank_id: str) -> BankProfile:
    """
-    Get bank profile (name, disposition + mission).
+    Get bank profile (name, disposition + background).
    Auto-creates bank with default values if not exists.

    Args:
@@ -46,13 +47,13 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
        bank_id: bank IDentifier

    Returns:
-        BankProfile with name, typed DispositionTraits, and mission
+        BankProfile with name, typed DispositionTraits, and background
    """
    async with acquire_with_retry(pool) as conn:
        # Try to get existing bank
        row = await conn.fetchrow(
            f"""
-            SELECT name, disposition, mission
+            SELECT name, disposition, background
            FROM {fq_table("banks")} WHERE bank_id = $1
            """,
            bank_id,
@@ -65,15 +66,13 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
                disposition_data = json.loads(disposition_data)

            return BankProfile(
-                name=row["name"],
-                disposition=DispositionTraits(**disposition_data),
-                mission=row["mission"] or "",
+                name=row["name"], disposition=DispositionTraits(**disposition_data), background=row["background"]
            )

        # Bank doesn't exist, create with defaults
        await conn.execute(
            f"""
-            INSERT INTO {fq_table("banks")} (bank_id, name, disposition, mission)
+            INSERT INTO {fq_table("banks")} (bank_id, name, disposition, background)
            VALUES ($1, $2, $3::jsonb, $4)
            ON CONFLICT (bank_id) DO NOTHING
            """,
@@ -83,7 +82,7 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
            "",
        )

-        return BankProfile(name=bank_id, disposition=DispositionTraits(**DEFAULT_DISPOSITION), mission="")
+        return BankProfile(name=bank_id, disposition=DispositionTraits(**DEFAULT_DISPOSITION), background="")


 async def update_bank_disposition(pool, bank_id: str, disposition: dict[str, int]) -> None:
@@ -111,121 +110,244 @@ async def update_bank_disposition(pool, bank_id: str, disposition: dict[str, int
        )


-async def set_bank_mission(pool, bank_id: str, mission: str) -> None:
+async def merge_bank_background(pool, llm_config, bank_id: str, new_info: str, update_disposition: bool = True) -> dict:
    """
-    Set bank mission (replacing any existing mission).
-
-    Args:
-        pool: Database connection pool
-        bank_id: bank IDentifier
-        mission: The mission text
-    """
-    # Ensure bank exists first
-    await get_bank_profile(pool, bank_id)
-
-    async with acquire_with_retry(pool) as conn:
-        await conn.execute(
-            f"""
-            UPDATE {fq_table("banks")}
-            SET mission = $2,
-                updated_at = NOW()
-            WHERE bank_id = $1
-            """,
-            bank_id,
-            mission,
-        )
-
-
-async def merge_bank_mission(pool, llm_config, bank_id: str, new_info: str) -> dict:
-    """
-    Merge new mission information with existing mission using LLM.
+    Merge new background information with existing background using LLM.
    Normalizes to first person ("I") and resolves conflicts.
+    Optionally infers disposition traits from the merged background.

    Args:
        pool: Database connection pool
-        llm_config: LLM configuration for mission merging
+        llm_config: LLM configuration for background merging
        bank_id: bank IDentifier
-        new_info: New mission information to add/merge
+        new_info: New background information to add/merge
+        update_disposition: If True, infer Big Five traits from background (default: True)

    Returns:
-        Dict with 'mission' (str) key
+        Dict with 'background' (str) and optionally 'disposition' (dict) keys
    """
    # Get current profile
    profile = await get_bank_profile(pool, bank_id)
-    current_mission = profile["mission"]
+    current_background = profile["background"]

-    # Use LLM to merge missions
-    result = await _llm_merge_mission(llm_config, current_mission, new_info)
+    # Use LLM to merge backgrounds and optionally infer disposition
+    result = await _llm_merge_background(llm_config, current_background, new_info, infer_disposition=update_disposition)

-    merged_mission = result["mission"]
+    merged_background = result["background"]
+    inferred_disposition = result.get("disposition")

    # Update in database
    async with acquire_with_retry(pool) as conn:
-        await conn.execute(
-            f"""
-            UPDATE {fq_table("banks")}
-            SET mission = $2,
-                updated_at = NOW()
-            WHERE bank_id = $1
-            """,
-            bank_id,
-            merged_mission,
-        )
+        if inferred_disposition:
+            # Update both background and disposition
+            await conn.execute(
+                f"""
+                UPDATE {fq_table("banks")}
+                SET background = $2,
+                    disposition = $3::jsonb,
+                    updated_at = NOW()
+                WHERE bank_id = $1
+                """,
+                bank_id,
+                merged_background,
+                json.dumps(inferred_disposition),
+            )
+        else:
+            # Update only background
+            await conn.execute(
+                f"""
+                UPDATE {fq_table("banks")}
+                SET background = $2,
+                    updated_at = NOW()
+                WHERE bank_id = $1
+                """,
+                bank_id,
+                merged_background,
+            )

-    return {"mission": merged_mission}
+    response = {"background": merged_background}
+    if inferred_disposition:
+        response["disposition"] = inferred_disposition
+
+    return response


-async def _llm_merge_mission(llm_config, current: str, new_info: str) -> dict:
+async def _llm_merge_background(llm_config, current: str, new_info: str, infer_disposition: bool = False) -> dict:
    """
-    Use LLM to intelligently merge mission information.
+    Use LLM to intelligently merge background information.
+    Optionally infer Big Five disposition traits from the merged background.

    Args:
        llm_config: LLM configuration to use
-        current: Current mission text
+        current: Current background text
        new_info: New information to merge
+        infer_disposition: If True, also infer disposition traits

    Returns:
-        Dict with 'mission' (str) key
+        Dict with 'background' (str) and optionally 'disposition' (dict) keys
    """
-    prompt = f"""You are helping maintain an agent's mission statement.
+    if infer_disposition:
+        prompt = f"""You are helping maintain a memory bank's background/profile and infer their disposition. You MUST respond with ONLY valid JSON.

-Current mission: {current if current else "(empty)"}
+Current background: {current if current else "(empty)"}

 New information to add: {new_info}

 Instructions:
-1. Merge the new information with the current mission
-2. If there are conflicts, the NEW information overwrites the old
+1. Merge the new information with the current background
+2. If there are conflicts (e.g., different birthplaces), the NEW information overwrites the old
+3. Keep additions that don't conflict
+4. Output in FIRST PERSON ("I") perspective
+5. Be concise - keep merged background under 500 characters
+6. Infer disposition traits from the merged background (each 1-5 integer):
+   - Skepticism: 1-5 (1=trusting, takes things at face value; 5=skeptical, questions everything)
+   - Literalism: 1-5 (1=flexible interpretation, reads between lines; 5=literal, exact interpretation)
+   - Empathy: 1-5 (1=detached, focuses on facts; 5=empathetic, considers emotional context)
+
+CRITICAL: You MUST respond with ONLY a valid JSON object. No markdown, no code blocks, no explanations. Just the JSON.
+
+Format:
+{{
+  "background": "the merged background text in first person",
+  "disposition": {{
+    "skepticism": 3,
+    "literalism": 3,
+    "empathy": 3
+  }}
+}}
+
+Trait inference examples:
+- "I'm a lawyer" → skepticism: 4, literalism: 5, empathy: 2
+- "I'm a therapist" → skepticism: 2, literalism: 2, empathy: 5
+- "I'm an engineer" → skepticism: 3, literalism: 4, empathy: 3
+- "I've been burned before by trusting people" → skepticism: 5, literalism: 3, empathy: 3
+- "I try to understand what people really mean" → skepticism: 3, literalism: 2, empathy: 4
+- "I take contracts very seriously" → skepticism: 4, literalism: 5, empathy: 2"""
+    else:
+        prompt = f"""You are helping maintain a memory bank's background/profile.
+
+Current background: {current if current else "(empty)"}
+
+New information to add: {new_info}
+
+Instructions:
+1. Merge the new information with the current background
+2. If there are conflicts (e.g., different birthplaces), the NEW information overwrites the old
 3. Keep additions that don't conflict
 4. Output in FIRST PERSON ("I") perspective
 5. Be concise - keep it under 500 characters
-6. Return ONLY the merged mission text, no explanations
+6. Return ONLY the merged background text, no explanations

-Merged mission:"""
+Merged background:"""

    try:
+        # Prepare messages
        messages = [{"role": "user", "content": prompt}]

+        if infer_disposition:
+            # Use structured output with Pydantic model for disposition inference
+            try:
+                parsed = await llm_config.call(
+                    messages=messages,
+                    response_format=BackgroundMergeResponse,
+                    scope="bank_background",
+                    temperature=0.3,
+                    max_completion_tokens=8192,
+                )
+                logger.info(f"Successfully got structured response: background={parsed.background[:100]}")
+
+                # Convert Pydantic model to dict format
+                return {"background": parsed.background, "disposition": parsed.disposition.model_dump()}
+            except Exception as e:
+                logger.warning(f"Structured output failed, falling back to manual parsing: {e}")
+                # Fall through to manual parsing below
+
+        # Manual parsing fallback or non-disposition merge
        content = await llm_config.call(
-            messages=messages, scope="bank_mission", temperature=0.3, max_completion_tokens=8192
+            messages=messages, scope="bank_background", temperature=0.3, max_completion_tokens=8192
        )

-        logger.info(f"LLM response for mission merge (first 500 chars): {content[:500]}")
+        logger.info(f"LLM response for background merge (first 500 chars): {content[:500]}")

-        merged = content.strip()
-        if not merged or merged.lower() in ["(empty)", "none", "n/a"]:
-            merged = new_info if new_info else ""
-        return {"mission": merged}
+        if infer_disposition:
+            # Parse JSON response - try multiple extraction methods
+            result = None
+
+            # Method 1: Direct parse
+            try:
+                result = json.loads(content)
+                logger.info("Successfully parsed JSON directly")
+            except json.JSONDecodeError:
+                pass
+
+            # Method 2: Extract from markdown code blocks
+            if result is None:
+                # Remove markdown code blocks
+                code_block_match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", content, re.DOTALL)
+                if code_block_match:
+                    try:
+                        result = json.loads(code_block_match.group(1))
+                        logger.info("Successfully extracted JSON from markdown code block")
+                    except json.JSONDecodeError:
+                        pass
+
+            # Method 3: Find nested JSON structure
+            if result is None:
+                # Look for JSON object with nested structure
+                json_match = re.search(
+                    r'\{[^{}]*"background"[^{}]*"disposition"[^{}]*\{[^{}]*\}[^{}]*\}', content, re.DOTALL
+                )
+                if json_match:
+                    try:
+                        result = json.loads(json_match.group())
+                        logger.info("Successfully extracted JSON using nested pattern")
+                    except json.JSONDecodeError:
+                        pass
+
+            # All parsing methods failed - use fallback
+            if result is None:
+                logger.warning(f"Failed to extract JSON from LLM response. Raw content: {content[:200]}")
+                # Fallback: use new_info as background with default disposition
+                return {
+                    "background": new_info if new_info else current if current else "",
+                    "disposition": DEFAULT_DISPOSITION.copy(),
+                }
+
+            # Validate disposition values
+            disposition = result.get("disposition", {})
+            for key in ["skepticism", "literalism", "empathy"]:
+                if key not in disposition:
+                    disposition[key] = 3  # Default to neutral
+                else:
+                    # Clamp to [1, 5] and convert to int
+                    disposition[key] = max(1, min(5, int(disposition[key])))
+
+            result["disposition"] = disposition
+
+            # Ensure background exists
+            if "background" not in result or not result["background"]:
+                result["background"] = new_info if new_info else ""
+
+            return result
+        else:
+            # Just background merge
+            merged = content
+            if not merged or merged.lower() in ["(empty)", "none", "n/a"]:
+                merged = new_info if new_info else ""
+            return {"background": merged}

    except Exception as e:
-        logger.error(f"Error merging mission with LLM: {e}")
+        logger.error(f"Error merging background with LLM: {e}")
        # Fallback: just append new info
        if current:
            merged = f"{current} {new_info}".strip()
        else:
            merged = new_info

-        return {"mission": merged}
+        result = {"background": merged}
+        if infer_disposition:
+            result["disposition"] = DEFAULT_DISPOSITION.copy()
+        return result


 async def list_banks(pool) -> list:
@@ -236,12 +358,12 @@ async def list_banks(pool) -> list:
        pool: Database connection pool

    Returns:
-        List of dicts with bank_id, name, disposition, mission, created_at, updated_at
+        List of dicts with bank_id, name, disposition, background, created_at, updated_at
    """
    async with acquire_with_retry(pool) as conn:
        rows = await conn.fetch(
            f"""
-            SELECT bank_id, name, disposition, mission, created_at, updated_at
+            SELECT bank_id, name, disposition, background, created_at, updated_at
            FROM {fq_table("banks")}
            ORDER BY updated_at DESC
            """
@@ -259,7 +381,7 @@ async def list_banks(pool) -> list:
                    "bank_id": row["bank_id"],
                    "name": row["name"],
                    "disposition": disposition_data,
-                    "mission": row["mission"] or "",
+                    "background": row["background"],
                    "created_at": row["created_at"].isoformat() if row["created_at"] else None,
                    "updated_at": row["updated_at"].isoformat() if row["updated_at"] else None,
                }
@@ -13,23 +13,16 @@ logger = logging.getLogger(__name__)


 async def process_entities_batch(
-    entity_resolver,
-    conn,
-    bank_id: str,
-    unit_ids: list[str],
-    facts: list[ProcessedFact],
-    log_buffer: list[str] = None,
-    user_entities_per_content: dict[int, list[dict]] = None,
+    entity_resolver, conn, bank_id: str, unit_ids: list[str], facts: list[ProcessedFact], log_buffer: list[str] = None
 ) -> list[EntityLink]:
    """
    Process entities for all facts and create entity links.

    This function:
    1. Extracts entity mentions from fact texts
-    2. Merges user-provided entities with LLM-extracted entities
-    3. Resolves entity names to canonical entities
-    4. Creates entity records in the database
-    5. Returns entity links ready for insertion
+    2. Resolves entity names to canonical entities
+    3. Creates entity records in the database
+    4. Returns entity links ready for insertion

    Args:
        entity_resolver: EntityResolver instance for entity resolution
@@ -38,7 +31,6 @@ async def process_entities_batch(
        unit_ids: List of unit IDs (same length as facts)
        facts: List of ProcessedFact objects
        log_buffer: Optional buffer for detailed logging
-        user_entities_per_content: Dict mapping content_index to list of user-provided entities

    Returns:
        List of EntityLink objects for batch insertion
@@ -49,35 +41,14 @@ async def process_entities_batch(
    if len(unit_ids) != len(facts):
        raise ValueError(f"Mismatch between unit_ids ({len(unit_ids)}) and facts ({len(facts)})")

-    user_entities_per_content = user_entities_per_content or {}
-
    # Extract data for link_utils function
    fact_texts = [fact.fact_text for fact in facts]
    # Use occurred_start if available, otherwise use mentioned_at for entity timestamps
    fact_dates = [fact.occurred_start if fact.occurred_start is not None else fact.mentioned_at for fact in facts]
-
-    # Convert EntityRef objects to dict format and merge with user-provided entities
-    entities_per_fact = []
-    for fact in facts:
-        # Start with LLM-extracted entities
-        llm_entities = [{"text": entity.name, "type": "CONCEPT"} for entity in (fact.entities or [])]
-
-        # Get user entities for this content (use content_index from fact)
-        user_entities = user_entities_per_content.get(fact.content_index, [])
-
-        # Merge with case-insensitive deduplication
-        seen_texts = {e["text"].lower() for e in llm_entities}
-        for user_entity in user_entities:
-            if user_entity["text"].lower() not in seen_texts:
-                llm_entities.append(
-                    {
-                        "text": user_entity["text"],
-                        "type": user_entity.get("type", "CONCEPT"),
-                    }
-                )
-                seen_texts.add(user_entity["text"].lower())
-
-        entities_per_fact.append(llm_entities)
+    # Convert EntityRef objects to dict format expected by link_utils
+    entities_per_fact = [
+        [{"text": entity.name, "type": "CONCEPT"} for entity in (fact.entities or [])] for fact in facts
+    ]

    # Use existing link_utils function for entity processing
    entity_links = await link_utils.extract_entities_batch_optimized(
@@ -14,47 +14,7 @@ from typing import Literal

 from pydantic import BaseModel, ConfigDict, Field, field_validator

-from ...config import get_config
 from ..llm_wrapper import LLMConfig, OutputTooLongError
-from ..response_models import TokenUsage
-
-
-def _infer_temporal_date(fact_text: str, event_date: datetime) -> str | None:
-    """
-    Infer a temporal date from fact text when LLM didn't provide occurred_start.
-
-    This is a fallback for when the LLM fails to extract temporal information
-    from relative time expressions like "last night", "yesterday", etc.
-    """
-    import re
-
-    fact_lower = fact_text.lower()
-
-    # Map relative time expressions to day offsets
-    temporal_patterns = {
-        r"\blast night\b": -1,
-        r"\byesterday\b": -1,
-        r"\btoday\b": 0,
-        r"\bthis morning\b": 0,
-        r"\bthis afternoon\b": 0,
-        r"\bthis evening\b": 0,
-        r"\btonigh?t\b": 0,
-        r"\btomorrow\b": 1,
-        r"\blast week\b": -7,
-        r"\bthis week\b": 0,
-        r"\bnext week\b": 7,
-        r"\blast month\b": -30,
-        r"\bthis month\b": 0,
-        r"\bnext month\b": 30,
-    }
-
-    for pattern, offset_days in temporal_patterns.items():
-        if re.search(pattern, fact_lower):
-            target_date = event_date + timedelta(days=offset_days)
-            return target_date.replace(hour=0, minute=0, second=0, microsecond=0).isoformat()
-
-    # If no relative time expression found, return None
-    return None


 def _sanitize_text(text: str) -> str:
@@ -111,38 +71,22 @@ class Fact(BaseModel):


 class CausalRelation(BaseModel):
-    """Causal relationship from this fact to a previous fact (stored format)."""
+    """Causal relationship between facts."""

-    target_fact_index: int = Field(description="Index of the related fact in the facts array (0-based).")
-    relation_type: Literal["caused_by"] = Field(
-        description="How this fact relates to the target: 'caused_by' = this fact was caused by the target"
+    target_fact_index: int = Field(
+        description="Index of the related fact in the facts array (0-based). "
+        "This creates a directed causal link to another fact in the extraction."
+    )
+    relation_type: Literal["causes", "caused_by", "enables", "prevents"] = Field(
+        description="Type of causal relationship: "
+        "'causes' = this fact directly causes the target fact, "
+        "'caused_by' = this fact was caused by the target fact, "
+        "'enables' = this fact enables/allows the target fact, "
+        "'prevents' = this fact prevents/blocks the target fact"
    )
    strength: float = Field(
-        description="Strength of relationship (0.0 to 1.0)",
-        ge=0.0,
-        le=1.0,
-        default=1.0,
-    )
-
-
-class FactCausalRelation(BaseModel):
-    """
-    Causal relationship from this fact to a PREVIOUS fact (embedded in each fact).
-
-    Uses index-based references but ONLY allows referencing facts that appear
-    BEFORE this fact in the list. This prevents hallucination of invalid indices.
-    """
-
-    target_index: int = Field(
-        description="Index of the PREVIOUS fact this relates to (0-based). "
-        "MUST be less than this fact's position in the list. "
-        "Example: if this is fact #5, target_index can only be 0, 1, 2, 3, or 4."
-    )
-    relation_type: Literal["caused_by"] = Field(
-        description="How this fact relates to the target fact: 'caused_by' = this fact was caused by the target fact"
-    )
-    strength: float = Field(
-        description="Strength of relationship (0.0 to 1.0). 1.0 = strong, 0.5 = moderate",
+        description="Strength of causal relationship (0.0 to 1.0). "
+        "1.0 = direct/strong causation, 0.5 = moderate, 0.3 = weak/indirect",
        ge=0.0,
        le=1.0,
        default=1.0,
@@ -150,67 +94,16 @@ class FactCausalRelation(BaseModel):


 class ExtractedFact(BaseModel):
-    """A single extracted fact."""
+    """A single extracted fact with 5 required dimensions for comprehensive capture."""

    model_config = ConfigDict(
        json_schema_mode="validation",
        json_schema_extra={"required": ["what", "when", "where", "who", "why", "fact_type"]},
    )

-    what: str = Field(description="Core fact - concise but complete (1-2 sentences)")
-    when: str = Field(description="When it happened. 'N/A' if unknown.")
-    where: str = Field(description="Location if relevant. 'N/A' if none.")
-    who: str = Field(description="People involved with relationships. 'N/A' if general.")
-    why: str = Field(description="Context/significance if important. 'N/A' if obvious.")
-
-    fact_kind: str = Field(default="conversation", description="'event' or 'conversation'")
-    occurred_start: str | None = Field(default=None, description="ISO timestamp for events")
-    occurred_end: str | None = Field(default=None, description="ISO timestamp for event end")
-    fact_type: Literal["world", "assistant"] = Field(description="'world' or 'assistant'")
-    entities: list[Entity] | None = Field(default=None, description="People, places, concepts")
-    causal_relations: list[FactCausalRelation] | None = Field(
-        default=None, description="Links to previous facts (target_index < this fact's index)"
-    )
-
-    @field_validator("entities", mode="before")
-    @classmethod
-    def ensure_entities_list(cls, v):
-        """Ensure entities is always a list (convert None to empty list)."""
-        if v is None:
-            return []
-        return v
-
-    def build_fact_text(self) -> str:
-        """Combine all dimensions into a single comprehensive fact string."""
-        parts = [self.what]
-
-        # Add 'who' if not N/A
-        if self.who and self.who.upper() != "N/A":
-            parts.append(f"Involving: {self.who}")
-
-        # Add 'why' if not N/A
-        if self.why and self.why.upper() != "N/A":
-            parts.append(self.why)
-
-        if len(parts) == 1:
-            return parts[0]
-
-        return " | ".join(parts)
-
-
-class FactExtractionResponse(BaseModel):
-    """Response containing all extracted facts (causal relations are embedded in each fact)."""
-
-    facts: list[ExtractedFact] = Field(description="List of extracted factual statements")
-
-
-class ExtractedFactVerbose(BaseModel):
-    """A single extracted fact with verbose field descriptions for detailed extraction."""
-
-    model_config = ConfigDict(
-        json_schema_mode="validation",
-        json_schema_extra={"required": ["what", "when", "where", "who", "why", "fact_type"]},
-    )
+    # ==========================================================================
+    # FIVE REQUIRED DIMENSIONS - LLM must think about each one
+    # ==========================================================================

    what: str = Field(
        description="WHAT happened - COMPLETE, DETAILED description with ALL specifics. "
@@ -253,11 +146,16 @@ class ExtractedFactVerbose(BaseModel):
        "NOT: 'User liked it' or 'To help user'"
    )

+    # ==========================================================================
+    # CLASSIFICATION
+    # ==========================================================================
+
    fact_kind: str = Field(
        default="conversation",
        description="'event' = specific datable occurrence (set occurred dates), 'conversation' = general info (no occurred dates)",
    )

+    # Temporal fields - optional
    occurred_start: str | None = Field(
        default=None,
        description="WHEN the event happened (ISO timestamp). Only for fact_kind='event'. Leave null for conversations.",
@@ -267,76 +165,59 @@ class ExtractedFactVerbose(BaseModel):
        description="WHEN the event ended (ISO timestamp). Only for events with duration. Leave null for conversations.",
    )

+    # Classification (CRITICAL - required)
+    # Note: LLM uses "assistant" but we convert to "bank" for storage
    fact_type: Literal["world", "assistant"] = Field(
        description="'world' = about the user/others (background, experiences). 'assistant' = experience with the assistant."
    )

+    # Entities - extracted from fact content
    entities: list[Entity] | None = Field(
        default=None,
        description="Named entities, objects, AND abstract concepts from the fact. Include: people names, organizations, places, significant objects (e.g., 'coffee maker', 'car'), AND abstract concepts/themes (e.g., 'friendship', 'career growth', 'loss', 'celebration'). Extract anything that could help link related facts together.",
    )
-
-    causal_relations: list[FactCausalRelation] | None = Field(
-        default=None,
-        description="Causal links to PREVIOUS facts only. target_index MUST be less than this fact's position. "
-        "Example: fact #3 can only reference facts 0, 1, or 2. Max 2 relations per fact.",
+    causal_relations: list[CausalRelation] | None = Field(
+        default=None, description="Causal links to other facts. Can be null."
    )

    @field_validator("entities", mode="before")
    @classmethod
    def ensure_entities_list(cls, v):
+        """Ensure entities is always a list (convert None to empty list)."""
        if v is None:
            return []
        return v

-
-class FactExtractionResponseVerbose(BaseModel):
-    """Response for verbose fact extraction."""
-
-    facts: list[ExtractedFactVerbose] = Field(description="List of extracted factual statements")
-
-
-class ExtractedFactNoCausal(BaseModel):
-    """A single extracted fact WITHOUT causal relations (for when causal extraction is disabled)."""
-
-    model_config = ConfigDict(
-        json_schema_mode="validation",
-        json_schema_extra={"required": ["what", "when", "where", "who", "why", "fact_type"]},
-    )
-
-    # Same fields as ExtractedFact but without causal_relations
-    what: str = Field(description="WHAT happened - COMPLETE, DETAILED description with ALL specifics.")
-    when: str = Field(description="WHEN it happened - include temporal information if mentioned.")
-    where: str = Field(description="WHERE it happened - SPECIFIC locations if applicable.")
-    who: str = Field(description="WHO is involved - ALL people/entities with relationships.")
-    why: str = Field(description="WHY it matters - emotional, contextual, and motivational details.")
-
-    fact_kind: str = Field(
-        default="conversation",
-        description="'event' = specific datable occurrence, 'conversation' = general info",
-    )
-    occurred_start: str | None = Field(default=None, description="WHEN the event happened (ISO timestamp).")
-    occurred_end: str | None = Field(default=None, description="WHEN the event ended (ISO timestamp).")
-    fact_type: Literal["world", "assistant"] = Field(
-        description="'world' = about the user/others. 'assistant' = experience with assistant."
-    )
-    entities: list[Entity] | None = Field(
-        default=None,
-        description="Named entities, objects, and concepts from the fact.",
-    )
-
-    @field_validator("entities", mode="before")
+    @field_validator("causal_relations", mode="before")
    @classmethod
-    def ensure_entities_list(cls, v):
+    def ensure_causal_relations_list(cls, v):
+        """Ensure causal_relations is always a list (convert None to empty list)."""
        if v is None:
            return []
        return v

+    def build_fact_text(self) -> str:
+        """Combine all dimensions into a single comprehensive fact string."""
+        parts = [self.what]

-class FactExtractionResponseNoCausal(BaseModel):
-    """Response for fact extraction without causal relations."""
+        # Add 'who' if not N/A
+        if self.who and self.who.upper() != "N/A":
+            parts.append(f"Involving: {self.who}")

-    facts: list[ExtractedFactNoCausal] = Field(description="List of extracted factual statements")
+        # Add 'why' if not N/A
+        if self.why and self.why.upper() != "N/A":
+            parts.append(self.why)
+
+        if len(parts) == 1:
+            return parts[0]
+
+        return " | ".join(parts)
+
+
+class FactExtractionResponse(BaseModel):
+    """Response containing all extracted facts."""
+
+    facts: list[ExtractedFact] = Field(description="List of extracted factual statements")


 def chunk_text(text: str, max_chars: int) -> list[str]:
@@ -428,140 +309,39 @@ def _chunk_conversation(turns: list[dict], max_chars: int) -> list[str]:
    return chunks if chunks else [json.dumps(turns, ensure_ascii=False)]


-# =============================================================================
-# FACT EXTRACTION PROMPTS
-# =============================================================================
+async def _extract_facts_from_chunk(
+    chunk: str,
+    chunk_index: int,
+    total_chunks: int,
+    event_date: datetime,
+    context: str,
+    llm_config: "LLMConfig",
+    agent_name: str = None,
+    extract_opinions: bool = False,
+) -> list[dict[str, str]]:
+    """
+    Extract facts from a single chunk (internal helper for parallel processing).

-# Base prompt template (shared by concise and custom modes)
-# 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.
+    Note: event_date parameter is kept for backward compatibility but not used in prompt.
+    The LLM extracts temporal information from the context string instead.
+    """
+    memory_bank_context = f"\n- Your name: {agent_name}" if agent_name and extract_opinions else ""

-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.
+    # Determine which fact types to extract based on the flag
+    # Note: We use "assistant" in the prompt but convert to "bank" for storage
+    if extract_opinions:
+        # Opinion extraction uses a separate prompt (not this one)
+        fact_types_instruction = "Extract ONLY 'opinion' type facts (formed opinions, beliefs, and perspectives). DO NOT extract 'world' or 'assistant' facts."
+    else:
+        fact_types_instruction = (
+            "Extract ONLY 'world' and 'assistant' type facts. DO NOT extract opinions - those are extracted separately."
+        )
+
+    prompt = f"""Extract facts from text into structured format with FOUR required dimensions - BE EXTREMELY DETAILED.

 {fact_types_instruction}

-{extraction_guidelines}

-══════════════════════════════════════════════════════════════════════════
-FACT FORMAT - BE CONCISE
-══════════════════════════════════════════════════════════════════════════
-
-1. **what**: Core fact - concise but complete (1-2 sentences max)
-2. **when**: Temporal info if mentioned. "N/A" if none. Use day name when known.
-3. **where**: Location if relevant. "N/A" if none.
-4. **who**: People involved with relationships. "N/A" if just general info.
-5. **why**: Context/significance ONLY if important. "N/A" if obvious.
-
-CONCISENESS: Capture the essence, not every word. One good sentence beats three mediocre ones.
-
-══════════════════════════════════════════════════════════════════════════
-COREFERENCE RESOLUTION
-══════════════════════════════════════════════════════════════════════════
-
-Link generic references to names when both appear:
- "my roommate" + "Emily" → use "Emily (user's roommate)"
- "the manager" + "Sarah" → use "Sarah (the manager)"
-
-══════════════════════════════════════════════════════════════════════════
-CLASSIFICATION
-══════════════════════════════════════════════════════════════════════════
-
-fact_kind:
- "event": Specific datable occurrence (set occurred_start/end)
- "conversation": Ongoing state, preference, trait (no dates)
-
-fact_type:
- "world": About user's life, other people, external events
- "assistant": Interactions with assistant (requests, recommendations)
-
-══════════════════════════════════════════════════════════════════════════
-TEMPORAL HANDLING
-══════════════════════════════════════════════════════════════════════════
-
-Use "Event Date" from input as reference for relative dates.
- "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
-
-══════════════════════════════════════════════════════════════════════════
-ENTITIES
-══════════════════════════════════════════════════════════════════════════
-
-Include: people names, organizations, places, key objects, abstract concepts (career, friendship, etc.)
-Always include "user" when fact is about the user.{examples}"""
-
-# Concise mode guidelines
-_CONCISE_GUIDELINES = """══════════════════════════════════════════════════════════════════════════
-SELECTIVITY - CRITICAL (Reduces 90% of unnecessary output)
-══════════════════════════════════════════════════════════════════════════
-
-ONLY extract facts that are:
-✅ Personal info: names, relationships, roles, background
-✅ Preferences: likes, dislikes, habits, interests (e.g., "Alice likes coffee")
-✅ Significant events: milestones, decisions, achievements, changes
-✅ Plans/goals: future intentions, deadlines, commitments
-✅ Expertise: skills, knowledge, certifications, experience
-✅ Important context: projects, problems, constraints
-✅ Sensory/emotional details: feelings, sensations, perceptions that provide context
-✅ Observations: descriptions of people, places, things with specific details
-
-DO NOT extract:
-❌ Generic greetings: "how are you", "hello", pleasantries without substance
-❌ Pure filler: "thanks", "sounds good", "ok", "got it", "sure"
-❌ Process chatter: "let me check", "one moment", "I'll look into it"
-❌ Repeated info: if already stated, don't extract again
-
-CONSOLIDATE related statements into ONE fact when possible."""
-
-# Concise mode examples
-_CONCISE_EXAMPLES = """
-
-══════════════════════════════════════════════════════════════════════════
-EXAMPLES
-══════════════════════════════════════════════════════════════════════════
-
-Example 1 - Selective extraction (Event Date: June 10, 2024):
-Input: "Hey! How's it going? Good morning! So I'm planning my wedding - want a small outdoor ceremony. Just got back from Emily's wedding, she married Sarah at a rooftop garden. It was nice weather. I grabbed a coffee on the way."
-
-Output: ONLY 2 facts (skip greetings, weather, coffee):
-1. what="User planning wedding, wants small outdoor ceremony", who="user", why="N/A", entities=["user", "wedding"]
-2. what="Emily married Sarah at rooftop garden", who="Emily (user's friend), Sarah", occurred_start="2024-06-09", entities=["Emily", "Sarah", "wedding"]
-
-Example 2 - Professional context:
-Input: "Alice has 5 years of Kubernetes experience and holds CKA certification. She's been leading the infrastructure team since March. By the way, she prefers dark roast coffee."
-
-Output: ONLY 2 facts (skip coffee preference - too trivial):
-1. what="Alice has 5 years Kubernetes experience, CKA certified", who="Alice", entities=["Alice", "Kubernetes", "CKA"]
-2. what="Alice leads infrastructure team since March", who="Alice", entities=["Alice", "infrastructure"]
-
-══════════════════════════════════════════════════════════════════════════
-QUALITY OVER QUANTITY
-══════════════════════════════════════════════════════════════════════════
-
-Ask: "Would this be useful to recall in 6 months?" If no, skip it."""
-
-# Assembled concise prompt (backward compatible - exact same output as before)
-CONCISE_FACT_EXTRACTION_PROMPT = _BASE_FACT_EXTRACTION_PROMPT.format(
-    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(
-    fact_types_instruction="{fact_types_instruction}",
-    extraction_guidelines="{custom_instructions}",
-    examples="",  # No examples for custom mode
-)
-
-
-# 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 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
@@ -655,100 +435,113 @@ FACT TYPE
  Include: what the user asked, what problem they wanted solved, what context they provided

 ══════════════════════════════════════════════════════════════════════════
-ENTITIES - EXTRACT EVERYTHING
+USER PREFERENCES (CRITICAL)
 ══════════════════════════════════════════════════════════════════════════

-Extract ALL of the following from the fact:
- People names (Emily, Alice, Dr. Smith)
- Organizations (Google, MIT, local coffee shop)
- Places (San Francisco, Brooklyn, Paris)
- Significant objects mentioned (coffee maker, new car, wedding dress)
- Abstract concepts/themes (friendship, career growth, loss, celebration)
+ALWAYS extract user preferences as separate facts! Watch for these keywords:
+- "enjoy", "like", "love", "prefer", "hate", "dislike", "favorite", "ideal", "dream", "want"

-ALWAYS include "user" when fact is about the user.
-Extract anything that could help link related facts together."""
-
-
-# Causal relationships section - appended when causal extraction is enabled
-CAUSAL_RELATIONSHIPS_SECTION = """
+Example: "I love Italian food and prefer outdoor dining"
+→ Fact 1: what="User loves Italian food", who="user", why="This is a food preference", entities=["user"]
+→ Fact 2: what="User prefers outdoor dining", who="user", why="This is a dining preference", entities=["user"]

 ══════════════════════════════════════════════════════════════════════════
-CAUSAL RELATIONSHIPS
+ENTITIES - INCLUDE PEOPLE, PLACES, OBJECTS, AND CONCEPTS (CRITICAL)
 ══════════════════════════════════════════════════════════════════════════

-Link facts with causal_relations (max 2 per fact). target_index must be < this fact's index.
-Type: "caused_by" (this fact was caused by the target fact)
+Extract entities that help link related facts together. Include:
+1. "user" - when the fact is about the user
+2. People names - Emily, Dr. Smith, etc.
+3. Organizations/Places - IKEA, Goodwill, New York, etc.
+4. Specific objects - coffee maker, toaster, car, laptop, kitchen, etc.
+5. Abstract concepts - themes, values, emotions, or ideas that capture the essence of the fact:
+   - "friendship" for facts about friends helping each other, bonding, loyalty
+   - "career growth" for facts about promotions, learning new skills, job changes
+   - "loss" or "grief" for facts about death, endings, saying goodbye
+   - "celebration" for facts about parties, achievements, milestones
+   - "trust" or "betrayal" for facts involving those themes

-Example: "Lost job → couldn't pay rent → moved apartment"
- Fact 0: Lost job, causal_relations: null
- Fact 1: Couldn't pay rent, causal_relations: [{target_index: 0, relation_type: "caused_by"}]
- Fact 2: Moved apartment, causal_relations: [{target_index: 1, relation_type: "caused_by"}]"""
+✅ CORRECT: entities=["user", "coffee maker", "Goodwill", "kitchen"] for "User donated their coffee maker to Goodwill"
+✅ CORRECT: entities=["user", "Emily", "friendship"] for "Emily helped user move to a new apartment"
+✅ CORRECT: entities=["user", "promotion", "career growth"] for "User got promoted to senior engineer"
+✅ CORRECT: entities=["user", "grandmother", "loss", "grief"] for "User's grandmother passed away last week"
+❌ WRONG: entities=["user", "Emily"] only - missing the "friendship" concept that links to other friendship facts!

+══════════════════════════════════════════════════════════════════════════
+EXAMPLES
+══════════════════════════════════════════════════════════════════════════

-async def _extract_facts_from_chunk(
-    chunk: str,
-    chunk_index: int,
-    total_chunks: int,
-    event_date: datetime,
-    context: str,
-    llm_config: "LLMConfig",
-    agent_name: str = None,
-) -> tuple[list[dict[str, str]], TokenUsage]:
-    """
-    Extract facts from a single chunk (internal helper for parallel processing).
+Example 1 - World Facts (Event Date: Tuesday, June 10, 2024):
+Input: "I'm planning my wedding and want a small outdoor ceremony. I just got back from my college roommate Emily's wedding - she married Sarah at a rooftop garden, it was so romantic!"
+
+Output facts:
+
+1. User's wedding preference
+   - what: "User wants a small outdoor ceremony for their wedding"
+   - who: "user"
+   - why: "User prefers intimate outdoor settings"
+   - fact_type: "world", fact_kind: "conversation"
+   - entities: ["user", "wedding", "outdoor ceremony"]
+
+2. User planning wedding
+   - what: "User is planning their own wedding"
+   - who: "user"
+   - why: "Inspired by Emily's ceremony"
+   - fact_type: "world", fact_kind: "conversation"
+   - entities: ["user", "wedding"]
+
+3. Emily's wedding (THE EVENT - note occurred_start AND occurred_end both set)
+   - what: "Emily got married to Sarah at a rooftop garden ceremony in the city"
+   - who: "Emily (user's college roommate), Sarah (Emily's partner)"
+   - why: "User found it romantic and beautiful"
+   - fact_type: "world", fact_kind: "event"
+   - occurred_start: "2024-06-09T00:00:00Z" (recently, user "just got back" - relative to Event Date June 10, 2024)
+   - occurred_end: "2024-06-09T23:59:59Z" (same day - point event)
+   - entities: ["user", "Emily", "Sarah", "wedding", "rooftop garden"]
+
+Example 2 - Assistant Facts (Context: March 5, 2024):
+Input: "User: My API is really slow when we have 1000+ concurrent users. What can I do?
+Assistant: I'd recommend implementing Redis for caching frequently-accessed data, which should reduce your database load by 70-80%."
+
+Output fact:
+   - what: "Assistant recommended implementing Redis for caching frequently-accessed data to improve API performance"
+   - when: "March 5, 2024 during conversation"
+   - who: "user, assistant"
+   - why: "User asked how to fix slow API performance with 1000+ concurrent users, expected 70-80% reduction in database load"
+   - fact_type: "assistant", fact_kind: "conversation"
+   - entities: ["user", "API", "Redis"]
+
+Example 3 - Kitchen Items with Concept Inference (Event Date: Thursday, May 30, 2024):
+Input: "I finally donated my old coffee maker to Goodwill. I upgraded to that new espresso machine last month and the old one was just taking up counter space."
+
+Output fact:
+   - what: "User donated their old coffee maker to Goodwill after upgrading to a new espresso machine"
+   - when: "Thursday, May 30, 2024"
+   - who: "user"
+   - why: "The old coffee maker was taking up counter space after the upgrade"
+   - fact_type: "world", fact_kind: "event"
+   - occurred_start: "2024-05-30T00:00:00Z" (uses Event Date year)
+   - occurred_end: "2024-05-30T23:59:59Z" (same day - point event)
+   - entities: ["user", "coffee maker", "Goodwill", "espresso machine", "kitchen"]
+
+Note: "kitchen" is inferred as a concept because coffee makers and espresso machines are kitchen appliances.
+This links the fact to other kitchen-related facts (toaster, faucet, kitchen mat, etc.) via the shared "kitchen" entity.
+
+Note how the "why" field captures the FULL STORY: what the user asked AND what outcome was expected!
+
+══════════════════════════════════════════════════════════════════════════
+WHAT TO EXTRACT vs SKIP
+══════════════════════════════════════════════════════════════════════════
+
+✅ EXTRACT: User preferences (ALWAYS as separate facts!), feelings, plans, events, relationships, achievements
+❌ SKIP: Greetings, filler ("thanks", "cool"), purely structural statements"""

-    Note: event_date parameter is kept for backward compatibility but not used in prompt.
-    The LLM extracts temporal information from the context string instead.
-    """
    import logging

    from openai import BadRequestError

    logger = logging.getLogger(__name__)

-    # 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
-
-    # 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
@@ -761,6 +554,7 @@ async def _extract_facts_from_chunk(
    # Format event_date with day of week for better temporal reasoning
    event_date_formatted = event_date.strftime("%A, %B %d, %Y")  # e.g., "Monday, June 10, 2024"
    user_message = f"""Extract facts from the following text chunk.
+{memory_bank_context}

 Chunk: {chunk_index + 1}/{total_chunks}
 Event Date: {event_date_formatted} ({event_date.isoformat()})
@@ -769,35 +563,16 @@ Context: {sanitized_context}
 Text:
 {sanitized_chunk}"""

-    usage = TokenUsage()  # Track cumulative usage across retries
    for attempt in range(max_retries):
        try:
-            # Use retain-specific overrides if set, otherwise fall back to global LLM config
-            max_retries = (
-                config.retain_llm_max_retries if config.retain_llm_max_retries is not None else config.llm_max_retries
-            )
-            initial_backoff = (
-                config.retain_llm_initial_backoff
-                if config.retain_llm_initial_backoff is not None
-                else config.llm_initial_backoff
-            )
-            max_backoff = (
-                config.retain_llm_max_backoff if config.retain_llm_max_backoff is not None else config.llm_max_backoff
-            )
-
-            extraction_response_json, call_usage = await llm_config.call(
+            extraction_response_json = await llm_config.call(
                messages=[{"role": "system", "content": prompt}, {"role": "user", "content": user_message}],
-                response_format=response_schema,
+                response_format=FactExtractionResponse,
                scope="memory_extract_facts",
                temperature=0.1,
-                max_completion_tokens=config.retain_max_completion_tokens,
-                max_retries=max_retries,
-                initial_backoff=initial_backoff,
-                max_backoff=max_backoff,
+                max_completion_tokens=65000,
                skip_validation=True,  # Get raw JSON, we'll validate leniently
-                return_usage=True,
            )
-            usage = usage + call_usage  # Aggregate usage across retries

            # Lenient parsing of facts from raw JSON
            chunk_facts = []
@@ -815,10 +590,9 @@ Text:
                        f"LLM returned non-dict JSON after {max_retries} attempts: {type(extraction_response_json).__name__}. "
                        f"Raw: {str(extraction_response_json)[:500]}"
                    )
-                    return [], usage
+                    return []

            raw_facts = extraction_response_json.get("facts", [])
-
            if not raw_facts:
                logger.debug(
                    f"LLM response missing 'facts' field or returned empty list. "
@@ -858,8 +632,7 @@ Text:

                # Critical field: fact_type
                # LLM uses "assistant" but we convert to "experience" for storage
-                original_fact_type = llm_fact.get("fact_type")
-                fact_type = original_fact_type
+                fact_type = llm_fact.get("fact_type")

                # Convert "assistant" → "experience" for storage
                if fact_type == "assistant":
@@ -876,10 +649,7 @@ Text:
                    else:
                        # Default to 'world' if we can't determine
                        fact_type = "world"
-                        logger.warning(
-                            f"Fact {i}: defaulting to fact_type='world' "
-                            f"(original fact_type={original_fact_type!r}, fact_kind={fact_kind!r})"
-                        )
+                        logger.warning(f"Fact {i}: defaulting to fact_type='world'")

                # Get fact_kind for temporal handling (but don't store it)
                fact_kind = llm_fact.get("fact_kind", "conversation")
@@ -906,18 +676,13 @@ Text:
                if fact_kind == "event":
                    occurred_start = get_value("occurred_start")
                    occurred_end = get_value("occurred_end")
-
-                    # If LLM didn't set temporal fields, try to extract them from the fact text
-                    if not occurred_start:
-                        fact_data["occurred_start"] = _infer_temporal_date(combined_text, event_date)
-                    else:
+                    if occurred_start:
                        fact_data["occurred_start"] = occurred_start
-
-                    # For point events: if occurred_end not set, default to occurred_start
-                    if occurred_end:
-                        fact_data["occurred_end"] = occurred_end
-                    elif fact_data.get("occurred_start"):
-                        fact_data["occurred_end"] = fact_data["occurred_start"]
+                        # For point events: if occurred_end not set, default to occurred_start
+                        if occurred_end:
+                            fact_data["occurred_end"] = occurred_end
+                        else:
+                            fact_data["occurred_end"] = occurred_start

                # Add entities if present (validate as Entity objects)
                # LLM sometimes returns strings instead of {"text": "..."} format
@@ -937,40 +702,17 @@ Text:
                    if validated_entities:
                        fact_data["entities"] = validated_entities

-                # Add per-fact causal relations (only if enabled in config)
-                if extract_causal_links:
+                # Add causal relations if present (validate as CausalRelation objects)
+                # Filter out invalid relations (missing required fields)
+                causal_relations = get_value("causal_relations")
+                if causal_relations:
                    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
-                            # New schema uses target_index
-                            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
-
-                            # Validate: target_index must be < current fact index
-                            if target_idx < 0 or target_idx >= i:
-                                logger.debug(
-                                    f"Invalid target_index {target_idx} for fact {i} (must be 0 to {i - 1}). Skipping."
-                                )
-                                continue
-
+                    for rel in causal_relations:
+                        if isinstance(rel, dict) and "target_fact_index" in rel and "relation_type" in rel:
                            try:
-                                validated_relations.append(
-                                    CausalRelation(
-                                        target_fact_index=target_idx,
-                                        relation_type=relation_type,
-                                        strength=strength,
-                                    )
-                                )
+                                validated_relations.append(CausalRelation.model_validate(rel))
                            except Exception as e:
-                                logger.debug(f"Invalid causal relation {rel}: {e}")
-
+                                logger.warning(f"Invalid causal relation {rel}: {e}")
                    if validated_relations:
                        fact_data["causal_relations"] = validated_relations

@@ -993,7 +735,7 @@ Text:
                )
                continue

-            return chunk_facts, usage
+            return chunk_facts

        except BadRequestError as e:
            last_error = e
@@ -1019,7 +761,8 @@ async def _extract_facts_with_auto_split(
    context: str,
    llm_config: LLMConfig,
    agent_name: str = None,
-) -> tuple[list[dict[str, str]], TokenUsage]:
+    extract_opinions: bool = False,
+) -> list[dict[str, str]]:
    """
    Extract facts from a chunk with automatic splitting if output exceeds token limits.

@@ -1034,9 +777,10 @@ async def _extract_facts_with_auto_split(
        context: Context about the conversation/document
        llm_config: LLM configuration to use
        agent_name: Optional agent name (memory owner)
+        extract_opinions: If True, extract ONLY opinions. If False, extract world and agent facts (no opinions)

    Returns:
-        Tuple of (facts list, token usage) extracted from the chunk (possibly from sub-chunks)
+        List of fact dictionaries extracted from the chunk (possibly from sub-chunks)
    """
    import logging

@@ -1052,6 +796,7 @@ async def _extract_facts_with_auto_split(
            context=context,
            llm_config=llm_config,
            agent_name=agent_name,
+            extract_opinions=extract_opinions,
        )
    except OutputTooLongError:
        # Output exceeded token limits - split the chunk in half and retry
@@ -1096,6 +841,7 @@ async def _extract_facts_with_auto_split(
                context=context,
                llm_config=llm_config,
                agent_name=agent_name,
+                extract_opinions=extract_opinions,
            ),
            _extract_facts_with_auto_split(
                chunk=second_half,
@@ -1105,6 +851,7 @@ async def _extract_facts_with_auto_split(
                context=context,
                llm_config=llm_config,
                agent_name=agent_name,
+                extract_opinions=extract_opinions,
            ),
        ]

@@ -1112,14 +859,12 @@ async def _extract_facts_with_auto_split(

        # Combine results from both halves
        all_facts = []
-        total_usage = TokenUsage()
-        for sub_facts, sub_usage in sub_results:
-            all_facts.extend(sub_facts)
-            total_usage = total_usage + sub_usage
+        for sub_result in sub_results:
+            all_facts.extend(sub_result)

        logger.info(f"Successfully extracted {len(all_facts)} facts from split chunk {chunk_index + 1}")

-        return all_facts, total_usage
+        return all_facts


 async def extract_facts_from_text(
@@ -1128,7 +873,8 @@ async def extract_facts_from_text(
    llm_config: LLMConfig,
    agent_name: str,
    context: str = "",
-) -> tuple[list[Fact], list[tuple[str, int]], TokenUsage]:
+    extract_opinions: bool = False,
+) -> tuple[list[Fact], list[tuple[str, int]]]:
    """
    Extract semantic facts from conversational or narrative text using LLM.

@@ -1144,24 +890,14 @@ async def extract_facts_from_text(
        context: Context about the conversation/document
        llm_config: LLM configuration to use
        agent_name: Agent name (memory owner)
+        extract_opinions: If True, extract ONLY opinions. If False, extract world and bank facts (no opinions)

    Returns:
-        Tuple of (facts, chunks, usage) where:
+        Tuple of (facts, chunks) where:
        - facts: List of Fact model instances
        - 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
-    total_chars = sum(len(c) for c in chunks)
-    if len(chunks) > 1:
-        logger.debug(
-            f"[FACT_EXTRACTION] Text chunked into {len(chunks)} chunks ({total_chars:,} chars total, "
-            f"chunk_size={config.retain_chunk_size:,}) - starting parallel LLM extraction"
-        )
-
+    chunks = chunk_text(text, max_chars=3000)
    tasks = [
        _extract_facts_with_auto_split(
            chunk=chunk,
@@ -1171,18 +907,17 @@ async def extract_facts_from_text(
            context=context,
            llm_config=llm_config,
            agent_name=agent_name,
+            extract_opinions=extract_opinions,
        )
        for i, chunk in enumerate(chunks)
    ]
    chunk_results = await asyncio.gather(*tasks)
    all_facts = []
    chunk_metadata = []  # [(chunk_text, fact_count), ...]
-    total_usage = TokenUsage()
-    for chunk, (chunk_facts, chunk_usage) in zip(chunks, chunk_results):
+    for chunk, chunk_facts in zip(chunks, chunk_results):
        all_facts.extend(chunk_facts)
        chunk_metadata.append((chunk, len(chunk_facts)))
-        total_usage = total_usage + chunk_usage
-    return all_facts, chunk_metadata, total_usage
+    return all_facts, chunk_metadata


 # ============================================================================
@@ -1202,8 +937,8 @@ SECONDS_PER_FACT = 10


 async def extract_facts_from_contents(
-    contents: list[RetainContent], llm_config, agent_name: str
-) -> tuple[list[ExtractedFactType], list[ChunkMetadata], TokenUsage]:
+    contents: list[RetainContent], llm_config, agent_name: str, extract_opinions: bool = False
+) -> tuple[list[ExtractedFactType], list[ChunkMetadata]]:
    """
    Extract facts from multiple content items in parallel.

@@ -1217,12 +952,13 @@ async def extract_facts_from_contents(
        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)
+        extract_opinions: If True, extract only opinions; otherwise world/bank facts

    Returns:
-        Tuple of (extracted_facts, chunks_metadata, usage)
+        Tuple of (extracted_facts, chunks_metadata)
    """
    if not contents:
-        return [], [], TokenUsage()
+        return [], []

    # Step 1: Create parallel fact extraction tasks
    fact_extraction_tasks = []
@@ -1235,6 +971,7 @@ async def extract_facts_from_contents(
            context=item.context,
            llm_config=llm_config,
            agent_name=agent_name,
+            extract_opinions=extract_opinions,
        )
        fact_extraction_tasks.append(task)

@@ -1244,15 +981,11 @@ async def extract_facts_from_contents(
    # Step 3: Flatten and convert to typed objects
    extracted_facts: list[ExtractedFactType] = []
    chunks_metadata: list[ChunkMetadata] = []
-    total_usage = TokenUsage()

    global_chunk_idx = 0
    global_fact_idx = 0

-    for content_index, (content, (facts_from_llm, chunks_from_llm, content_usage)) in enumerate(
-        zip(contents, all_fact_results)
-    ):
-        total_usage = total_usage + content_usage
+    for content_index, (content, (facts_from_llm, chunks_from_llm)) in enumerate(zip(contents, all_fact_results)):
        chunk_start_idx = global_chunk_idx

        # Convert chunk tuples to ChunkMetadata objects
@@ -1297,7 +1030,6 @@ async def extract_facts_from_contents(
                        # mentioned_at: always the event_date (when the conversation/document occurred)
                        mentioned_at=content.event_date,
                        metadata=content.metadata,
-                        tags=content.tags,
                    )

                    extracted_facts.append(extracted_fact)
@@ -1307,7 +1039,7 @@ async def extract_facts_from_contents(
    # Step 4: Add time offsets to preserve ordering within each content
    _add_temporal_offsets(extracted_facts, contents)

-    return extracted_facts, chunks_metadata, total_usage
+    return extracted_facts, chunks_metadata


 def _parse_datetime(date_str: str):
@@ -41,10 +41,10 @@ async def insert_facts_batch(
    contexts = []
    fact_types = []
    confidence_scores = []
+    access_counts = []
    metadata_jsons = []
    chunk_ids = []
    document_ids = []
-    tags_list = []

    for fact in facts:
        fact_texts.append(fact.fact_text)
@@ -60,35 +60,21 @@ async def insert_facts_batch(
        fact_types.append(fact.fact_type)
        # confidence_score is only for opinion facts
        confidence_scores.append(1.0 if fact.fact_type == "opinion" else None)
+        access_counts.append(0)  # Initial access count
        metadata_jsons.append(json.dumps(fact.metadata))
        chunk_ids.append(fact.chunk_id)
        # Use per-fact document_id if available, otherwise fallback to batch-level document_id
        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 []))

    # Batch insert all facts
-    # Note: tags are passed as JSON strings and converted back to varchar[] via jsonb_array_elements_text + array_agg
    results = await conn.fetch(
        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
+                                 context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id)
+        SELECT $1, * FROM unnest(
+            $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
+            $8::text[], $9::text[], $10::float[], $11::int[], $12::jsonb[], $13::text[], $14::text[]
+        )
        RETURNING id
        """,
        bank_id,
@@ -101,10 +87,10 @@ async def insert_facts_batch(
        contexts,
        fact_types,
        confidence_scores,
+        access_counts,
        metadata_jsons,
        chunk_ids,
        document_ids,
-        tags_list,
    )

    unit_ids = [str(row["id"]) for row in results]
@@ -123,7 +109,7 @@ async def ensure_bank_exists(conn, bank_id: str) -> None:
    """
    await conn.execute(
        f"""
-        INSERT INTO {fq_table("banks")} (bank_id, disposition, mission)
+        INSERT INTO {fq_table("banks")} (bank_id, disposition, background)
        VALUES ($1, $2::jsonb, $3)
        ON CONFLICT (bank_id) DO UPDATE
        SET updated_at = NOW()
@@ -135,13 +121,7 @@ async def ensure_bank_exists(conn, bank_id: str) -> None:


 async def handle_document_tracking(
-    conn,
-    bank_id: str,
-    document_id: str,
-    combined_content: str,
-    is_first_batch: bool,
-    retain_params: dict | None = None,
-    document_tags: list[str] | None = None,
+    conn, bank_id: str, document_id: str, combined_content: str, is_first_batch: bool, retain_params: dict | None = None
 ) -> None:
    """
    Handle document tracking in the database.
@@ -153,7 +133,6 @@ async def handle_document_tracking(
        combined_content: Combined content text from all content items
        is_first_batch: Whether this is the first batch (for chunked operations)
        retain_params: Optional parameters passed during retain (context, event_date, etc.)
-        document_tags: Optional list of tags to associate with the document
    """
    import hashlib

@@ -170,14 +149,13 @@ async def handle_document_tracking(
    # Insert document (or update if exists from concurrent operations)
    await conn.execute(
        f"""
-        INSERT INTO {fq_table("documents")} (id, bank_id, original_text, content_hash, metadata, retain_params, tags)
-        VALUES ($1, $2, $3, $4, $5, $6, $7)
+        INSERT INTO {fq_table("documents")} (id, bank_id, original_text, content_hash, metadata, retain_params)
+        VALUES ($1, $2, $3, $4, $5, $6)
        ON CONFLICT (id, bank_id) DO UPDATE
        SET original_text = EXCLUDED.original_text,
            content_hash = EXCLUDED.content_hash,
            metadata = EXCLUDED.metadata,
            retain_params = EXCLUDED.retain_params,
-            tags = EXCLUDED.tags,
            updated_at = NOW()
        """,
        document_id,
@@ -186,5 +164,4 @@ async def handle_document_tracking(
        content_hash,
        json.dumps({}),  # Empty metadata dict
        json.dumps(retain_params) if retain_params else None,
-        document_tags or [],
    )
@@ -479,18 +479,14 @@ async def create_temporal_links_batch_per_fact(

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

        return len(links)
@@ -648,18 +644,14 @@ async def create_semantic_links_batch(

        if all_links:
            insert_start = time_mod.time()
-            # Batch inserts to avoid timeout on large batches
-            BATCH_SIZE = 1000
-            for batch_start in range(0, len(all_links), BATCH_SIZE):
-                batch = all_links[batch_start : batch_start + BATCH_SIZE]
-                await conn.executemany(
-                    f"""
-                    INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
-                    VALUES ($1, $2, $3, $4, $5)
-                    ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
-                    """,
-                    batch,
-                )
+            await conn.executemany(
+                f"""
+                INSERT INTO {fq_table("memory_links")} (from_unit_id, to_unit_id, link_type, weight, entity_id)
+                VALUES ($1, $2, $3, $4, $5)
+                ON CONFLICT (from_unit_id, to_unit_id, link_type, COALESCE(entity_id, '00000000-0000-0000-0000-000000000000'::uuid)) DO NOTHING
+                """,
+                all_links,
+            )
            _log(
                log_buffer, f"      [8.3] Insert {len(all_links)} semantic links: {time_mod.time() - insert_start:.3f}s"
            )
@@ -754,14 +746,17 @@ async def create_causal_links_batch(
        causal_relations_per_fact: List of causal relations for each fact.
            Each element is a list of dicts with:
            - target_fact_index: Index into unit_ids for the target fact
-            - relation_type: "caused_by"
+            - relation_type: "causes", "caused_by", "enables", or "prevents"
            - strength: Float in [0.0, 1.0] representing relationship strength

    Returns:
        Number of causal links created

-    Causal link type:
-    - "caused_by": This fact was caused by the target fact
+    Causal link types:
+    - "causes": This fact directly causes the target fact (forward causation)
+    - "caused_by": This fact was caused by the target fact (backward causation)
+    - "enables": This fact enables/allows the target fact (enablement)
+    - "prevents": This fact prevents/blocks the target fact (prevention)
    """
    if not unit_ids or not causal_relations_per_fact:
        return 0
@@ -784,8 +779,8 @@ async def create_causal_links_batch(
                relation_type = relation["relation_type"]
                strength = relation.get("strength", 1.0)

-                # Validate relation_type - only "caused_by" is supported (DB constraint)
-                valid_types = {"caused_by"}
+                # Validate relation_type - must match database constraint
+                valid_types = {"causes", "caused_by", "enables", "prevents"}
                if relation_type not in valid_types:
                    logger.error(
                        f"Invalid relation_type '{relation_type}' (type: {type(relation_type).__name__}) "
@@ -0,0 +1,252 @@
+"""
+Observation regeneration for retain pipeline.
+
+Regenerates entity observations as part of the retain transaction.
+"""
+
+import logging
+import time
+import uuid
+from datetime import UTC, datetime
+
+from ..memory_engine import fq_table
+from ..search import observation_utils
+from . import embedding_utils
+from .types import EntityLink
+
+logger = logging.getLogger(__name__)
+
+
+def utcnow():
+    """Get current UTC time."""
+    return datetime.now(UTC)
+
+
+# Simple dataclass-like container for facts (avoid importing from memory_engine)
+class MemoryFactForObservation:
+    def __init__(self, id: str, text: str, fact_type: str, context: str, occurred_start: str | None):
+        self.id = id
+        self.text = text
+        self.fact_type = fact_type
+        self.context = context
+        self.occurred_start = occurred_start
+
+
+async def regenerate_observations_batch(
+    conn, embeddings_model, llm_config, bank_id: str, entity_links: list[EntityLink], log_buffer: list[str] = None
+) -> None:
+    """
+    Regenerate observations for top entities in this batch.
+
+    Called INSIDE the retain transaction for atomicity - if observations
+    fail, the entire retain batch is rolled back.
+
+    Args:
+        conn: Database connection (from the retain transaction)
+        embeddings_model: Embeddings model for generating observation embeddings
+        llm_config: LLM configuration for observation extraction
+        bank_id: Bank identifier
+        entity_links: Entity links from this batch
+        log_buffer: Optional log buffer for timing
+    """
+    TOP_N_ENTITIES = 5
+    MIN_FACTS_THRESHOLD = 5
+
+    if not entity_links:
+        return
+
+    # Count mentions per entity in this batch
+    entity_mention_counts: dict[str, int] = {}
+    for link in entity_links:
+        if link.entity_id:
+            entity_id = str(link.entity_id)
+            entity_mention_counts[entity_id] = entity_mention_counts.get(entity_id, 0) + 1
+
+    if not entity_mention_counts:
+        return
+
+    # Sort by mention count descending and take top N
+    sorted_entities = sorted(entity_mention_counts.items(), key=lambda x: x[1], reverse=True)
+    entities_to_process = [e[0] for e in sorted_entities[:TOP_N_ENTITIES]]
+
+    obs_start = time.time()
+
+    # Convert to UUIDs
+    entity_uuids = [uuid.UUID(eid) if isinstance(eid, str) else eid for eid in entities_to_process]
+
+    # Batch query for entity names
+    entity_rows = await conn.fetch(
+        f"""
+        SELECT id, canonical_name FROM {fq_table("entities")}
+        WHERE id = ANY($1) AND bank_id = $2
+        """,
+        entity_uuids,
+        bank_id,
+    )
+    entity_names = {row["id"]: row["canonical_name"] for row in entity_rows}
+
+    # Batch query for fact counts
+    fact_counts = await conn.fetch(
+        f"""
+        SELECT ue.entity_id, COUNT(*) as cnt
+        FROM {fq_table("unit_entities")} ue
+        JOIN {fq_table("memory_units")} mu ON ue.unit_id = mu.id
+        WHERE ue.entity_id = ANY($1) AND mu.bank_id = $2
+        GROUP BY ue.entity_id
+        """,
+        entity_uuids,
+        bank_id,
+    )
+    entity_fact_counts = {row["entity_id"]: row["cnt"] for row in fact_counts}
+
+    # Filter entities that meet the threshold
+    entities_with_names = []
+    for entity_id in entities_to_process:
+        entity_uuid = uuid.UUID(entity_id) if isinstance(entity_id, str) else entity_id
+        if entity_uuid not in entity_names:
+            continue
+        fact_count = entity_fact_counts.get(entity_uuid, 0)
+        if fact_count >= MIN_FACTS_THRESHOLD:
+            entities_with_names.append((entity_id, entity_names[entity_uuid]))
+
+    if not entities_with_names:
+        return
+
+    # Process entities SEQUENTIALLY (asyncpg doesn't allow concurrent queries on same connection)
+    # We must use the same connection to stay in the retain transaction
+    total_observations = 0
+
+    for entity_id, entity_name in entities_with_names:
+        try:
+            obs_ids = await _regenerate_entity_observations(
+                conn, embeddings_model, llm_config, bank_id, entity_id, entity_name
+            )
+            total_observations += len(obs_ids)
+        except Exception as e:
+            logger.error(f"[OBSERVATIONS] Error processing entity {entity_id}: {e}")
+
+    obs_time = time.time() - obs_start
+    if log_buffer is not None:
+        log_buffer.append(
+            f"[11] Observations: {total_observations} observations for {len(entities_with_names)} entities in {obs_time:.3f}s"
+        )
+
+
+async def _regenerate_entity_observations(
+    conn, embeddings_model, llm_config, bank_id: str, entity_id: str, entity_name: str
+) -> list[str]:
+    """
+    Regenerate observations for a single entity.
+
+    Uses the provided connection (part of retain transaction).
+
+    Args:
+        conn: Database connection (from the retain transaction)
+        embeddings_model: Embeddings model
+        llm_config: LLM configuration
+        bank_id: Bank identifier
+        entity_id: Entity UUID
+        entity_name: Canonical name of the entity
+
+    Returns:
+        List of created observation IDs
+    """
+    entity_uuid = uuid.UUID(entity_id) if isinstance(entity_id, str) else entity_id
+
+    # Get all facts mentioning this entity (exclude observations themselves)
+    rows = await conn.fetch(
+        f"""
+        SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.fact_type
+        FROM {fq_table("memory_units")} mu
+        JOIN {fq_table("unit_entities")} ue ON mu.id = ue.unit_id
+        WHERE mu.bank_id = $1
+          AND ue.entity_id = $2
+          AND mu.fact_type IN ('world', 'experience')
+        ORDER BY mu.occurred_start DESC
+        LIMIT 50
+        """,
+        bank_id,
+        entity_uuid,
+    )
+
+    if not rows:
+        return []
+
+    # Convert to fact objects for observation extraction
+    facts = []
+    for row in rows:
+        occurred_start = row["occurred_start"].isoformat() if row["occurred_start"] else None
+        facts.append(
+            MemoryFactForObservation(
+                id=str(row["id"]),
+                text=row["text"],
+                fact_type=row["fact_type"],
+                context=row["context"],
+                occurred_start=occurred_start,
+            )
+        )
+
+    # Extract observations using LLM
+    observations = await observation_utils.extract_observations_from_facts(llm_config, entity_name, facts)
+
+    if not observations:
+        return []
+
+    # Delete old observations for this entity
+    await conn.execute(
+        f"""
+        DELETE FROM {fq_table("memory_units")}
+        WHERE id IN (
+            SELECT mu.id
+            FROM {fq_table("memory_units")} mu
+            JOIN {fq_table("unit_entities")} ue ON mu.id = ue.unit_id
+            WHERE mu.bank_id = $1
+              AND mu.fact_type = 'observation'
+              AND ue.entity_id = $2
+        )
+        """,
+        bank_id,
+        entity_uuid,
+    )
+
+    # Generate embeddings for new observations
+    embeddings = await embedding_utils.generate_embeddings_batch(embeddings_model, observations)
+
+    # Insert new observations
+    current_time = utcnow()
+    created_ids = []
+
+    for obs_text, embedding in zip(observations, embeddings):
+        result = await conn.fetchrow(
+            f"""
+            INSERT INTO {fq_table("memory_units")} (
+                bank_id, text, embedding, context, event_date,
+                occurred_start, occurred_end, mentioned_at,
+                fact_type, access_count
+            )
+            VALUES ($1, $2, $3, $4, $5, $6, $7, $8, 'observation', 0)
+            RETURNING id
+            """,
+            bank_id,
+            obs_text,
+            str(embedding),
+            f"observation about {entity_name}",
+            current_time,
+            current_time,
+            current_time,
+            current_time,
+        )
+        obs_id = str(result["id"])
+        created_ids.append(obs_id)
+
+        # Link observation to entity
+        await conn.execute(
+            f"""
+            INSERT INTO {fq_table("unit_entities")} (unit_id, entity_id)
+            VALUES ($1, $2)
+            """,
+            uuid.UUID(obs_id),
+            entity_uuid,
+        )
+
+    return created_ids
@@ -18,7 +18,6 @@ def utcnow():
    return datetime.now(UTC)


-from ..response_models import TokenUsage
 from . import (
    chunk_storage,
    deduplication,
@@ -27,8 +26,9 @@ from . import (
    fact_extraction,
    fact_storage,
    link_creation,
+    observation_regeneration,
 )
-from .types import EntityLink, ExtractedFact, ProcessedFact, RetainContent, RetainContentDict
+from .types import ExtractedFact, ProcessedFact, RetainContent, RetainContentDict

 logger = logging.getLogger(__name__)

@@ -38,6 +38,7 @@ async def retain_batch(
    embeddings_model,
    llm_config,
    entity_resolver,
+    task_backend,
    format_date_fn,
    duplicate_checker_fn,
    bank_id: str,
@@ -46,8 +47,7 @@ async def retain_batch(
    is_first_batch: bool = True,
    fact_type_override: str | None = None,
    confidence_score: float | None = None,
-    document_tags: list[str] | None = None,
-) -> tuple[list[list[str]], TokenUsage]:
+) -> list[list[str]]:
    """
    Process a batch of content through the retain pipeline.

@@ -56,6 +56,7 @@ async def retain_batch(
        embeddings_model: Embeddings model for generating embeddings
        llm_config: LLM configuration for fact extraction
        entity_resolver: Entity resolver for entity processing
+        task_backend: Task backend for background jobs
        format_date_fn: Function to format datetime to readable string
        duplicate_checker_fn: Function to check for duplicate facts
        bank_id: Bank identifier
@@ -64,10 +65,9 @@ async def retain_batch(
        is_first_batch: Whether this is the first batch
        fact_type_override: Override fact type for all facts
        confidence_score: Confidence score for opinions
-        document_tags: Tags applied to all items in this batch

    Returns:
-        Tuple of (unit ID lists, token usage for fact extraction)
+        List of unit ID lists (one list per content item)
    """
    start_time = time.time()
    total_chars = sum(len(item.get("content", "")) for item in contents_dicts)
@@ -86,86 +86,31 @@ async def retain_batch(
    # Convert dicts to RetainContent objects
    contents = []
    for item in contents_dicts:
-        # Merge item-level tags with document-level tags
-        item_tags = item.get("tags", []) or []
-        merged_tags = list(set(item_tags + (document_tags or [])))
        content = RetainContent(
            content=item["content"],
            context=item.get("context", ""),
            event_date=item.get("event_date") or utcnow(),
            metadata=item.get("metadata", {}),
-            entities=item.get("entities", []),
-            tags=merged_tags,
        )
        contents.append(content)

    # Step 1: Extract facts from all contents
    step_start = time.time()
+    extract_opinions = fact_type_override == "opinion"

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

    if not extracted_facts:
-        # Still need to create document if document_id was provided
-        async with acquire_with_retry(pool) as conn:
-            async with conn.transaction():
-                await fact_storage.ensure_bank_exists(conn, bank_id)
-
-                # Handle document tracking even with no facts
-                if document_id:
-                    combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
-                    retain_params = {}
-                    if contents_dicts:
-                        first_item = contents_dicts[0]
-                        if first_item.get("context"):
-                            retain_params["context"] = first_item["context"]
-                        if first_item.get("event_date"):
-                            retain_params["event_date"] = (
-                                first_item["event_date"].isoformat()
-                                if hasattr(first_item["event_date"], "isoformat")
-                                else str(first_item["event_date"])
-                            )
-                        if first_item.get("metadata"):
-                            retain_params["metadata"] = first_item["metadata"]
-                    await fact_storage.handle_document_tracking(
-                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params, document_tags
-                    )
-                else:
-                    # Check for per-item document_ids
-                    from collections import defaultdict
-
-                    contents_by_doc = defaultdict(list)
-                    for idx, content_dict in enumerate(contents_dicts):
-                        doc_id = content_dict.get("document_id")
-                        if doc_id:
-                            contents_by_doc[doc_id].append((idx, content_dict))
-
-                    for doc_id, doc_contents in contents_by_doc.items():
-                        combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
-                        retain_params = {}
-                        if doc_contents:
-                            first_item = doc_contents[0][1]
-                            if first_item.get("context"):
-                                retain_params["context"] = first_item["context"]
-                            if first_item.get("event_date"):
-                                retain_params["event_date"] = (
-                                    first_item["event_date"].isoformat()
-                                    if hasattr(first_item["event_date"], "isoformat")
-                                    else str(first_item["event_date"])
-                                )
-                            if first_item.get("metadata"):
-                                retain_params["metadata"] = first_item["metadata"]
-                        await fact_storage.handle_document_tracking(
-                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params, document_tags
-                        )
-
        total_time = time.time() - start_time
        logger.info(
-            f"RETAIN_BATCH COMPLETE: 0 facts extracted from {len(contents)} contents in {total_time:.3f}s (document tracked, no facts)"
+            f"RETAIN_BATCH COMPLETE: 0 facts extracted from {len(contents)} contents in {total_time:.3f}s (nothing to store)"
        )
-        return [[] for _ in contents], usage
+        return [[] for _ in contents]

    # Apply fact_type_override if provided
    if fact_type_override:
@@ -224,7 +169,7 @@ async def retain_batch(
                        retain_params["metadata"] = first_item["metadata"]

                await fact_storage.handle_document_tracking(
-                    conn, bank_id, document_id, combined_content, is_first_batch, retain_params, document_tags
+                    conn, bank_id, document_id, combined_content, is_first_batch, retain_params
                )
                document_ids_added.append(document_id)
                doc_id_mapping[None] = document_id  # For backwards compatibility
@@ -268,13 +213,7 @@ async def retain_batch(
                                    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,
-                                document_tags,
+                                conn, bank_id, actual_doc_id, combined_content, is_first_batch, retain_params
                            )
                            document_ids_added.append(actual_doc_id)

@@ -351,7 +290,7 @@ async def retain_batch(
            non_duplicate_facts = deduplication.filter_duplicates(processed_facts, is_duplicate_flags)

            if not non_duplicate_facts:
-                return [[] for _ in contents], usage
+                return [[] for _ in contents]

            # Insert facts (document_id is now stored per-fact)
            step_start = time.time()
@@ -360,18 +299,8 @@ async def retain_batch(

            # Process entities
            step_start = time.time()
-            # Build map of content_index -> user entities for merging
-            user_entities_per_content = {
-                idx: content.entities for idx, content in enumerate(contents) if content.entities
-            }
            entity_links = await entity_processing.process_entities_batch(
-                entity_resolver,
-                conn,
-                bank_id,
-                unit_ids,
-                non_duplicate_facts,
-                log_buffer,
-                user_entities_per_content=user_entities_per_content,
+                entity_resolver, conn, bank_id, unit_ids, non_duplicate_facts, log_buffer
            )
            log_buffer.append(f"[6] Process entities: {len(entity_links)} links in {time.time() - step_start:.3f}s")

@@ -401,9 +330,17 @@ async def retain_batch(
            causal_link_count = await link_creation.create_causal_links_batch(conn, unit_ids, non_duplicate_facts)
            log_buffer.append(f"[10] Causal links: {causal_link_count} links in {time.time() - step_start:.3f}s")

+            # Regenerate observations INSIDE transaction for atomicity
+            await observation_regeneration.regenerate_observations_batch(
+                conn, embeddings_model, llm_config, bank_id, entity_links, log_buffer
+            )
+
            # Map results back to original content items
            result_unit_ids = _map_results_to_contents(contents, extracted_facts, is_duplicate_flags, unit_ids)

+        # Trigger background tasks AFTER transaction commits (opinion reinforcement only)
+        await _trigger_background_tasks(task_backend, bank_id, unit_ids, non_duplicate_facts)
+
        # Log final summary
        total_time = time.time() - start_time
        log_buffer.append(f"{'=' * 60}")
@@ -414,7 +351,7 @@ async def retain_batch(

        logger.info("\n" + "\n".join(log_buffer) + "\n")

-        return result_unit_ids, usage
+        return result_unit_ids


 def _map_results_to_contents(
@@ -445,3 +382,24 @@ def _map_results_to_contents(
        result_unit_ids.append(content_unit_ids)

    return result_unit_ids
+
+
+async def _trigger_background_tasks(
+    task_backend,
+    bank_id: str,
+    unit_ids: list[str],
+    facts: list[ProcessedFact],
+) -> None:
+    """Trigger opinion reinforcement as background task (after transaction commits)."""
+    # Trigger opinion reinforcement if there are entities
+    fact_entities = [[e.name for e in fact.entities] for fact in facts]
+    if any(fact_entities):
+        await task_backend.submit_task(
+            {
+                "type": "reinforce_opinion",
+                "bank_id": bank_id,
+                "created_unit_ids": unit_ids,
+                "unit_texts": [fact.fact_text for fact in facts],
+                "unit_entities": fact_entities,
+            }
+        )
@@ -20,8 +20,6 @@ class RetainContentDict(TypedDict, total=False):
        event_date: When the content occurred (optional, defaults to now)
        metadata: Custom key-value metadata (optional)
        document_id: Document ID for this content item (optional)
-        entities: User-provided entities to merge with extracted entities (optional)
-        tags: Visibility scope tags for this content item (optional)
    """

    content: str  # Required
@@ -29,8 +27,6 @@ class RetainContentDict(TypedDict, total=False):
    event_date: datetime
    metadata: dict[str, str]
    document_id: str
-    entities: list[dict[str, str]]  # [{"text": "...", "type": "..."}]
-    tags: list[str]  # Visibility scope tags


 def _now_utc() -> datetime:
@@ -50,8 +46,6 @@ class RetainContent:
    context: str = ""
    event_date: datetime = field(default_factory=_now_utc)
    metadata: dict[str, str] = field(default_factory=dict)
-    entities: list[dict[str, str]] = field(default_factory=list)  # User-provided entities
-    tags: list[str] = field(default_factory=list)  # Visibility scope tags


@dataclass
@@ -86,10 +80,10 @@ class CausalRelation:
    """
    Causal relationship between facts.

-    Represents how one fact was caused by another.
+    Represents how one fact causes, enables, or prevents another.
    """

-    relation_type: str  # "caused_by"
+    relation_type: str  # "causes", "enables", "prevents", "caused_by"
    target_fact_index: int  # Index of the target fact in the batch
    strength: float = 1.0  # Strength of the causal relationship

@@ -116,7 +110,6 @@ class ExtractedFact:
    context: str = ""
    mentioned_at: datetime | None = None
    metadata: dict[str, str] = field(default_factory=dict)
-    tags: list[str] = field(default_factory=list)  # Visibility scope tags


@dataclass
@@ -159,12 +152,6 @@ class ProcessedFact:
    # DB fields (set after insertion)
    unit_id: UUID | None = None

-    # Track which content this fact came from (for user entity merging)
-    content_index: int = 0
-
-    # Visibility scope tags
-    tags: list[str] = field(default_factory=list)
-
    @property
    def is_duplicate(self) -> bool:
        """Check if this fact was marked as a duplicate."""
@@ -207,8 +194,6 @@ class ProcessedFact:
            entities=entities,
            causal_relations=extracted_fact.causal_relations,
            chunk_id=chunk_id,
-            content_index=extracted_fact.content_index,
-            tags=extracted_fact.tags,
        )


@@ -240,7 +225,6 @@ class RetainBatch:
    document_id: str | None = None
    fact_type_override: str | None = None
    confidence_score: float | None = None
-    document_tags: list[str] = field(default_factory=list)  # Tags applied to all items

    # Extracted data (populated during processing)
    extracted_facts: list[ExtractedFact] = field(default_factory=list)
@@ -11,8 +11,7 @@ from abc import ABC, abstractmethod

 from ..db_utils import acquire_with_retry
 from ..memory_engine import fq_table
-from .tags import TagsMatch, filter_results_by_tags
-from .types import MPFPTimings, RetrievalResult
+from .types import RetrievalResult

 logger = logging.getLogger(__name__)

@@ -43,10 +42,7 @@ class GraphRetriever(ABC):
        query_text: str | None = None,
        semantic_seeds: list[RetrievalResult] | None = None,
        temporal_seeds: list[RetrievalResult] | None = None,
-        adjacency=None,  # TypedAdjacency, optional pre-loaded graph
-        tags: list[str] | None = None,  # Visibility scope tags for filtering
-        tags_match: TagsMatch = "any",  # How to match tags: 'any' (OR) or 'all' (AND)
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
+    ) -> list[RetrievalResult]:
        """
        Retrieve relevant facts via graph traversal.

@@ -59,11 +55,9 @@ class GraphRetriever(ABC):
            query_text: Original query text (optional, for some strategies)
            semantic_seeds: Pre-computed semantic entry points (from semantic retrieval)
            temporal_seeds: Pre-computed temporal entry points (from temporal retrieval)
-            adjacency: Pre-loaded typed adjacency graph (optional, for MPFP)
-            tags: Optional list of tags for visibility filtering (OR matching)

        Returns:
-            Tuple of (List of RetrievalResult with activation scores, optional timing info)
+            List of RetrievalResult objects with activation scores set
        """
        pass

@@ -117,10 +111,7 @@ class BFSGraphRetriever(GraphRetriever):
        query_text: str | None = None,
        semantic_seeds: list[RetrievalResult] | None = None,
        temporal_seeds: list[RetrievalResult] | None = None,
-        adjacency=None,  # Not used by BFS
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
+    ) -> list[RetrievalResult]:
        """
        Retrieve facts using BFS spreading activation.

@@ -131,14 +122,11 @@ class BFSGraphRetriever(GraphRetriever):
        4. Return visited nodes up to budget

        Note: BFS finds its own entry points via embedding search.
-        The semantic_seeds, temporal_seeds, and adjacency parameters are accepted
+        The semantic_seeds and temporal_seeds parameters are accepted
        for interface compatibility but not used.
        """
        async with acquire_with_retry(pool) as conn:
-            results = await self._retrieve_with_conn(
-                conn, query_embedding_str, bank_id, fact_type, budget, tags=tags, tags_match=tags_match
-            )
-            return results, None
+            return await self._retrieve_with_conn(conn, query_embedding_str, bank_id, fact_type, budget)

    async def _retrieve_with_conn(
        self,
@@ -147,46 +135,33 @@ class BFSGraphRetriever(GraphRetriever):
        bank_id: str,
        fact_type: str,
        budget: int,
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
    ) -> list[RetrievalResult]:
        """Internal implementation with connection."""
-        from .tags import build_tags_where_clause_simple
-
-        tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
-        params = [query_embedding_str, bank_id, fact_type, self.entry_point_threshold, self.entry_point_limit]
-        if tags:
-            params.append(tags)

        # Step 1: Find entry points
        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,
+                   mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
                   1 - (embedding <=> $1::vector) AS similarity
            FROM {fq_table("memory_units")}
            WHERE bank_id = $2
              AND embedding IS NOT NULL
              AND fact_type = $3
              AND (1 - (embedding <=> $1::vector)) >= $4
-              {tags_clause}
            ORDER BY embedding <=> $1::vector
            LIMIT $5
            """,
-            *params,
+            query_embedding_str,
+            bank_id,
+            fact_type,
+            self.entry_point_threshold,
+            self.entry_point_limit,
        )

        if not entry_points:
-            logger.debug(
-                f"[BFS] No entry points found for fact_type={fact_type} (tags={tags}, tags_match={tags_match})"
-            )
            return []

-        logger.debug(
-            f"[BFS] Found {len(entry_points)} entry points for fact_type={fact_type} "
-            f"(tags={tags}, tags_match={tags_match})"
-        )
-
        # Step 2: BFS spreading activation
        visited = set()
        results = []
@@ -216,8 +191,8 @@ 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.embedding, mu.fact_type,
-                           mu.document_id, mu.chunk_id, mu.tags,
+                           mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type,
+                           mu.document_id, mu.chunk_id,
                           ml.weight, ml.link_type, ml.from_unit_id
                    FROM {fq_table("memory_links")} ml
                    JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
@@ -257,8 +232,4 @@ class BFSGraphRetriever(GraphRetriever):
                            neighbor_result = RetrievalResult.from_db_row(dict(n))
                            queue.append((neighbor_result, new_activation))

-        # Apply tags filtering (BFS may traverse into memories that don't match tags criteria)
-        if tags:
-            results = filter_results_by_tags(results, tags, match=tags_match)
-
        return results
@@ -1,391 +0,0 @@
-"""
-Link Expansion graph retrieval.
-
-A simple, fast graph retrieval that expands from seeds via:
-1. Entity links: Find facts sharing entities with seeds (filtered by entity frequency)
-2. Causal links: Find facts causally linked to seeds (top-k by weight)
-
-Characteristics:
- 2-3 DB queries (seed finding + parallel entity/causal expansion)
- Sublinear: only touches connected facts via indexes
- No iteration, no propagation, no normalization
- Target: <100ms
-"""
-
-import logging
-import time
-
-from ..db_utils import acquire_with_retry
-from ..memory_engine import fq_table
-from .graph_retrieval import GraphRetriever
-from .tags import TagsMatch, filter_results_by_tags
-from .types import MPFPTimings, RetrievalResult
-
-logger = logging.getLogger(__name__)
-
-
-async def _find_semantic_seeds(
-    conn,
-    query_embedding_str: str,
-    bank_id: str,
-    fact_type: str,
-    limit: int = 20,
-    threshold: float = 0.3,
-    tags: list[str] | None = None,
-    tags_match: TagsMatch = "any",
-) -> list[RetrievalResult]:
-    """Find semantic seeds via embedding search."""
-    from .tags import build_tags_where_clause_simple
-
-    tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
-    params = [query_embedding_str, bank_id, fact_type, threshold, limit]
-    if tags:
-        params.append(tags)
-
-    rows = 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)) >= $4
-          {tags_clause}
-        ORDER BY embedding <=> $1::vector
-        LIMIT $5
-        """,
-        *params,
-    )
-    return [RetrievalResult.from_db_row(dict(r)) for r in rows]
-
-
-class LinkExpansionRetriever(GraphRetriever):
-    """
-    Graph retrieval via direct link expansion from seeds.
-
-    Expands through entity co-occurrence and causal links in a single query.
-    Fast and simple alternative to MPFP.
-    """
-
-    def __init__(
-        self,
-        max_entity_frequency: int = 500,
-        causal_weight_threshold: float = 0.3,
-        causal_limit_per_seed: int = 10,
-    ):
-        """
-        Initialize link expansion retriever.
-
-        Args:
-            max_entity_frequency: Skip entities appearing in more than this many facts
-            causal_weight_threshold: Minimum weight for causal links
-            causal_limit_per_seed: Max causal links to follow per seed
-        """
-        self.max_entity_frequency = max_entity_frequency
-        self.causal_weight_threshold = causal_weight_threshold
-        self.causal_limit_per_seed = causal_limit_per_seed
-
-    @property
-    def name(self) -> str:
-        return "link_expansion"
-
-    async def retrieve(
-        self,
-        pool,
-        query_embedding_str: str,
-        bank_id: str,
-        fact_type: str,
-        budget: int,
-        query_text: str | None = None,
-        semantic_seeds: list[RetrievalResult] | None = None,
-        temporal_seeds: list[RetrievalResult] | None = None,
-        adjacency=None,
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
-        """
-        Retrieve facts by expanding links from seeds.
-
-        Args:
-            pool: Database connection pool
-            query_embedding_str: Query embedding (unused, kept for interface)
-            bank_id: Memory bank ID
-            fact_type: Fact type to filter
-            budget: Maximum results to return
-            query_text: Original query text (unused)
-            semantic_seeds: Pre-computed semantic entry points
-            temporal_seeds: Pre-computed temporal entry points
-            adjacency: Unused, kept for interface compatibility
-            tags: Optional list of tags for visibility filtering (OR matching)
-
-        Returns:
-            Tuple of (results, timings)
-        """
-        start_time = time.time()
-        timings = MPFPTimings(fact_type=fact_type)
-
-        # Use single connection for all queries to reduce pool pressure
-        # (queries are fast ~50ms each, connection acquisition is the bottleneck)
-        async with acquire_with_retry(pool) as conn:
-            # Find seeds if not provided
-            if semantic_seeds:
-                all_seeds = list(semantic_seeds)
-            else:
-                seeds_start = time.time()
-                all_seeds = await _find_semantic_seeds(
-                    conn,
-                    query_embedding_str,
-                    bank_id,
-                    fact_type,
-                    limit=20,
-                    threshold=0.3,
-                    tags=tags,
-                    tags_match=tags_match,
-                )
-                timings.seeds_time = time.time() - seeds_start
-                logger.debug(
-                    f"[LinkExpansion] Found {len(all_seeds)} semantic seeds for fact_type={fact_type} "
-                    f"(tags={tags}, tags_match={tags_match})"
-                )
-
-            # Add temporal seeds if provided
-            if temporal_seeds:
-                all_seeds.extend(temporal_seeds)
-
-            if not all_seeds:
-                return [], timings
-
-            seed_ids = list({s.id for s in all_seeds})
-            timings.pattern_count = len(seed_ids)
-
-            # Run entity and causal expansion sequentially on same connection
-            query_start = time.time()
-
-            # For observations, traverse through source_memory_ids to find entity connections.
-            # Observations don't have direct unit_entities - they inherit entities via their
-            # source world/experience facts.
-            #
-            # Path: observation → source_memory_ids → world fact → entities →
-            #       ALL world facts with those entities → their observations (excluding seeds)
-            if fact_type == "observation":
-                # Debug: Check what source_memory_ids exist on seed observations
-                debug_sources = await conn.fetch(
-                    f"""
-                    SELECT id, source_memory_ids
-                    FROM {fq_table("memory_units")}
-                    WHERE id = ANY($1::uuid[])
-                    """,
-                    seed_ids,
-                )
-                source_ids_found = []
-                for row in debug_sources:
-                    if row["source_memory_ids"]:
-                        source_ids_found.extend(row["source_memory_ids"])
-                logger.debug(
-                    f"[LinkExpansion] observation graph: {len(seed_ids)} seeds, "
-                    f"{len(source_ids_found)} source_memory_ids found"
-                )
-
-                entity_rows = await conn.fetch(
-                    f"""
-                    WITH seed_sources AS (
-                        -- Get source memory IDs from seed observations
-                        SELECT DISTINCT unnest(source_memory_ids) AS source_id
-                        FROM {fq_table("memory_units")}
-                        WHERE id = ANY($1::uuid[])
-                          AND source_memory_ids IS NOT NULL
-                    ),
-                    source_entities AS (
-                        -- Get entities from those source memories (filtered by frequency)
-                        SELECT DISTINCT ue.entity_id
-                        FROM seed_sources ss
-                        JOIN {fq_table("unit_entities")} ue ON ss.source_id = ue.unit_id
-                        JOIN {fq_table("entities")} e ON ue.entity_id = e.id
-                        WHERE e.mention_count < $2
-                    ),
-                    all_connected_sources AS (
-                        -- Find ALL world facts sharing those entities (don't exclude seed sources)
-                        -- The exclusion happens at the observation level, not the source level
-                        SELECT DISTINCT other_ue.unit_id AS source_id
-                        FROM source_entities se
-                        JOIN {fq_table("unit_entities")} other_ue ON se.entity_id = other_ue.entity_id
-                    )
-                    -- Find observations derived from connected source memories
-                    -- Only exclude the actual seed observations
-                    SELECT
-                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                        mu.occurred_end, mu.mentioned_at, mu.embedding,
-                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                        COUNT(DISTINCT cs.source_id)::float AS score
-                    FROM all_connected_sources cs
-                    JOIN {fq_table("memory_units")} mu
-                        ON mu.source_memory_ids @> ARRAY[cs.source_id]
-                    WHERE mu.fact_type = 'observation'
-                      AND mu.id != ALL($1::uuid[])
-                    GROUP BY mu.id
-                    ORDER BY score DESC
-                    LIMIT $3
-                    """,
-                    seed_ids,
-                    self.max_entity_frequency,
-                    budget,
-                )
-                logger.debug(f"[LinkExpansion] observation graph: found {len(entity_rows)} connected observations")
-            else:
-                # For world/experience facts, use direct entity lookup
-                entity_rows = await conn.fetch(
-                    f"""
-                    SELECT
-                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                        mu.occurred_end, mu.mentioned_at, mu.embedding,
-                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                        COUNT(*)::float AS score
-                    FROM {fq_table("unit_entities")} seed_ue
-                    JOIN {fq_table("entities")} e ON seed_ue.entity_id = e.id
-                    JOIN {fq_table("unit_entities")} other_ue ON seed_ue.entity_id = other_ue.entity_id
-                    JOIN {fq_table("memory_units")} mu ON other_ue.unit_id = mu.id
-                    WHERE seed_ue.unit_id = ANY($1::uuid[])
-                      AND e.mention_count < $2
-                      AND mu.id != ALL($1::uuid[])
-                      AND mu.fact_type = $3
-                    GROUP BY mu.id
-                    ORDER BY score DESC
-                    LIMIT $4
-                    """,
-                    seed_ids,
-                    self.max_entity_frequency,
-                    fact_type,
-                    budget,
-                )
-
-            causal_rows = await conn.fetch(
-                f"""
-                SELECT DISTINCT ON (mu.id)
-                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                    mu.occurred_end, mu.mentioned_at, mu.embedding,
-                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                    ml.weight + 1.0 AS score
-                FROM {fq_table("memory_links")} ml
-                JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
-                WHERE ml.from_unit_id = ANY($1::uuid[])
-                  AND ml.link_type IN ('causes', 'caused_by', 'enables', 'prevents')
-                  AND ml.weight >= $2
-                  AND mu.fact_type = $3
-                ORDER BY mu.id, ml.weight DESC
-                LIMIT $4
-                """,
-                seed_ids,
-                self.causal_weight_threshold,
-                fact_type,
-                budget,
-            )
-
-            # Fallback: semantic/temporal/entity links from memory_links table
-            # These are secondary to entity links (via unit_entities) and causal links
-            # Weight is halved (0.5x) to prioritize primary link types
-            # Check both directions: seeds -> others AND others -> seeds
-            fallback_rows = await conn.fetch(
-                f"""
-                WITH outgoing AS (
-                    -- Links FROM seeds TO other facts
-                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                           mu.occurred_end, mu.mentioned_at, mu.embedding,
-                           mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                           ml.weight
-                    FROM {fq_table("memory_links")} ml
-                    JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
-                    WHERE ml.from_unit_id = ANY($1::uuid[])
-                      AND ml.link_type IN ('semantic', 'temporal', 'entity')
-                      AND ml.weight >= $2
-                      AND mu.fact_type = $3
-                      AND mu.id != ALL($1::uuid[])
-                ),
-                incoming AS (
-                    -- Links FROM other facts TO seeds (reverse direction)
-                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                           mu.occurred_end, mu.mentioned_at, mu.embedding,
-                           mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                           ml.weight
-                    FROM {fq_table("memory_links")} ml
-                    JOIN {fq_table("memory_units")} mu ON ml.from_unit_id = mu.id
-                    WHERE ml.to_unit_id = ANY($1::uuid[])
-                      AND ml.link_type IN ('semantic', 'temporal', 'entity')
-                      AND ml.weight >= $2
-                      AND mu.fact_type = $3
-                      AND mu.id != ALL($1::uuid[])
-                ),
-                combined AS (
-                    SELECT * FROM outgoing
-                    UNION ALL
-                    SELECT * FROM incoming
-                )
-                SELECT DISTINCT ON (id)
-                    id, text, context, event_date, occurred_start,
-                    occurred_end, mentioned_at, embedding,
-                    fact_type, document_id, chunk_id, tags,
-                    (MAX(weight) * 0.5) AS score
-                FROM combined
-                GROUP BY id, text, context, event_date, occurred_start,
-                         occurred_end, mentioned_at, embedding,
-                         fact_type, document_id, chunk_id, tags
-                ORDER BY id, score DESC
-                LIMIT $4
-                """,
-                seed_ids,
-                self.causal_weight_threshold,
-                fact_type,
-                budget,
-            )
-
-            timings.edge_load_time = time.time() - query_start
-            timings.db_queries = 3
-            timings.edge_count = len(entity_rows) + len(causal_rows) + len(fallback_rows)
-
-        # Merge results, taking max score per fact
-        # Priority: entity links (unit_entities) > causal links > fallback links
-        score_map: dict[str, float] = {}
-        row_map: dict[str, dict] = {}
-
-        for row in entity_rows:
-            fact_id = str(row["id"])
-            score_map[fact_id] = max(score_map.get(fact_id, 0), row["score"])
-            row_map[fact_id] = dict(row)
-
-        for row in causal_rows:
-            fact_id = str(row["id"])
-            score_map[fact_id] = max(score_map.get(fact_id, 0), row["score"])
-            if fact_id not in row_map:
-                row_map[fact_id] = dict(row)
-
-        for row in fallback_rows:
-            fact_id = str(row["id"])
-            score_map[fact_id] = max(score_map.get(fact_id, 0), row["score"])
-            if fact_id not in row_map:
-                row_map[fact_id] = dict(row)
-
-        # Sort by score and limit
-        sorted_ids = sorted(score_map.keys(), key=lambda x: score_map[x], reverse=True)[:budget]
-        rows = [row_map[fact_id] for fact_id in sorted_ids]
-
-        # Convert to results
-        results = []
-        for row in rows:
-            result = RetrievalResult.from_db_row(dict(row))
-            result.activation = row["score"]
-            results.append(result)
-
-        # Apply tags filtering (graph expansion may reach untagged memories)
-        if tags:
-            results = filter_results_by_tags(results, tags, match=tags_match)
-
-        timings.result_count = len(results)
-        timings.traverse = time.time() - start_time
-
-        logger.debug(
-            f"LinkExpansion: {len(results)} results from {len(seed_ids)} seeds "
-            f"in {timings.traverse * 1000:.1f}ms (query: {timings.edge_load_time * 1000:.1f}ms)"
-        )
-
-        return results, timings
@@ -9,7 +9,6 @@ propagation from Approximate PPR.

 Key properties:
 - Sublinear in graph size (threshold pruning bounds active nodes)
- Lazy edge loading: only loads edges for frontier nodes, not entire graph
 - Predefined patterns capture different retrieval intents
 - All patterns run in parallel, results fused via RRF
 - No LLM in the loop during traversal
@@ -23,8 +22,7 @@ from dataclasses import dataclass, field
 from ..db_utils import acquire_with_retry
 from ..memory_engine import fq_table
 from .graph_retrieval import GraphRetriever
-from .tags import TagsMatch
-from .types import MPFPTimings, RetrievalResult
+from .types import RetrievalResult

 logger = logging.getLogger(__name__)

@@ -43,27 +41,11 @@ class EdgeTarget:


@dataclass
-class EdgeCache:
-    """
-    Cache for lazily-loaded edges.
+class TypedAdjacency:
+    """Adjacency lists split by edge type."""

-    Grows per-hop as edges are loaded for frontier nodes.
-    Shared across patterns to avoid redundant loads.
-    Loads ALL edge types at once to minimize DB queries.
-    Thread-safe via asyncio lock to prevent redundant concurrent loads.
-    """
-
-    # edge_type -> from_node_id -> list of EdgeTarget
+    # edge_type -> from_node_id -> list of (to_node_id, weight)
    graphs: dict[str, dict[str, list[EdgeTarget]]] = field(default_factory=dict)
-    # Track which nodes have been fully loaded (all edge types)
-    _fully_loaded: set[str] = field(default_factory=set)
-    # Timing stats
-    db_queries: int = 0
-    edge_load_time: float = 0.0
-    # Detailed hop timing for debugging
-    hop_details: list[dict] = field(default_factory=list)
-    # Lock to prevent redundant concurrent loads
-    _lock: asyncio.Lock = field(default_factory=asyncio.Lock)

    def get_neighbors(self, edge_type: str, node_id: str) -> list[EdgeTarget]:
        """Get neighbors for a node via a specific edge type."""
@@ -81,31 +63,6 @@ class EdgeCache:

        return [EdgeTarget(node_id=n.node_id, weight=n.weight / total) for n in neighbors]

-    def is_fully_loaded(self, node_id: str) -> bool:
-        """Check if all edges for this node have been loaded."""
-        return node_id in self._fully_loaded
-
-    def get_uncached(self, node_ids: list[str]) -> list[str]:
-        """Get node IDs that haven't been fully loaded yet."""
-        return [n for n in node_ids if not self.is_fully_loaded(n)]
-
-    def add_all_edges(self, edges_by_type: dict[str, dict[str, list[EdgeTarget]]], all_queried: list[str]):
-        """
-        Add loaded edges to the cache (all edge types at once).
-
-        Args:
-            edges_by_type: Dict mapping edge_type -> from_node_id -> list of EdgeTarget
-            all_queried: All node IDs that were queried (marks them as fully loaded)
-        """
-        for edge_type, edges in edges_by_type.items():
-            if edge_type not in self.graphs:
-                self.graphs[edge_type] = {}
-            for node_id, neighbors in edges.items():
-                self.graphs[edge_type][node_id] = neighbors
-
-        # Mark all queried nodes as fully loaded (even if they have no edges)
-        self._fully_loaded.update(all_queried)
-

@dataclass
 class PatternResult:
@@ -152,249 +109,66 @@ class SeedNode:


 # -----------------------------------------------------------------------------
-# Lazy Edge Loading
+# Core Algorithm
 # -----------------------------------------------------------------------------


-async def load_all_edges_for_frontier(
-    pool,
-    node_ids: list[str],
-    top_k_per_type: int = 20,
-) -> dict[str, dict[str, list[EdgeTarget]]]:
-    """
-    Load top-k edges per (node, edge_type) for frontier nodes.
-
-    Uses a LATERAL join to efficiently fetch only the top-k edges per type,
-    avoiding loading hundreds of entity edges when only 20 are needed.
-
-    Requires composite index: (from_unit_id, link_type, weight DESC)
-
-    Args:
-        pool: Database connection pool
-        node_ids: Frontier node IDs to load edges for
-        top_k_per_type: Max edges to load per (node, link_type) pair
-
-    Returns:
-        Dict mapping edge_type -> from_node_id -> list of EdgeTarget
-    """
-    if not node_ids:
-        return {}
-
-    async with acquire_with_retry(pool) as conn:
-        # Use LATERAL join to get top-k per (from_node, link_type)
-        # This leverages the composite index for efficient early termination
-        rows = await conn.fetch(
-            f"""
-            WITH frontier(node_id) AS (SELECT unnest($1::uuid[]))
-            SELECT f.node_id as from_unit_id, lt.link_type, edges.to_unit_id, edges.weight
-            FROM frontier f
-            CROSS JOIN (VALUES ('semantic'), ('temporal'), ('entity'), ('causes'), ('caused_by')) AS lt(link_type)
-            CROSS JOIN LATERAL (
-                SELECT ml.to_unit_id, ml.weight
-                FROM {fq_table("memory_links")} ml
-                WHERE ml.from_unit_id = f.node_id
-                  AND ml.link_type = lt.link_type
-                  AND ml.weight >= 0.1
-                ORDER BY ml.weight DESC
-                LIMIT $2
-            ) edges
-            """,
-            node_ids,
-            top_k_per_type,
-        )
-
-    # Group by edge_type -> from_node -> neighbors
-    result: dict[str, dict[str, list[EdgeTarget]]] = defaultdict(lambda: defaultdict(list))
-    for row in rows:
-        edge_type = row["link_type"]
-        from_id = str(row["from_unit_id"])
-        to_id = str(row["to_unit_id"])
-        weight = row["weight"]
-        result[edge_type][from_id].append(EdgeTarget(node_id=to_id, weight=weight))
-
-    # Convert nested defaultdicts to regular dicts
-    return {edge_type: dict(edges) for edge_type, edges in result.items()}
-
-
-# -----------------------------------------------------------------------------
-# Core Algorithm (Async with Lazy Loading)
-# -----------------------------------------------------------------------------
-
-
-@dataclass
-class PatternState:
-    """State for a pattern traversal between hops."""
-
-    pattern: list[str]
-    hop_index: int
-    scores: dict[str, float]
-    frontier: dict[str, float]
-
-
-def _init_pattern_state(seeds: list[SeedNode], pattern: list[str]) -> PatternState:
-    """Initialize pattern state from seeds."""
-    if not seeds:
-        return PatternState(pattern=pattern, hop_index=0, scores={}, frontier={})
-
-    total_seed_score = sum(s.score for s in seeds)
-    if total_seed_score == 0:
-        total_seed_score = len(seeds)
-
-    frontier = {s.node_id: s.score / total_seed_score for s in seeds}
-    return PatternState(pattern=pattern, hop_index=0, scores={}, frontier=frontier)
-
-
-def _execute_hop(state: PatternState, cache: EdgeCache, config: MPFPConfig) -> set[str]:
-    """
-    Execute ONE hop of traversal, return frontier nodes for next hop.
-
-    This is a pure function that uses cached edges (no DB access).
-    Returns set of uncached nodes needed for next hop.
-    """
-    if state.hop_index >= len(state.pattern):
-        return set()
-
-    edge_type = state.pattern[state.hop_index]
-
-    # Collect active nodes above threshold
-    active_nodes = [node_id for node_id, mass in state.frontier.items() if mass >= config.threshold]
-    if not active_nodes:
-        state.frontier = {}
-        return set()
-
-    # Propagate mass using cached edges
-    next_frontier: dict[str, float] = {}
-    uncached_for_next: set[str] = set()
-
-    for node_id, mass in state.frontier.items():
-        if mass < config.threshold:
-            continue
-
-        # Keep α portion for this node
-        state.scores[node_id] = state.scores.get(node_id, 0) + config.alpha * mass
-
-        # Push (1-α) to neighbors
-        push_mass = (1 - config.alpha) * mass
-        neighbors = cache.get_normalized_neighbors(edge_type, node_id, config.top_k_neighbors)
-
-        for neighbor in neighbors:
-            next_frontier[neighbor.node_id] = next_frontier.get(neighbor.node_id, 0) + push_mass * neighbor.weight
-            # Track if we'll need edges for this node in the next hop
-            if not cache.is_fully_loaded(neighbor.node_id):
-                uncached_for_next.add(neighbor.node_id)
-
-    state.frontier = next_frontier
-    state.hop_index += 1
-
-    return uncached_for_next
-
-
-def _finalize_pattern(state: PatternState, config: MPFPConfig) -> PatternResult:
-    """Finalize pattern by adding remaining frontier mass to scores."""
-    for node_id, mass in state.frontier.items():
-        if mass >= config.threshold:
-            state.scores[node_id] = state.scores.get(node_id, 0) + mass
-
-    return PatternResult(pattern=state.pattern, scores=state.scores)
-
-
-async def mpfp_traverse_hop_synchronized(
-    pool,
-    pattern_jobs: list[tuple[list[SeedNode], list[str]]],
-    config: MPFPConfig,
-    cache: EdgeCache,
-) -> list[PatternResult]:
-    """
-    Execute ALL patterns with hop-synchronized edge loading.
-
-    Instead of running each pattern independently (causing multiple DB queries),
-    this function:
-    1. Runs hop 1 for ALL patterns (using pre-warmed seed edges)
-    2. Collects ALL unique hop-2 frontier nodes across patterns
-    3. Pre-warms hop-2 edges in ONE query
-    4. Runs hop 2 for ALL patterns
-
-    This reduces DB queries from O(patterns * hops) to O(hops).
-
-    Args:
-        pool: Database connection pool
-        pattern_jobs: List of (seeds, pattern) tuples
-        config: Algorithm parameters
-        cache: Shared edge cache (should be pre-warmed with seed edges)
-
-    Returns:
-        List of PatternResult for each pattern
-    """
-    import time
-
-    # Initialize all pattern states
-    states = [_init_pattern_state(seeds, pattern) for seeds, pattern in pattern_jobs]
-
-    # Determine max hops (all patterns should be same length, but be safe)
-    max_hops = max((len(p) for _, p in pattern_jobs), default=0)
-
-    # Detailed timing for debugging
-    hop_times: list[dict] = []
-
-    # Execute hop-by-hop across ALL patterns
-    for hop in range(max_hops):
-        hop_start = time.time()
-        hop_timing = {"hop": hop, "patterns_executed": 0, "uncached_count": 0, "load_time": 0.0}
-
-        # Execute this hop for all patterns, collect uncached nodes for next hop
-        all_uncached: set[str] = set()
-        exec_start = time.time()
-        for state in states:
-            if state.hop_index < len(state.pattern):
-                uncached = _execute_hop(state, cache, config)
-                all_uncached.update(uncached)
-                hop_timing["patterns_executed"] += 1
-        hop_timing["exec_time"] = time.time() - exec_start
-
-        # Pre-warm edges for ALL uncached nodes before next hop
-        hop_timing["uncached_count"] = len(all_uncached)
-        if all_uncached:
-            uncached_list = list(all_uncached - cache._fully_loaded)
-            hop_timing["uncached_after_filter"] = len(uncached_list)
-            if uncached_list:
-                load_start = time.time()
-                edges_by_type = await load_all_edges_for_frontier(pool, uncached_list, config.top_k_neighbors)
-                hop_timing["load_time"] = time.time() - load_start
-                cache.edge_load_time += hop_timing["load_time"]
-                cache.db_queries += 1
-                cache.add_all_edges(edges_by_type, uncached_list)
-                hop_timing["edges_loaded"] = sum(
-                    len(neighbors) for edges in edges_by_type.values() for neighbors in edges.values()
-                )
-
-        hop_timing["total_time"] = time.time() - hop_start
-        hop_times.append(hop_timing)
-
-    # Store hop timing details in cache for logging
-    cache.hop_details = hop_times
-
-    # Finalize all patterns
-    return [_finalize_pattern(state, config) for state in states]
-
-
-async def mpfp_traverse_async(
-    pool,
+def mpfp_traverse(
    seeds: list[SeedNode],
    pattern: list[str],
+    adjacency: TypedAdjacency,
    config: MPFPConfig,
-    cache: EdgeCache,
 ) -> PatternResult:
    """
-    Async Forward Push traversal with lazy edge loading.
+    Forward Push traversal following a meta-path pattern.

-    NOTE: For better performance with multiple patterns, use mpfp_traverse_hop_synchronized().
-    This function is kept for single-pattern use cases.
+    Args:
+        seeds: Entry point nodes with initial scores
+        pattern: Sequence of edge types to follow
+        adjacency: Typed adjacency structure
+        config: Algorithm parameters
+
+    Returns:
+        PatternResult with accumulated scores per node
    """
    if not seeds:
        return PatternResult(pattern=pattern, scores={})

-    results = await mpfp_traverse_hop_synchronized(pool, [(seeds, pattern)], config, cache)
-    return results[0] if results else PatternResult(pattern=pattern, scores={})
+    scores: dict[str, float] = {}
+
+    # Initialize frontier with seed masses (normalized)
+    total_seed_score = sum(s.score for s in seeds)
+    if total_seed_score == 0:
+        total_seed_score = len(seeds)  # fallback to uniform
+
+    frontier: dict[str, float] = {s.node_id: s.score / total_seed_score for s in seeds}
+
+    # Follow pattern hop by hop
+    for edge_type in pattern:
+        next_frontier: dict[str, float] = {}
+
+        for node_id, mass in frontier.items():
+            if mass < config.threshold:
+                continue
+
+            # Keep α portion for this node
+            scores[node_id] = scores.get(node_id, 0) + config.alpha * mass
+
+            # Push (1-α) to neighbors
+            push_mass = (1 - config.alpha) * mass
+            neighbors = adjacency.get_normalized_neighbors(edge_type, node_id, config.top_k_neighbors)
+
+            for neighbor in neighbors:
+                next_frontier[neighbor.node_id] = next_frontier.get(neighbor.node_id, 0) + push_mass * neighbor.weight
+
+        frontier = next_frontier
+
+    # Final frontier nodes get their remaining mass
+    for node_id, mass in frontier.items():
+        if mass >= config.threshold:
+            scores[node_id] = scores.get(node_id, 0) + mass
+
+    return PatternResult(pattern=pattern, scores=scores)


 def rrf_fusion(
@@ -436,6 +210,38 @@ def rrf_fusion(
 # -----------------------------------------------------------------------------


+async def load_typed_adjacency(pool, bank_id: str) -> TypedAdjacency:
+    """
+    Load all edges for a bank, split by edge type.
+
+    Single query, then organize in-memory for fast traversal.
+    """
+    async with acquire_with_retry(pool) as conn:
+        rows = await conn.fetch(
+            f"""
+            SELECT ml.from_unit_id, ml.to_unit_id, ml.link_type, ml.weight
+            FROM {fq_table("memory_links")} ml
+            JOIN {fq_table("memory_units")} mu ON ml.from_unit_id = mu.id
+            WHERE mu.bank_id = $1
+              AND ml.weight >= 0.1
+            ORDER BY ml.from_unit_id, ml.weight DESC
+            """,
+            bank_id,
+        )
+
+    graphs: dict[str, dict[str, list[EdgeTarget]]] = defaultdict(lambda: defaultdict(list))
+
+    for row in rows:
+        from_id = str(row["from_unit_id"])
+        to_id = str(row["to_unit_id"])
+        link_type = row["link_type"]
+        weight = row["weight"]
+
+        graphs[link_type][from_id].append(EdgeTarget(node_id=to_id, weight=weight))
+
+    return TypedAdjacency(graphs=dict(graphs))
+
+
 async def fetch_memory_units_by_ids(
    pool,
    node_ids: list[str],
@@ -449,7 +255,7 @@ async def fetch_memory_units_by_ids(
        rows = await conn.fetch(
            f"""
            SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, embedding, fact_type, document_id, chunk_id, tags
+                   mentioned_at, access_count, embedding, fact_type, document_id, chunk_id
            FROM {fq_table("memory_units")}
            WHERE id = ANY($1::uuid[])
              AND fact_type = $2
@@ -468,10 +274,10 @@ async def fetch_memory_units_by_ids(

 class MPFPGraphRetriever(GraphRetriever):
    """
-    Graph retrieval using Meta-Path Forward Push with lazy edge loading.
+    Graph retrieval using Meta-Path Forward Push.

    Runs predefined patterns in parallel from semantic and temporal seeds,
-    loading edges on-demand per hop instead of loading entire graph upfront.
+    then fuses results via RRF.
    """

    def __init__(self, config: MPFPConfig | None = None):
@@ -481,13 +287,8 @@ class MPFPGraphRetriever(GraphRetriever):
        Args:
            config: Algorithm configuration (uses defaults if None)
        """
-        if config is None:
-            # Read top_k_neighbors from global config
-            from ...config import get_config
-
-            global_config = get_config()
-            config = MPFPConfig(top_k_neighbors=global_config.mpfp_top_k_neighbors)
-        self.config = config
+        self.config = config or MPFPConfig()
+        self._adjacency_cache: dict[str, TypedAdjacency] = {}

    @property
    def name(self) -> str:
@@ -503,12 +304,9 @@ class MPFPGraphRetriever(GraphRetriever):
        query_text: str | None = None,
        semantic_seeds: list[RetrievalResult] | None = None,
        temporal_seeds: list[RetrievalResult] | None = None,
-        adjacency=None,  # Ignored - kept for interface compatibility
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
+    ) -> list[RetrievalResult]:
        """
-        Retrieve facts using MPFP algorithm with lazy edge loading.
+        Retrieve facts using MPFP algorithm.

        Args:
            pool: Database connection pool
@@ -519,15 +317,12 @@ class MPFPGraphRetriever(GraphRetriever):
            query_text: Original query text (optional)
            semantic_seeds: Pre-computed semantic entry points
            temporal_seeds: Pre-computed temporal entry points
-            adjacency: Ignored (kept for interface compatibility)
-            tags: Optional list of tags for visibility filtering (OR matching)

        Returns:
-            Tuple of (List of RetrievalResult with activation scores, MPFPTimings)
+            List of RetrievalResult with activation scores
        """
-        import time
-
-        timings = MPFPTimings(fact_type=fact_type)
+        # Load typed adjacency (could cache per bank_id with TTL)
+        adjacency = await load_typed_adjacency(pool, bank_id)

        # Convert seeds to SeedNode format
        semantic_seed_nodes = self._convert_seeds(semantic_seeds, "similarity")
@@ -535,88 +330,54 @@ class MPFPGraphRetriever(GraphRetriever):

        # If no semantic seeds provided, fall back to finding our own
        if not semantic_seed_nodes:
-            seeds_start = time.time()
-            semantic_seed_nodes = await self._find_semantic_seeds(
-                pool, query_embedding_str, bank_id, fact_type, tags=tags, tags_match=tags_match
-            )
-            timings.seeds_time = time.time() - seeds_start
-            logger.debug(
-                f"[MPFP] Found {len(semantic_seed_nodes)} semantic seeds for fact_type={fact_type} (tags={tags}, tags_match={tags_match})"
-            )
+            semantic_seed_nodes = await self._find_semantic_seeds(pool, query_embedding_str, bank_id, fact_type)

-        # Collect all pattern jobs
-        pattern_jobs = []
+        # Run all patterns in parallel
+        tasks = []

        # Patterns from semantic seeds
        for pattern in self.config.patterns_semantic:
            if semantic_seed_nodes:
-                pattern_jobs.append((semantic_seed_nodes, pattern))
+                tasks.append(
+                    asyncio.to_thread(
+                        mpfp_traverse,
+                        semantic_seed_nodes,
+                        pattern,
+                        adjacency,
+                        self.config,
+                    )
+                )

        # Patterns from temporal seeds
        for pattern in self.config.patterns_temporal:
            if temporal_seed_nodes:
-                pattern_jobs.append((temporal_seed_nodes, pattern))
+                tasks.append(
+                    asyncio.to_thread(
+                        mpfp_traverse,
+                        temporal_seed_nodes,
+                        pattern,
+                        adjacency,
+                        self.config,
+                    )
+                )

-        if not pattern_jobs:
-            logger.debug(
-                f"[MPFP] No pattern jobs (semantic_seeds={len(semantic_seed_nodes)}, temporal_seeds={len(temporal_seed_nodes)})"
-            )
-            return [], timings
+        if not tasks:
+            return []

-        timings.pattern_count = len(pattern_jobs)
-
-        # Shared edge cache across all patterns
-        cache = EdgeCache()
-
-        # Pre-warm cache with ALL seed node edges BEFORE running patterns
-        # This prevents redundant DB queries at hop 1
-        all_seed_ids = list({s.node_id for seeds, _ in pattern_jobs for s in seeds})
-        if all_seed_ids:
-            import time as time_module
-
-            prewarm_start = time_module.time()
-            edges_by_type = await load_all_edges_for_frontier(pool, all_seed_ids, self.config.top_k_neighbors)
-            cache.edge_load_time += time_module.time() - prewarm_start
-            cache.db_queries += 1
-            cache.add_all_edges(edges_by_type, all_seed_ids)
-
-        # Run all patterns with HOP-SYNCHRONIZED edge loading
-        # This batches hop-2 edge loads across ALL patterns into ONE query
-        # Reduces DB queries from O(patterns * hops) to O(hops)
-        step_start = time.time()
-        pattern_results = await mpfp_traverse_hop_synchronized(pool, pattern_jobs, self.config, cache)
-        timings.traverse = time.time() - step_start
-
-        # Record edge loading stats from cache
-        timings.edge_count = sum(len(neighbors) for g in cache.graphs.values() for neighbors in g.values())
-        timings.db_queries = cache.db_queries
-        timings.edge_load_time = cache.edge_load_time
-        timings.hop_details = cache.hop_details
+        # Gather pattern results
+        pattern_results = await asyncio.gather(*tasks)

        # Fuse results
-        step_start = time.time()
        fused = rrf_fusion(pattern_results, top_k=budget)
-        timings.fusion = time.time() - step_start

        if not fused:
-            logger.debug(f"[MPFP] No fused results after RRF fusion (pattern_count={len(pattern_results)})")
-            return [], timings
+            return []

-        # Get top result IDs
+        # Get top result IDs (don't exclude seeds - they may be highly relevant)
        result_ids = [node_id for node_id, score in fused][:budget]

        # Fetch full details
-        step_start = time.time()
        results = await fetch_memory_units_by_ids(pool, result_ids, fact_type)
-        timings.fetch = time.time() - step_start
-
-        # Filter results by tags (graph traversal may have picked up unfiltered memories)
-        if tags:
-            from .tags import filter_results_by_tags
-
-            results = filter_results_by_tags(results, tags, match=tags_match)
-
-        timings.result_count = len(results)

        # Add activation scores from fusion
        score_map = {node_id: score for node_id, score in fused}
@@ -626,7 +387,7 @@ class MPFPGraphRetriever(GraphRetriever):
        # Sort by activation
        results.sort(key=lambda r: r.activation or 0, reverse=True)

-        return results, timings
+        return results

    def _convert_seeds(
        self,
@@ -654,17 +415,8 @@ class MPFPGraphRetriever(GraphRetriever):
        fact_type: str,
        limit: int = 20,
        threshold: float = 0.3,
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
    ) -> list[SeedNode]:
        """Fallback: find semantic seeds via embedding search."""
-        from .tags import build_tags_where_clause_simple
-
-        tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
-        params = [query_embedding_str, bank_id, fact_type, threshold, limit]
-        if tags:
-            params.append(tags)
-
        async with acquire_with_retry(pool) as conn:
            rows = await conn.fetch(
                f"""
@@ -674,11 +426,14 @@ class MPFPGraphRetriever(GraphRetriever):
                  AND embedding IS NOT NULL
                  AND fact_type = $3
                  AND (1 - (embedding <=> $1::vector)) >= $4
-                  {tags_clause}
                ORDER BY embedding <=> $1::vector
                LIMIT $5
                """,
-                *params,
+                query_embedding_str,
+                bank_id,
+                fact_type,
+                threshold,
+                limit,
            )

        return [SeedNode(node_id=str(r["id"]), score=r["similarity"]) for r in rows]
@@ -0,0 +1,125 @@
+"""
+Observation utilities for generating entity observations from facts.
+
+Observations are objective facts synthesized from multiple memory facts
+about an entity, without personality influence.
+"""
+
+import logging
+
+from pydantic import BaseModel, Field
+
+from ..response_models import MemoryFact
+
+logger = logging.getLogger(__name__)
+
+
+class Observation(BaseModel):
+    """An observation about an entity."""
+
+    observation: str = Field(description="The observation text - a factual statement about the entity")
+
+
+class ObservationExtractionResponse(BaseModel):
+    """Response containing extracted observations."""
+
+    observations: list[Observation] = Field(default_factory=list, description="List of observations about the entity")
+
+
+def format_facts_for_observation_prompt(facts: list[MemoryFact]) -> str:
+    """Format facts as text for observation extraction prompt."""
+    import json
+
+    if not facts:
+        return "[]"
+    formatted = []
+    for fact in facts:
+        fact_obj = {"text": fact.text}
+
+        # Add context if available
+        if fact.context:
+            fact_obj["context"] = fact.context
+
+        # Add occurred_start if available
+        if fact.occurred_start:
+            fact_obj["occurred_at"] = fact.occurred_start
+
+        formatted.append(fact_obj)
+
+    return json.dumps(formatted, indent=2)
+
+
+def build_observation_prompt(
+    entity_name: str,
+    facts_text: str,
+) -> str:
+    """Build the observation extraction prompt for the LLM."""
+    return f"""Based on the following facts about "{entity_name}", generate a list of key observations.
+
+FACTS ABOUT {entity_name.upper()}:
+{facts_text}
+
+Your task: Synthesize the facts into clear, objective observations about {entity_name}.
+
+GUIDELINES:
+1. Each observation should be a factual statement about {entity_name}
+2. Combine related facts into single observations where appropriate
+3. Be objective - do not add opinions, judgments, or interpretations
+4. Focus on what we KNOW about {entity_name}, not what we assume
+5. Include observations about: identity, characteristics, roles, relationships, activities
+6. Write in third person (e.g., "John is..." not "I think John is...")
+7. If there are conflicting facts, note the most recent or most supported one
+
+EXAMPLES of good observations:
+- "John works at Google as a software engineer"
+- "John is detail-oriented and methodical in his approach"
+- "John collaborates frequently with Sarah on the AI project"
+- "John joined the company in 2023"
+
+EXAMPLES of bad observations (avoid these):
+- "John seems like a good person" (opinion/judgment)
+- "John probably likes his job" (assumption)
+- "I believe John is reliable" (first-person opinion)
+
+Generate 3-7 observations based on the available facts. If there are very few facts, generate fewer observations."""
+
+
+def get_observation_system_message() -> str:
+    """Get the system message for observation extraction."""
+    return "You are an objective observer synthesizing facts about an entity. Generate clear, factual observations without opinions or personality influence. Be concise and accurate."
+
+
+async def extract_observations_from_facts(llm_config, entity_name: str, facts: list[MemoryFact]) -> list[str]:
+    """
+    Extract observations from facts about an entity using LLM.
+
+    Args:
+        llm_config: LLM configuration to use
+        entity_name: Name of the entity to generate observations about
+        facts: List of facts mentioning the entity
+
+    Returns:
+        List of observation strings
+    """
+    if not facts:
+        return []
+
+    facts_text = format_facts_for_observation_prompt(facts)
+    prompt = build_observation_prompt(entity_name, facts_text)
+
+    try:
+        result = await llm_config.call(
+            messages=[
+                {"role": "system", "content": get_observation_system_message()},
+                {"role": "user", "content": prompt},
+            ],
+            response_format=ObservationExtractionResponse,
+            scope="memory_extract_observation",
+        )
+
+        observations = [op.observation for op in result.observations]
+        return observations
+
+    except Exception as e:
+        logger.warning(f"Failed to extract observations for {entity_name}: {str(e)}")
+        return []
@@ -44,7 +44,7 @@ class CrossEncoderReranker:
            await cross_encoder.initialize()
        self._initialized = True

-    async def rerank(self, query: str, candidates: list[MergedCandidate]) -> list[ScoredResult]:
+    def rerank(self, query: str, candidates: list[MergedCandidate]) -> list[ScoredResult]:
        """
        Rerank candidates using cross-encoder scores.

@@ -85,7 +85,7 @@ class CrossEncoderReranker:
            pairs.append([query, doc_text])

        # Get cross-encoder scores
-        scores = await self.cross_encoder.predict(pairs)
+        scores = self.cross_encoder.predict(pairs)

        # Normalize scores using sigmoid to [0, 1] range
        # Cross-encoder returns logits which can be negative
@@ -0,0 +1,159 @@
+"""
+Scoring functions for memory search and retrieval.
+
+Includes recency weighting, frequency weighting, temporal proximity,
+and similarity calculations used in memory activation and ranking.
+"""
+
+from datetime import datetime
+
+
+def cosine_similarity(vec1: list[float], vec2: list[float]) -> float:
+    """
+    Calculate cosine similarity between two vectors.
+
+    Args:
+        vec1: First vector
+        vec2: Second vector
+
+    Returns:
+        Similarity score between 0 and 1
+    """
+    if len(vec1) != len(vec2):
+        raise ValueError("Vectors must have same dimension")
+
+    dot_product = sum(a * b for a, b in zip(vec1, vec2))
+    magnitude1 = sum(a * a for a in vec1) ** 0.5
+    magnitude2 = sum(b * b for b in vec2) ** 0.5
+
+    if magnitude1 == 0 or magnitude2 == 0:
+        return 0.0
+
+    return dot_product / (magnitude1 * magnitude2)
+
+
+def calculate_recency_weight(days_since: float, half_life_days: float = 365.0) -> float:
+    """
+    Calculate recency weight using logarithmic decay.
+
+    This provides much better differentiation over long time periods compared to
+    exponential decay. Uses a log-based decay where the half-life parameter controls
+    when memories reach 50% weight.
+
+    Examples:
+        - Today (0 days): 1.0
+        - 1 year (365 days): ~0.5 (with default half_life=365)
+        - 2 years (730 days): ~0.33
+        - 5 years (1825 days): ~0.17
+        - 10 years (3650 days): ~0.09
+
+    This ensures that 2-year-old and 5-year-old memories have meaningfully
+    different weights, unlike exponential decay which makes them both ~0.
+
+    Args:
+        days_since: Number of days since the memory was created
+        half_life_days: Number of days for weight to reach 0.5 (default: 1 year)
+
+    Returns:
+        Weight between 0 and 1
+    """
+    import math
+
+    # Logarithmic decay: 1 / (1 + log(1 + days_since/half_life))
+    # This decays much slower than exponential, giving better long-term differentiation
+    normalized_age = days_since / half_life_days
+    return 1.0 / (1.0 + math.log1p(normalized_age))
+
+
+def calculate_frequency_weight(access_count: int, max_boost: float = 2.0) -> float:
+    """
+    Calculate frequency weight based on access count.
+
+    Frequently accessed memories are weighted higher.
+    Uses logarithmic scaling to avoid over-weighting.
+
+    Args:
+        access_count: Number of times the memory was accessed
+        max_boost: Maximum multiplier for frequently accessed memories
+
+    Returns:
+        Weight between 1.0 and max_boost
+    """
+    import math
+
+    if access_count <= 0:
+        return 1.0
+
+    # Logarithmic scaling: log(access_count + 1) / log(10)
+    # This gives: 0 accesses = 1.0, 9 accesses ~= 1.5, 99 accesses ~= 2.0
+    normalized = math.log(access_count + 1) / math.log(10)
+    return 1.0 + min(normalized, max_boost - 1.0)
+
+
+def calculate_temporal_anchor(occurred_start: datetime, occurred_end: datetime) -> datetime:
+    """
+    Calculate a single temporal anchor point from a temporal range.
+
+    Used for spreading activation - we need a single representative date
+    to calculate temporal proximity between facts. This simplifies the
+    range-to-range distance problem.
+
+    Strategy: Use midpoint of the range for balanced representation.
+
+    Args:
+        occurred_start: Start of temporal range
+        occurred_end: End of temporal range
+
+    Returns:
+        Single datetime representing the temporal anchor (midpoint)
+
+    Examples:
+        - Point event (July 14): start=July 14, end=July 14 → anchor=July 14
+        - Month range (February): start=Feb 1, end=Feb 28 → anchor=Feb 14
+        - Year range (2023): start=Jan 1, end=Dec 31 → anchor=July 1
+    """
+    # Calculate midpoint
+    time_delta = occurred_end - occurred_start
+    midpoint = occurred_start + (time_delta / 2)
+    return midpoint
+
+
+def calculate_temporal_proximity(anchor_a: datetime, anchor_b: datetime, half_life_days: float = 30.0) -> float:
+    """
+    Calculate temporal proximity between two temporal anchors.
+
+    Used for spreading activation to determine how "close" two facts are
+    in time. Uses logarithmic decay so that temporal similarity doesn't
+    drop off too quickly.
+
+    Args:
+        anchor_a: Temporal anchor of first fact
+        anchor_b: Temporal anchor of second fact
+        half_life_days: Number of days for proximity to reach 0.5
+                       (default: 30 days = 1 month)
+
+    Returns:
+        Proximity score in [0, 1] where:
+        - 1.0 = same day
+        - 0.5 = ~half_life days apart
+        - 0.0 = very distant in time
+
+    Examples:
+        - Same day: 1.0
+        - 1 week apart (half_life=30): ~0.7
+        - 1 month apart (half_life=30): ~0.5
+        - 1 year apart (half_life=30): ~0.2
+    """
+    import math
+
+    days_apart = abs((anchor_a - anchor_b).days)
+
+    if days_apart == 0:
+        return 1.0
+
+    # Logarithmic decay: 1 / (1 + log(1 + days_apart/half_life))
+    # Similar to calculate_recency_weight but for proximity between events
+    normalized_distance = days_apart / half_life_days
+    proximity = 1.0 / (1.0 + math.log1p(normalized_distance))
+
+    return proximity
@@ -1,172 +0,0 @@
-"""
-Tags filtering utilities for retrieval.
-
-Provides SQL building functions for filtering memories by tags.
-Supports four matching modes via TagsMatch enum:
- "any": OR matching, includes untagged memories (default, backward compatible)
- "all": AND matching, includes untagged memories
- "any_strict": OR matching, excludes untagged memories
- "all_strict": AND matching, excludes untagged memories
-
-OR matching (any/any_strict): Memory matches if ANY of its tags overlap with request tags
-AND matching (all/all_strict): Memory matches if ALL request tags are present in its tags
-"""
-
-from typing import Literal
-
-TagsMatch = Literal["any", "all", "any_strict", "all_strict"]
-
-
-def _parse_tags_match(match: TagsMatch) -> tuple[str, bool]:
-    """
-    Parse TagsMatch into operator and include_untagged flag.
-
-    Returns:
-        Tuple of (operator, include_untagged)
-        - operator: "&&" for any/any_strict, "@>" for all/all_strict
-        - include_untagged: True for any/all, False for any_strict/all_strict
-    """
-    if match == "any":
-        return "&&", True
-    elif match == "all":
-        return "@>", True
-    elif match == "any_strict":
-        return "&&", False
-    elif match == "all_strict":
-        return "@>", False
-    else:
-        # Default to "any" behavior
-        return "&&", True
-
-
-def build_tags_where_clause(
-    tags: list[str] | None,
-    param_offset: int = 1,
-    table_alias: str = "",
-    match: TagsMatch = "any",
-) -> tuple[str, list, int]:
-    """
-    Build a SQL WHERE clause for filtering by tags.
-
-    Supports four matching modes:
-    - "any" (default): OR matching, includes untagged memories
-    - "all": AND matching, includes untagged memories
-    - "any_strict": OR matching, excludes untagged memories
-    - "all_strict": AND matching, excludes untagged memories
-
-    Args:
-        tags: List of tags to filter by. If None or empty, returns empty clause (no filtering).
-        param_offset: Starting parameter number for SQL placeholders (default 1).
-        table_alias: Optional table alias prefix (e.g., "mu." for "memory_units mu").
-        match: Matching mode. Defaults to "any".
-
-    Returns:
-        Tuple of (sql_clause, params, next_param_offset):
-        - sql_clause: SQL WHERE clause string
-        - params: List of parameter values to bind
-        - next_param_offset: Next available parameter number
-
-    Example:
-        >>> clause, params, next_offset = build_tags_where_clause(['user_a'], 3, 'mu.', 'any_strict')
-        >>> print(clause)  # "AND mu.tags IS NOT NULL AND mu.tags != '{}' AND mu.tags && $3"
-    """
-    if not tags:
-        return "", [], param_offset
-
-    column = f"{table_alias}tags" if table_alias else "tags"
-    operator, include_untagged = _parse_tags_match(match)
-
-    if include_untagged:
-        # Include untagged memories (NULL or empty array) OR matching tags
-        clause = f"AND ({column} IS NULL OR {column} = '{{}}' OR {column} {operator} ${param_offset})"
-    else:
-        # Strict: only memories with matching tags (exclude NULL and empty)
-        clause = f"AND {column} IS NOT NULL AND {column} != '{{}}' AND {column} {operator} ${param_offset}"
-
-    return clause, [tags], param_offset + 1
-
-
-def build_tags_where_clause_simple(
-    tags: list[str] | None,
-    param_num: int,
-    table_alias: str = "",
-    match: TagsMatch = "any",
-) -> str:
-    """
-    Build a simple SQL WHERE clause for tags filtering.
-
-    This is a convenience version that returns just the clause string,
-    assuming the caller will add the tags array to their params list.
-
-    Args:
-        tags: List of tags to filter by. If None or empty, returns empty string.
-        param_num: Parameter number to use in the clause.
-        table_alias: Optional table alias prefix.
-        match: Matching mode. Defaults to "any".
-
-    Returns:
-        SQL clause string or empty string.
-    """
-    if not tags:
-        return ""
-
-    column = f"{table_alias}tags" if table_alias else "tags"
-    operator, include_untagged = _parse_tags_match(match)
-
-    if include_untagged:
-        # Include untagged memories (NULL or empty array) OR matching tags
-        return f"AND ({column} IS NULL OR {column} = '{{}}' OR {column} {operator} ${param_num})"
-    else:
-        # Strict: only memories with matching tags (exclude NULL and empty)
-        return f"AND {column} IS NOT NULL AND {column} != '{{}}' AND {column} {operator} ${param_num}"
-
-
-def filter_results_by_tags(
-    results: list,
-    tags: list[str] | None,
-    match: TagsMatch = "any",
-) -> list:
-    """
-    Filter retrieval results by tags in Python (for post-processing).
-
-    Used when SQL filtering isn't possible (e.g., graph traversal results).
-
-    Args:
-        results: List of RetrievalResult objects with a 'tags' attribute.
-        tags: List of tags to filter by. If None or empty, returns all results.
-        match: Matching mode. Defaults to "any".
-
-    Returns:
-        Filtered list of results.
-    """
-    if not tags:
-        return results
-
-    _, include_untagged = _parse_tags_match(match)
-    is_any_match = match in ("any", "any_strict")
-
-    tags_set = set(tags)
-    filtered = []
-
-    for result in results:
-        result_tags = getattr(result, "tags", None)
-
-        # Check if untagged
-        is_untagged = result_tags is None or len(result_tags) == 0
-
-        if is_untagged:
-            if include_untagged:
-                filtered.append(result)
-            # else: skip untagged
-        else:
-            result_tags_set = set(result_tags)
-            if is_any_match:
-                # Any overlap
-                if result_tags_set & tags_set:
-                    filtered.append(result)
-            else:
-                # All tags must be present
-                if tags_set <= result_tags_set:
-                    filtered.append(result)
-
-    return filtered
@@ -3,13 +3,31 @@ Think operation utilities for formulating answers based on agent and world facts
 """

 import logging
+import re
 from datetime import datetime

+from pydantic import BaseModel, Field
+
 from ..response_models import DispositionTraits, MemoryFact

 logger = logging.getLogger(__name__)


+class Opinion(BaseModel):
+    """An opinion formed by the bank."""
+
+    opinion: str = Field(description="The opinion or perspective with reasoning included")
+    confidence: float = Field(description="Confidence score for this opinion (0.0 to 1.0, where 1.0 is very confident)")
+
+
+class OpinionExtractionResponse(BaseModel):
+    """Response containing extracted opinions."""
+
+    opinions: list[Opinion] = Field(
+        default_factory=list, description="List of opinions formed with their supporting reasons and confidence scores"
+    )
+
+
 def describe_trait_level(value: int) -> str:
    """Convert trait value (1-5) to descriptive text."""
    levels = {1: "very low", 2: "low", 3: "moderate", 4: "high", 5: "very high"}
@@ -75,46 +93,17 @@ def format_facts_for_prompt(facts: list[MemoryFact]) -> str:
    return json.dumps(formatted, indent=2)


-def format_entity_summaries_for_prompt(entities: dict) -> str:
-    """Format entity summaries for inclusion in the reflect prompt.
-
-    Args:
-        entities: Dict mapping entity name to EntityState objects
-
-    Returns:
-        Formatted string with entity summaries, or empty string if no summaries
-    """
-    if not entities:
-        return ""
-
-    summaries = []
-    for name, state in entities.items():
-        # Get summary from observations (summary is stored as single observation)
-        if state.observations:
-            summary_text = state.observations[0].text
-            summaries.append(f"## {name}\n{summary_text}")
-
-    if not summaries:
-        return ""
-
-    return "\n\n".join(summaries)
-
-
 def build_think_prompt(
    agent_facts_text: str,
    world_facts_text: str,
+    opinion_facts_text: str,
    query: str,
    name: str,
    disposition: DispositionTraits,
    background: str,
    context: str | None = None,
-    entity_summaries_text: str | None = None,
 ) -> str:
-    """Build the think prompt for the LLM.
-
-    Note: opinion_facts_text parameter removed - opinions are now stored as mental models
-    and included via entity_summaries_text.
-    """
+    """Build the think prompt for the LLM."""
    disposition_desc = build_disposition_description(disposition)

    name_section = f"""
@@ -136,14 +125,6 @@ Your background:
 ADDITIONAL CONTEXT:
 {context}

-"""
-
-    entity_section = ""
-    if entity_summaries_text:
-        entity_section = f"""
-KEY PEOPLE, PLACES & THINGS I KNOW ABOUT:
-{entity_summaries_text}
-
 """

    return f"""Here's what I know and have experienced:
@@ -154,11 +135,14 @@ MY IDENTITY & EXPERIENCES:
 WHAT I KNOW ABOUT THE WORLD:
 {world_facts_text}

-{entity_section}{context_section}{name_section}{disposition_desc}{background_section}
+MY EXISTING OPINIONS & BELIEFS:
+{opinion_facts_text}
+
+{context_section}{name_section}{disposition_desc}{background_section}

 QUESTION: {query}

-Based on everything I know, believe, and who I am (including my name, disposition and background), here's what I genuinely think about this question. I'll draw on my experiences, knowledge, and personal traits to give you my honest perspective."""
+Based on everything I know, believe, and who I am (including my name, disposition and background), here's what I genuinely think about this question. I'll draw on my experiences, knowledge, opinions, and personal traits to give you my honest perspective."""


 def get_system_message(disposition: DispositionTraits) -> str:
@@ -188,7 +172,117 @@ def get_system_message(disposition: DispositionTraits) -> str:
        " ".join(instructions) if instructions else "Balance your disposition traits when interpreting information."
    )

-    return f"You are a person with your own thoughts, experiences, opinions, and disposition. When asked a question, respond naturally like a human would - share what you genuinely think based on what you know and have experienced. {disposition_instruction} Be direct, express your views confidently, and use 'I think', 'I believe', 'in my view', etc. Respond in plain text without markdown formatting. IMPORTANT: Detect the language of the question and respond in the SAME language. Do not translate to English if the question is in another language."
+    return f"You are a person with your own thoughts, experiences, opinions, and disposition. When asked a question, respond naturally like a human would - share what you genuinely think based on what you know and have experienced. {disposition_instruction} Be direct, express your views confidently, and use 'I think', 'I believe', 'in my view', etc. Respond in plain text without markdown formatting."
+
+
+async def extract_opinions_from_text(llm_config, text: str, query: str) -> list[Opinion]:
+    """
+    Extract opinions with reasons and confidence from text using LLM.
+
+    Args:
+        llm_config: LLM configuration to use
+        text: Text to extract opinions from
+        query: The original query that prompted this response
+
+    Returns:
+        List of Opinion objects with text and confidence
+    """
+    extraction_prompt = f"""Extract any NEW opinions or perspectives from the answer below and rewrite them in FIRST-PERSON as if YOU are stating the opinion directly.
+
+ORIGINAL QUESTION:
+{query}
+
+ANSWER PROVIDED:
+{text}
+
+Your task: Find opinions in the answer and rewrite them AS IF YOU ARE THE ONE SAYING THEM.
+
+An opinion is a judgment, viewpoint, or conclusion that goes beyond just stating facts.
+
+IMPORTANT: Do NOT extract statements like:
+- "I don't have enough information"
+- "The facts don't contain information about X"
+- "I cannot answer because..."
+
+ONLY extract actual opinions about substantive topics.
+
+CRITICAL FORMAT REQUIREMENTS:
+1. **ALWAYS start with first-person phrases**: "I think...", "I believe...", "In my view...", "I've come to believe...", "Previously I thought... but now..."
+2. **NEVER use third-person**: Do NOT say "The speaker thinks..." or "They believe..." - always use "I"
+3. Include the reasoning naturally within the statement
+4. Provide a confidence score (0.0 to 1.0)
+
+CORRECT Examples (✓ FIRST-PERSON):
+- "I think Alice is more reliable because she consistently delivers on time and writes clean code"
+- "Previously I thought all engineers were equal, but now I feel that experience and track record really matter"
+- "I believe reliability is best measured by consistent output over time"
+- "I've come to believe that track records are more important than potential"
+
+WRONG Examples (✗ THIRD-PERSON - DO NOT USE):
+- "The speaker thinks Alice is more reliable"
+- "They believe reliability matters"
+- "It is believed that Alice is better"
+
+If no genuine opinions are expressed (e.g., the response just says "I don't know"), return an empty list."""
+
+    try:
+        result = await llm_config.call(
+            messages=[
+                {
+                    "role": "system",
+                    "content": "You are converting opinions from text into first-person statements. Always use 'I think', 'I believe', 'I feel', etc. NEVER use third-person like 'The speaker' or 'They'.",
+                },
+                {"role": "user", "content": extraction_prompt},
+            ],
+            response_format=OpinionExtractionResponse,
+            scope="memory_extract_opinion",
+        )
+
+        # Format opinions with confidence score and convert to first-person
+        formatted_opinions = []
+        for op in result.opinions:
+            # Convert third-person to first-person if needed
+            opinion_text = op.opinion
+
+            # Replace common third-person patterns with first-person
+            def singularize_verb(verb):
+                if verb.endswith("es"):
+                    return verb[:-1]  # believes -> believe
+                elif verb.endswith("s"):
+                    return verb[:-1]  # thinks -> think
+                return verb
+
+            # Pattern: "The speaker/user [verb]..." -> "I [verb]..."
+            match = re.match(
+                r"^(The speaker|The user|They|It is believed) (believes?|thinks?|feels?|says|asserts?|considers?)(\s+that)?(.*)$",
+                opinion_text,
+                re.IGNORECASE,
+            )
+            if match:
+                verb = singularize_verb(match.group(2))
+                that_part = match.group(3) or ""  # Keep " that" if present
+                rest = match.group(4)
+                opinion_text = f"I {verb}{that_part}{rest}"
+
+            # If still doesn't start with first-person, prepend "I believe that "
+            first_person_starters = [
+                "I think",
+                "I believe",
+                "I feel",
+                "In my view",
+                "I've come to believe",
+                "Previously I",
+            ]
+            if not any(opinion_text.startswith(starter) for starter in first_person_starters):
+                opinion_text = "I believe that " + opinion_text[0].lower() + opinion_text[1:]
+
+            formatted_opinions.append(Opinion(opinion=opinion_text, confidence=op.confidence))
+
+        return formatted_opinions
+
+    except Exception as e:
+        logger.warning(f"Failed to extract opinions: {str(e)}")
+        return []


 async def reflect(
@@ -196,6 +290,7 @@ async def reflect(
    query: str,
    experience_facts: list[str] = None,
    world_facts: list[str] = None,
+    opinion_facts: list[str] = None,
    name: str = "Assistant",
    disposition: DispositionTraits = None,
    background: str = "",
@@ -212,6 +307,7 @@ async def reflect(
        query: Question to answer
        experience_facts: List of experience/agent fact strings
        world_facts: List of world fact strings
+        opinion_facts: List of opinion fact strings
        name: Name of the agent/persona
        disposition: Disposition traits (defaults to neutral)
        background: Background information
@@ -232,15 +328,18 @@ async def reflect(

    agent_results = to_memory_facts(experience_facts or [], "experience")
    world_results = to_memory_facts(world_facts or [], "world")
+    opinion_results = to_memory_facts(opinion_facts or [], "opinion")

    # Format facts for prompt
    agent_facts_text = format_facts_for_prompt(agent_results)
    world_facts_text = format_facts_for_prompt(world_results)
+    opinion_facts_text = format_facts_for_prompt(opinion_results)

    # Build prompt
    prompt = build_think_prompt(
        agent_facts_text=agent_facts_text,
        world_facts_text=world_facts_text,
+        opinion_facts_text=opinion_facts_text,
        query=query,
        name=name,
        disposition=disposition,
@@ -11,13 +11,6 @@ from typing import Any, Literal
 from pydantic import BaseModel, Field


-class TemporalConstraint(BaseModel):
-    """Detected temporal constraint from query analysis."""
-
-    start: datetime | None = Field(default=None, description="Start of temporal range")
-    end: datetime | None = Field(default=None, description="End of temporal range")
-
-
 class QueryInfo(BaseModel):
    """Information about the search query."""

@@ -26,11 +19,6 @@ class QueryInfo(BaseModel):
    timestamp: datetime = Field(description="When the query was executed")
    budget: int = Field(description="Maximum nodes to explore")
    max_tokens: int = Field(description="Maximum tokens to return in results")
-    tags: list[str] | None = Field(default=None, description="Tags filter applied to recall")
-    tags_match: str | None = Field(default=None, description="Tags matching mode: any, all, any_strict, all_strict")
-    temporal_constraint: TemporalConstraint | None = Field(
-        default=None, description="Detected temporal range from query"
-    )


 class EntryPoint(BaseModel):
@@ -85,6 +73,7 @@ class NodeVisit(BaseModel):
    text: str = Field(description="Memory unit text content")
    context: str = Field(description="Memory unit context")
    event_date: datetime | None = Field(default=None, description="When the memory occurred")
+    access_count: int = Field(description="Number of times accessed before this search")

    # How this node was reached
    is_entry_point: bool = Field(description="Whether this is an entry point")
@@ -22,7 +22,6 @@ from .trace import (
    SearchPhaseMetrics,
    SearchSummary,
    SearchTrace,
-    TemporalConstraint,
    WeightComponents,
 )

@@ -46,14 +45,7 @@ class SearchTracer:
        json_output = trace.to_json()
    """

-    def __init__(
-        self,
-        query: str,
-        budget: int,
-        max_tokens: int,
-        tags: list[str] | None = None,
-        tags_match: str | None = None,
-    ):
+    def __init__(self, query: str, budget: int, max_tokens: int):
        """
        Initialize tracer.

@@ -61,14 +53,10 @@ class SearchTracer:
            query: Search query text
            budget: Maximum nodes to explore
            max_tokens: Maximum tokens to return in results
-            tags: Tags filter applied to recall
-            tags_match: Tags matching mode (any, all, any_strict, all_strict)
        """
        self.query_text = query
        self.budget = budget
        self.max_tokens = max_tokens
-        self.tags = tags
-        self.tags_match = tags_match

        # Trace data
        self.query_embedding: list[float] | None = None
@@ -78,9 +66,6 @@ class SearchTracer:
        self.pruned: list[PruningDecision] = []
        self.phase_metrics: list[SearchPhaseMetrics] = []

-        # Temporal constraint detected from query
-        self.temporal_constraint: TemporalConstraint | None = None
-
        # New 4-way retrieval tracking
        self.retrieval_results: list[RetrievalMethodResults] = []
        self.rrf_merged: list[RRFMergeResult] = []
@@ -103,11 +88,6 @@ class SearchTracer:
        """Record the query embedding."""
        self.query_embedding = embedding

-    def record_temporal_constraint(self, start: datetime | None, end: datetime | None):
-        """Record the detected temporal constraint from query analysis."""
-        if start is not None or end is not None:
-            self.temporal_constraint = TemporalConstraint(start=start, end=end)
-
    def add_entry_point(self, node_id: str, text: str, similarity: float, rank: int):
        """
        Record an entry point.
@@ -136,6 +116,7 @@ class SearchTracer:
        text: str,
        context: str,
        event_date: datetime | None,
+        access_count: int,
        is_entry_point: bool,
        parent_node_id: str | None,
        link_type: Literal["temporal", "semantic", "entity"] | None,
@@ -154,6 +135,7 @@ class SearchTracer:
            text: Memory unit text
            context: Memory unit context
            event_date: When the memory occurred
+            access_count: Access count before this search
            is_entry_point: Whether this is an entry point
            parent_node_id: Node that led here (None for entry points)
            link_type: Type of link from parent
@@ -192,6 +174,7 @@ class SearchTracer:
            text=text,
            context=context,
            event_date=event_date,
+            access_count=access_count,
            is_entry_point=is_entry_point,
            parent_node_id=parent_node_id,
            link_type=link_type,
@@ -330,8 +313,8 @@ class SearchTracer:
                RetrievalResult(
                    rank=rank,
                    node_id=doc_id,
-                    text=data.get("text") or "",
-                    context=data.get("context") or "",
+                    text=data.get("text", ""),
+                    context=data.get("context", ""),
                    event_date=data.get("event_date"),
                    fact_type=data.get("fact_type") or fact_type,
                    score=score,
@@ -445,9 +428,6 @@ class SearchTracer:
            timestamp=datetime.now(UTC),
            budget=self.budget,
            max_tokens=self.max_tokens,
-            tags=self.tags,
-            tags_match=self.tags_match,
-            temporal_constraint=self.temporal_constraint,
        )

        # Create summary
@@ -10,24 +10,6 @@ from datetime import datetime
 from typing import Any


-@dataclass
-class MPFPTimings:
-    """Timing breakdown for a single MPFP retrieval call."""
-
-    fact_type: str
-    edge_count: int = 0  # Total edges loaded
-    db_queries: int = 0  # Number of DB queries for edge loading
-    edge_load_time: float = 0.0  # Time spent loading edges from DB
-    traverse: float = 0.0  # Total traversal time (includes edge loading)
-    pattern_count: int = 0  # Number of patterns executed
-    fusion: float = 0.0  # Time for RRF fusion
-    fetch: float = 0.0  # Time to fetch memory unit details
-    seeds_time: float = 0.0  # Time to find semantic seeds (if fallback used)
-    result_count: int = 0  # Number of results returned
-    # Detailed per-hop timing: list of {hop, exec_time, uncached, load_time, edges_loaded, total_time}
-    hop_details: list[dict] = field(default_factory=list)
-
-
@dataclass
 class RetrievalResult:
    """
@@ -46,8 +28,8 @@ class RetrievalResult:
    mentioned_at: datetime | None = None
    document_id: str | None = None
    chunk_id: str | None = None
+    access_count: int = 0
    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)
    similarity: float | None = None  # Semantic retrieval
@@ -70,8 +52,8 @@ class RetrievalResult:
            mentioned_at=row.get("mentioned_at"),
            document_id=row.get("document_id"),
            chunk_id=row.get("chunk_id"),
+            access_count=row.get("access_count", 0),
            embedding=row.get("embedding"),
-            tags=row.get("tags"),
            similarity=row.get("similarity"),
            bm25_score=row.get("bm25_score"),
            activation=row.get("activation"),
@@ -154,8 +136,8 @@ class ScoredResult:
            "mentioned_at": self.retrieval.mentioned_at,
            "document_id": self.retrieval.document_id,
            "chunk_id": self.retrieval.chunk_id,
+            "access_count": self.retrieval.access_count,
            "embedding": self.retrieval.embedding,
-            "tags": self.retrieval.tags,
            "semantic_similarity": self.retrieval.similarity,
            "bm25_score": self.retrieval.bm25_score,
        }
@@ -1,40 +1,31 @@
 """
-Task backend for distributed task processing.
+Abstract task backend for running async tasks.

-This provides an abstraction for task storage and execution:
- BrokerTaskBackend: Uses PostgreSQL as broker (production)
- SyncTaskBackend: Executes tasks immediately (testing/embedded)
+This provides an abstraction that can be adapted to different execution models:
+- AsyncIO queue (default implementation)
+- Pub/Sub architectures (future)
+- Message brokers (future)
 """

-import json
+import asyncio
 import logging
 from abc import ABC, abstractmethod
 from collections.abc import Awaitable, Callable
-from typing import TYPE_CHECKING, Any
-
-if TYPE_CHECKING:
-    import asyncpg
+from typing import Any

 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 TaskBackend(ABC):
    """
    Abstract base class for task execution backends.

    Implementations must:
    1. Store/publish task events (as serializable dicts)
-    2. Execute tasks through a provided executor callback (optional)
+    2. Execute tasks through a provided executor callback

    The backend treats tasks as pure dictionaries that can be serialized
-    and stored in the database. The executor (typically MemoryEngine.execute_task)
+    and sent over the network. The executor (typically MemoryEngine.execute_task)
    receives the dict and routes it to the appropriate handler.
    """

@@ -55,7 +46,7 @@ class TaskBackend(ABC):
    @abstractmethod
    async def initialize(self):
        """
-        Initialize the backend (e.g., connect to database).
+        Initialize the backend (e.g., start workers, connect to broker).
        """
        pass

@@ -72,7 +63,7 @@ class TaskBackend(ABC):
    @abstractmethod
    async def shutdown(self):
        """
-        Shutdown the backend gracefully.
+        Shutdown the backend gracefully (e.g., stop workers, close connections).
        """
        pass

@@ -102,8 +93,9 @@ class SyncTaskBackend(TaskBackend):
    """
    Synchronous task backend that executes tasks immediately.

-    This is useful for tests and embedded/CLI usage where we don't want
-    background workers. Tasks are executed inline rather than being queued.
+    This is useful for embedded/CLI usage where we don't want background
+    workers that prevent clean exit. Tasks are executed inline rather than
+    being queued.
    """

    async def initialize(self):
@@ -129,129 +121,130 @@ class SyncTaskBackend(TaskBackend):
        logger.debug("SyncTaskBackend shutdown")


-class BrokerTaskBackend(TaskBackend):
+class AsyncIOQueueBackend(TaskBackend):
    """
-    Task backend using PostgreSQL as broker.
+    Task backend implementation using asyncio queues.

-    submit_task() stores task_payload in async_operations table.
-    Actual polling and execution is handled separately by WorkerPoller.
-
-    This backend is used by the API to store tasks. Workers poll
-    the database separately to claim and execute tasks.
+    This is the default implementation that uses in-process asyncio queues
+    and a periodic consumer worker.
    """

-    def __init__(
-        self,
-        pool_getter: Callable[[], "asyncpg.Pool"],
-        schema: str | None = None,
-        schema_getter: Callable[[], str | None] | None = None,
-    ):
+    def __init__(self, batch_size: int = 100, batch_interval: float = 1.0):
        """
-        Initialize the broker task backend.
+        Initialize AsyncIO queue backend.

        Args:
-            pool_getter: Callable that returns the asyncpg connection pool
-            schema: Database schema for multi-tenant support (optional, static)
-            schema_getter: Callable that returns current schema dynamically (optional).
-                          If set, takes precedence over static schema for submit_task.
+            batch_size: Maximum number of tasks to process in one batch
+            batch_interval: Maximum time (seconds) to wait before processing batch
        """
        super().__init__()
-        self._pool_getter = pool_getter
-        self._schema = schema
-        self._schema_getter = schema_getter
+        self._queue: asyncio.Queue | None = None
+        self._worker_task: asyncio.Task | None = None
+        self._shutdown_event: asyncio.Event | None = None
+        self._batch_size = batch_size
+        self._batch_interval = batch_interval

    async def initialize(self):
-        """Initialize the backend."""
+        """Initialize the queue and start the worker."""
+        if self._initialized:
+            return
+
+        self._queue = asyncio.Queue()
+        self._shutdown_event = asyncio.Event()
+        self._worker_task = asyncio.create_task(self._worker())
        self._initialized = True
-        logger.info("BrokerTaskBackend initialized")
+        logger.info("AsyncIOQueueBackend initialized")

    async def submit_task(self, task_dict: dict[str, Any]):
        """
-        Store task payload in async_operations table.
-
-        The task_dict should contain an 'operation_id' if updating an existing
-        operation record, otherwise a new operation will be created.
+        Submit a task by putting it in the queue.

        Args:
-            task_dict: Task dictionary to store (must be JSON serializable)
+            task_dict: Task dictionary to execute
        """
        if not self._initialized:
            await self.initialize()

-        pool = self._pool_getter()
-        operation_id = task_dict.get("operation_id")
+        await self._queue.put(task_dict)
        task_type = task_dict.get("type", "unknown")
-        bank_id = task_dict.get("bank_id")
-        payload_json = json.dumps(task_dict)
+        task_id = task_dict.get("id")

-        schema = self._schema_getter() if self._schema_getter else self._schema
-        table = fq_table("async_operations", schema)
-
-        if operation_id:
-            # Update existing operation with task payload
-            await pool.execute(
-                f"""
-                UPDATE {table}
-                SET task_payload = $1::jsonb, updated_at = now()
-                WHERE operation_id = $2
-                """,
-                payload_json,
-                operation_id,
-            )
-            logger.debug(f"Updated task payload for operation {operation_id}")
-        else:
-            # Insert new operation (for tasks without pre-created records)
-            # e.g., access_count_update tasks
-            import uuid
-
-            new_id = uuid.uuid4()
-            await pool.execute(
-                f"""
-                INSERT INTO {table} (operation_id, bank_id, operation_type, status, task_payload)
-                VALUES ($1, $2, $3, 'pending', $4::jsonb)
-                """,
-                new_id,
-                bank_id,
-                task_type,
-                payload_json,
-            )
-            logger.debug(f"Created new operation {new_id} for task type {task_type}")
-
-    async def shutdown(self):
-        """Shutdown the backend."""
-        self._initialized = False
-        logger.info("BrokerTaskBackend shutdown")
-
-    async def wait_for_pending_tasks(self, timeout: float = 120.0):
+    async def wait_for_pending_tasks(self, timeout: float = 5.0):
        """
-        Wait for pending tasks to be processed.
+        Wait for all pending tasks in the queue to be processed.

-        In the broker model, this polls the database to check if tasks
-        for this process have been completed. This is useful in tests
-        when worker_enabled=True (API processes its own tasks).
+        This is useful in tests to ensure background tasks complete before assertions.

        Args:
            timeout: Maximum time to wait in seconds
        """
-        import asyncio
-
-        pool = self._pool_getter()
-        schema = self._schema_getter() if self._schema_getter else self._schema
-        table = fq_table("async_operations", schema)
+        if not self._initialized or self._queue is None:
+            return

+        # Wait for queue to be empty and give worker time to process
        start_time = asyncio.get_event_loop().time()
        while asyncio.get_event_loop().time() - start_time < timeout:
-            # Check if there are any pending tasks with payloads
-            count = await pool.fetchval(
-                f"""
-                SELECT COUNT(*) FROM {table}
-                WHERE status = 'pending' AND task_payload IS NOT NULL
-                """
-            )
+            if self._queue.empty():
+                # Queue is empty, give worker a bit more time to finish any in-flight task
+                await asyncio.sleep(0.3)
+                # Check again - if still empty, we're done
+                if self._queue.empty():
+                    return
+            else:
+                # Queue not empty, wait a bit
+                await asyncio.sleep(0.1)

-            if count == 0:
-                return
+    async def shutdown(self):
+        """Shutdown the worker and drain the queue."""
+        if not self._initialized:
+            return

-            await asyncio.sleep(0.5)
+        logger.info("Shutting down AsyncIOQueueBackend...")

-        logger.warning(f"Timeout waiting for pending tasks after {timeout}s")
+        # Signal shutdown
+        self._shutdown_event.set()
+
+        # Cancel worker
+        if self._worker_task is not None:
+            self._worker_task.cancel()
+            try:
+                await self._worker_task
+            except asyncio.CancelledError:
+                pass  # Worker cancelled successfully
+
+        self._initialized = False
+        logger.info("AsyncIOQueueBackend shutdown complete")
+
+    async def _worker(self):
+        """
+        Background worker that processes tasks in batches.
+
+        Collects tasks for up to batch_interval seconds or batch_size items,
+        then processes them.
+        """
+        while not self._shutdown_event.is_set():
+            try:
+                # Collect tasks for batching
+                tasks = []
+                deadline = asyncio.get_event_loop().time() + self._batch_interval
+
+                while len(tasks) < self._batch_size and asyncio.get_event_loop().time() < deadline:
+                    try:
+                        remaining_time = max(0.1, deadline - asyncio.get_event_loop().time())
+                        task_dict = await asyncio.wait_for(self._queue.get(), timeout=remaining_time)
+                        tasks.append(task_dict)
+                    except TimeoutError:
+                        break
+
+                # Process batch
+                if tasks:
+                    # Execute tasks concurrently
+                    await asyncio.gather(
+                        *[self._execute_task(task_dict) for task_dict in tasks], return_exceptions=True
+                    )
+
+            except asyncio.CancelledError:
+                break
+            except Exception as e:
+                logger.error(f"Worker error: {e}")
+                await asyncio.sleep(1)  # Backoff on error
@@ -19,6 +19,7 @@ async def extract_facts(
    context: str = "",
    llm_config: "LLMConfig" = None,
    agent_name: str = None,
+    extract_opinions: bool = False,
 ) -> tuple[list["Fact"], list[tuple[str, int]]]:
    """
    Extract semantic facts from text using LLM.
@@ -35,6 +36,7 @@ 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
+        extract_opinions: If True, extract ONLY opinions. If False, extract world and agent facts (no opinions)

    Returns:
        Tuple of (facts, chunks) where:
@@ -47,12 +49,13 @@ async def extract_facts(
    if not text or not text.strip():
        return [], []

-    facts, chunks, _ = await extract_facts_from_text(
+    facts, chunks = await extract_facts_from_text(
        text,
        event_date,
        context=context,
        llm_config=llm_config,
        agent_name=agent_name,
+        extract_opinions=extract_opinions,
    )

    if not facts:
@@ -62,3 +65,154 @@ async def extract_facts(
        return [], chunks

    return facts, chunks
+
+
+def cosine_similarity(vec1: list[float], vec2: list[float]) -> float:
+    """
+    Calculate cosine similarity between two vectors.
+
+    Args:
+        vec1: First vector
+        vec2: Second vector
+
+    Returns:
+        Similarity score between 0 and 1
+    """
+    if len(vec1) != len(vec2):
+        raise ValueError("Vectors must have same dimension")
+
+    dot_product = sum(a * b for a, b in zip(vec1, vec2))
+    magnitude1 = sum(a * a for a in vec1) ** 0.5
+    magnitude2 = sum(b * b for b in vec2) ** 0.5
+
+    if magnitude1 == 0 or magnitude2 == 0:
+        return 0.0
+
+    return dot_product / (magnitude1 * magnitude2)
+
+
+def calculate_recency_weight(days_since: float, half_life_days: float = 365.0) -> float:
+    """
+    Calculate recency weight using logarithmic decay.
+
+    This provides much better differentiation over long time periods compared to
+    exponential decay. Uses a log-based decay where the half-life parameter controls
+    when memories reach 50% weight.
+
+    Examples:
+        - Today (0 days): 1.0
+        - 1 year (365 days): ~0.5 (with default half_life=365)
+        - 2 years (730 days): ~0.33
+        - 5 years (1825 days): ~0.17
+        - 10 years (3650 days): ~0.09
+
+    This ensures that 2-year-old and 5-year-old memories have meaningfully
+    different weights, unlike exponential decay which makes them both ~0.
+
+    Args:
+        days_since: Number of days since the memory was created
+        half_life_days: Number of days for weight to reach 0.5 (default: 1 year)
+
+    Returns:
+        Weight between 0 and 1
+    """
+    import math
+
+    # Logarithmic decay: 1 / (1 + log(1 + days_since/half_life))
+    # This decays much slower than exponential, giving better long-term differentiation
+    normalized_age = days_since / half_life_days
+    return 1.0 / (1.0 + math.log1p(normalized_age))
+
+
+def calculate_frequency_weight(access_count: int, max_boost: float = 2.0) -> float:
+    """
+    Calculate frequency weight based on access count.
+
+    Frequently accessed memories are weighted higher.
+    Uses logarithmic scaling to avoid over-weighting.
+
+    Args:
+        access_count: Number of times the memory was accessed
+        max_boost: Maximum multiplier for frequently accessed memories
+
+    Returns:
+        Weight between 1.0 and max_boost
+    """
+    import math
+
+    if access_count <= 0:
+        return 1.0
+
+    # Logarithmic scaling: log(access_count + 1) / log(10)
+    # This gives: 0 accesses = 1.0, 9 accesses ~= 1.5, 99 accesses ~= 2.0
+    normalized = math.log(access_count + 1) / math.log(10)
+    return 1.0 + min(normalized, max_boost - 1.0)
+
+
+def calculate_temporal_anchor(occurred_start: datetime, occurred_end: datetime) -> datetime:
+    """
+    Calculate a single temporal anchor point from a temporal range.
+
+    Used for spreading activation - we need a single representative date
+    to calculate temporal proximity between facts. This simplifies the
+    range-to-range distance problem.
+
+    Strategy: Use midpoint of the range for balanced representation.
+
+    Args:
+        occurred_start: Start of temporal range
+        occurred_end: End of temporal range
+
+    Returns:
+        Single datetime representing the temporal anchor (midpoint)
+
+    Examples:
+        - Point event (July 14): start=July 14, end=July 14 → anchor=July 14
+        - Month range (February): start=Feb 1, end=Feb 28 → anchor=Feb 14
+        - Year range (2023): start=Jan 1, end=Dec 31 → anchor=July 1
+    """
+    # Calculate midpoint
+    time_delta = occurred_end - occurred_start
+    midpoint = occurred_start + (time_delta / 2)
+    return midpoint
+
+
+def calculate_temporal_proximity(anchor_a: datetime, anchor_b: datetime, half_life_days: float = 30.0) -> float:
+    """
+    Calculate temporal proximity between two temporal anchors.
+
+    Used for spreading activation to determine how "close" two facts are
+    in time. Uses logarithmic decay so that temporal similarity doesn't
+    drop off too quickly.
+
+    Args:
+        anchor_a: Temporal anchor of first fact
+        anchor_b: Temporal anchor of second fact
+        half_life_days: Number of days for proximity to reach 0.5
+                       (default: 30 days = 1 month)
+
+    Returns:
+        Proximity score in [0, 1] where:
+        - 1.0 = same day
+        - 0.5 = ~half_life days apart
+        - 0.0 = very distant in time
+
+    Examples:
+        - Same day: 1.0
+        - 1 week apart (half_life=30): ~0.7
+        - 1 month apart (half_life=30): ~0.5
+        - 1 year apart (half_life=30): ~0.2
+    """
+    import math
+
+    days_apart = abs((anchor_a - anchor_b).days)
+
+    if days_apart == 0:
+        return 1.0
+
+    # Logarithmic decay: 1 / (1 + log(1 + days_apart/half_life))
+    # Similar to calculate_recency_weight but for proximity between events
+    normalized_distance = days_apart / half_life_days
+    proximity = 1.0 / (1.0 + math.log1p(normalized_distance))
+
+    return proximity
@@ -21,10 +21,6 @@ from hindsight_api.extensions.context import DefaultExtensionContext, ExtensionC
 from hindsight_api.extensions.http import HttpExtension
 from hindsight_api.extensions.loader import load_extension
 from hindsight_api.extensions.operation_validator import (
-    # Consolidation operation
-    ConsolidateContext,
-    ConsolidateResult,
-    # Core operations
    OperationValidationError,
    OperationValidatorExtension,
    RecallContext,
@@ -37,7 +33,6 @@ from hindsight_api.extensions.operation_validator import (
 )
 from hindsight_api.extensions.tenant import (
    AuthenticationError,
-    Tenant,
    TenantContext,
    TenantExtension,
 )
@@ -52,7 +47,7 @@ __all__ = [
    "DefaultExtensionContext",
    # HTTP Extension
    "HttpExtension",
-    # Operation Validator - Core
+    # Operation Validator
    "OperationValidationError",
    "OperationValidatorExtension",
    "RecallContext",
@@ -62,14 +57,10 @@ __all__ = [
    "RetainContext",
    "RetainResult",
    "ValidationResult",
-    # Operation Validator - Consolidation
-    "ConsolidateContext",
-    "ConsolidateResult",
    # Tenant/Auth
    "ApiKeyTenantExtension",
    "AuthenticationError",
    "RequestContext",
-    "Tenant",
    "TenantContext",
    "TenantExtension",
 ]
@@ -1,7 +1,6 @@
 """Built-in tenant extension implementations."""

-from hindsight_api.config import get_config
-from hindsight_api.extensions.tenant import AuthenticationError, Tenant, TenantContext, TenantExtension
+from hindsight_api.extensions.tenant import AuthenticationError, TenantContext, TenantExtension
 from hindsight_api.models import RequestContext


@@ -11,13 +10,11 @@ class ApiKeyTenantExtension(TenantExtension):

    This is a simple implementation that:
    1. Validates the API key matches HINDSIGHT_API_TENANT_API_KEY
-    2. Returns the configured schema (HINDSIGHT_API_DATABASE_SCHEMA, default 'public')
-       for all authenticated requests
+    2. Returns 'public' as the schema for all authenticated requests

    Configuration:
        HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
        HINDSIGHT_API_TENANT_API_KEY=your-secret-key
-        HINDSIGHT_API_DATABASE_SCHEMA=your-schema (optional, defaults to 'public')

    For multi-tenant setups with separate schemas per tenant, implement a custom
    TenantExtension that looks up the schema based on the API key or token claims.
@@ -30,11 +27,7 @@ class ApiKeyTenantExtension(TenantExtension):
            raise ValueError("HINDSIGHT_API_TENANT_API_KEY is required when using ApiKeyTenantExtension")

    async def authenticate(self, context: RequestContext) -> TenantContext:
-        """Validate API key and return configured schema context."""
+        """Validate API key and return public schema context."""
        if context.api_key != self.expected_api_key:
            raise AuthenticationError("Invalid API key")
-        return TenantContext(schema_name=get_config().database_schema)
-
-    async def list_tenants(self) -> list[Tenant]:
-        """Return configured schema for single-tenant setup."""
-        return [Tenant(schema=get_config().database_schema)]
+        return TenantContext(schema_name="public")
@@ -96,25 +96,9 @@ class DefaultExtensionContext(ExtensionContext):

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

-        # Prefer getting URL from memory engine (handles pg0 case where URL is set after init)
-        db_url = self._database_url
-        if self._memory_engine is not None:
-            engine_url = getattr(self._memory_engine, "db_url", None)
-            if engine_url:
-                db_url = engine_url
-
-        run_migrations(db_url, schema=schema)
-
-        # 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:
-            embeddings = getattr(self._memory_engine, "embeddings", None)
-            if embeddings is not None:
-                dimension = getattr(embeddings, "dimension", None)
-                if dimension is not None:
-                    ensure_embedding_dimension(db_url, dimension, schema=schema)
+        run_migrations(self._database_url, schema=schema)

    def get_memory_engine(self) -> "MemoryEngineInterface":
        """Get the memory engine interface."""
@@ -1,4 +1,4 @@
-"""Operation Validator Extension for validating retain/recall/reflect/consolidate operations."""
+"""Operation Validator Extension for validating retain/recall/reflect operations."""

 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
@@ -17,9 +17,8 @@ if TYPE_CHECKING:
 class OperationValidationError(Exception):
    """Raised when an operation fails validation."""

-    def __init__(self, reason: str, status_code: int = 403):
+    def __init__(self, reason: str):
        self.reason = reason
-        self.status_code = status_code
        super().__init__(f"Operation validation failed: {reason}")


@@ -29,7 +28,6 @@ class ValidationResult:

    allowed: bool
    reason: str | None = None
-    status_code: int = 403  # Default to Forbidden

    @classmethod
    def accept(cls) -> "ValidationResult":
@@ -37,9 +35,9 @@ class ValidationResult:
        return cls(allowed=True)

    @classmethod
-    def reject(cls, reason: str, status_code: int = 403) -> "ValidationResult":
-        """Create a rejected validation result with a reason and HTTP status code."""
-        return cls(allowed=False, reason=reason, status_code=status_code)
+    def reject(cls, reason: str) -> "ValidationResult":
+        """Create a rejected validation result with a reason."""
+        return cls(allowed=False, reason=reason)


 # =============================================================================
@@ -97,19 +95,6 @@ class ReflectContext:
    context: str | None = None


-# =============================================================================
-# Consolidation Pre-operation Context
-# =============================================================================
-
-
-@dataclass
-class ConsolidateContext:
-    """Context for a consolidation operation validation (pre-operation)."""
-
-    bank_id: str
-    request_context: "RequestContext"
-
-
 # =============================================================================
 # Post-operation Contexts (includes results)
 # =============================================================================
@@ -177,28 +162,9 @@ class ReflectResultContext:
    error: str | None = None


-# =============================================================================
-# Consolidation Post-operation Context
-# =============================================================================
-
-
-@dataclass
-class ConsolidateResult:
-    """Result context for post-consolidation hook."""
-
-    bank_id: str
-    request_context: "RequestContext"
-    # Result
-    processed: int = 0
-    created: int = 0
-    updated: int = 0
-    success: bool = True
-    error: str | None = None
-
-
 class OperationValidatorExtension(Extension, ABC):
    """
-    Validates and hooks into retain/recall/reflect/consolidate operations.
+    Validates and hooks into retain/recall/reflect operations.

    This extension allows implementing custom logic such as:
    - Rate limiting (pre-operation)
@@ -217,13 +183,9 @@ class OperationValidatorExtension(Extension, ABC):
        -> config = {"max_requests": "100"}

    Hook execution order:
-        1. validate_* (pre-operation)
+        1. validate_retain/validate_recall/validate_reflect (pre-operation)
        2. [operation executes]
-        3. on_*_complete (post-operation)
-
-    Supported operations:
-        - retain, recall, reflect (core memory operations)
-        - consolidate (mental models consolidation)
+        3. on_retain_complete/on_recall_complete/on_reflect_complete (post-operation)
    """

    # =========================================================================
@@ -361,44 +323,3 @@ class OperationValidatorExtension(Extension, ABC):
                - error: Error message (if failed)
        """
        pass
-
-    # =========================================================================
-    # Consolidation - Pre-operation validation hook (optional - override to implement)
-    # =========================================================================
-
-    async def validate_consolidate(self, ctx: ConsolidateContext) -> ValidationResult:
-        """
-        Validate a consolidation operation before execution.
-
-        Override to implement custom validation logic for consolidation.
-
-        Args:
-            ctx: Context containing:
-                - bank_id: Bank identifier
-                - request_context: Request context with auth info
-
-        Returns:
-            ValidationResult indicating whether the operation is allowed.
-        """
-        return ValidationResult.accept()
-
-    # =========================================================================
-    # Consolidation - Post-operation hook (optional - override to implement)
-    # =========================================================================
-
-    async def on_consolidate_complete(self, result: ConsolidateResult) -> None:
-        """
-        Called after a consolidation operation completes (success or failure).
-
-        Override to implement post-operation logic such as usage tracking or audit logging.
-
-        Args:
-            result: Result context containing:
-                - bank_id: Bank identifier
-                - processed: Number of memories processed
-                - created: Number of mental models created
-                - updated: Number of mental models updated
-                - success: Whether the operation succeeded
-                - error: Error message (if failed)
-        """
-        pass
@@ -28,18 +28,6 @@ class TenantContext:
    schema_name: str


-@dataclass
-class Tenant:
-    """
-    Represents a tenant for worker discovery.
-
-    Used by list_tenants() to return tenant information including
-    the PostgreSQL schema name for database operations.
-    """
-
-    schema: str
-
-
 class TenantExtension(Extension, ABC):
    """
    Extension for multi-tenancy and API key authentication.
@@ -73,17 +61,3 @@ class TenantExtension(Extension, ABC):
            AuthenticationError: If authentication fails.
        """
        ...
-
-    @abstractmethod
-    async def list_tenants(self) -> list[Tenant]:
-        """
-        List all tenants that should be processed by workers.
-
-        This method is used by the worker to discover all tenants that need
-        task polling. Workers will poll for pending tasks in each tenant's schema.
-
-        Returns:
-            List of Tenant objects containing schema information.
-            For single-tenant setups, return [Tenant(schema="public")].
-        """
-        ...
@@ -23,7 +23,7 @@ import uvicorn
 from . import MemoryEngine
 from .api import create_app
 from .banner import print_banner
-from .config import DEFAULT_WORKERS, ENV_WORKERS, HindsightConfig, get_config
+from .config import HindsightConfig, get_config
 from .daemon import (
    DEFAULT_DAEMON_PORT,
    DEFAULT_IDLE_TIMEOUT,
@@ -31,7 +31,6 @@ from .daemon import (
    IdleTimeoutMiddleware,
    daemonize,
 )
-from .extensions import DefaultExtensionContext, OperationValidatorExtension, TenantExtension, load_extension

 # Filter deprecation warnings from third-party libraries
 warnings.filterwarnings("ignore", message="websockets.legacy is deprecated")
@@ -95,12 +94,7 @@ def main():

    # Development options
    parser.add_argument("--reload", action="store_true", help="Enable auto-reload on code changes (development only)")
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=int(os.getenv(ENV_WORKERS, str(DEFAULT_WORKERS))),
-        help=f"Number of worker processes (env: {ENV_WORKERS}, default: {DEFAULT_WORKERS})",
-    )
+    parser.add_argument("--workers", type=int, default=1, help="Number of worker processes (default: 1)")

    # Access log options
    parser.add_argument("--access-log", action="store_true", help="Enable access log")
@@ -170,93 +164,23 @@ def main():
    if args.log_level != config.log_level:
        config = HindsightConfig(
            database_url=config.database_url,
-            database_schema=config.database_schema,
            llm_provider=config.llm_provider,
            llm_api_key=config.llm_api_key,
            llm_model=config.llm_model,
            llm_base_url=config.llm_base_url,
-            llm_max_concurrent=config.llm_max_concurrent,
-            llm_max_retries=config.llm_max_retries,
-            llm_initial_backoff=config.llm_initial_backoff,
-            llm_max_backoff=config.llm_max_backoff,
-            llm_timeout=config.llm_timeout,
-            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,
-            retain_llm_provider=config.retain_llm_provider,
-            retain_llm_api_key=config.retain_llm_api_key,
-            retain_llm_model=config.retain_llm_model,
-            retain_llm_base_url=config.retain_llm_base_url,
-            retain_llm_max_concurrent=config.retain_llm_max_concurrent,
-            retain_llm_max_retries=config.retain_llm_max_retries,
-            retain_llm_initial_backoff=config.retain_llm_initial_backoff,
-            retain_llm_max_backoff=config.retain_llm_max_backoff,
-            retain_llm_timeout=config.retain_llm_timeout,
-            reflect_llm_provider=config.reflect_llm_provider,
-            reflect_llm_api_key=config.reflect_llm_api_key,
-            reflect_llm_model=config.reflect_llm_model,
-            reflect_llm_base_url=config.reflect_llm_base_url,
-            reflect_llm_max_concurrent=config.reflect_llm_max_concurrent,
-            reflect_llm_max_retries=config.reflect_llm_max_retries,
-            reflect_llm_initial_backoff=config.reflect_llm_initial_backoff,
-            reflect_llm_max_backoff=config.reflect_llm_max_backoff,
-            reflect_llm_timeout=config.reflect_llm_timeout,
-            consolidation_llm_provider=config.consolidation_llm_provider,
-            consolidation_llm_api_key=config.consolidation_llm_api_key,
-            consolidation_llm_model=config.consolidation_llm_model,
-            consolidation_llm_base_url=config.consolidation_llm_base_url,
-            consolidation_llm_max_concurrent=config.consolidation_llm_max_concurrent,
-            consolidation_llm_max_retries=config.consolidation_llm_max_retries,
-            consolidation_llm_initial_backoff=config.consolidation_llm_initial_backoff,
-            consolidation_llm_max_backoff=config.consolidation_llm_max_backoff,
-            consolidation_llm_timeout=config.consolidation_llm_timeout,
            embeddings_provider=config.embeddings_provider,
            embeddings_local_model=config.embeddings_local_model,
-            embeddings_local_force_cpu=config.embeddings_local_force_cpu,
            embeddings_tei_url=config.embeddings_tei_url,
-            embeddings_openai_base_url=config.embeddings_openai_base_url,
-            embeddings_cohere_base_url=config.embeddings_cohere_base_url,
            reranker_provider=config.reranker_provider,
            reranker_local_model=config.reranker_local_model,
-            reranker_local_force_cpu=config.reranker_local_force_cpu,
-            reranker_local_max_concurrent=config.reranker_local_max_concurrent,
            reranker_tei_url=config.reranker_tei_url,
-            reranker_tei_batch_size=config.reranker_tei_batch_size,
-            reranker_tei_max_concurrent=config.reranker_tei_max_concurrent,
-            reranker_max_candidates=config.reranker_max_candidates,
-            reranker_cohere_base_url=config.reranker_cohere_base_url,
            host=args.host,
            port=args.port,
            log_level=args.log_level,
-            log_format=config.log_format,
            mcp_enabled=config.mcp_enabled,
            graph_retriever=config.graph_retriever,
-            mpfp_top_k_neighbors=config.mpfp_top_k_neighbors,
-            recall_max_concurrent=config.recall_max_concurrent,
-            recall_connection_budget=config.recall_connection_budget,
-            retain_max_completion_tokens=config.retain_max_completion_tokens,
-            retain_chunk_size=config.retain_chunk_size,
-            retain_extract_causal_links=config.retain_extract_causal_links,
-            retain_extraction_mode=config.retain_extraction_mode,
-            retain_custom_instructions=config.retain_custom_instructions,
-            enable_observations=config.enable_observations,
-            consolidation_batch_size=config.consolidation_batch_size,
-            consolidation_max_tokens=config.consolidation_max_tokens,
            skip_llm_verification=config.skip_llm_verification,
            lazy_reranker=config.lazy_reranker,
-            run_migrations_on_startup=config.run_migrations_on_startup,
-            db_pool_min_size=config.db_pool_min_size,
-            db_pool_max_size=config.db_pool_max_size,
-            db_command_timeout=config.db_command_timeout,
-            db_acquire_timeout=config.db_acquire_timeout,
-            worker_enabled=config.worker_enabled,
-            worker_id=config.worker_id,
-            worker_poll_interval_ms=config.worker_poll_interval_ms,
-            worker_max_retries=config.worker_max_retries,
-            worker_batch_size=config.worker_batch_size,
-            worker_http_port=config.worker_http_port,
-            reflect_max_iterations=config.reflect_max_iterations,
-            mental_model_refresh_concurrency=config.mental_model_refresh_concurrency,
        )
    config.configure_logging()
    if not args.daemon:
@@ -267,35 +191,8 @@ def main():
    signal.signal(signal.SIGINT, _signal_handler)
    signal.signal(signal.SIGTERM, _signal_handler)

-    # Load operation validator extension if configured
-    operation_validator = load_extension("OPERATION_VALIDATOR", OperationValidatorExtension)
-    if operation_validator:
-        import logging
-
-        logging.info(f"Loaded operation validator: {operation_validator.__class__.__name__}")
-
-    # Load tenant extension if configured
-    tenant_extension = load_extension("TENANT", TenantExtension)
-    if tenant_extension:
-        import logging
-
-        logging.info(f"Loaded tenant extension: {tenant_extension.__class__.__name__}")
-
    # Create MemoryEngine (reads configuration from environment)
-    _memory = MemoryEngine(
-        operation_validator=operation_validator,
-        tenant_extension=tenant_extension,
-        run_migrations=config.run_migrations_on_startup,
-    )
-
-    # Set extension context on tenant extension (needed for schema provisioning)
-    if tenant_extension:
-        extension_context = DefaultExtensionContext(
-            database_url=config.database_url,
-            memory_engine=_memory,
-        )
-        tenant_extension.set_context(extension_context)
-        logging.info("Extension context set on tenant extension")
+    _memory = MemoryEngine()

    # Create FastAPI app
    app = create_app(
@@ -313,27 +210,14 @@ def main():
        app = idle_middleware

    # Prepare uvicorn config
-    # When using workers or reload, we must use import string so each worker can import the app
-    use_import_string = args.workers > 1 or args.reload
-    # Check for uvloop availability
-    try:
-        import uvloop  # noqa: F401
-
-        loop_impl = "uvloop"
-        print("uvloop available, will use for event loop")
-    except ImportError:
-        loop_impl = "asyncio"
-        print("uvloop not installed, using default asyncio event loop")
-
    uvicorn_config = {
-        "app": "hindsight_api.server:app" if use_import_string else app,
+        "app": app,
        "host": args.host,
        "port": args.port,
        "log_level": args.log_level,
        "access_log": args.access_log,
        "proxy_headers": args.proxy_headers,
        "ws": "wsproto",  # Use wsproto instead of websockets to avoid deprecation warnings
-        "loop": loop_impl,  # Explicitly set event loop implementation
    }

    # Add optional parameters if provided
@@ -366,7 +250,6 @@ def main():
    # Start idle checker in daemon mode
    if idle_middleware is not None:
        # Start the idle checker in a background thread with its own event loop
-        import logging
        import threading

        def run_idle_checker():
@@ -377,12 +260,12 @@ def main():
                loop = asyncio.new_event_loop()
                asyncio.set_event_loop(loop)
                loop.run_until_complete(idle_middleware._check_idle())
-            except Exception as e:
-                logging.error(f"Idle checker error: {e}", exc_info=True)
+            except Exception:
+                pass

        threading.Thread(target=run_idle_checker, daemon=True).start()

-    uvicorn.run(**uvicorn_config)
+    uvicorn.run(**uvicorn_config)  # type: ignore[invalid-argument-type] - dict kwargs


 if __name__ == "__main__":
@@ -44,6 +44,7 @@ import os
 import sys

 from mcp.server.fastmcp import FastMCP
+from mcp.types import Icon

 from hindsight_api.config import (
    DEFAULT_MCP_LOCAL_BANK_ID,
@@ -52,7 +53,6 @@ from hindsight_api.config import (
    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
@@ -85,6 +85,9 @@ def create_local_mcp_server(bank_id: str, memory=None) -> FastMCP:
    """
    # Import here to avoid slow startup if just checking --help
    from hindsight_api import MemoryEngine
+    from hindsight_api.engine.memory_engine import Budget
+    from hindsight_api.engine.response_models import VALID_RECALL_FACT_TYPES
+    from hindsight_api.models import RequestContext

    # Create memory engine with pg0 embedded database if not provided
    if memory is None:
@@ -102,17 +105,55 @@ def create_local_mcp_server(bank_id: str, memory=None) -> FastMCP:

    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
-    )
+    @mcp.tool(description=retain_description)
+    async def retain(content: str, context: str = "general") -> dict:
+        """
+        Args:
+            content: The fact/memory to store (be specific and include relevant details)
+            context: Category for the memory (e.g., 'preferences', 'work', 'hobbies', 'family'). Default: 'general'
+        """
+        import asyncio

-    register_mcp_tools(mcp, memory, config)
+        async def _retain():
+            try:
+                await memory.retain_batch_async(
+                    bank_id=bank_id,
+                    contents=[{"content": content, "context": context}],
+                    request_context=RequestContext(),
+                )
+            except Exception as e:
+                logger.error(f"Error storing memory: {e}", exc_info=True)
+
+        # Fire and forget - don't block on memory storage
+        asyncio.create_task(_retain())
+        return {"status": "accepted", "message": "Memory storage initiated"}
+
+    @mcp.tool(description=recall_description)
+    async def recall(query: str, max_tokens: int = 4096, budget: str = "low") -> dict:
+        """
+        Args:
+            query: Natural language search query (e.g., "user's food preferences", "what projects is user working on")
+            max_tokens: Maximum tokens to return in results (default: 4096)
+            budget: Search budget level - "low", "mid", or "high" (default: "low")
+        """
+        try:
+            # Map string budget to enum
+            budget_map = {"low": Budget.LOW, "mid": Budget.MID, "high": Budget.HIGH}
+            budget_enum = budget_map.get(budget.lower(), Budget.LOW)
+
+            search_result = await memory.recall_async(
+                bank_id=bank_id,
+                query=query,
+                fact_type=list(VALID_RECALL_FACT_TYPES),
+                budget=budget_enum,
+                max_tokens=max_tokens,
+                request_context=RequestContext(),
+            )
+
+            return search_result.model_dump()
+        except Exception as e:
+            logger.error(f"Error searching: {e}", exc_info=True)
+            return {"error": str(e), "results": []}

    return mcp

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Nicolò Boschi	bdef0d80f4	refactor deamin	2025-12-22 22:00:54 +01:00
Nicolò Boschi	7db1eb1b3a	refactor deamin	2025-12-22 21:50:19 +01:00
Nicolò Boschi	c96ae58c28	refactor deamin	2025-12-22 21:43:35 +01:00
Nicolò Boschi	d349376df7	refactor deamin	2025-12-22 21:30:54 +01:00
Nicolò Boschi	b403b94795	feat: refactor hindsight-embed architecture	2025-12-22 18:20:00 +01:00
Nicolò Boschi	fe5d3a999b	feat: refactor hindsight-embed architecture	2025-12-22 18:19:55 +01:00