test: add client tests for ReflectResponse parsing

Added comprehensive tests in hindsight-clients/python/tests to verify: - v0.4.0+ format with empty based_on object - v0.4.0+ format with null based_on - v0.4.0+ format with populated facts - v0.3.0 format (list) correctly fails validation - Missing based_on field handling These tests document the v0.3.0 -> v0.4.0 breaking change where based_on changed from list to object.
fix: remove client imports from API test
2026-02-12 10:06:20 +01:00 · 2026-02-12 10:05:08 +01:00 · 2026-02-11 17:51:09 +01:00 · 2026-02-11 17:06:59 +01:00 · 2026-02-11 11:24:26 +01:00 · 2026-02-11 10:41:41 +01:00
667 changed files with 98571 additions and 26948 deletions
@@ -2,7 +2,7 @@
 # Copy this file to .env and fill in your values

 # LLM Configuration (Required)
-# Supported providers: openai, groq, ollama, gemini, anthropic, lmstudio
+# 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
@@ -13,6 +13,13 @@ HINDSIGHT_API_LLM_BASE_URL=https://api.openai.com/v1
 # 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
@@ -26,6 +33,7 @@ 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)
@@ -42,3 +50,18 @@ HINDSIGHT_API_LOG_LEVEL=info
 # HINDSIGHT_API_RERANKER_LOCAL_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2
 # For TEI provider:
 # HINDSIGHT_API_RERANKER_TEI_URL=http://localhost:8081
+
+# Observability & Tracing (Optional - disabled by default)
+# Enable OpenTelemetry tracing for LLM calls (GenAI semantic conventions)
+# HINDSIGHT_API_OTEL_TRACES_ENABLED=true
+#
+# Local development with Grafana LGTM stack (recommended - see scripts/dev/grafana/README.md)
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
+#
+# Cloud backends (Grafana Cloud, Langfuse, DataDog, etc.)
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT=https://your-backend-url
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer your-token"
+#
+# Custom service name and environment (optional, defaults: hindsight-api, development)
+# HINDSIGHT_API_OTEL_SERVICE_NAME=hindsight-production
+# HINDSIGHT_API_OTEL_DEPLOYMENT_ENVIRONMENT=production
@@ -139,6 +139,104 @@ jobs:
        path: hindsight-clients/typescript/*.tgz
        retention-days: 1

+  release-openclaw-integration:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm ci
+
+    - name: Build
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm run build
+
+    - name: Publish to npm
+      working-directory: ./hindsight-integrations/openclaw
+      run: |
+        set +e
+        OUTPUT=$(npm publish --access public 2>&1)
+        EXIT_CODE=$?
+        echo "$OUTPUT"
+        if [ $EXIT_CODE -ne 0 ]; then
+          if echo "$OUTPUT" | grep -q "cannot publish over"; then
+            echo "Package version already published, skipping..."
+            exit 0
+          fi
+          exit $EXIT_CODE
+        fi
+      env:
+        NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+    - name: Pack for GitHub release
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: openclaw-integration
+        path: hindsight-integrations/openclaw/*.tgz
+        retention-days: 1
+
+  release-ai-sdk-integration:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm ci
+
+    - name: Build
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm run build
+
+    - name: Publish to npm
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: |
+        set +e
+        OUTPUT=$(npm publish --access public 2>&1)
+        EXIT_CODE=$?
+        echo "$OUTPUT"
+        if [ $EXIT_CODE -ne 0 ]; then
+          if echo "$OUTPUT" | grep -q "cannot publish over"; then
+            echo "Package version already published, skipping..."
+            exit 0
+          fi
+          exit $EXIT_CODE
+        fi
+      env:
+        NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+    - name: Pack for GitHub release
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: ai-sdk-integration
+        path: hindsight-integrations/ai-sdk/*.tgz
+        retention-days: 1
+
  release-control-plane:
    runs-on: ubuntu-latest
    environment: npm
@@ -242,6 +340,7 @@ jobs:
        retention-days: 1

  release-docker-images:
+    name: Release Docker (${{ matrix.image_name }}${{ matrix.tag_suffix }})
    runs-on: ubuntu-latest
    permissions:
      contents: read
@@ -251,10 +350,28 @@ jobs:
        include:
          - target: api-only
            image_name: hindsight-api
+            tag_suffix: ""
+            build_args: ""
+          - target: api-only
+            image_name: hindsight-api
+            tag_suffix: "-slim"
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
          - target: cp-only
            image_name: hindsight-control-plane
+            tag_suffix: ""
+            build_args: ""
          - target: standalone
            image_name: hindsight
+            tag_suffix: ""
+            build_args: ""
+          - target: standalone
+            image_name: hindsight
+            tag_suffix: "-slim"
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false

    steps:
    - uses: actions/checkout@v4
@@ -292,6 +409,9 @@ jobs:
      uses: docker/metadata-action@v5
      with:
        images: ghcr.io/${{ github.repository_owner }}/${{ matrix.image_name }}
+        flavor: |
+          latest=auto
+          suffix=${{ matrix.tag_suffix }}
        tags: |
          type=semver,pattern={{version}},value=${{ steps.get_version.outputs.VERSION }}
          type=semver,pattern={{major}}.{{minor}},value=${{ steps.get_version.outputs.VERSION }}
@@ -317,7 +437,7 @@ jobs:
    # - name: Smoke test - verify container starts
    #   env:
    #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
-    #   run: ./scripts/docker-smoke-test.sh "${{ matrix.image_name }}:test" "${{ matrix.target }}"
+    #   run: ./docker/test-image.sh "${{ matrix.image_name }}:test" "${{ matrix.target }}"

    # Build multi-platform and push to release tags
    - name: Build and push release images
@@ -326,6 +446,7 @@ jobs:
        context: .
        file: docker/standalone/Dockerfile
        target: ${{ matrix.target }}
+        build-args: ${{ matrix.build_args }}
        push: true
        platforms: linux/amd64,linux/arm64
        tags: ${{ steps.meta.outputs.tags }}
@@ -366,7 +487,7 @@ jobs:

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

@@ -389,6 +510,18 @@ jobs:
        name: typescript-client
        path: ./artifacts/typescript-client

+    - name: Download OpenClaw Integration
+      uses: actions/download-artifact@v4
+      with:
+        name: openclaw-integration
+        path: ./artifacts/openclaw-integration
+
+    - name: Download AI SDK Integration
+      uses: actions/download-artifact@v4
+      with:
+        name: ai-sdk-integration
+        path: ./artifacts/ai-sdk-integration
+
    - name: Download Control Plane
      uses: actions/download-artifact@v4
      with:
@@ -430,6 +563,10 @@ jobs:
        cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
        # TypeScript client
        cp artifacts/typescript-client/*.tgz release-assets/ || true
+        # OpenClaw Integration
+        cp artifacts/openclaw-integration/*.tgz release-assets/ || true
+        # AI SDK Integration
+        cp artifacts/ai-sdk-integration/*.tgz release-assets/ || true
        # Control Plane
        cp artifacts/control-plane/*.tgz release-assets/ || true
        # Rust CLI binaries
@@ -9,42 +9,11 @@ concurrency:
  cancel-in-progress: true

 jobs:
-  build-python-packages:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        include:
-          - name: hindsight-all
-            path: hindsight
-          - name: hindsight-api
-            path: hindsight-api
-          - name: hindsight-client
-            path: hindsight-clients/python
-          - name: hindsight-embed
-            path: hindsight-embed
-
-    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: Build ${{ matrix.name }}
-      working-directory: ./${{ matrix.path }}
-      run: uv build
-
  build-api-python-versions:
    runs-on: ubuntu-latest
    strategy:
      matrix:
-        python-version: ['3.11', '3.12', '3.13']
+        python-version: ['3.11', '3.12', '3.13', '3.14']

    steps:
    - uses: actions/checkout@v4
@@ -82,6 +51,52 @@ jobs:
    - name: Build TypeScript client
      run: npm run build --workspace=hindsight-clients/typescript

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

@@ -262,16 +277,35 @@ jobs:
      run: helm lint helm/hindsight

  build-docker-images:
+    name: Build Docker (${{ matrix.name }})
    runs-on: ubuntu-latest
    strategy:
      matrix:
        include:
          - target: api-only
            name: api
+            variant: full
+            build_args: ""
+          - target: api-only
+            name: api-slim
+            variant: slim
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
          - target: cp-only
            name: control-plane
+            variant: full
+            build_args: ""
          - target: standalone
            name: standalone
+            variant: full
+            build_args: ""
+          - target: standalone
+            name: standalone-slim
+            variant: slim
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false

    steps:
    - uses: actions/checkout@v4
@@ -290,20 +324,31 @@ jobs:
    - name: Set up Docker Buildx
      uses: docker/setup-buildx-action@v3

-    - name: Build ${{ matrix.name }} image
+    - name: Build ${{ matrix.name }} image (${{ matrix.variant }})
      uses: docker/build-push-action@v6
      with:
        context: .
        file: docker/standalone/Dockerfile
        target: ${{ matrix.target }}
+        build-args: ${{ matrix.build_args }}
        push: false
-        load: false
+        load: ${{ matrix.variant == 'slim' }}
+        tags: hindsight-${{ matrix.name }}:test
+        # Removed GitHub Actions cache (type=gha) - it frequently returns 502 errors
+        # causing buildx to fail with "failed to parse error response 502"
+        # Build will be slower but more reliable

-    # TODO: Re-enable smoke test when disk space issue is resolved
-    # - name: Smoke test - verify container starts
-    #   env:
-    #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
-    #   run: ./scripts/docker-smoke-test.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"
+    # Only test slim variants to save disk space (they're much smaller)
+    # Slim variants require external embedding providers
+    - name: Smoke test - verify container starts
+      if: matrix.variant == 'slim'
+      env:
+        GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
+        HINDSIGHT_API_EMBEDDINGS_PROVIDER: openai
+        HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+        HINDSIGHT_API_RERANKER_PROVIDER: cohere
+        HINDSIGHT_API_COHERE_API_KEY: ${{ secrets.COHERE_API_KEY }}
+      run: ./docker/test-image.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"

  test-api:
    runs-on: ubuntu-latest
@@ -726,9 +771,9 @@ jobs:
  test-embed:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_EMBED_LLM_PROVIDER: groq
-      HINDSIGHT_EMBED_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_EMBED_LLM_MODEL: openai/gpt-oss-20b
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      # Prefer CPU-only PyTorch in CI
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu

@@ -759,10 +804,62 @@ jobs:
          ${{ runner.os }}-huggingface-embed-
          ${{ runner.os }}-huggingface-

+    - name: Run unit and integration tests
+      working-directory: ./hindsight-embed
+      run: uv run pytest tests/ -v
+
    - name: Run smoke test
      working-directory: ./hindsight-embed
      run: ./test.sh

+  test-hindsight-all:
+    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
+      # For test_server_integration.py compatibility
+      HINDSIGHT_LLM_PROVIDER: groq
+      HINDSIGHT_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_LLM_MODEL: openai/gpt-oss-20b
+      # Prefer CPU-only PyTorch in CI
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: 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 hindsight-all
+      working-directory: ./hindsight
+      run: uv build
+
+    - name: Install dependencies
+      working-directory: ./hindsight
+      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-all-${{ hashFiles('hindsight/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-all-
+          ${{ runner.os }}-huggingface-
+
+    - name: Run unit tests
+      working-directory: ./hindsight
+      run: uv run pytest tests/ -v
+
  test-doc-examples:
    runs-on: ubuntu-latest
    needs: test-rust-cli
@@ -875,6 +972,78 @@ jobs:
        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: |
+          ${{ runner.os }}-huggingface-
+
+    - name: Install hindsight-dev dependencies
+      working-directory: ./hindsight-dev
+      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
+
+    - name: Install current hindsight-api
+      working-directory: ./hindsight-api
+      run: uv sync --frozen --index-strategy unsafe-best-match
+
+    - name: Pre-download models
+      working-directory: ./hindsight-api
+      run: |
+        uv run python -c "
+        from sentence_transformers import SentenceTransformer, CrossEncoder
+        print('Downloading embedding model...')
+        SentenceTransformer('BAAI/bge-small-en-v1.5')
+        print('Downloading cross-encoder model...')
+        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
+        print('Models downloaded successfully')
+        "
+
+    - name: Run upgrade tests
+      working-directory: ./hindsight-dev
+      run: uv run pytest upgrade_tests/ -v --tb=short
+
+    - name: Show upgrade test logs
+      if: always()
+      run: |
+        echo "=== Upgrade Test Server Logs ==="
+        for log in /tmp/upgrade-test-*.log; do
+          if [ -f "$log" ]; then
+            echo ""
+            echo "--- $log ---"
+            tail -500 "$log"
+          fi
+        done
+
  verify-generated-files:
    runs-on: ubuntu-latest
    env:
@@ -945,4 +1114,52 @@ jobs:
          git diff --stat
          exit 1
        fi
-        echo "✓ All generated files are up to date"
+        echo "✓ All generated files are up to date"
+
+  check-openapi-compatibility:
+    runs-on: ubuntu-latest
+    env:
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        fetch-depth: 0  # Fetch full git history to access base branch
+
+    - 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: Install hindsight-dev dependencies
+      run: |
+        cd hindsight-dev && uv sync --frozen --index-strategy unsafe-best-match
+
+    - name: Check OpenAPI compatibility with base branch
+      run: |
+        # Get the base branch (usually main)
+        BASE_BRANCH="${{ github.base_ref }}"
+
+        if [ -z "$BASE_BRANCH" ]; then
+          echo "⚠️  Warning: No base branch found (not a PR?). Skipping compatibility check."
+          exit 0
+        fi
+
+        echo "Checking OpenAPI compatibility against base branch: $BASE_BRANCH"
+
+        # Extract the old OpenAPI spec from base branch
+        git show "origin/$BASE_BRANCH:hindsight-docs/static/openapi.json" > /tmp/old-openapi.json
+
+        if [ ! -s /tmp/old-openapi.json ]; then
+          echo "⚠️  Warning: Could not find OpenAPI spec in base branch. Skipping compatibility check."
+          exit 0
+        fi
+
+        # Check compatibility using our tool
+        cd hindsight-dev
+        uv run check-openapi-compatibility /tmp/old-openapi.json ../hindsight-docs/static/openapi.json
@@ -45,9 +45,14 @@ 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.md
+# Changelog is now tracked in hindsight-docs/src/pages/changelog.md
+# CHANGELOG.md
+
+blog-post*
@@ -7,8 +7,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 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")
- **Opinion facts**: Beliefs with confidence scores ("Paris is beautiful" - 0.9 confidence)
- **Observations**: Complex mental models derived from reflection
+- **Mental models**: Consolidated knowledge synthesized from facts ("User prefers functional programming patterns")

 ## Development Commands

@@ -46,6 +45,7 @@ cd hindsight-control-plane && npm run dev
 ./scripts/dev/start-docs.sh
 ```

+
 ### Generating Clients/OpenAPI
 ```bash
 # Regenerate OpenAPI spec after API changes (REQUIRED after changing endpoints)
@@ -101,7 +101,7 @@ cd hindsight-control-plane && npm run dev
 Main operations:
 - **Retain**: Store memories, extracts facts/entities/relationships
 - **Recall**: Retrieve memories via 4 parallel strategies (semantic, BM25, graph, temporal) + reranking
- **Reflect**: Deep analysis forming new opinions/observations (disposition-aware)
+- **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.
@@ -93,6 +93,34 @@ uv run ty check hindsight_api  # Type check
 3. Run tests to ensure nothing breaks
 4. Submit a PR with a clear description of changes

+## Release Process
+
+The project uses `scripts/release.sh` for creating releases. This script automates the entire release workflow:
+
+1. Bumps version in all components (API, clients, CLI, control plane, Helm)
+2. **Regenerates OpenAPI spec and client SDKs** (Python, TypeScript, Rust)
+3. Updates documentation versioning
+4. Creates a commit and git tag
+5. Pushes to GitHub (triggers CI/CD to publish packages)
+
+### Usage
+
+```bash
+./scripts/release.sh <version>
+```
+
+**Example:**
+```bash
+./scripts/release.sh 0.5.0
+```
+
+### Important for Developers
+
+- During development, version bumps in `__init__.py` do NOT require client regeneration
+- Clients are only regenerated during releases
+- Do not manually run `./scripts/generate-clients.sh` unless testing generation changes
+- Client version comments will reflect the API version from the latest release
+
 ## Reporting Issues

 Open an issue on GitHub with:
@@ -1,6 +1,6 @@
 <div align="center">

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

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

@@ -17,76 +17,76 @@

 ## What is Hindsight?

-Hindsight™ is an agent memory system built to create smarter agents that learn over time. It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.
+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 addresses common challenges that have frustrated AI engineers building agents to automate tasks and assist users with conversational interfaces. Many of these challenges stem directly from a lack of memory.

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

-## 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.
-
---
+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.

 ## Memory Performance & Accuracy

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

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

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

-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.
+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)
+
+---
+
+> 🤖 **Using a coding agent?** Install the Hindsight documentation skill for instant access to docs while you code:
+> ```bash
+> npx skills add https://github.com/vectorize-io/hindsight --skill hindsight-docs
+> ```
+> Works with Claude Code, Cursor, and other AI coding assistants.
+
+---

-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

 ### Docker (recommended)

 ```bash
-export OPENAI_API_KEY=your-key
+export OPENAI_API_KEY=sk-xxx

 docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
  -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
-  -e HINDSIGHT_API_LLM_MODEL=o3-mini \
  -v $HOME/.hindsight-docker:/home/hindsight/.pg0 \
  ghcr.io/vectorize-io/hindsight:latest
 ```

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

-API: http://localhost:8888  
-UI: http://localhost:9999

-Install client:
+
+### Docker (external PostgreSQL)
+
+```bash
+export OPENAI_API_KEY=sk-xxx
+export HINDSIGHT_DB_PASSWORD=choose-a-password
+cd docker/docker-compose
+docker compose up 
+```
+
+
+>API: http://localhost:8888
+>UI: http://localhost:9999
+
+### Client

 ```bash
 pip install hindsight-client -U
@@ -94,7 +94,7 @@ pip install hindsight-client -U
 npm install @vectorize-io/hindsight-client
 ```

-Python example:
+#### Python

 ```python
 from hindsight_client import Hindsight
@@ -111,7 +111,29 @@ client.recall(bank_id="my-bank", query="What does Alice do?")
 client.reflect(bank_id="my-bank", query="Tell me about Alice")
 ```

-### Python (embedded, no Docker)
+#### Node.js / TypeScript
+
+```bash
+npm install @vectorize-io/hindsight-client
+```
+
+```javascript
+const { HindsightClient } = require('@vectorize-io/hindsight-client');
+
+const main = async () => {
+  const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });
+
+  await client.retain('my-bank', 'Alice loves hiking in Yosemite');
+
+  const results = await client.recall('my-bank', 'What does Alice like?');
+  console.log(results);
+}
+
+main();
+```
+
+
+### Python Embedded (no server required)

 ```bash
 pip install hindsight-all -U
@@ -131,25 +153,48 @@ with HindsightServer(
    results = client.recall(bank_id="my-bank", query="Where does Alice work?")
 ```

-### Node.js / TypeScript

-```bash
-npm install @vectorize-io/hindsight-client
-```
+---

-```javascript
-const { HindsightClient } = require('@vectorize-io/hindsight-client');
+## Use Cases

-const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });

-await client.retain('my-bank', 'Alice loves hiking in Yosemite');
-await client.recall('my-bank', 'What does Alice like?');
-```
+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.
@@ -208,7 +253,7 @@ 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 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. 
+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.

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

@@ -0,0 +1,54 @@
+# Docker Compose file for Hindsight with PostgreSQL and pgvector
+#
+# Make sure to set the required environment variables before running:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - Configure LLM provider variables as needed (see below in the hindsight service)
+#
+# Usage:
+# docker compose up -d
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
+
+services:
+  db:
+    # Use a PostgreSQL-Image with pgvector extension pre-installed
+    # see https://hub.docker.com/r/pgvector/pgvector 
+    image: pgvector/pgvector:pg${HINDSIGHT_DB_VERSION:-18}
+    container_name: hindsight-db
+    restart: always
+    # Expose PostgreSQL port
+    # ports:
+      # - "5432:5432"
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      - HINDSIGHT_API_LLM_API_KEY=${OPENAI_API_KEY?Please set the OPENAI_API_KEY env variable}
+      - HINDSIGHT_API_DATABASE_URL=postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+    depends_on:
+      - db
+    networks:
+      - hindsight-net
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:
@@ -8,6 +8,7 @@
 #                                     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
+#                                     NOTE: tiktoken encodings are ALWAYS preloaded (required for air-gapped deployments)
 #
 # Examples:
 #   docker build -t hindsight .                                          # Both (standalone)
@@ -167,18 +168,57 @@ USER hindsight

 ENV PATH="/app/api/.venv/bin:${PATH}"

+# Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
+# Tiktoken is a core runtime dependency, not an optional ML model
+RUN MAX_RETRIES=3; \
+    RETRY_DELAY=5; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading tiktoken encoding..."; \
+      /app/api/.venv/bin/python -c "\
+import tiktoken; \
+print('Downloading cl100k_base encoding...'); \
+tiktoken.get_encoding('cl100k_base'); \
+print('Tiktoken encoding cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ]; then \
+      echo "ERROR: Failed to download tiktoken encoding after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi
+
 # Pre-download ML models to avoid runtime download (conditional)
 # Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
+# Includes retry logic with exponential backoff for transient network failures
 ARG PRELOAD_ML_MODELS
 ARG INCLUDE_LOCAL_MODELS
+ENV HF_HUB_DOWNLOAD_TIMEOUT=600
 RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
-    /app/api/.venv/bin/python -c "\
+    MAX_RETRIES=3; \
+    RETRY_DELAY=10; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading ML models..."; \
+      /app/api/.venv/bin/python -c "\
+import os; os.environ['HF_HUB_DOWNLOAD_TIMEOUT'] = '600'; \
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Models cached successfully')"; \
+print('Models cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ] && ! /app/api/.venv/bin/python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-small-en-v1.5')" 2>/dev/null; then \
+      echo "ERROR: Failed to download models after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi; \
    elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
    else echo "Skipping ML model preload"; fi

@@ -190,6 +230,10 @@ ENV HINDSIGHT_API_LOG_LEVEL=info
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=false
 ENV PYTHONUNBUFFERED=1
+# Suppress verbose transformers/HuggingFace model loading warnings
+ENV TRANSFORMERS_VERBOSITY=error
+ENV HF_HUB_VERBOSITY=error
+ENV TOKENIZERS_PARALLELISM=false

 CMD ["/app/start-all.sh"]

@@ -275,18 +319,57 @@ USER hindsight

 ENV PATH="/app/api/.venv/bin:${PATH}"

+# Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
+# Tiktoken is a core runtime dependency, not an optional ML model
+RUN MAX_RETRIES=3; \
+    RETRY_DELAY=5; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading tiktoken encoding..."; \
+      /app/api/.venv/bin/python -c "\
+import tiktoken; \
+print('Downloading cl100k_base encoding...'); \
+tiktoken.get_encoding('cl100k_base'); \
+print('Tiktoken encoding cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ]; then \
+      echo "ERROR: Failed to download tiktoken encoding after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi
+
 # Pre-download ML models to avoid runtime download (conditional)
 # Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
+# Includes retry logic with exponential backoff for transient network failures
 ARG PRELOAD_ML_MODELS
 ARG INCLUDE_LOCAL_MODELS
+ENV HF_HUB_DOWNLOAD_TIMEOUT=600
 RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
-    /app/api/.venv/bin/python -c "\
+    MAX_RETRIES=3; \
+    RETRY_DELAY=10; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading ML models..."; \
+      /app/api/.venv/bin/python -c "\
+import os; os.environ['HF_HUB_DOWNLOAD_TIMEOUT'] = '600'; \
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Models cached successfully')"; \
+print('Models cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ] && ! /app/api/.venv/bin/python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-small-en-v1.5')" 2>/dev/null; then \
+      echo "ERROR: Failed to download models after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi; \
    elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
    else echo "Skipping ML model preload"; fi

@@ -300,6 +383,10 @@ ENV HINDSIGHT_CP_DATAPLANE_API_URL=http://localhost:8888
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=true
 ENV PYTHONUNBUFFERED=1
+# Suppress verbose transformers/HuggingFace model loading warnings
+ENV TRANSFORMERS_VERBOSITY=error
+ENV HF_HUB_VERBOSITY=error
+ENV TOKENIZERS_PARALLELISM=false

 CMD ["/app/start-all.sh"]

@@ -6,28 +6,40 @@
 # Can be run locally or in CI pipelines.
 #
 # Usage:
-#   ./scripts/docker-smoke-test.sh <image> [target]
+#   ./docker/test-image.sh <image> [target]
 #
 # Arguments:
 #   image   - Docker image to test (e.g., hindsight-api:test, ghcr.io/vectorize-io/hindsight:latest)
 #   target  - Optional: 'cp-only' for control plane, otherwise assumes API image (default: api)
 #
 # Environment variables:
-#   GROQ_API_KEY              - Required for API/standalone images (LLM verification)
-#   HINDSIGHT_API_LLM_PROVIDER - LLM provider (default: groq)
-#   HINDSIGHT_API_LLM_MODEL    - LLM model (default: llama-3.3-70b-versatile)
-#   SMOKE_TEST_TIMEOUT        - Timeout in seconds (default: 120)
-#   SMOKE_TEST_CONTAINER_NAME - Container name (default: hindsight-smoke-test)
+#   GROQ_API_KEY                                - Required for API/standalone images (LLM verification)
+#   HINDSIGHT_API_LLM_PROVIDER                  - LLM provider (default: groq)
+#   HINDSIGHT_API_LLM_MODEL                     - LLM model (default: llama-3.3-70b-versatile)
+#   HINDSIGHT_API_EMBEDDINGS_PROVIDER           - Embeddings provider (optional, for slim images: openai, cohere, tei)
+#   HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY     - OpenAI API key for embeddings (optional)
+#   HINDSIGHT_API_RERANKER_PROVIDER             - Reranker provider (optional, for slim images: cohere, tei)
+#   HINDSIGHT_API_COHERE_API_KEY                - Cohere API key for reranking (optional)
+#   SMOKE_TEST_TIMEOUT                          - Timeout in seconds (default: 120)
+#   SMOKE_TEST_CONTAINER_NAME                   - Container name (default: hindsight-smoke-test)
 #
 # Examples:
-#   # Test a locally built image
-#   ./scripts/docker-smoke-test.sh hindsight-api:test
+#   # Test a locally built full image
+#   ./docker/test-image.sh hindsight-api:test
 #
 #   # Test a released image
-#   ./scripts/docker-smoke-test.sh ghcr.io/vectorize-io/hindsight:latest
+#   ./docker/test-image.sh ghcr.io/vectorize-io/hindsight:latest
 #
 #   # Test control plane image
-#   ./scripts/docker-smoke-test.sh hindsight-control-plane:test cp-only
+#   ./docker/test-image.sh hindsight-control-plane:test cp-only
+#
+#   # Test slim image with external providers
+#   export GROQ_API_KEY=gsk_xxx
+#   export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
+#   export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=sk-xxx
+#   export HINDSIGHT_API_RERANKER_PROVIDER=cohere
+#   export HINDSIGHT_API_COHERE_API_KEY=xxx
+#   ./docker/test-image.sh hindsight-slim:test
 #
 # Exit codes:
 #   0 - Success (container healthy)
@@ -108,12 +120,32 @@ if [ "$TARGET" = "cp-only" ]; then
        -p "${HEALTH_PORT}:${HEALTH_PORT}" \
        "$IMAGE"
 else
-    docker run -d --name "$CONTAINER_NAME" \
-        -e HINDSIGHT_API_LLM_PROVIDER="$LLM_PROVIDER" \
-        -e HINDSIGHT_API_LLM_API_KEY="${GROQ_API_KEY}" \
-        -e HINDSIGHT_API_LLM_MODEL="$LLM_MODEL" \
-        -p "${HEALTH_PORT}:${HEALTH_PORT}" \
-        "$IMAGE"
+    # Build docker run command with required and optional env vars
+    DOCKER_CMD="docker run -d --name $CONTAINER_NAME"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_PROVIDER=$LLM_PROVIDER"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_API_KEY=${GROQ_API_KEY}"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_MODEL=$LLM_MODEL"
+
+    # Add optional embeddings provider config
+    if [ -n "${HINDSIGHT_API_EMBEDDINGS_PROVIDER:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_EMBEDDINGS_PROVIDER=${HINDSIGHT_API_EMBEDDINGS_PROVIDER}"
+    fi
+    if [ -n "${HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=${HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY}"
+    fi
+
+    # Add optional reranker provider config
+    if [ -n "${HINDSIGHT_API_RERANKER_PROVIDER:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_RERANKER_PROVIDER=${HINDSIGHT_API_RERANKER_PROVIDER}"
+    fi
+    if [ -n "${HINDSIGHT_API_COHERE_API_KEY:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_COHERE_API_KEY=${HINDSIGHT_API_COHERE_API_KEY}"
+    fi
+
+    DOCKER_CMD="$DOCKER_CMD -p ${HEALTH_PORT}:${HEALTH_PORT}"
+    DOCKER_CMD="$DOCKER_CMD $IMAGE"
+
+    eval $DOCKER_CMD
 fi

 # Wait for health endpoint
@@ -0,0 +1,51 @@
+#!/bin/bash
+#
+# Local Test Script for Slim Docker Images
+#
+# This script makes it easy to test slim images locally with external providers.
+# It expects API keys to be set in environment variables.
+#
+# Usage:
+#   export GROQ_API_KEY=gsk_xxx
+#   export OPENAI_API_KEY=sk-xxx
+#   export COHERE_API_KEY=xxx
+#   ./docker/test-slim-local.sh
+#
+# Or inline:
+#   GROQ_API_KEY=gsk_xxx OPENAI_API_KEY=sk_xxx COHERE_API_KEY=xxx ./docker/test-slim-local.sh
+#
+
+set -euo pipefail
+
+# Check for required API keys
+if [ -z "${GROQ_API_KEY:-}" ]; then
+    echo "❌ Error: GROQ_API_KEY environment variable is required"
+    echo "Set it with: export GROQ_API_KEY=gsk_xxx"
+    exit 1
+fi
+
+if [ -z "${OPENAI_API_KEY:-}" ]; then
+    echo "❌ Error: OPENAI_API_KEY environment variable is required"
+    echo "Set it with: export OPENAI_API_KEY=sk-xxx"
+    exit 1
+fi
+
+if [ -z "${COHERE_API_KEY:-}" ]; then
+    echo "❌ Error: COHERE_API_KEY environment variable is required"
+    echo "Set it with: export COHERE_API_KEY=xxx"
+    exit 1
+fi
+
+# Configuration
+IMAGE="${1:-hindsight-slim:test}"
+echo "Testing image: $IMAGE"
+echo ""
+
+# Set up external providers
+export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
+export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=$OPENAI_API_KEY
+export HINDSIGHT_API_RERANKER_PROVIDER=cohere
+export HINDSIGHT_API_COHERE_API_KEY=$COHERE_API_KEY
+
+# Run the test
+exec "$(dirname "$0")/test-image.sh" "$IMAGE" standalone
@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.3.0
-appVersion: "0.3.0"
+version: 0.4.10
+appVersion: "0.4.10"
 keywords:
  - ai
  - memory
@@ -127,6 +127,38 @@ API URL for control plane
 {{- printf "http://%s-api:%d" (include "hindsight.fullname" .) (.Values.api.service.port | int) }}
 {{- end }}

+{{/*
+TEI reranker labels
+*/}}
+{{- define "hindsight.tei.reranker.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: tei-reranker
+{{- end }}
+
+{{/*
+TEI reranker selector labels
+*/}}
+{{- define "hindsight.tei.reranker.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: tei-reranker
+{{- end }}
+
+{{/*
+TEI embedding labels
+*/}}
+{{- define "hindsight.tei.embedding.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: tei-embedding
+{{- end }}
+
+{{/*
+TEI embedding selector labels
+*/}}
+{{- define "hindsight.tei.embedding.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: tei-embedding
+{{- end }}
+
 {{/*
 Get the name of the secret to use
 */}}
@@ -33,7 +33,7 @@ spec:
      - name: api
        securityContext:
          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag | default .Values.version }}"
+        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag | default .Values.version | default .Chart.AppVersion }}"
        imagePullPolicy: {{ .Values.api.image.pullPolicy }}
        ports:
        - name: http
@@ -60,10 +60,25 @@ spec:
        - name: HINDSIGHT_API_WORKER_ENABLED
          value: "false"
        {{- end }}
+        {{- /* Explicitly set port to override K8s service discovery env var (HINDSIGHT_API_PORT) */}}
+        - name: HINDSIGHT_API_PORT
+          value: {{ .Values.api.service.targetPort | quote }}
        {{- range $key, $value := .Values.api.env }}
        - name: {{ $key }}
          value: {{ $value | quote }}
        {{- end }}
+        {{- if .Values.tei.reranker.enabled }}
+        - name: HINDSIGHT_API_RERANKER_PROVIDER
+          value: "tei"
+        - name: HINDSIGHT_API_RERANKER_TEI_URL
+          value: "http://{{ include "hindsight.fullname" . }}-tei-reranker:{{ .Values.tei.reranker.port }}"
+        {{- end }}
+        {{- if .Values.tei.embedding.enabled }}
+        - name: HINDSIGHT_API_EMBEDDINGS_PROVIDER
+          value: "tei"
+        - name: HINDSIGHT_API_EMBEDDINGS_TEI_URL
+          value: "http://{{ include "hindsight.fullname" . }}-tei-embedding:{{ .Values.tei.embedding.port }}"
+        {{- end }}
        {{- /* Only use api.secrets when not using existingSecret (for chart-managed secrets) */}}
        {{- if not .Values.existingSecret }}
        {{- range $key, $value := .Values.api.secrets }}
@@ -84,7 +99,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.api.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -33,7 +33,7 @@ spec:
      - name: control-plane
        securityContext:
          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag | default .Values.version }}"
+        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag | default .Values.version | default .Chart.AppVersion }}"
        imagePullPolicy: {{ .Values.controlPlane.image.pullPolicy }}
        ports:
        - name: http
@@ -71,7 +71,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.controlPlane.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -0,0 +1,56 @@
+{{- if and .Values.api.enabled .Values.api.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-api
+  labels:
+    {{- include "hindsight.api.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.api.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.api.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.api.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.api.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.api.selectorLabels" . | nindent 6 }}
+{{- end }}
+---
+{{- if and .Values.controlPlane.enabled .Values.controlPlane.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-control-plane
+  labels:
+    {{- include "hindsight.controlPlane.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.controlPlane.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.controlPlane.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.controlPlane.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.controlPlane.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.controlPlane.selectorLabels" . | nindent 6 }}
+{{- end }}
+---
+{{- if and .Values.worker.enabled .Values.worker.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-worker
+  labels:
+    {{- include "hindsight.worker.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.worker.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.worker.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.worker.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.worker.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.worker.selectorLabels" . | nindent 6 }}
+{{- end }}
@@ -0,0 +1,76 @@
+{{- if .Values.tei.embedding.enabled }}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-embedding
+  labels:
+    {{- include "hindsight.tei.embedding.labels" . | nindent 4 }}
+spec:
+  replicas: {{ .Values.tei.embedding.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      {{- with .Values.podAnnotations }}
+      annotations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      labels:
+        {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: tei-embedding
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.tei.embedding.image.repository }}:{{ .Values.tei.embedding.image.tag }}"
+        imagePullPolicy: {{ .Values.tei.embedding.image.pullPolicy }}
+        args:
+          - "--model-id"
+          - {{ .Values.tei.embedding.model | quote }}
+          - "--hostname"
+          - "0.0.0.0"
+          {{- range .Values.tei.embedding.args }}
+          - {{ . | quote }}
+          {{- end }}
+        ports:
+        - name: http
+          containerPort: {{ .Values.tei.embedding.port }}
+          protocol: TCP
+        env:
+        - name: PORT
+          value: {{ .Values.tei.embedding.port | quote }}
+        {{- range $key, $value := .Values.tei.embedding.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.tei.embedding.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.tei.embedding.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.tei.embedding.resources | nindent 10 }}
+        volumeMounts:
+        - name: model-cache
+          mountPath: /data
+      volumes:
+      - name: model-cache
+        emptyDir: {}
+      {{- 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 }}
@@ -0,0 +1,17 @@
+{{- if .Values.tei.embedding.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-embedding
+  labels:
+    {{- include "hindsight.tei.embedding.labels" . | nindent 4 }}
+spec:
+  type: ClusterIP
+  ports:
+  - port: {{ .Values.tei.embedding.port }}
+    targetPort: http
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 4 }}
+{{- end }}
@@ -0,0 +1,76 @@
+{{- if .Values.tei.reranker.enabled }}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-reranker
+  labels:
+    {{- include "hindsight.tei.reranker.labels" . | nindent 4 }}
+spec:
+  replicas: {{ .Values.tei.reranker.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      {{- with .Values.podAnnotations }}
+      annotations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      labels:
+        {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: tei-reranker
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.tei.reranker.image.repository }}:{{ .Values.tei.reranker.image.tag }}"
+        imagePullPolicy: {{ .Values.tei.reranker.image.pullPolicy }}
+        args:
+          - "--model-id"
+          - {{ .Values.tei.reranker.model | quote }}
+          - "--hostname"
+          - "0.0.0.0"
+          {{- range .Values.tei.reranker.args }}
+          - {{ . | quote }}
+          {{- end }}
+        ports:
+        - name: http
+          containerPort: {{ .Values.tei.reranker.port }}
+          protocol: TCP
+        env:
+        - name: PORT
+          value: {{ .Values.tei.reranker.port | quote }}
+        {{- range $key, $value := .Values.tei.reranker.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.tei.reranker.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.tei.reranker.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.tei.reranker.resources | nindent 10 }}
+        volumeMounts:
+        - name: model-cache
+          mountPath: /data
+      volumes:
+      - name: model-cache
+        emptyDir: {}
+      {{- 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 }}
@@ -0,0 +1,17 @@
+{{- if .Values.tei.reranker.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-reranker
+  labels:
+    {{- include "hindsight.tei.reranker.labels" . | nindent 4 }}
+spec:
+  type: ClusterIP
+  ports:
+  - port: {{ .Values.tei.reranker.port }}
+    targetPort: http
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 4 }}
+{{- end }}
@@ -32,7 +32,7 @@ spec:
      - name: worker
        securityContext:
          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version }}"
+        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version | default .Chart.AppVersion }}"
        imagePullPolicy: {{ .Values.worker.image.pullPolicy }}
        command: ["hindsight-worker"]
        ports:
@@ -99,7 +99,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.worker.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -1,7 +1,8 @@
 # Default values for hindsight

-# Chart version - use this to set a consistent image tag across all components
-version: "0.1.1"
+# Global version override - use this to set a consistent image tag across all components
+# If not set, defaults to Chart.appVersion from Chart.yaml
+# version: ""

 # Use an existing secret instead of creating one from values
 # When set, all keys from this secret are injected as environment variables via envFrom
@@ -57,6 +58,15 @@ api:
    timeoutSeconds: 3
    failureThreshold: 3

+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
  # Environment variables
  env:
    #HINDSIGHT_API_LLM_PROVIDER: "groq"
@@ -75,7 +85,7 @@ worker:
  image:
    repository: ghcr.io/vectorize-io/hindsight-api
    pullPolicy: IfNotPresent
-    # tag defaults to .Values.version if not specified
+    # tag: ""  # defaults to .Values.version, then Chart.appVersion if not specified

  service:
    # Service for metrics scraping (headless for StatefulSet)
@@ -121,6 +131,15 @@ worker:
    # HTTP port for metrics/health (matches service.targetPort)
    HINDSIGHT_API_WORKER_HTTP_PORT: "8889"

+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
  # Secret environment variables (inherited from api.secrets if not specified)
  secrets: {}

@@ -164,6 +183,15 @@ controlPlane:
    timeoutSeconds: 3
    failureThreshold: 3

+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
  # Environment variables
  env:
    NODE_ENV: "production"
@@ -262,9 +290,87 @@ nodeSelector: {}
 # Tolerations
 tolerations: []

-# Affinity
+# Affinity (applied to all components unless overridden per-component)
 affinity: {}

+# TEI (Text Embeddings Inference) - optional standalone deployments
+# for reranking and/or embedding models
+tei:
+  reranker:
+    enabled: false
+    replicaCount: 1
+    image:
+      repository: ghcr.io/huggingface/text-embeddings-inference
+      tag: cpu-1.8.3
+      pullPolicy: IfNotPresent
+    model: "cross-encoder/ms-marco-MiniLM-L-6-v2"
+    port: 8090
+    args:
+      - "--auto-truncate"
+    env:
+      PAYLOAD_LIMIT: "10000000"
+      MAX_CLIENT_BATCH_SIZE: "256"
+    resources:
+      limits:
+        cpu: 2000m
+        memory: 2Gi
+      requests:
+        cpu: 500m
+        memory: 1Gi
+    livenessProbe:
+      httpGet:
+        path: /health
+        port: 8090
+      initialDelaySeconds: 30
+      periodSeconds: 10
+      timeoutSeconds: 5
+      failureThreshold: 6
+    readinessProbe:
+      httpGet:
+        path: /health
+        port: 8090
+      initialDelaySeconds: 15
+      periodSeconds: 5
+      timeoutSeconds: 3
+      failureThreshold: 3
+
+  embedding:
+    enabled: false
+    replicaCount: 1
+    image:
+      repository: ghcr.io/huggingface/text-embeddings-inference
+      tag: cpu-1.8.3
+      pullPolicy: IfNotPresent
+    model: "sentence-transformers/all-MiniLM-L6-v2"
+    port: 8091
+    args: []
+    env:
+      PAYLOAD_LIMIT: "10000000"
+      MAX_CLIENT_BATCH_SIZE: "256"
+    resources:
+      limits:
+        cpu: 2000m
+        memory: 2Gi
+      requests:
+        cpu: 500m
+        memory: 1Gi
+    livenessProbe:
+      httpGet:
+        path: /health
+        port: 8091
+      initialDelaySeconds: 30
+      periodSeconds: 10
+      timeoutSeconds: 5
+      failureThreshold: 6
+    readinessProbe:
+      httpGet:
+        path: /health
+        port: 8091
+      initialDelaySeconds: 15
+      periodSeconds: 5
+      timeoutSeconds: 3
+      failureThreshold: 3
+
 # Autoscaling
 autoscaling:
  enabled: false
@@ -46,4 +46,4 @@ __all__ = [
    "RemoteTEICrossEncoder",
    "LLMConfig",
 ]
-__version__ = "0.1.0"
+__version__ = "0.4.10"
@@ -11,6 +11,7 @@ from collections.abc import Sequence
 import sqlalchemy as sa
 from alembic import op
 from pgvector.sqlalchemy import Vector
+from sqlalchemy import text
 from sqlalchemy.dialects import postgresql

 # revision identifiers, used by Alembic.
@@ -23,8 +24,21 @@ depends_on: str | Sequence[str] | None = None
 def upgrade() -> None:
    """Upgrade schema - create all tables from scratch."""

-    # Enable required extensions
-    op.execute("CREATE EXTENSION IF NOT EXISTS vector")
+    # Note: pgvector extension is installed globally BEFORE migrations run
+    # See migrations.py:run_migrations() - this ensures the extension is available
+    # to all schemas, not just the one being migrated
+
+    # We keep this here as a fallback for backwards compatibility
+    # This may fail if user lacks permissions, which is fine if extension already exists
+    try:
+        op.execute("CREATE EXTENSION IF NOT EXISTS vector")
+    except Exception:
+        # Extension might already exist or user lacks permissions - verify it exists
+        conn = op.get_bind()
+        result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).fetchone()
+        if not result:
+            # Extension truly doesn't exist - re-raise the error
+            raise

    # Create banks table
    op.create_table(
@@ -0,0 +1,41 @@
+"""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)")
@@ -0,0 +1,134 @@
+"""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")
@@ -0,0 +1,113 @@
+"""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
@@ -0,0 +1,194 @@
+"""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
@@ -0,0 +1,50 @@
+"""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'))
+    """)
@@ -0,0 +1,47 @@
+"""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
+    """)
@@ -0,0 +1,53 @@
+"""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")
@@ -0,0 +1,134 @@
+"""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'
+    """)
@@ -0,0 +1,41 @@
+"""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")
@@ -0,0 +1,50 @@
+"""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")
@@ -0,0 +1,60 @@
+"""Fix mental_models primary key to be scoped per bank
+
+Revision ID: w8r9s0t1u2v3
+Revises: v7q8r9s0t1u2
+Create Date: 2026-02-05
+
+This migration fixes a critical bank isolation bug where mental_models.id was
+globally unique across all banks instead of being scoped per bank. This caused
+conflicts when different banks tried to use the same custom ID.
+
+CRITICAL FIX: Changes primary key from (id) to (bank_id, id) to ensure proper isolation.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "w8r9s0t1u2v3"
+down_revision: str | Sequence[str] | None = "v7q8r9s0t1u2"
+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 primary key from (id) to (bank_id, id) for proper bank isolation."""
+    schema = _get_schema_prefix()
+
+    # Drop the old primary key constraint (just id)
+    # Note: The constraint might be named differently on different DBs
+    # Try both old names (pinned_reflections_pkey from original, mental_models_pkey from rename)
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS pinned_reflections_pkey")
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS mental_models_pkey")
+
+    # Create the new composite primary key (bank_id, id)
+    # This ensures IDs are scoped per bank, not globally
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT mental_models_pkey PRIMARY KEY (bank_id, id)
+    """)
+
+
+def downgrade() -> None:
+    """Revert mental_models primary key from (bank_id, id) to (id)."""
+    schema = _get_schema_prefix()
+
+    # Drop the composite primary key
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS mental_models_pkey")
+
+    # Restore the old primary key (just id)
+    # WARNING: This downgrade will fail if there are duplicate IDs across banks
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT mental_models_pkey PRIMARY KEY (id)
+    """)
@@ -6,7 +6,6 @@ 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

@@ -46,14 +45,14 @@ def create_app(
        # Both HTTP and MCP
        app = create_app(memory, mcp_api_enabled=True)
    """
-    mcp_app = None
+    mcp_servers = None

-    # Create MCP app first if enabled (we need its lifespan for chaining)
+    # Create MCP servers first if enabled (we need their lifespans for chaining)
    if mcp_api_enabled:
        try:
-            from .mcp import create_mcp_app
+            from .mcp import MCPMiddleware, create_mcp_servers

-            mcp_app = create_mcp_app(memory=memory)
+            mcp_servers = create_mcp_servers(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]")
@@ -70,30 +69,41 @@ 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
+    # Add MCP middleware and chain its lifespan if enabled
+    if mcp_servers is not None:
+        multi_bank_server, single_bank_server, multi_bank_starlette_app, single_bank_starlette_app = mcp_servers

        # 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")
+            """Chain both MCP lifespans with the main app lifespan."""
+            # Start both MCP lifespans (multi-bank and single-bank)
+            async with multi_bank_starlette_app.router.lifespan_context(multi_bank_starlette_app):
+                async with single_bank_starlette_app.router.lifespan_context(single_bank_starlette_app):
+                    logger.info("MCP lifespans started (multi-bank and single-bank)")
+                    # Then start the original app lifespan
+                    async with original_lifespan(app_instance):
+                        yield
+                logger.info("MCP lifespans 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)
+        # Add MCP as a wrapping middleware — intercepts /mcp* requests directly,
+        # passes everything else through to the FastAPI app. No Starlette Mount
+        # means no 307 redirect for /mcp (no trailing slash).
+        app.add_middleware(
+            MCPMiddleware,
+            memory=memory,
+            prefix=mcp_mount_path,
+            multi_bank_app=multi_bank_starlette_app,
+            single_bank_app=single_bank_starlette_app,
+            multi_bank_server=multi_bank_server,
+            single_bank_server=single_bank_server,
+        )
+
        logger.info(f"MCP server enabled at {mcp_mount_path}/")

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

 import json
 import logging
@@ -8,7 +8,10 @@ from contextvars import ContextVar
 from fastmcp import FastMCP

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

 # Configure logging from HINDSIGHT_API_LOG_LEVEL environment variable
@@ -30,21 +33,49 @@ logger = logging.getLogger(__name__)
 # Default bank_id from environment variable
 DEFAULT_BANK_ID = os.environ.get("HINDSIGHT_MCP_BANK_ID", "default")

+# Legacy MCP authentication token (for backwards compatibility)
+# If set, this token is checked first before TenantExtension auth
+MCP_AUTH_TOKEN = os.environ.get("HINDSIGHT_API_MCP_AUTH_TOKEN")
+
 # Context variable to hold the current bank_id
 _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)
+
+# Context variables for tenant_id and api_key_id (set by authenticate, used by usage metering)
+_current_tenant_id: ContextVar[str | None] = ContextVar("current_tenant_id", default=None)
+_current_api_key_id: ContextVar[str | None] = ContextVar("current_api_key_id", default=None)
+

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


-def create_mcp_server(memory: MemoryEngine) -> FastMCP:
+def get_current_api_key() -> str | None:
+    """Get the current API key from context."""
+    return _current_api_key.get()
+
+
+def get_current_tenant_id() -> str | None:
+    """Get the current tenant_id from context."""
+    return _current_tenant_id.get()
+
+
+def get_current_api_key_id() -> str | None:
+    """Get the current api_key_id from context."""
+    return _current_api_key_id.get()
+
+
+def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
    """
    Create and configure the Hindsight MCP server.

    Args:
        memory: MemoryEngine instance (required)
+        multi_bank: If True, expose all tools with bank_id parameters (default).
+                   If False, only expose bank-scoped tools without bank_id parameters.

    Returns:
        Configured FastMCP server instance with stateless_http enabled
@@ -52,218 +83,102 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
    # Use stateless_http=True for Claude Code compatibility
    mcp = FastMCP("hindsight-mcp-server", stateless_http=True)

-    @mcp.tool()
-    async def retain(
-        content: str,
-        context: str = "general",
-        async_processing: bool = True,
-        bank_id: str | None = None,
-    ) -> str:
-        """
-        Store important information to long-term memory.
+    # Configure and register tools using shared module
+    config = MCPToolsConfig(
+        bank_id_resolver=get_current_bank_id,
+        api_key_resolver=get_current_api_key,  # Propagate API key for tenant auth
+        tenant_id_resolver=get_current_tenant_id,  # Propagate tenant_id for usage metering
+        api_key_id_resolver=get_current_api_key_id,  # Propagate api_key_id for usage metering
+        include_bank_id_param=multi_bank,
+        tools=None
+        if multi_bank
+        else {
+            "retain",
+            "recall",
+            "reflect",
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        },  # Scoped tools for single-bank mode (excludes bank management: list_banks, create_bank)
+        retain_fire_and_forget=False,  # HTTP MCP supports sync/async modes
+    )

-        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
+    register_mcp_tools(mcp, memory, config)

-        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'
-            async_processing: If True, queue for background processing and return immediately. If False, wait for completion. Default: True
-            bank_id: Optional bank to store in (defaults to session bank). Use for cross-bank operations.
-        """
-        try:
-            target_bank = bank_id or get_current_bank_id()
-            if target_bank is None:
-                return "Error: No bank_id configured"
-            contents = [{"content": content, "context": context}]
-            if async_processing:
-                # Queue for background processing and return immediately
-                result = await memory.submit_async_retain(
-                    bank_id=target_bank, contents=contents, request_context=RequestContext()
-                )
-                return f"Memory queued for background processing (operation_id: {result.get('operation_id', 'N/A')})"
-            else:
-                # Wait for completion
-                await memory.retain_batch_async(
-                    bank_id=target_bank,
-                    contents=contents,
-                    request_context=RequestContext(),
-                )
-                return f"Memory stored successfully in bank '{target_bank}'"
-        except Exception as e:
-            logger.error(f"Error storing memory: {e}", exc_info=True)
-            return f"Error: {str(e)}"
-
-    @mcp.tool()
-    async def recall(query: str, max_tokens: int = 4096, bank_id: str | None = None) -> 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_tokens: Maximum tokens in the response (default: 4096)
-            bank_id: Optional bank to search in (defaults to session bank). Use for cross-bank operations.
-        """
-        try:
-            target_bank = bank_id or get_current_bank_id()
-            if target_bank is None:
-                return "Error: No bank_id configured"
-            from hindsight_api.engine.memory_engine import Budget
-
-            recall_result = await memory.recall_async(
-                bank_id=target_bank,
-                query=query,
-                fact_type=list(VALID_RECALL_FACT_TYPES),
-                budget=Budget.HIGH,
-                max_tokens=max_tokens,
-                request_context=RequestContext(),
-            )
-
-            # Use model's JSON serialization
-            return recall_result.model_dump_json(indent=2)
-        except Exception as e:
-            logger.error(f"Error searching: {e}", exc_info=True)
-            return f'{{"error": "{e}", "results": []}}'
-
-    @mcp.tool()
-    async def reflect(query: str, context: str | None = None, budget: str = "low", bank_id: str | None = None) -> str:
-        """
-        Generate thoughtful analysis by synthesizing stored memories with the bank's personality.
-
-        WHEN TO USE THIS TOOL:
-        Use reflect when you need reasoned analysis, not just fact retrieval. This tool
-        thinks through the question using everything the bank knows and its personality traits.
-
-        EXAMPLES OF GOOD QUERIES:
-        - "What patterns have emerged in how I approach debugging?"
-        - "Based on my past decisions, what architectural style do I prefer?"
-        - "What might be the best approach for this problem given what you know about me?"
-        - "How should I prioritize these tasks based on my goals?"
-
-        HOW IT DIFFERS FROM RECALL:
-        - recall: Returns raw facts matching your search (fast lookup)
-        - reflect: Reasons across memories to form a synthesized answer (deeper analysis)
-
-        Use recall for "what did I say about X?" and reflect for "what should I do about X?"
-
-        Args:
-            query: The question or topic to reflect on
-            context: Optional context about why this reflection is needed
-            budget: Search budget - 'low', 'mid', or 'high' (default: 'low')
-            bank_id: Optional bank to reflect in (defaults to session bank). Use for cross-bank operations.
-        """
-        try:
-            target_bank = bank_id or get_current_bank_id()
-            if target_bank is None:
-                return "Error: No bank_id configured"
-            from hindsight_api.engine.memory_engine import Budget
-
-            # Map string budget to enum
-            budget_map = {"low": Budget.LOW, "mid": Budget.MID, "high": Budget.HIGH}
-            budget_enum = budget_map.get(budget.lower(), Budget.LOW)
-
-            reflect_result = await memory.reflect_async(
-                bank_id=target_bank,
-                query=query,
-                budget=budget_enum,
-                context=context,
-                request_context=RequestContext(),
-            )
-
-            return reflect_result.model_dump_json(indent=2)
-        except Exception as e:
-            logger.error(f"Error reflecting: {e}", exc_info=True)
-            return f'{{"error": "{e}", "text": ""}}'
-
-    @mcp.tool()
-    async def list_banks() -> str:
-        """
-        List all available memory banks.
-
-        Use this tool to discover what memory banks exist in the system.
-        Each bank is an isolated memory store (like a separate "brain").
-
-        Returns:
-            JSON list of banks with their IDs, names, dispositions, and missions.
-        """
-        try:
-            banks = await memory.list_banks(request_context=RequestContext())
-            return json.dumps({"banks": banks}, indent=2)
-        except Exception as e:
-            logger.error(f"Error listing banks: {e}", exc_info=True)
-            return f'{{"error": "{e}", "banks": []}}'
-
-    @mcp.tool()
-    async def create_bank(bank_id: str, name: str | None = None, mission: str | None = None) -> str:
-        """
-        Create a new memory bank or get an existing one.
-
-        Memory banks are isolated stores - each one is like a separate "brain" for a user/agent.
-        Banks are auto-created with default settings if they don't exist.
-
-        Args:
-            bank_id: Unique identifier for the bank (e.g., 'user-123', 'agent-alpha')
-            name: Optional human-friendly name for the bank
-            mission: Optional mission describing who the agent is and what they're trying to accomplish
-        """
-        try:
-            # get_bank_profile auto-creates bank if it doesn't exist
-            profile = await memory.get_bank_profile(bank_id, request_context=RequestContext())
-
-            # Update name/mission if provided
-            if name is not None or mission is not None:
-                await memory.update_bank(
-                    bank_id,
-                    name=name,
-                    mission=mission,
-                    request_context=RequestContext(),
-                )
-                # Fetch updated profile
-                profile = await memory.get_bank_profile(bank_id, request_context=RequestContext())
-
-            # Serialize disposition if it's a Pydantic model
-            if "disposition" in profile and hasattr(profile["disposition"], "model_dump"):
-                profile["disposition"] = profile["disposition"].model_dump()
-            return json.dumps(profile, indent=2)
-        except Exception as e:
-            logger.error(f"Error creating bank: {e}", exc_info=True)
-            return f'{{"error": "{e}"}}'
+    # Load and register additional tools from MCP extension if configured
+    mcp_extension = load_extension("MCP", MCPExtension)
+    if mcp_extension:
+        logger.info(f"Loading MCP extension: {mcp_extension.__class__.__name__}")
+        mcp_extension.register_tools(mcp, memory)

    return mcp


 class MCPMiddleware:
-    """ASGI middleware that extracts bank_id from header or path and sets context.
+    """ASGI middleware that intercepts MCP requests and routes to appropriate MCP server.

-    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)
+    This middleware wraps the main FastAPI app and intercepts requests matching the
+    configured prefix (default: /mcp). Non-MCP requests pass through to the inner app.

-    For Claude Code, configure with:
+    Authentication:
+        1. If HINDSIGHT_API_MCP_AUTH_TOKEN is set (legacy), validates against that token
+        2. Otherwise, uses TenantExtension.authenticate_mcp() from the MemoryEngine
+           - DefaultTenantExtension: no auth required (local dev)
+           - ApiKeyTenantExtension: validates against env var
+
+    Two modes based on URL structure:
+
+    1. Multi-bank mode (for /mcp/ root endpoint):
+       - Exposes all tools: retain, recall, reflect, list_banks, create_bank
+       - All tools include optional bank_id parameter for cross-bank operations
+       - Bank ID from: X-Bank-Id header or HINDSIGHT_MCP_BANK_ID env var
+
+    2. Single-bank mode (for /mcp/{bank_id}/ endpoints):
+       - Exposes bank-scoped tools only: retain, recall, reflect
+       - No bank_id parameter (comes from URL)
+       - No bank management tools (list_banks, create_bank)
+       - Recommended for agent isolation
+
+    Examples:
+        # Single-bank mode (recommended for agent isolation)
+        claude mcp add --transport http my-agent http://localhost:8888/mcp/my-agent-bank/ \\
+            --header "Authorization: Bearer <token>"
+
+        # Multi-bank mode (for cross-bank operations)
        claude mcp add --transport http hindsight http://localhost:8888/mcp \\
-            --header "X-Bank-Id: my-bank"
+            --header "X-Bank-Id: my-bank" --header "Authorization: Bearer <token>"
    """

-    def __init__(self, app, memory: MemoryEngine):
+    def __init__(
+        self,
+        app,
+        memory: MemoryEngine,
+        prefix: str = "/mcp",
+        multi_bank_app=None,
+        single_bank_app=None,
+        multi_bank_server=None,
+        single_bank_server=None,
+    ):
        self.app = app
+        self.prefix = prefix
        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
+        self.tenant_extension = memory._tenant_extension
+
+        if multi_bank_app and single_bank_app:
+            # Pre-created servers (used when called via add_middleware from create_app)
+            self.multi_bank_app = multi_bank_app
+            self.single_bank_app = single_bank_app
+            self.multi_bank_server = multi_bank_server
+            self.single_bank_server = single_bank_server
+        else:
+            # Create servers internally (for direct construction / tests)
+            self.multi_bank_server = create_mcp_server(memory, multi_bank=True)
+            self.multi_bank_app = self.multi_bank_server.http_app(path="/")
+            self.single_bank_server = create_mcp_server(memory, multi_bank=False)
+            self.single_bank_app = self.single_bank_server.http_app(path="/")

    def _get_header(self, scope: dict, name: str) -> str | None:
        """Extract a header value from ASGI scope."""
@@ -275,36 +190,70 @@ class MCPMiddleware:

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

        path = scope.get("path", "")

-        # Strip any mount prefix (e.g., /mcp) that FastAPI might not have stripped
-        root_path = scope.get("root_path", "")
-        if root_path and path.startswith(root_path):
-            path = path[len(root_path) :] or "/"
+        # Check if this is an MCP request (matches prefix)
+        if not (path == self.prefix or path.startswith(self.prefix + "/")):
+            # Not an MCP request — pass through to the inner app
+            await self.app(scope, receive, send)
+            return

-        # 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 = "/"
+        # Strip prefix from path
+        path = path[len(self.prefix) :] or "/"
+
+        # 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: check legacy MCP_AUTH_TOKEN first, then TenantExtension
+        tenant_context = None
+        auth_tenant_id: str | None = None
+        auth_api_key_id: str | None = None
+        if MCP_AUTH_TOKEN:
+            # Legacy authentication mode - validate against static 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
+            # Legacy mode doesn't use tenant schemas
+            tenant_context = None
+        else:
+            # Use TenantExtension.authenticate_mcp() for auth
+            try:
+                auth_context = RequestContext(api_key=auth_token)
+                tenant_context = await self.tenant_extension.authenticate_mcp(auth_context)
+                # Capture tenant_id and api_key_id set by authenticate() for usage metering
+                auth_tenant_id = auth_context.tenant_id
+                auth_api_key_id = auth_context.api_key_id
+            except AuthenticationError as e:
+                await self._send_error(send, 401, str(e))
+                return
+
+        # Set schema from tenant context so downstream DB queries use the correct schema
+        schema_token = (
+            _current_schema.set(tenant_context.schema_name) if tenant_context and tenant_context.schema_name else None
+        )

        # Try to get bank_id from header first (for Claude Code compatibility)
        bank_id = self._get_header(scope, "X-Bank-Id")
-
-        # MCP endpoint paths that should not be treated as bank_ids
-        MCP_ENDPOINTS = {"sse", "messages"}
+        bank_id_from_path = False

        # 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:
+            if parts[0]:
                # First segment looks like a bank_id
                bank_id = parts[0]
+                bank_id_from_path = True
                new_path = "/" + parts[1] if len(parts) > 1 else "/"

        # Fall back to default bank_id
@@ -312,17 +261,37 @@ class MCPMiddleware:
            bank_id = DEFAULT_BANK_ID
            logger.debug(f"Using default bank_id: {bank_id}")

-        # Set bank_id context
-        token = _current_bank_id.set(bank_id)
+        # Select the appropriate MCP app based on how bank_id was provided:
+        # - Path-based bank_id → single-bank app (no bank_id param, scoped tools)
+        # - Header/env bank_id → multi-bank app (bank_id param, all tools)
+        target_app = self.single_bank_app if bank_id_from_path else self.multi_bank_app
+
+        # Set bank_id, api_key, tenant_id, and api_key_id context
+        bank_id_token = _current_bank_id.set(bank_id)
+        # Store the auth token for tenant extension to validate
+        api_key_token = _current_api_key.set(auth_token) if auth_token else None
+        # Store tenant_id and api_key_id from authentication for usage metering
+        tenant_id_token = _current_tenant_id.set(auth_tenant_id) if auth_tenant_id else None
+        api_key_id_token = _current_api_key_id.set(auth_api_key_id) if auth_api_key_id else None
        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 if using path-based routing.
+            # Only rewrite SSE (text/event-stream) responses to avoid corrupting tool results
+            # that might contain the literal string "data: /messages".
+            is_sse_response = False
+
            async def send_wrapper(message):
-                if message["type"] == "http.response.body":
+                nonlocal is_sse_response
+                if message["type"] == "http.response.start":
+                    for header_name, header_value in message.get("headers", []):
+                        if header_name == b"content-type" and b"text/event-stream" in header_value:
+                            is_sse_response = True
+                            break
+                if message["type"] == "http.response.body" and bank_id_from_path and is_sse_response:
                    body = message.get("body", b"")
                    if body and b"/messages" in body:
                        # Rewrite /messages to /{bank_id}/messages in SSE endpoint event
@@ -330,9 +299,17 @@ class MCPMiddleware:
                        message = {**message, "body": body}
                await send(message)

-            await self.mcp_app(new_scope, receive, send_wrapper)
+            await target_app(new_scope, receive, send_wrapper)
        finally:
-            _current_bank_id.reset(token)
+            _current_bank_id.reset(bank_id_token)
+            if api_key_token is not None:
+                _current_api_key.reset(api_key_token)
+            if tenant_id_token is not None:
+                _current_tenant_id.reset(tenant_id_token)
+            if api_key_id_token is not None:
+                _current_api_key_id.reset(api_key_id_token)
+            if schema_token is not None:
+                _current_schema.reset(schema_token)

    async def _send_error(self, send, status: int, message: str):
        """Send an error response."""
@@ -352,19 +329,19 @@ class MCPMiddleware:
        )


-def create_mcp_app(memory: MemoryEngine):
-    """
-    Create an ASGI app that handles MCP requests.
+def create_mcp_servers(memory: MemoryEngine):
+    """Create multi-bank and single-bank MCP servers and their Starlette apps.

-    Bank ID can be provided via:
-    1. X-Bank-Id header: claude mcp add --transport http hindsight http://localhost:8888/mcp --header "X-Bank-Id: my-bank"
-    2. URL path: /mcp/{bank_id}/
-    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback, default: "default")
-
-    Args:
-        memory: MemoryEngine instance
+    Returns the servers and apps separately so lifespans can be chained before
+    the middleware wraps the main app.

    Returns:
-        ASGI application
+        Tuple of (multi_bank_server, single_bank_server, multi_bank_app, single_bank_app)
    """
-    return MCPMiddleware(None, memory)
+    multi_bank_server = create_mcp_server(memory, multi_bank=True)
+    multi_bank_app = multi_bank_server.http_app(path="/")
+
+    single_bank_server = create_mcp_server(memory, multi_bank=False)
+    single_bank_app = single_bank_server.http_app(path="/")
+
+    return multi_bank_server, single_bank_server, multi_bank_app, single_bank_app
@@ -4,6 +4,8 @@ Banner display for Hindsight API startup.
 Shows the logo and tagline with gradient colors.
 """

+from .utils import mask_network_location
+
 # Gradient colors: #0074d9 -> #009296
 GRADIENT_START = (0, 116, 217)  # #0074d9
 GRADIENT_END = (0, 146, 150)  # #009296
@@ -83,11 +85,14 @@ def print_startup_info(
    embeddings_provider: str,
    reranker_provider: str,
    mcp_enabled: bool = False,
+    version: str | None = None,
 ):
    """Print styled startup information."""
    print(color_start("Starting Hindsight API..."))
+    if version:
+        print(f"  {dim('Version:')} {color(f'v{version}', 0.1)}")
    print(f"  {dim('URL:')} {color(f'http://{host}:{port}', 0.2)}")
-    print(f"  {dim('Database:')} {color(database_url, 0.4)}")
+    print(f"  {dim('Database:')} {color(mask_network_location(database_url), 0.4)}")
    print(f"  {dim('LLM:')} {color(f'{llm_provider} / {llm_model}', 0.6)}")
    print(f"  {dim('Embeddings:')} {color(embeddings_provider, 0.8)}")
    print(f"  {dim('Reranker:')} {color(reranker_provider, 1.0)}")
@@ -20,11 +20,15 @@ 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"

@@ -33,34 +37,69 @@ 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_LOCAL_TRUST_REMOTE_CODE = "HINDSIGHT_API_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE"
 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"
+# Cohere configuration (separate for embeddings and reranker)
+ENV_EMBEDDINGS_COHERE_API_KEY = "HINDSIGHT_API_EMBEDDINGS_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_API_KEY = "HINDSIGHT_API_RERANKER_COHERE_API_KEY"
 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)
+# Deprecated: Legacy shared Cohere API key (for backward compatibility)
+ENV_COHERE_API_KEY = "HINDSIGHT_API_COHERE_API_KEY"
+
+# LiteLLM configuration (separate for embeddings and reranker)
+ENV_EMBEDDINGS_LITELLM_API_BASE = "HINDSIGHT_API_EMBEDDINGS_LITELLM_API_BASE"
+ENV_EMBEDDINGS_LITELLM_API_KEY = "HINDSIGHT_API_EMBEDDINGS_LITELLM_API_KEY"
+ENV_EMBEDDINGS_LITELLM_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_MODEL"
+ENV_RERANKER_LITELLM_API_BASE = "HINDSIGHT_API_RERANKER_LITELLM_API_BASE"
+ENV_RERANKER_LITELLM_API_KEY = "HINDSIGHT_API_RERANKER_LITELLM_API_KEY"
+ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"
+
+# Deprecated: Legacy shared LiteLLM config (for backward compatibility)
 ENV_LITELLM_API_BASE = "HINDSIGHT_API_LITELLM_API_BASE"
 ENV_LITELLM_API_KEY = "HINDSIGHT_API_LITELLM_API_KEY"
-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_LOCAL_TRUST_REMOTE_CODE = "HINDSIGHT_API_RERANKER_LOCAL_TRUST_REMOTE_CODE"
 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"
@@ -82,16 +121,29 @@ 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"

-# Observation thresholds
-ENV_OBSERVATION_MIN_FACTS = "HINDSIGHT_API_OBSERVATION_MIN_FACTS"
-ENV_OBSERVATION_TOP_ENTITIES = "HINDSIGHT_API_OBSERVATION_TOP_ENTITIES"
+# OpenTelemetry tracing configuration
+ENV_OTEL_TRACES_ENABLED = "HINDSIGHT_API_OTEL_TRACES_ENABLED"
+ENV_OTEL_EXPORTER_OTLP_ENDPOINT = "HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT"
+ENV_OTEL_EXPORTER_OTLP_HEADERS = "HINDSIGHT_API_OTEL_EXPORTER_OTLP_HEADERS"
+ENV_OTEL_SERVICE_NAME = "HINDSIGHT_API_OTEL_SERVICE_NAME"
+ENV_OTEL_DEPLOYMENT_ENVIRONMENT = "HINDSIGHT_API_OTEL_DEPLOYMENT_ENVIRONMENT"
+
+# 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_OBSERVATIONS_ASYNC = "HINDSIGHT_API_RETAIN_OBSERVATIONS_ASYNC"
+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"
@@ -111,27 +163,57 @@ 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"
+ENV_WORKER_MAX_SLOTS = "HINDSIGHT_API_WORKER_MAX_SLOTS"
+ENV_WORKER_CONSOLIDATION_MAX_SLOTS = "HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLOTS"

 # 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"
+
+# Provider-specific default models
+PROVIDER_DEFAULT_MODELS = {
+    "openai": "o3-mini",
+    "anthropic": "claude-haiku-4-5-20251001",
+    "gemini": "gemini-2.5-flash",
+    "groq": "openai/gpt-oss-120b",
+    "ollama": "gemma3:12b",
+    "lmstudio": "local-model",
+    "vertexai": "gemini-2.0-flash-001",
+    "openai-codex": "gpt-5.2-codex",
+    "claude-code": "claude-sonnet-4-5-20250929",
+    "mock": "mock-model",
+}
+DEFAULT_LLM_MODEL = "o3-mini"  # Fallback if provider not in table
 DEFAULT_LLM_MAX_CONCURRENT = 32
+DEFAULT_LLM_MAX_RETRIES = 10  # Max retry attempts for LLM API calls
+DEFAULT_LLM_INITIAL_BACKOFF = 1.0  # Initial backoff in seconds for retry exponential backoff
+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_LOCAL_TRUST_REMOTE_CODE = False  # Security: disabled by default, required for some models
 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_LOCAL_TRUST_REMOTE_CODE = (
+    False  # Security: disabled by default, required for some models like jina-reranker-v2
+)
 DEFAULT_RERANKER_TEI_BATCH_SIZE = 128
 DEFAULT_RERANKER_TEI_MAX_CONCURRENT = 8
 DEFAULT_RERANKER_MAX_CANDIDATES = 300
@@ -159,17 +241,18 @@ DEFAULT_RECALL_CONNECTION_BUDGET = 4  # Max concurrent DB connections per recall
 DEFAULT_MCP_LOCAL_BANK_ID = "mcp"
 DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY = 8  # Max concurrent mental model refreshes

-# Observation thresholds
-DEFAULT_OBSERVATION_MIN_FACTS = 5  # Min facts required to generate entity observations
-DEFAULT_OBSERVATION_TOP_ENTITIES = 5  # Max entities to process per retain batch
-
 # 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" or "verbose"
-RETAIN_EXTRACTION_MODES = ("concise", "verbose")  # Allowed extraction modes
-DEFAULT_RETAIN_OBSERVATIONS_ASYNC = False  # Run observation generation async (after retain completes)
+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
@@ -185,12 +268,18 @@ 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
+DEFAULT_WORKER_MAX_SLOTS = 10  # Total concurrent tasks per worker
+DEFAULT_WORKER_CONSOLIDATION_MAX_SLOTS = 2  # Max concurrent consolidation tasks per worker

 # Reflect agent settings
 DEFAULT_REFLECT_MAX_ITERATIONS = 10  # Max tool call iterations before forcing response

+# OpenTelemetry tracing configuration
+DEFAULT_OTEL_TRACES_ENABLED = False  # Disabled by default for backward compatibility
+DEFAULT_OTEL_SERVICE_NAME = "hindsight-api"
+DEFAULT_OTEL_DEPLOYMENT_ENVIRONMENT = "development"
+
 # Default MCP tool descriptions (can be customized via env vars)
 DEFAULT_MCP_RETAIN_DESCRIPTION = """Store important information to long-term memory.

@@ -257,12 +346,18 @@ def _validate_extraction_mode(mode: str) -> str:
    return mode_lower


+def _get_default_model_for_provider(provider: str) -> str:
+    """Get the default model for a given provider."""
+    return PROVIDER_DEFAULT_MODELS.get(provider.lower(), DEFAULT_LLM_MODEL)
+
+
@dataclass
 class HindsightConfig:
    """Configuration container for Hindsight API."""

    # Database
    database_url: str
+    database_schema: str

    # LLM (default, used as fallback for per-operation config)
    llm_provider: str
@@ -270,34 +365,77 @@ class HindsightConfig:
    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_local_trust_remote_code: bool
    embeddings_tei_url: str | None
    embeddings_openai_base_url: str | None
+    embeddings_cohere_api_key: str | None
+    embeddings_cohere_model: str
    embeddings_cohere_base_url: str | None
+    embeddings_litellm_api_base: str
+    embeddings_litellm_api_key: str | None
+    embeddings_litellm_model: str

    # Reranker
    reranker_provider: str
    reranker_local_model: str
+    reranker_local_force_cpu: bool
+    reranker_local_max_concurrent: int
+    reranker_local_trust_remote_code: bool
    reranker_tei_url: str | None
    reranker_tei_batch_size: int
    reranker_tei_max_concurrent: int
    reranker_max_candidates: int
+    reranker_cohere_api_key: str | None
+    reranker_cohere_model: str
    reranker_cohere_base_url: str | None
+    reranker_litellm_api_base: str
+    reranker_litellm_api_key: str | None
+    reranker_litellm_model: str

    # Server
    host: str
@@ -313,16 +451,17 @@ class HindsightConfig:
    recall_connection_budget: int
    mental_model_refresh_concurrency: int

-    # Observation thresholds
-    observation_min_facts: int
-    observation_top_entities: int
-
    # Retain settings
    retain_max_completion_tokens: int
    retain_chunk_size: int
    retain_extract_causal_links: bool
    retain_extraction_mode: str
-    retain_observations_async: bool
+    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
@@ -342,50 +481,184 @@ class HindsightConfig:
    worker_id: str | None
    worker_poll_interval_ms: int
    worker_max_retries: int
-    worker_batch_size: int
    worker_http_port: int
+    worker_max_slots: int
+    worker_consolidation_max_slots: int

    # Reflect agent settings
    reflect_max_iterations: int

+    # OpenTelemetry tracing configuration
+    otel_traces_enabled: bool
+    otel_exporter_otlp_endpoint: str | None
+    otel_exporter_otlp_headers: str | None
+    otel_service_name: str
+    otel_deployment_environment: str
+
+    def validate(self) -> None:
+        """Validate configuration values and raise errors for invalid combinations."""
+        # RETAIN_MAX_COMPLETION_TOKENS must be greater than RETAIN_CHUNK_SIZE
+        # to ensure the LLM has enough output capacity to extract facts from chunks
+        if self.retain_max_completion_tokens <= self.retain_chunk_size:
+            raise ValueError(
+                f"Invalid configuration: HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS "
+                f"({self.retain_max_completion_tokens}) must be greater than "
+                f"HINDSIGHT_API_RETAIN_CHUNK_SIZE ({self.retain_chunk_size}). "
+                f"\n\nYou have two options to fix this:"
+                f"\n  1. Increase HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS to a value > {self.retain_chunk_size}"
+                f"\n  2. Use a model that supports at least {self.retain_max_completion_tokens} output tokens"
+                f"\n     (current model: {self.retain_llm_model or self.llm_model}, "
+                f"provider: {self.retain_llm_provider or self.llm_provider})"
+            )
+
    @classmethod
    def from_env(cls) -> "HindsightConfig":
        """Create configuration from environment variables."""
-        return cls(
+        # Get provider first to determine default model
+        llm_provider = os.getenv(ENV_LLM_PROVIDER, DEFAULT_LLM_PROVIDER)
+        llm_model = os.getenv(ENV_LLM_MODEL) or _get_default_model_for_provider(llm_provider)
+
+        config = 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_provider=llm_provider,
            llm_api_key=os.getenv(ENV_LLM_API_KEY),
-            llm_model=os.getenv(ENV_LLM_MODEL, DEFAULT_LLM_MODEL),
+            llm_model=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_model=os.getenv(ENV_RETAIN_LLM_MODEL)
+            or (
+                _get_default_model_for_provider(os.getenv(ENV_RETAIN_LLM_PROVIDER))
+                if os.getenv(ENV_RETAIN_LLM_PROVIDER)
+                else 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_model=os.getenv(ENV_REFLECT_LLM_MODEL)
+            or (
+                _get_default_model_for_provider(os.getenv(ENV_REFLECT_LLM_PROVIDER))
+                if os.getenv(ENV_REFLECT_LLM_PROVIDER)
+                else 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 (
+                _get_default_model_for_provider(os.getenv(ENV_CONSOLIDATION_LLM_PROVIDER))
+                if os.getenv(ENV_CONSOLIDATION_LLM_PROVIDER)
+                else 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_local_trust_remote_code=os.getenv(
+                ENV_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE, str(DEFAULT_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE)
+            ).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,
+            # Cohere embeddings (with backward-compatible fallback to shared API key)
+            embeddings_cohere_api_key=os.getenv(ENV_EMBEDDINGS_COHERE_API_KEY) or os.getenv(ENV_COHERE_API_KEY),
+            embeddings_cohere_model=os.getenv(ENV_EMBEDDINGS_COHERE_MODEL, DEFAULT_EMBEDDINGS_COHERE_MODEL),
            embeddings_cohere_base_url=os.getenv(ENV_EMBEDDINGS_COHERE_BASE_URL) or None,
+            # LiteLLM embeddings (with backward-compatible fallback to shared config)
+            embeddings_litellm_api_base=os.getenv(ENV_EMBEDDINGS_LITELLM_API_BASE)
+            or os.getenv(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE),
+            embeddings_litellm_api_key=os.getenv(ENV_EMBEDDINGS_LITELLM_API_KEY) or os.getenv(ENV_LITELLM_API_KEY),
+            embeddings_litellm_model=os.getenv(ENV_EMBEDDINGS_LITELLM_MODEL, DEFAULT_EMBEDDINGS_LITELLM_MODEL),
            # 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_local_trust_remote_code=os.getenv(
+                ENV_RERANKER_LOCAL_TRUST_REMOTE_CODE, str(DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE)
+            ).lower()
+            in ("true", "1"),
            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))),
+            # Cohere reranker (with backward-compatible fallback to shared API key)
+            reranker_cohere_api_key=os.getenv(ENV_RERANKER_COHERE_API_KEY) or os.getenv(ENV_COHERE_API_KEY),
+            reranker_cohere_model=os.getenv(ENV_RERANKER_COHERE_MODEL, DEFAULT_RERANKER_COHERE_MODEL),
            reranker_cohere_base_url=os.getenv(ENV_RERANKER_COHERE_BASE_URL) or None,
+            # LiteLLM reranker (with backward-compatible fallback to shared config)
+            reranker_litellm_api_base=os.getenv(ENV_RERANKER_LITELLM_API_BASE)
+            or os.getenv(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE),
+            reranker_litellm_api_key=os.getenv(ENV_RERANKER_LITELLM_API_KEY) or os.getenv(ENV_LITELLM_API_KEY),
+            reranker_litellm_model=os.getenv(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL),
            # Server
            host=os.getenv(ENV_HOST, DEFAULT_HOST),
            port=int(os.getenv(ENV_PORT, DEFAULT_PORT)),
@@ -405,11 +678,6 @@ class HindsightConfig:
            # Optimization flags
            skip_llm_verification=os.getenv(ENV_SKIP_LLM_VERIFICATION, "false").lower() == "true",
            lazy_reranker=os.getenv(ENV_LAZY_RERANKER, "false").lower() == "true",
-            # Observation thresholds
-            observation_min_facts=int(os.getenv(ENV_OBSERVATION_MIN_FACTS, str(DEFAULT_OBSERVATION_MIN_FACTS))),
-            observation_top_entities=int(
-                os.getenv(ENV_OBSERVATION_TOP_ENTITIES, str(DEFAULT_OBSERVATION_TOP_ENTITIES))
-            ),
            # Retain settings
            retain_max_completion_tokens=int(
                os.getenv(ENV_RETAIN_MAX_COMPLETION_TOKENS, str(DEFAULT_RETAIN_MAX_COMPLETION_TOKENS))
@@ -422,10 +690,15 @@ class HindsightConfig:
            retain_extraction_mode=_validate_extraction_mode(
                os.getenv(ENV_RETAIN_EXTRACTION_MODE, DEFAULT_RETAIN_EXTRACTION_MODE)
            ),
-            retain_observations_async=os.getenv(
-                ENV_RETAIN_OBSERVATIONS_ASYNC, str(DEFAULT_RETAIN_OBSERVATIONS_ASYNC)
-            ).lower()
-            == "true",
+            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
@@ -438,11 +711,23 @@ class HindsightConfig:
            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))),
+            worker_max_slots=int(os.getenv(ENV_WORKER_MAX_SLOTS, str(DEFAULT_WORKER_MAX_SLOTS))),
+            worker_consolidation_max_slots=int(
+                os.getenv(ENV_WORKER_CONSOLIDATION_MAX_SLOTS, str(DEFAULT_WORKER_CONSOLIDATION_MAX_SLOTS))
+            ),
            # Reflect agent settings
            reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
+            # OpenTelemetry tracing configuration
+            otel_traces_enabled=os.getenv(ENV_OTEL_TRACES_ENABLED, str(DEFAULT_OTEL_TRACES_ENABLED)).lower()
+            in ("true", "1", "yes"),
+            otel_exporter_otlp_endpoint=os.getenv(ENV_OTEL_EXPORTER_OTLP_ENDPOINT) or None,
+            otel_exporter_otlp_headers=os.getenv(ENV_OTEL_EXPORTER_OTLP_HEADERS) or None,
+            otel_service_name=os.getenv(ENV_OTEL_SERVICE_NAME, DEFAULT_OTEL_SERVICE_NAME),
+            otel_deployment_environment=os.getenv(ENV_OTEL_DEPLOYMENT_ENVIRONMENT, DEFAULT_OTEL_DEPLOYMENT_ENVIRONMENT),
        )
+        config.validate()
+        return config

    def get_llm_base_url(self) -> str:
        """Get the LLM base URL, with provider-specific defaults."""
@@ -497,7 +782,7 @@ class HindsightConfig:

    def log_config(self) -> None:
        """Log the current configuration (without sensitive values)."""
-        logger.info(f"Database: {self.database_url}")
+        logger.info(f"Database: {self.database_url} (schema: {self.database_schema})")
        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
@@ -507,6 +792,10 @@ class HindsightConfig:
            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}")
@@ -1,11 +1,10 @@
 """
 Daemon mode support for Hindsight API.

-Provides idle timeout and lockfile management for running as a background daemon.
+Provides idle timeout for running as a background daemon.
 """

 import asyncio
-import fcntl
 import logging
 import os
 import sys
@@ -15,10 +14,11 @@ from pathlib import Path
 logger = logging.getLogger(__name__)

 # Default daemon configuration
-DEFAULT_DAEMON_PORT = 8889
+DEFAULT_DAEMON_PORT = 8888
 DEFAULT_IDLE_TIMEOUT = 0  # 0 = no auto-exit (hindsight-embed passes its own timeout)
-LOCKFILE_PATH = Path.home() / ".hindsight" / "daemon.lock"
-DAEMON_LOG_PATH = Path.home() / ".hindsight" / "daemon.log"
+
+# Allow override via environment variable for profile-specific logs
+DAEMON_LOG_PATH = Path(os.getenv("HINDSIGHT_API_DAEMON_LOG", str(Path.home() / ".hindsight" / "daemon.log")))


 class IdleTimeoutMiddleware:
@@ -52,82 +52,10 @@ 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)
-                os._exit(0)
+                # Send SIGTERM to ourselves to trigger graceful shutdown
+                import signal

-
-class DaemonLock:
-    """
-    File-based lock to prevent multiple daemon instances.
-
-    Uses fcntl.flock for atomic locking on Unix systems.
-    """
-
-    def __init__(self, lockfile: Path = LOCKFILE_PATH):
-        self.lockfile = lockfile
-        self._fd = None
-
-    def acquire(self) -> bool:
-        """
-        Try to acquire the daemon lock.
-
-        Returns True if lock acquired, False if another daemon is running.
-        """
-        self.lockfile.parent.mkdir(parents=True, exist_ok=True)
-
-        try:
-            self._fd = open(self.lockfile, "w")
-            fcntl.flock(self._fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
-            # Write PID for debugging
-            self._fd.write(str(os.getpid()))
-            self._fd.flush()
-            return True
-        except (IOError, OSError):
-            # Lock is held by another process
-            if self._fd:
-                self._fd.close()
-                self._fd = None
-            return False
-
-    def release(self):
-        """Release the daemon lock."""
-        if self._fd:
-            try:
-                fcntl.flock(self._fd.fileno(), fcntl.LOCK_UN)
-                self._fd.close()
-            except Exception:
-                pass
-            finally:
-                self._fd = None
-                # Remove lockfile
-                try:
-                    self.lockfile.unlink()
-                except Exception:
-                    pass
-
-    def is_locked(self) -> bool:
-        """Check if the lock is held by another process."""
-        if not self.lockfile.exists():
-            return False
-
-        try:
-            fd = open(self.lockfile, "r")
-            fcntl.flock(fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
-            # We got the lock, so no one else has it
-            fcntl.flock(fd.fileno(), fcntl.LOCK_UN)
-            fd.close()
-            return False
-        except (IOError, OSError):
-            return True
-
-    def get_pid(self) -> int | None:
-        """Get the PID of the daemon holding the lock."""
-        if not self.lockfile.exists():
-            return None
-        try:
-            with open(self.lockfile, "r") as f:
-                return int(f.read().strip())
-        except (ValueError, IOError):
-            return None
+                os.kill(os.getpid(), signal.SIGTERM)


 def daemonize():
@@ -136,16 +64,21 @@ def daemonize():

    Uses double-fork technique to properly detach from terminal.
    """
-    # First fork
-    pid = os.fork()
-    if pid > 0:
-        # Parent exits
-        sys.exit(0)
+    # First fork - detach from parent
+    try:
+        pid = os.fork()
+        if pid > 0:
+            sys.exit(0)
+    except OSError as e:
+        sys.stderr.write(f"fork #1 failed: {e}\n")
+        sys.exit(1)

-    # Create new session
+    # Decouple from parent environment
+    os.chdir("/")
    os.setsid()
+    os.umask(0)

-    # Second fork to prevent zombie processes
+    # Second fork - prevent zombie
    pid = os.fork()
    if pid > 0:
        sys.exit(0)
@@ -178,27 +111,3 @@ def check_daemon_running(port: int = DEFAULT_DAEMON_PORT) -> bool:
        return result == 0
    except Exception:
        return False
-
-
-def stop_daemon(port: int = DEFAULT_DAEMON_PORT) -> bool:
-    """Stop a running daemon by sending SIGTERM to the process."""
-    lock = DaemonLock()
-    pid = lock.get_pid()
-
-    if pid is None:
-        return False
-
-    try:
-        import signal
-
-        os.kill(pid, signal.SIGTERM)
-        # Wait for process to exit
-        for _ in range(50):  # Wait up to 5 seconds
-            time.sleep(0.1)
-            try:
-                os.kill(pid, 0)  # Check if process exists
-            except OSError:
-                return True  # Process exited
-        return False
-    except OSError:
-        return False
@@ -0,0 +1,5 @@
+"""Consolidation engine for automatic learning creation from memories."""
+
+from .consolidator import run_consolidation_job
+
+__all__ = ["run_consolidation_job"]
@@ -0,0 +1,85 @@
+"""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 code blocks or additional text. However, the "text" field within each observation should use markdown formatting (headers, lists, bold, etc.) for clarity and readability.
+
+## 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 temporal markers showing change
+   Example: "Alex used to love pizza but now hates it" OR "Alex's pizza preference changed from love to hate"
+3. UPDATE: New state replacing old state → update showing the transition with "used to", "now", "changed from X to Y"
+
+## CRITICAL RULES:
+- NEVER merge facts about DIFFERENT people
+- NEVER merge unrelated topics (food preferences vs work vs hobbies)
+- When merging contradictions, the "text" field MUST capture BOTH states with temporal markers:
+  * Use "used to X, now Y" OR "changed from X to Y" OR "X but now Y"
+  * DO NOT just state the new fact - you MUST show the change
+- Keep observations focused on ONE specific topic per person
+- The "text" field MUST contain durable knowledge, not ephemeral state
+- Do NOT include "tags" in output - tags are handled automatically"""
+
+CONSOLIDATION_USER_PROMPT = """Analyze this new fact and consolidate into knowledge.
+{mission_section}
+NEW FACT: {fact_text}
+
+EXISTING OBSERVATIONS (JSON array with source memories and dates):
+{observations_text}
+
+Each observation includes:
+- id: unique identifier for updating
+- text: the observation content
+- proof_count: number of supporting memories
+- tags: visibility scope (handled automatically)
+- created_at/updated_at: when observation was created/modified
+- occurred_start/occurred_end: temporal range of source facts
+- source_memories: array of supporting facts with their text and dates
+
+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 (the "text" field should use markdown formatting for structure):
+[
+  {{"action": "update", "learning_id": "uuid-from-observations", "text": "## Updated Knowledge\n\n**Key point**: details here\n\n- Supporting detail 1\n- Supporting detail 2", "reason": "..."}},
+  {{"action": "create", "text": "## New Durable Knowledge\n\nDescription with **emphasis** and proper structure", "reason": "..."}}
+]
+
+Return [] if fact contains no durable knowledge.
+
+IMPORTANT: Format the "text" field with markdown for better readability:
+- Use headers, lists, bold/italic, tables where appropriate
+- CRITICAL: Add blank lines before and after block elements (tables, code blocks, lists)
+- Ensure proper spacing for markdown to render correctly"""
@@ -9,6 +9,7 @@ Configuration via environment variables - see hindsight_api.config for all env v
 import asyncio
 import logging
 import os
+import warnings
 from abc import ABC, abstractmethod
 from concurrent.futures import ThreadPoolExecutor

@@ -20,21 +21,21 @@ from ..config import (
    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_LOCAL_TRUST_REMOTE_CODE,
    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_API_KEY,
    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_LOCAL_TRUST_REMOTE_CODE,
    ENV_RERANKER_PROVIDER,
    ENV_RERANKER_TEI_BATCH_SIZE,
    ENV_RERANKER_TEI_MAX_CONCURRENT,
@@ -99,7 +100,13 @@ class LocalSTCrossEncoder(CrossEncoderModel):
    _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):
+    def __init__(
+        self,
+        model_name: str | None = None,
+        max_concurrent: int = 4,
+        force_cpu: bool = False,
+        trust_remote_code: bool = False,
+    ):
        """
        Initialize local SentenceTransformers cross-encoder.

@@ -108,8 +115,15 @@ class LocalSTCrossEncoder(CrossEncoderModel):
                       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
+            trust_remote_code: Allow loading models with custom code (security risk).
+                              Required for some models like jina-reranker-v2-base-multilingual.
+                              Default: False (disabled for security)
        """
        self.model_name = model_name or DEFAULT_RERANKER_LOCAL_MODEL
+        self.force_cpu = force_cpu
+        self.trust_remote_code = trust_remote_code
        self._model = None
        LocalSTCrossEncoder._max_concurrent = max_concurrent

@@ -130,18 +144,56 @@ class LocalSTCrossEncoder(CrossEncoderModel):
                "Install it with: pip install sentence-transformers"
            )

-        # Note: We use CPU even when GPU/MPS is available because:
-        # 1. The reranker model (MiniLM) is tiny (~22M params)
-        # 2. Batch sizes are small (~100-200 pairs)
-        # 3. Data transfer overhead to GPU outweighs compute benefit
-        # 4. CPU inference is actually faster for this workload
        logger.info(f"Reranker: initializing local provider with model {self.model_name}")
-        # Disable lazy loading (meta tensors) which causes issues with newer transformers/accelerate.
-        # Setting low_cpu_mem_usage=False and device_map=None ensures tensors are fully materialized.
-        self._model = CrossEncoder(
-            self.model_name,
-            model_kwargs={"low_cpu_mem_usage": False, "device_map": None},
-        )
+
+        # 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}")
+
+        # Suppress verbose transformers warnings during model loading
+        # This suppresses the "UNEXPECTED" warnings from CrossEncoder which are harmless
+        # but look alarming to users (e.g., "embeddings.position_ids | UNEXPECTED")
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=UserWarning)
+            warnings.filterwarnings("ignore", message=".*was not found in model state dict.*")
+            warnings.filterwarnings("ignore", message=".*UNEXPECTED.*")
+
+            # Also suppress transformers library logging temporarily
+            transformers_logger = logging.getLogger("transformers")
+            original_level = transformers_logger.level
+            transformers_logger.setLevel(logging.ERROR)
+
+            try:
+                self._model = CrossEncoder(
+                    self.model_name,
+                    device=device,
+                    model_kwargs={"low_cpu_mem_usage": False},
+                    trust_remote_code=self.trust_remote_code,
+                )
+            finally:
+                # Restore original logging level
+                transformers_logger.setLevel(original_level)

        # Initialize shared executor (limited workers naturally limits concurrency)
        if LocalSTCrossEncoder._executor is None:
@@ -153,6 +205,11 @@ class LocalSTCrossEncoder(CrossEncoderModel):
        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]:
        """
        Score query-document pairs for relevance.
@@ -170,11 +227,11 @@ class LocalSTCrossEncoder(CrossEncoderModel):

        # Use dedicated executor - limited workers naturally limits concurrency
        loop = asyncio.get_event_loop()
-        scores = await loop.run_in_executor(
+        return await loop.run_in_executor(
            LocalSTCrossEncoder._executor,
-            lambda: self._model.predict(pairs, show_progress_bar=False),
+            self._predict_sync,
+            pairs,
        )
-        return scores.tolist() if hasattr(scores, "tolist") else list(scores)


 class RemoteTEICrossEncoder(CrossEncoderModel):
@@ -584,7 +641,7 @@ class FlashRankCrossEncoder(CrossEncoderModel):
            return

        try:
-            from flashrank import Ranker  # type: ignore[import-untyped]
+            from flashrank import Ranker
        except ImportError:
            raise ImportError("flashrank is required for FlashRankCrossEncoder. Install it with: pip install flashrank")

@@ -611,7 +668,7 @@ class FlashRankCrossEncoder(CrossEncoderModel):

    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
        """Synchronous predict - processes each query group."""
-        from flashrank import RerankRequest  # type: ignore[import-untyped]
+        from flashrank import RerankRequest

        if not pairs:
            return []
@@ -773,45 +830,53 @@ class LiteLLMCrossEncoder(CrossEncoderModel):

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

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

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

    if provider == "tei":
-        url = os.environ.get(ENV_RERANKER_TEI_URL)
+        url = config.reranker_tei_url
        if not url:
            raise ValueError(f"{ENV_RERANKER_TEI_URL} is required when {ENV_RERANKER_PROVIDER} is 'tei'")
-        batch_size = int(os.environ.get(ENV_RERANKER_TEI_BATCH_SIZE, str(DEFAULT_RERANKER_TEI_BATCH_SIZE)))
-        max_concurrent = int(os.environ.get(ENV_RERANKER_TEI_MAX_CONCURRENT, str(DEFAULT_RERANKER_TEI_MAX_CONCURRENT)))
-        return RemoteTEICrossEncoder(base_url=url, batch_size=batch_size, max_concurrent=max_concurrent)
+        return RemoteTEICrossEncoder(
+            base_url=url,
+            batch_size=config.reranker_tei_batch_size,
+            max_concurrent=config.reranker_tei_max_concurrent,
+        )
    elif provider == "local":
-        model = os.environ.get(ENV_RERANKER_LOCAL_MODEL)
-        model_name = model or DEFAULT_RERANKER_LOCAL_MODEL
-        max_concurrent = int(
-            os.environ.get(ENV_RERANKER_LOCAL_MAX_CONCURRENT, str(DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT))
+        return LocalSTCrossEncoder(
+            model_name=config.reranker_local_model,
+            max_concurrent=config.reranker_local_max_concurrent,
+            force_cpu=config.reranker_local_force_cpu,
+            trust_remote_code=config.reranker_local_trust_remote_code,
        )
-        return LocalSTCrossEncoder(model_name=model_name, max_concurrent=max_concurrent)
    elif provider == "cohere":
-        api_key = os.environ.get(ENV_COHERE_API_KEY)
+        api_key = config.reranker_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)
+            raise ValueError(f"{ENV_RERANKER_COHERE_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'cohere'")
+        return CohereCrossEncoder(
+            api_key=api_key,
+            model=config.reranker_cohere_model,
+            base_url=config.reranker_cohere_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)
+        return LiteLLMCrossEncoder(
+            api_base=config.reranker_litellm_api_base,
+            api_key=config.reranker_litellm_api_key,
+            model=config.reranker_litellm_model,
+        )
    elif provider == "rrf":
        return RRFPassthroughCrossEncoder()
    else:
@@ -0,0 +1,5 @@
+"""Directives module for hard rules injected into prompts."""
+
+from .models import Directive
+
+__all__ = ["Directive"]
@@ -0,0 +1,37 @@
+"""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
@@ -11,6 +11,7 @@ Configuration via environment variables - see hindsight_api.config for all env v

 import logging
 import os
+import warnings
 from abc import ABC, abstractmethod

 import httpx
@@ -18,22 +19,21 @@ import httpx
 from ..config import (
    DEFAULT_EMBEDDINGS_COHERE_MODEL,
    DEFAULT_EMBEDDINGS_LITELLM_MODEL,
+    DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU,
    DEFAULT_EMBEDDINGS_LOCAL_MODEL,
+    DEFAULT_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
    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_COHERE_API_KEY,
+    ENV_EMBEDDINGS_LOCAL_FORCE_CPU,
    ENV_EMBEDDINGS_LOCAL_MODEL,
+    ENV_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
    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,
 )

@@ -92,15 +92,22 @@ class LocalSTEmbeddings(Embeddings):
    The embedding dimension is auto-detected from the model.
    """

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

        Args:
            model_name: Name of the SentenceTransformer model to use.
                       Default: BAAI/bge-small-en-v1.5
+            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
+                      Default: False
+            trust_remote_code: Allow loading models with custom code (security risk).
+                              Required for some models with custom architectures.
+                              Default: False (disabled for security)
        """
        self.model_name = model_name or DEFAULT_EMBEDDINGS_LOCAL_MODEL
+        self.force_cpu = force_cpu
+        self.trust_remote_code = trust_remote_code
        self._model = None
        self._dimension: int | None = None

@@ -128,12 +135,53 @@ class LocalSTEmbeddings(Embeddings):
            )

        logger.info(f"Embeddings: initializing local provider with model {self.model_name}")
-        # Disable lazy loading (meta tensors) which causes issues with newer transformers/accelerate
-        # Setting low_cpu_mem_usage=False and device_map=None ensures tensors are fully materialized
-        self._model = SentenceTransformer(
-            self.model_name,
-            model_kwargs={"low_cpu_mem_usage": False, "device_map": None},
-        )
+
+        # 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}")
+
+        # Suppress verbose transformers warnings during model loading
+        # This suppresses the "UNEXPECTED" warnings from BertModel which are harmless
+        # but look alarming to users (e.g., "embeddings.position_ids | UNEXPECTED")
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=UserWarning)
+            warnings.filterwarnings("ignore", message=".*was not found in model state dict.*")
+            warnings.filterwarnings("ignore", message=".*UNEXPECTED.*")
+
+            # Also suppress transformers library logging temporarily
+            transformers_logger = logging.getLogger("transformers")
+            original_level = transformers_logger.level
+            transformers_logger.setLevel(logging.ERROR)
+
+            try:
+                self._model = SentenceTransformer(
+                    self.model_name,
+                    device=device,
+                    model_kwargs={"low_cpu_mem_usage": False},
+                    trust_remote_code=self.trust_remote_code,
+                )
+            finally:
+                # Restore original logging level
+                transformers_logger.setLevel(original_level)

        self._dimension = self._model.get_sentence_embedding_dimension()
        logger.info(f"Embeddings: local provider initialized (dim: {self._dimension})")
@@ -150,6 +198,7 @@ class LocalSTEmbeddings(Embeddings):
        """
        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]

@@ -516,7 +565,7 @@ class CohereEmbeddings(Embeddings):
                model=self.model,
                input_type=self.input_type,
            )
-            if response.embeddings:
+            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})")
@@ -673,24 +722,29 @@ class LiteLLMEmbeddings(Embeddings):

 def create_embeddings_from_env() -> Embeddings:
    """
-    Create an Embeddings instance based on environment variables.
+    Create an Embeddings instance based on configuration.

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

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

    if provider == "tei":
-        url = os.environ.get(ENV_EMBEDDINGS_TEI_URL)
+        url = config.embeddings_tei_url
        if not url:
            raise ValueError(f"{ENV_EMBEDDINGS_TEI_URL} is required when {ENV_EMBEDDINGS_PROVIDER} is 'tei'")
        return RemoteTEIEmbeddings(base_url=url)
    elif provider == "local":
-        model = os.environ.get(ENV_EMBEDDINGS_LOCAL_MODEL)
-        model_name = model or DEFAULT_EMBEDDINGS_LOCAL_MODEL
-        return LocalSTEmbeddings(model_name=model_name)
+        return LocalSTEmbeddings(
+            model_name=config.embeddings_local_model,
+            force_cpu=config.embeddings_local_force_cpu,
+            trust_remote_code=config.embeddings_local_trust_remote_code,
+        )
    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)
@@ -703,17 +757,20 @@ def create_embeddings_from_env() -> Embeddings:
        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)
+        api_key = config.embeddings_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)
+            raise ValueError(f"{ENV_EMBEDDINGS_COHERE_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'cohere'")
+        return CohereEmbeddings(
+            api_key=api_key,
+            model=config.embeddings_cohere_model,
+            base_url=config.embeddings_cohere_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)
+        return LiteLLMEmbeddings(
+            api_base=config.embeddings_litellm_api_base,
+            api_key=config.embeddings_litellm_api_key,
+            model=config.embeddings_litellm_model,
+        )
    else:
        raise ValueError(
            f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai', 'cohere', 'litellm'"
@@ -442,49 +442,6 @@ class MemoryEngineInterface(ABC):
        """
        ...

-    @abstractmethod
-    async def get_entity_observations(
-        self,
-        bank_id: str,
-        entity_id: str,
-        *,
-        limit: int = 10,
-        request_context: "RequestContext",
-    ) -> list[Any]:
-        """
-        Get observations for an entity.
-
-        Args:
-            bank_id: The memory bank ID.
-            entity_id: The entity ID.
-            limit: Maximum observations.
-            request_context: Request context for authentication.
-
-        Returns:
-            List of EntityObservation objects.
-        """
-        ...
-
-    @abstractmethod
-    async def regenerate_entity_observations(
-        self,
-        bank_id: str,
-        entity_id: str,
-        entity_name: str,
-        *,
-        request_context: "RequestContext",
-    ) -> None:
-        """
-        Regenerate observations for an entity.
-
-        Args:
-            bank_id: The memory bank ID.
-            entity_id: The entity ID.
-            entity_name: The entity's canonical name.
-            request_context: Request context for authentication.
-        """
-        ...
-
    # =========================================================================
    # Statistics & Operations
    # =========================================================================
@@ -0,0 +1,146 @@
+"""
+Abstract interface for LLM providers.
+
+This module defines the interface that all LLM providers must implement,
+enabling support for multiple LLM backends (OpenAI, Anthropic, Gemini, Codex, etc.)
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+from .response_models import LLMToolCallResult, TokenUsage
+
+
+class LLMInterface(ABC):
+    """
+    Abstract interface for LLM providers.
+
+    All LLM provider implementations must inherit from this class and implement
+    the required methods.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """
+        Initialize LLM provider.
+
+        Args:
+            provider: Provider name (e.g., "openai", "codex", "anthropic", "gemini").
+            api_key: API key or authentication token.
+            base_url: Base URL for the API.
+            model: Model name.
+            reasoning_effort: Reasoning effort level for supported providers.
+            **kwargs: Additional provider-specific parameters.
+        """
+        self.provider = provider.lower()
+        self.api_key = api_key
+        self.base_url = base_url
+        self.model = model
+        self.reasoning_effort = reasoning_effort
+
+    @abstractmethod
+    async def verify_connection(self) -> None:
+        """
+        Verify that the LLM provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        pass
+
+    @abstractmethod
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (OpenAI only).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        pass
+
+    @abstractmethod
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        pass
+
+    @abstractmethod
+    async def cleanup(self) -> None:
+        """Clean up resources (close connections, etc.)."""
+        pass
+
+
+class OutputTooLongError(Exception):
+    """
+    Bridge exception raised when LLM output exceeds token limits.
+
+    This wraps provider-specific errors (e.g., OpenAI's LengthFinishReasonError)
+    to allow callers to handle output length issues without depending on
+    provider-specific implementations.
+    """
+
+    pass
@@ -1,16 +1,12 @@
 """
 Mental models module for Hindsight.

-Mental models are synthesized summaries that represent understanding. They come
-in different subtypes based on how they were created:
+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).

- Structural: Derived from the bank's mission (e.g., "Be a PM for engineering team")
-  These are created upfront based on what any agent with this role would need.
-
- Emergent: Discovered from data patterns (named entities, temporal clusters, etc.)
-  These surface organically as facts are retained.
-
- Pinned: User-defined models that persist across refreshes.
+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
@@ -1,311 +0,0 @@
-"""
-Emergent mental model detection and promotion.
-
-Emergent models are discovered from data patterns:
- Named entity extraction (people, projects, systems)
- Temporal clustering (events with multiple references)
- Causal patterns ("Because X, we do Y")
- Behavioral anchors ("After X, we started Y")
- Reference frequency (anything mentioned repeatedly)
-
-When a pattern is detected, it goes through a mission filter to check relevance,
-and if relevant, is promoted to a mental model.
-"""
-
-import logging
-from typing import TYPE_CHECKING
-
-from pydantic import BaseModel, Field
-
-from .models import EmergentCandidate
-
-if TYPE_CHECKING:
-    from ..llm_wrapper import LLMConfig
-
-logger = logging.getLogger(__name__)
-
-
-class MissionFilterCandidate(BaseModel):
-    """Result of mission filtering for a single candidate."""
-
-    name: str
-    promote: bool = Field(description="True if this is a specific named entity worth tracking")
-    reason: str = Field(description="Brief explanation for the decision")
-
-
-class MissionFilterResponse(BaseModel):
-    """Response from LLM for mission filtering."""
-
-    candidates: list[MissionFilterCandidate] = Field(description="Filtering decision for each candidate")
-
-
-def build_mission_filter_prompt(mission: str, candidates: list[EmergentCandidate]) -> str:
-    """Build the prompt for filtering candidates by mission relevance."""
-    candidate_list = "\n".join(
-        [f"- {c.name} (mentions: {c.mention_count}, method: {c.detection_method})" for c in candidates]
-    )
-
-    return f"""Filter these detected entities. For each one, decide: promote=true or promote=false.
-
-MISSION: {mission}
-
-DETECTED ENTITIES:
-{candidate_list}
-
-=== DECISION RULES ===
-
-Set promote=true ONLY for specific, named entities:
- Person names: "John", "Maria", "Alice Chen", "Dr. Smith"
- Named organizations: "Google", "Acme Corp", "Frontend Team"
- Named places: "Central Park Zoo", "NYC Office", "Building A"
- Named projects: "Project Phoenix", "Auth Service v2"
-
-Set promote=false for EVERYTHING ELSE, including:
- Common English words: user, support, help, family, kids, parents, friends, people, team, photo, nature, park, office, home, work, school, joy, love, hope, fear, anger, gratitude, kindness, passion, motivation, inspiration, encouragement, positivity, energy, community, connection, commitment, collaboration, growth, impact, difference, success, progress, change, education, volunteering, veterans, homeless, shelter, meeting, project, system, process, event
- Generic categories (even capitalized): Users, Customers, Team, Family, Kids, Veterans, Community
- Abstract concepts: motivation, inspiration, gratitude, commitment, resilience
-
-THE TEST: Is this a specific name you'd find in a contact list or org chart?
- "John" → YES (promote=true)
- "kids" → NO (promote=false)
- "community" → NO (promote=false)
- "Maria" → YES (promote=true)
- "park" → NO (promote=false)
-
-When in doubt, set promote=false."""
-
-
-def get_mission_filter_system_message() -> str:
-    """System message for mission filtering."""
-    return """You filter entities for promotion. Output JSON with 'candidates' array.
-
-Rules:
- promote=true ONLY for specific names (people, organizations, named places/projects)
- promote=false for common words, generic categories, abstract concepts
-
-Examples:
- "John" → promote=true (person name)
- "kids" → promote=false (generic category)
- "community" → promote=false (abstract concept)
- "Google" → promote=true (organization name)
- "motivation" → promote=false (abstract concept)
-
-When in doubt, promote=false. Most entities should be rejected."""
-
-
-async def filter_candidates_by_mission(
-    llm_config: "LLMConfig",
-    mission: str,
-    candidates: list[EmergentCandidate],
-) -> list[EmergentCandidate]:
-    """
-    Filter emergent candidates to keep only specific, named entities.
-
-    Args:
-        llm_config: LLM configuration
-        mission: The bank's mission (used for context)
-        candidates: List of detected candidates
-
-    Returns:
-        Filtered list of candidates that are specific named entities
-    """
-    if not candidates:
-        return []
-
-    if not mission:
-        # No mission = no filtering, keep all candidates
-        logger.debug("[EMERGENT] No mission set, skipping filter")
-        return candidates
-
-    prompt = build_mission_filter_prompt(mission, candidates)
-
-    try:
-        result = await llm_config.call(
-            messages=[
-                {"role": "system", "content": get_mission_filter_system_message()},
-                {"role": "user", "content": prompt},
-            ],
-            response_format=MissionFilterResponse,
-            scope="mental_model_mission_filter",
-        )
-
-        # Build name -> promote map
-        promote_map = {c.name: c.promote for c in result.candidates}
-
-        # Filter candidates
-        filtered = []
-        for candidate in candidates:
-            if candidate.name in promote_map:
-                if promote_map[candidate.name]:
-                    filtered.append(candidate)
-                    logger.debug(f"[EMERGENT] Promoting '{candidate.name}'")
-                else:
-                    logger.debug(f"[EMERGENT] Rejecting '{candidate.name}'")
-            else:
-                # Candidate not in response - reject by default
-                logger.debug(f"[EMERGENT] '{candidate.name}' not in response, rejecting")
-
-        logger.info(f"[EMERGENT] Mission filter: {len(filtered)}/{len(candidates)} candidates promoted")
-        return filtered
-
-    except Exception as e:
-        logger.warning(f"[EMERGENT] Mission filter failed, rejecting all candidates: {e}")
-        return []
-
-
-async def evaluate_emergent_models(
-    llm_config: "LLMConfig",
-    models: list[dict],
-) -> list[str]:
-    """
-    Evaluate existing emergent models to check if they should be kept.
-
-    This re-evaluates emergent models using the same filtering criteria
-    as new candidates. Models that are generic/abstract will be removed.
-
-    Args:
-        llm_config: LLM configuration
-        models: List of existing emergent model dicts with 'name', 'id'
-
-    Returns:
-        List of model IDs that should be REMOVED (no longer valid)
-    """
-    if not models:
-        return []
-
-    # Convert existing models to candidates for evaluation
-    candidates = [
-        EmergentCandidate(
-            name=m["name"],
-            detection_method="existing_emergent_model",
-            mention_count=0,
-        )
-        for m in models
-    ]
-
-    # Build a simple prompt for re-evaluation
-    names_list = "\n".join([f"- {m['name']}" for m in models])
-    prompt = f"""Re-evaluate these existing mental models. For each one, decide: promote=true (keep) or promote=false (remove).
-
-EXISTING MODELS:
-{names_list}
-
-=== DECISION RULES ===
-
-Set promote=true ONLY for specific, named entities:
- Person names: "John", "Maria", "Alice Chen", "Dr. Smith"
- Named organizations: "Google", "Acme Corp", "Frontend Team"
- Named places: "Central Park Zoo", "NYC Office", "Building A"
- Named projects: "Project Phoenix", "Auth Service v2"
-
-Set promote=false for EVERYTHING ELSE, including:
- Common English words: user, support, help, family, kids, parents, friends, people, team, photo, nature, park, office, home, work, school, joy, love, hope, fear, anger, gratitude, kindness, passion, motivation, inspiration, encouragement, positivity, energy, community, connection, commitment, collaboration, growth, impact, difference, success, progress, change, education, volunteering, veterans, homeless, shelter, meeting, project, system, process, event
- Generic categories (even capitalized): Users, Customers, Team, Family, Kids, Veterans, Community
- Abstract concepts: motivation, inspiration, gratitude, commitment, resilience
-
-THE TEST: Is this a specific name you'd find in a contact list or org chart?
- "John" → YES (promote=true)
- "kids" → NO (promote=false)
- "community" → NO (promote=false)
-
-When in doubt, set promote=false."""
-
-    try:
-        result = await llm_config.call(
-            messages=[
-                {"role": "system", "content": get_mission_filter_system_message()},
-                {"role": "user", "content": prompt},
-            ],
-            response_format=MissionFilterResponse,
-            scope="mental_model_emergent_evaluation",
-        )
-
-        # Build name -> promote map
-        promote_map = {c.name: c.promote for c in result.candidates}
-
-        # Find models to remove
-        models_to_remove = []
-        for model in models:
-            name = model["name"]
-            if name in promote_map:
-                if not promote_map[name]:
-                    models_to_remove.append(model["id"])
-                else:
-                    logger.debug(f"[EMERGENT] Keeping '{name}'")
-            else:
-                # Model not in response - remove to be safe
-                logger.info(f"[EMERGENT] '{name}' not in evaluation response, marking for removal")
-                models_to_remove.append(model["id"])
-
-        logger.info(f"[EMERGENT] Evaluation: {len(models_to_remove)}/{len(models)} emergent models marked for removal")
-        return models_to_remove
-
-    except Exception as e:
-        logger.warning(f"[EMERGENT] Evaluation failed, keeping all models: {e}")
-        return []
-
-
-async def detect_entity_candidates(
-    pool,
-    bank_id: str,
-    min_mentions: int = 5,
-    top_percent: int = 20,
-) -> list[EmergentCandidate]:
-    """
-    Detect entities that are candidates for promotion to mental models.
-
-    Args:
-        pool: Database connection pool
-        bank_id: Bank identifier
-        min_mentions: Minimum mention count to consider
-        top_percent: Only consider top X% by mention count
-
-    Returns:
-        List of entity candidates
-    """
-    from ..db_utils import acquire_with_retry
-    from ..memory_engine import fq_table
-
-    candidates = []
-
-    async with acquire_with_retry(pool) as conn:
-        # Get entities that meet criteria and don't already have mental models
-        rows = await conn.fetch(
-            f"""
-            WITH ranked AS (
-                SELECT
-                    e.id,
-                    e.canonical_name,
-                    e.mention_count,
-                    PERCENT_RANK() OVER (ORDER BY e.mention_count DESC) as rank_pct
-                FROM {fq_table("entities")} e
-                LEFT JOIN {fq_table("mental_models")} mm
-                    ON mm.entity_id = e.id AND mm.bank_id = e.bank_id
-                WHERE e.bank_id = $1
-                  AND e.mention_count >= $2
-                  AND mm.id IS NULL  -- Not already a mental model
-            )
-            SELECT id, canonical_name, mention_count
-            FROM ranked
-            WHERE rank_pct <= $3
-            ORDER BY mention_count DESC
-            LIMIT 50
-            """,
-            bank_id,
-            min_mentions,
-            top_percent / 100.0,
-        )
-
-        for row in rows:
-            candidates.append(
-                EmergentCandidate(
-                    name=row["canonical_name"],
-                    detection_method="named_entity_extraction",
-                    mention_count=row["mention_count"],
-                    entity_id=str(row["id"]),
-                    relevance_score=0.0,
-                )
-            )
-
-    logger.debug(f"[EMERGENT] Detected {len(candidates)} entity candidates")
-    return candidates
@@ -9,12 +9,14 @@ from pydantic import BaseModel, Field


 class MentalModelSubtype(str, Enum):
-    """Subtype of mental model - how it was created."""
+    """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
+    """

-    STRUCTURAL = "structural"  # Derived from mission, created upfront
-    EMERGENT = "emergent"  # Discovered from data patterns
-    LEARNED = "learned"  # Formed through reflection
-    PINNED = "pinned"  # User-defined topic, observations LLM-generated
    DIRECTIVE = "directive"  # User-defined hard rules, observations user-provided


@@ -49,50 +51,3 @@ class MentalModel(BaseModel):
    created_at: datetime = Field(
        default_factory=lambda: datetime.now(timezone.utc), description="When this model was created"
    )
-
-
-class StructuralModelTemplate(BaseModel):
-    """
-    A template for a structural mental model.
-
-    Generated by LLM based on the bank's mission. Represents what any agent
-    with this role would need to track.
-    """
-
-    id: str = Field(default="", description="Existing model ID to keep, or empty for new models")
-    name: str = Field(description="Human-readable name")
-    description: str = Field(description="What this model should track")
-    initial_probes: list[str] = Field(default_factory=list, description="Initial search queries to populate this model")
-
-
-class StructuralModelDerivationResponse(BaseModel):
-    """Response from LLM for structural model derivation."""
-
-    templates: list[StructuralModelTemplate] = Field(description="Structural model templates derived from the mission")
-
-
-class EmergentCandidate(BaseModel):
-    """
-    A candidate for promotion to emergent mental model.
-
-    Detected through pattern analysis of facts.
-    """
-
-    name: str = Field(description="Name of the detected pattern/entity")
-    detection_method: str = Field(description="How this candidate was detected")
-    mention_count: int = Field(default=0, description="How many times referenced")
-    entity_id: str | None = Field(default=None, description="Entity ID if detected as entity")
-    relevance_score: float = Field(default=0.0, description="Score from mission filter (0-1)")
-
-
-class ResearchResult(BaseModel):
-    """
-    Result from the research endpoint.
-
-    Contains the answer along with the mental models and facts used.
-    """
-
-    answer: str = Field(description="The synthesized answer")
-    mental_models_used: list[str] = Field(default_factory=list, description="IDs of mental models that contributed")
-    facts_used: list[str] = Field(default_factory=list, description="Fact IDs that contributed")
-    question_type: str | None = Field(default=None, description="Detected question type (WHO, WHAT, HOW, etc.)")
@@ -1,228 +0,0 @@
-"""
-Structural mental model derivation from bank mission.
-
-Structural models are derived from the bank's mission - they represent what
-any agent with this role would need to track. For example:
-
-Mission: "Be a PM for engineering team"
-Structural models:
-  - Team Structure (who's on the team, roles)
-  - Project Overview (current projects, status)
-  - Processes (how releases work, how decisions are made)
-  - Key Systems (what we own, dependencies)
-"""
-
-import logging
-from typing import TYPE_CHECKING
-
-from pydantic import BaseModel, Field
-
-from .models import StructuralModelTemplate
-
-if TYPE_CHECKING:
-    from ..llm_wrapper import LLMConfig
-
-logger = logging.getLogger(__name__)
-
-
-class StructuralDerivationResponse(BaseModel):
-    """Response from LLM for structural model derivation."""
-
-    templates: list[StructuralModelTemplate] = Field(description="Structural model templates derived from the mission")
-
-
-class StructuralRelevanceResult(BaseModel):
-    """Result of evaluating a structural model's relevance to the mission."""
-
-    name: str
-    relevant: bool
-    reason: str
-
-
-class StructuralRelevanceResponse(BaseModel):
-    """Response from LLM for structural model relevance evaluation."""
-
-    models: list[StructuralRelevanceResult] = Field(description="Relevance evaluation for each model")
-
-
-def build_structural_derivation_prompt(mission: str, existing_models: list[dict] | None = None) -> str:
-    """Build the prompt for deriving structural models from a mission."""
-    existing_section = ""
-    if existing_models:
-        model_list = "\n".join([f"- id='{m['id']}' name='{m['name']}': {m['description']}" for m in existing_models])
-        existing_section = f"""
-EXISTING STRUCTURAL MODELS:
-{model_list}
-
-IMPORTANT: If keeping an existing model, you MUST return its EXACT 'id' value.
-Models not included in your output will be REMOVED.
-"""
-
-    return f"""Given this agent mission, identify the KEY THINGS to track to achieve it.
-
-MISSION: {mission}
-{existing_section}
-IMPORTANT CONSTRAINTS:
- Return 0-3 structural models MAXIMUM (less is better!)
- Only include models for SPECIFIC, CONCRETE things the agent needs to track
- Each model must be DIRECTLY tied to achieving the mission
- If the mission is simple, return 0 models (empty array is fine)
- If existing models are provided and you want to keep one, use its EXACT id
- Do NOT create near-duplicates (e.g., don't create "topic-map" if "topic-connections" exists)
-
-GOOD examples (specific, actionable):
- Mission: "Be a PM for engineering team" → "Team Members" (track who's on the team)
- Mission: "Track customer feedback" → "Customer Issues" (track specific complaints/requests)
- Mission: "Manage project X" → "Project X Milestones" (track progress)
-
-BAD examples (too generic, don't create these):
- "Processes", "Workflows", "Key Systems", "Important Events"
- "Communication", "Collaboration", "Progress", "Status"
- Generic role-based models not tied to the specific mission
-
-For each model:
-1. id: Use EXACT existing id if keeping a model, or leave empty for new models
-2. name: Short, specific name (e.g., "Team Members", "Sprint Goals")
-3. description: One line describing what to track
-4. initial_probes: 2-3 search queries to find relevant information
-
-Return ONLY the models that should exist. Existing models not in your output will be deleted."""
-
-
-def get_structural_derivation_system_message() -> str:
-    """System message for structural model derivation."""
-    return """You identify the key things to track for a mission. Be VERY selective.
-
-Rules:
- Maximum 3 models (prefer fewer)
- Only SPECIFIC, CONCRETE things - not generic categories
- Each must DIRECTLY help achieve the mission
- Empty array is valid if no models are truly needed
- If existing models are shown and you want to keep one, return its EXACT id
- Never create duplicates - if a similar model exists, keep the existing one
-
-Output JSON with 'templates' array (can be empty)."""
-
-
-def _normalize_id(text: str) -> str:
-    """Normalize a string to a canonical form for comparison.
-
-    Removes common suffixes, pluralization, and normalizes separators.
-    """
-    # Lowercase and normalize separators
-    normalized = text.lower().replace(" ", "-").replace("_", "-")
-
-    # Remove common suffixes that indicate the same concept
-    suffixes_to_remove = ["-map", "-list", "-overview", "-tracker", "-s"]
-    for suffix in suffixes_to_remove:
-        if normalized.endswith(suffix) and len(normalized) > len(suffix):
-            normalized = normalized[: -len(suffix)]
-
-    return normalized
-
-
-def _find_similar_existing_id(new_id: str, existing_models: list[dict]) -> str | None:
-    """Find an existing model ID that is similar to the new ID.
-
-    Returns the existing ID if a similar one is found, None otherwise.
-    """
-    if not existing_models:
-        return None
-
-    new_normalized = _normalize_id(new_id)
-
-    for model in existing_models:
-        existing_id = model.get("id", "")
-        existing_normalized = _normalize_id(existing_id)
-
-        # Check if one is a prefix of the other (normalized)
-        if new_normalized.startswith(existing_normalized) or existing_normalized.startswith(new_normalized):
-            return existing_id
-
-        # Check if they're the same when normalized
-        if new_normalized == existing_normalized:
-            return existing_id
-
-    return None
-
-
-async def derive_structural_models(
-    llm_config: "LLMConfig",
-    mission: str,
-    existing_models: list[dict] | None = None,
-) -> tuple[list[StructuralModelTemplate], list[str]]:
-    """
-    Derive structural model templates from a bank's mission.
-
-    This combines derivation and evaluation in one call. The LLM sees existing
-    models and decides which to keep. Any existing model not in the output
-    will be marked for removal.
-
-    Args:
-        llm_config: LLM configuration for calling the model
-        mission: The bank's mission (e.g., "Be a PM for engineering team")
-        existing_models: Optional list of existing model dicts with 'name', 'description', 'id'
-
-    Returns:
-        Tuple of (templates to create/keep, IDs of existing models to remove)
-
-    Raises:
-        Exception: If LLM call fails
-    """
-    prompt = build_structural_derivation_prompt(mission, existing_models)
-
-    result = await llm_config.call(
-        messages=[
-            {"role": "system", "content": get_structural_derivation_system_message()},
-            {"role": "user", "content": prompt},
-        ],
-        response_format=StructuralDerivationResponse,
-        scope="mental_model_structural_derivation",
-    )
-
-    templates = result.templates
-    logger.info(f"[STRUCTURAL] LLM returned {len(templates)} structural models")
-
-    # Build set of existing IDs for quick lookup
-    existing_ids = {m["id"] for m in existing_models} if existing_models else set()
-
-    # Process templates: validate IDs, deduplicate, assign stable IDs
-    processed_templates: list[StructuralModelTemplate] = []
-    kept_existing_ids: set[str] = set()
-
-    for template in templates:
-        # If LLM returned an ID, check if it's a valid existing ID
-        if template.id and template.id in existing_ids:
-            # LLM is keeping an existing model
-            kept_existing_ids.add(template.id)
-            processed_templates.append(template)
-            logger.info(f"[STRUCTURAL] Keeping existing model: {template.id}")
-        else:
-            # New model or LLM didn't return a valid ID
-            # Generate ID from name
-            generated_id = template.name.lower().replace(" ", "-").replace("_", "-")
-
-            # Check for similar existing models to prevent near-duplicates
-            similar_id = _find_similar_existing_id(generated_id, existing_models)
-            if similar_id and similar_id not in kept_existing_ids:
-                # Use the existing similar model instead of creating a new one
-                logger.info(f"[STRUCTURAL] Detected near-duplicate: '{generated_id}' matches existing '{similar_id}'")
-                template.id = similar_id
-                kept_existing_ids.add(similar_id)
-            else:
-                template.id = generated_id
-
-            processed_templates.append(template)
-
-    # Find existing models to remove (not kept in LLM output)
-    models_to_remove = []
-    if existing_models:
-        for model in existing_models:
-            if model["id"] not in kept_existing_ids:
-                logger.info(f"[STRUCTURAL] Marking '{model['name']}' (id={model['id']}) for removal")
-                models_to_remove.append(model["id"])
-
-    if models_to_remove:
-        logger.info(f"[STRUCTURAL] {len(models_to_remove)} existing models will be removed")
-
-    return processed_templates, models_to_remove
@@ -0,0 +1,14 @@
+"""
+LLM provider implementations.
+
+This package contains concrete implementations of the LLMInterface for various providers.
+"""
+
+from .anthropic_llm import AnthropicLLM
+from .claude_code_llm import ClaudeCodeLLM
+from .codex_llm import CodexLLM
+from .gemini_llm import GeminiLLM
+from .mock_llm import MockLLM
+from .openai_compatible_llm import OpenAICompatibleLLM
+
+__all__ = ["AnthropicLLM", "ClaudeCodeLLM", "CodexLLM", "GeminiLLM", "MockLLM", "OpenAICompatibleLLM"]
@@ -0,0 +1,477 @@
+"""
+Anthropic LLM provider using the Anthropic Python SDK.
+
+This provider enables using Claude models from Anthropic with support for:
+- Structured JSON output
+- Tool/function calling with proper format conversion
+- Extended thinking mode
+- Retry logic with exponential backoff
+"""
+
+import asyncio
+import json
+import logging
+import time
+from typing import Any
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class AnthropicLLM(LLMInterface):
+    """
+    LLM provider using Anthropic's Claude models.
+
+    Supports structured output, tool calling, and extended thinking mode.
+    Handles format conversion between OpenAI-style messages and Anthropic's format.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        timeout: float = 300.0,
+        **kwargs: Any,
+    ):
+        """
+        Initialize Anthropic LLM provider.
+
+        Args:
+            provider: Provider name (should be "anthropic").
+            api_key: Anthropic API key.
+            base_url: Base URL for the API (optional, uses Anthropic default if empty).
+            model: Model name (e.g., "claude-sonnet-4-20250514").
+            reasoning_effort: Reasoning effort level (not used by Anthropic).
+            timeout: Request timeout in seconds.
+            **kwargs: Additional provider-specific parameters.
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        if not self.api_key:
+            raise ValueError("API key is required for Anthropic provider")
+
+        # Import and initialize Anthropic client
+        try:
+            from anthropic import AsyncAnthropic
+
+            client_kwargs: dict[str, Any] = {"api_key": self.api_key}
+            if self.base_url:
+                client_kwargs["base_url"] = self.base_url
+            if timeout:
+                client_kwargs["timeout"] = timeout
+
+            self._client = AsyncAnthropic(**client_kwargs)
+            logger.info(f"Anthropic client initialized for model: {self.model}")
+        except ImportError as e:
+            raise RuntimeError("Anthropic SDK not installed. Run: uv add anthropic or pip install anthropic") from e
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Anthropic provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            test_messages = [{"role": "user", "content": "test"}]
+            await self.call(
+                messages=test_messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+                scope="verification",
+                max_retries=0,
+            )
+            logger.info("Anthropic connection verified successfully")
+        except Exception as e:
+            logger.error(f"Anthropic connection verification failed: {e}")
+            raise RuntimeError(f"Failed to verify Anthropic connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported by Anthropic).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        from anthropic import APIConnectionError, APIStatusError, RateLimitError
+
+        start_time = time.time()
+
+        # Convert OpenAI-style messages to Anthropic format
+        system_prompt = None
+        anthropic_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                if system_prompt:
+                    system_prompt += "\n\n" + content
+                else:
+                    system_prompt = content
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            if system_prompt:
+                system_prompt += schema_msg
+            else:
+                system_prompt = schema_msg
+
+        # Prepare parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "max_tokens": max_completion_tokens if max_completion_tokens is not None else 4096,
+        }
+
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.messages.create(**call_params)
+
+                # Anthropic response content is a list of blocks
+                content = ""
+                for block in response.content:
+                    if block.type == "text":
+                        content += block.text
+
+                if response_format is not None:
+                    # Models may wrap JSON in markdown code blocks
+                    clean_content = content
+                    if "```json" in content:
+                        clean_content = content.split("```json")[1].split("```")[0].strip()
+                    elif "```" in content:
+                        clean_content = content.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_content)
+                    except json.JSONDecodeError:
+                        # Fallback to parsing raw content if markdown stripping failed
+                        json_data = json.loads(content)
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Record metrics and log slow calls
+                duration = time.time() - start_time
+                input_tokens = response.usage.input_tokens or 0 if response.usage else 0
+                output_tokens = response.usage.output_tokens or 0 if response.usage else 0
+                total_tokens = input_tokens + output_tokens
+
+                # Record LLM metrics
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import _serialize_for_span, get_span_recorder
+
+                finish_reason = response.stop_reason if hasattr(response, "stop_reason") else None
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=total_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except json.JSONDecodeError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    logger.warning("Anthropic returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Anthropic returned invalid JSON after {max_retries + 1} attempts")
+                    raise
+
+            except (APIConnectionError, RateLimitError, APIStatusError) as e:
+                # Fast fail on 401/403
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    logger.error(f"Anthropic auth error (HTTP {e.status_code}), not retrying: {str(e)}")
+                    raise
+
+                last_exception = e
+                if attempt < max_retries:
+                    # Check if it's a rate limit or server error
+                    should_retry = isinstance(e, (APIConnectionError, RateLimitError)) or (
+                        isinstance(e, APIStatusError) and e.status_code >= 500
+                    )
+
+                    if should_retry:
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                        await asyncio.sleep(backoff + jitter)
+                        continue
+
+                logger.error(f"Anthropic API error after {max_retries + 1} attempts: {str(e)}")
+                raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Anthropic call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        from anthropic import APIConnectionError, APIStatusError
+
+        start_time = time.time()
+
+        # Convert OpenAI tool format to Anthropic format
+        anthropic_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            anthropic_tools.append(
+                {
+                    "name": func.get("name", ""),
+                    "description": func.get("description", ""),
+                    "input_schema": func.get("parameters", {"type": "object", "properties": {}}),
+                }
+            )
+
+        # Convert messages - handle tool results
+        system_prompt = None
+        anthropic_messages = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt = (system_prompt + "\n\n" + content) if system_prompt else content
+            elif role == "tool":
+                # Anthropic uses tool_result blocks
+                anthropic_messages.append(
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "tool_result", "tool_use_id": msg.get("tool_call_id", ""), "content": content}
+                        ],
+                    }
+                )
+            elif role == "assistant" and msg.get("tool_calls"):
+                # Convert assistant tool calls
+                tool_use_blocks = []
+                for tc in msg["tool_calls"]:
+                    tool_use_blocks.append(
+                        {
+                            "type": "tool_use",
+                            "id": tc.get("id", ""),
+                            "name": tc.get("function", {}).get("name", ""),
+                            "input": json.loads(tc.get("function", {}).get("arguments", "{}")),
+                        }
+                    )
+                anthropic_messages.append({"role": "assistant", "content": tool_use_blocks})
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "tools": anthropic_tools,
+            "max_tokens": max_completion_tokens or 4096,
+        }
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.messages.create(**call_params)
+
+                # Extract content and tool calls
+                content_parts = []
+                tool_calls: list[LLMToolCall] = []
+
+                for block in response.content:
+                    if block.type == "text":
+                        content_parts.append(block.text)
+                    elif block.type == "tool_use":
+                        tool_calls.append(LLMToolCall(id=block.id, name=block.name, arguments=block.input or {}))
+
+                content = "".join(content_parts) if content_parts else None
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Extract token usage
+                input_tokens = response.usage.input_tokens or 0
+                output_tokens = response.usage.output_tokens or 0
+
+                # Record metrics
+                metrics = get_metrics_collector()
+                duration = time.time() - start_time
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except (APIConnectionError, APIStatusError) as e:
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic tool call failed")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close Anthropic client connections)."""
+        if hasattr(self, "_client") and self._client:
+            await self._client.close()
@@ -0,0 +1,510 @@
+"""
+Claude Code LLM provider using Claude Agent SDK.
+
+This provider enables using Claude Pro/Max subscriptions for API calls
+via the Claude CLI authentication. It uses the Claude Agent SDK which
+automatically handles authentication via `claude auth login` credentials.
+"""
+
+import asyncio
+import json
+import logging
+import time
+from typing import Any
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class ClaudeCodeLLM(LLMInterface):
+    """
+    LLM provider using Claude Code authentication.
+
+    Authenticates using Claude Pro/Max credentials via `claude auth login`
+    and makes API calls through the Claude Agent SDK.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,  # Will be ignored, uses CLI auth
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Claude Code LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Verify Claude Agent SDK is available
+        try:
+            self._verify_claude_code_available()
+            logger.info("Claude Code: Using Claude Agent SDK (authentication via claude auth login)")
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to initialize Claude Code provider: {e}\n\n"
+                "To set up Claude Code authentication:\n"
+                "1. Install Claude Code CLI: npm install -g @anthropics/claude-code\n"
+                "2. Login with your Pro/Max plan: claude auth login\n"
+                "3. Verify authentication: claude --version\n\n"
+                "Or use a different provider (anthropic, openai, gemini) with API keys."
+            ) from e
+
+        # Metrics collector is imported at module level
+
+    def _verify_claude_code_available(self) -> None:
+        """
+        Verify that Claude Agent SDK can be imported and is properly configured.
+
+        Raises:
+            ImportError: If Claude Agent SDK is not installed.
+            RuntimeError: If Claude Code is not authenticated.
+        """
+        try:
+            # Import Claude Agent SDK
+            # Reduce Claude Agent SDK logging verbosity
+            import logging as sdk_logging
+
+            from claude_agent_sdk import query  # noqa: F401
+
+            sdk_logging.getLogger("claude_agent_sdk").setLevel(sdk_logging.WARNING)
+            sdk_logging.getLogger("claude_agent_sdk._internal").setLevel(sdk_logging.WARNING)
+
+            logger.debug("Claude Agent SDK imported successfully")
+        except ImportError as e:
+            raise ImportError(
+                "Claude Agent SDK not installed. Run: uv add claude-agent-sdk or pip install claude-agent-sdk"
+            ) from e
+
+        # SDK will automatically check for authentication when first used
+        # No need to verify here - let it fail gracefully on first call with helpful error
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Claude Code provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            test_messages = [{"role": "user", "content": "test"}]
+            await self.call(
+                messages=test_messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+                scope="verification",
+                max_retries=0,
+            )
+            logger.info("Claude Code connection verified successfully")
+        except Exception as e:
+            logger.error(f"Claude Code connection verification failed: {e}")
+            raise RuntimeError(f"Failed to verify Claude Code connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response (ignored by Claude Agent SDK).
+            temperature: Sampling temperature (ignored by Claude Agent SDK).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with estimated token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits (not supported by Claude Agent SDK).
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        from claude_agent_sdk import AssistantMessage, ClaudeAgentOptions, TextBlock, query
+
+        start_time = time.time()
+
+        # Build system prompt
+        system_prompt = ""
+        user_content = ""
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt += ("\n\n" + content) if system_prompt else content
+            elif role == "user":
+                user_content += ("\n\n" + content) if user_content else content
+            elif role == "assistant":
+                # Claude Agent SDK doesn't support multi-turn easily in query()
+                # For now, prepend assistant messages to user content
+                user_content += f"\n\n[Previous assistant response: {content}]"
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_instruction = (
+                f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}\n\n"
+                "Respond with ONLY the JSON, no markdown formatting."
+            )
+            user_content += schema_instruction
+
+        # Configure SDK options
+        options = ClaudeAgentOptions(
+            system_prompt=system_prompt if system_prompt else None,
+            max_turns=1,  # Single-turn for API-style interactions
+            allowed_tools=[],  # Disable tools for standard LLM calls
+        )
+
+        # Call Claude Agent SDK
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                # Collect streaming response
+                full_text = ""
+
+                async for message in query(prompt=user_content, options=options):
+                    if isinstance(message, AssistantMessage):
+                        for block in message.content:
+                            if isinstance(block, TextBlock):
+                                full_text += block.text
+
+                # Handle structured output
+                if response_format is not None:
+                    # Models may wrap JSON in markdown
+                    clean_text = full_text
+                    if "```json" in full_text:
+                        clean_text = full_text.split("```json")[1].split("```")[0].strip()
+                    elif "```" in full_text:
+                        clean_text = full_text.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_text)
+                    except json.JSONDecodeError as e:
+                        logger.warning(f"Claude Code JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = e
+                            continue
+                        raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = full_text
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+
+                # Estimate token usage (Claude Agent SDK doesn't report exact counts)
+                # Use character count / 4 as rough estimate (1 token ≈ 4 characters)
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(full_text) // 4
+
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=result if isinstance(result, str) else json.dumps(result),
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    duration=duration,
+                    finish_reason=None,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=estimated_input,
+                        output_tokens=estimated_output,
+                        total_tokens=estimated_input + estimated_output,
+                    )
+                    return result, token_usage
+
+                return result
+
+            except Exception as e:
+                last_exception = e
+
+                # Check for authentication errors
+                error_str = str(e).lower()
+                if "auth" in error_str or "login" in error_str or "credential" in error_str:
+                    logger.error(f"Claude Code authentication error: {e}")
+                    raise RuntimeError(
+                        f"Claude Code authentication failed: {e}\n\n"
+                        "Run 'claude auth login' to authenticate with Claude Pro/Max."
+                    ) from e
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Claude Code error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Claude Code error after {max_retries + 1} attempts: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Claude Code call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support using Claude Agent SDK.
+
+        This implementation uses ClaudeSDKClient (not query()) because custom tools via
+        SDK MCP servers are only supported with the client. Tools are converted from OpenAI
+        format to SDK MCP tools, and tool names are formatted as mcp__hindsight_tools__{name}.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response (not used by Claude Agent SDK).
+            temperature: Sampling temperature (not used by Claude Agent SDK).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools (not used by Claude Agent SDK).
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        from claude_agent_sdk import (
+            AssistantMessage,
+            ClaudeAgentOptions,
+            ClaudeSDKClient,
+            SdkMcpTool,
+            TextBlock,
+            ToolUseBlock,
+            create_sdk_mcp_server,
+        )
+
+        start_time = time.time()
+
+        # Convert OpenAI tool format to Claude Agent SDK SdkMcpTool format
+        sdk_tools: list[SdkMcpTool] = []
+        tool_names: list[str] = []
+
+        for tool in tools:
+            func = tool.get("function", {})
+            tool_name = func.get("name", "")
+            tool_description = func.get("description", "")
+            parameters = func.get("parameters", {})
+
+            # Create a handler with proper closure to avoid transport issues
+            def make_handler(name: str):
+                async def handler(args: dict[str, Any]) -> dict[str, Any]:
+                    # Return immediately with success - tool execution happens externally
+                    return {
+                        "content": [
+                            {
+                                "type": "text",
+                                "text": f"[Tool {name} called successfully]",
+                            }
+                        ]
+                    }
+
+                return handler
+
+            sdk_tools.append(
+                SdkMcpTool(
+                    name=tool_name,
+                    description=tool_description,
+                    input_schema=parameters,
+                    handler=make_handler(tool_name),
+                )
+            )
+            tool_names.append(tool_name)
+
+        # Create an MCP server with the tools
+        mcp_server = create_sdk_mcp_server(
+            name="hindsight_tools",
+            version="1.0.0",
+            tools=sdk_tools if sdk_tools else None,
+        )
+
+        # Build system prompt and user content from messages
+        system_prompt = ""
+        user_content = ""
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt += ("\n\n" + content) if system_prompt else content
+            elif role == "user":
+                user_content += ("\n\n" + content) if user_content else content
+            elif role == "assistant":
+                # Include previous assistant messages as context
+                user_content += f"\n\n[Previous assistant response: {content}]"
+            elif role == "tool":
+                # Tool results are already in tool_results_map, append to user context
+                tool_call_id = msg.get("tool_call_id", "")
+                user_content += f"\n\n[Tool result for {tool_call_id}: {content}]"
+
+        # Format tool names for SDK MCP servers: mcp__{server_name}__{tool_name}
+        # This is required by the Claude Agent SDK for MCP server tools
+        allowed_tool_names = [f"mcp__hindsight_tools__{name}" for name in tool_names]
+
+        # Configure SDK options with MCP server
+        options = ClaudeAgentOptions(
+            system_prompt=system_prompt if system_prompt else None,
+            max_turns=1,  # Single-turn for API-style interactions
+            mcp_servers={"hindsight_tools": mcp_server} if sdk_tools else {},
+            allowed_tools=allowed_tool_names if allowed_tool_names else [],
+        )
+
+        # Call Claude Agent SDK with retry logic
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                full_text = ""
+                tool_calls: list[LLMToolCall] = []
+
+                # Use ClaudeSDKClient for tool calling support
+                # Note: query() does NOT support custom tools, only ClaudeSDKClient does
+                async with ClaudeSDKClient(options=options) as client:
+                    # Send the query
+                    await client.query(user_content)
+
+                    # Receive response
+                    async for message in client.receive_response():
+                        if isinstance(message, AssistantMessage):
+                            for block in message.content:
+                                if isinstance(block, TextBlock):
+                                    full_text += block.text
+                                elif isinstance(block, ToolUseBlock):
+                                    # SDK returns tool names with MCP prefix (mcp__hindsight_tools__{name})
+                                    # Strip the prefix to return original tool name expected by caller
+                                    tool_name = block.name
+                                    if tool_name.startswith("mcp__hindsight_tools__"):
+                                        tool_name = tool_name.replace("mcp__hindsight_tools__", "", 1)
+
+                                    tool_calls.append(
+                                        LLMToolCall(
+                                            id=block.id,
+                                            name=tool_name,
+                                            arguments=block.input,
+                                        )
+                                    )
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+
+                # Estimate token usage (Claude Agent SDK doesn't report exact counts)
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(full_text) // 4
+
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    success=True,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, time={duration:.3f}s"
+                    )
+
+                return LLMToolCallResult(
+                    content=full_text if full_text else None,
+                    tool_calls=tool_calls,
+                    finish_reason="tool_calls" if tool_calls else "stop",
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                )
+
+            except Exception as e:
+                last_exception = e
+
+                # Check for authentication errors
+                error_str = str(e).lower()
+                if "auth" in error_str or "login" in error_str or "credential" in error_str:
+                    logger.error(f"Claude Code authentication error: {e}")
+                    raise RuntimeError(
+                        f"Claude Code authentication failed: {e}\n\n"
+                        "Run 'claude auth login' to authenticate with Claude Pro/Max."
+                    ) from e
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Claude Code tool call error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Claude Code tool call error after {max_retries + 1} attempts: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Claude Code tool call failed after all retries")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (no HTTP client to close for Claude Agent SDK)."""
+        pass
@@ -0,0 +1,621 @@
+"""
+OpenAI Codex LLM provider using ChatGPT Plus/Pro OAuth authentication.
+
+This provider enables using ChatGPT Plus/Pro subscriptions for API calls
+without separate OpenAI Platform API credits. It uses OAuth tokens from
+~/.codex/auth.json and communicates with the ChatGPT backend API.
+"""
+
+import asyncio
+import json
+import logging
+import os
+import time
+import uuid
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class CodexLLM(LLMInterface):
+    """
+    LLM provider using OpenAI Codex OAuth authentication.
+
+    Authenticates using ChatGPT Plus/Pro credentials stored in ~/.codex/auth.json
+    and makes API calls to chatgpt.com/backend-api/codex/responses.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,  # Will be ignored, reads from ~/.codex/auth.json
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Codex LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Load Codex OAuth credentials
+        try:
+            self.access_token, self.account_id = self._load_codex_auth()
+            logger.info(f"Loaded Codex OAuth credentials for account: {self.account_id}")
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to load Codex OAuth credentials from ~/.codex/auth.json: {e}\n\n"
+                "To set up Codex authentication:\n"
+                "1. Install Codex CLI: npm install -g @openai/codex\n"
+                "2. Login: codex auth login\n"
+                "3. Verify: ls ~/.codex/auth.json\n\n"
+                "Or use a different provider (openai, anthropic, gemini) with API keys."
+            ) from e
+
+        # Use ChatGPT backend API endpoint
+        if not self.base_url:
+            self.base_url = "https://chatgpt.com/backend-api"
+
+        # Normalize model name (strip openai/ prefix if present)
+        if self.model.startswith("openai/"):
+            self.model = self.model[len("openai/") :]
+
+        # Map reasoning effort to Codex reasoning summary format
+        # Codex supports: "auto", "concise", "detailed"
+        self.reasoning_summary = self._map_reasoning_effort(reasoning_effort)
+
+        # HTTP client for SSE streaming
+        self._client = httpx.AsyncClient(timeout=120.0)
+
+    def _load_codex_auth(self) -> tuple[str, str]:
+        """
+        Load OAuth credentials from ~/.codex/auth.json.
+
+        Returns:
+            Tuple of (access_token, account_id).
+
+        Raises:
+            FileNotFoundError: If auth file doesn't exist.
+            ValueError: If auth file is invalid.
+        """
+        auth_file = Path.home() / ".codex" / "auth.json"
+
+        if not auth_file.exists():
+            raise FileNotFoundError(
+                f"Codex auth file not found: {auth_file}\nRun 'codex auth login' to authenticate with ChatGPT Plus/Pro."
+            )
+
+        with open(auth_file) as f:
+            data = json.load(f)
+
+        # Validate auth structure
+        auth_mode = data.get("auth_mode")
+        if auth_mode != "chatgpt":
+            raise ValueError(f"Expected auth_mode='chatgpt', got: {auth_mode}")
+
+        tokens = data.get("tokens", {})
+        access_token = tokens.get("access_token")
+        account_id = tokens.get("account_id")
+
+        if not access_token:
+            raise ValueError("No access_token found in Codex auth file. Run 'codex auth login' again.")
+
+        return access_token, account_id
+
+    def _map_reasoning_effort(self, effort: str) -> str:
+        """
+        Map standard reasoning effort to Codex reasoning summary format.
+
+        Args:
+            effort: Standard effort level ("low", "medium", "high", "xhigh").
+
+        Returns:
+            Codex reasoning summary: "concise", "detailed", or "auto".
+        """
+        mapping = {
+            "low": "concise",
+            "medium": "auto",
+            "high": "detailed",
+            "xhigh": "detailed",
+        }
+        return mapping.get(effort.lower(), "auto")
+
+    async def verify_connection(self) -> None:
+        """Verify Codex connection by making a simple test call."""
+        try:
+            logger.info(f"Verifying Codex LLM: model={self.model}, account={self.account_id}...")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=10,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"Codex LLM verified: {self.model}")
+        except Exception as e:
+            raise RuntimeError(f"Codex LLM connection verification failed for {self.model}: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """Make API call to Codex backend with SSE streaming."""
+        start_time = time.time()
+
+        # Prepare system instructions
+        system_instruction = ""
+        user_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction += ("\n\n" + content) if system_instruction else content
+            else:
+                user_messages.append(msg)
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            system_instruction += schema_msg
+
+        # gpt-5.2-codex only supports "detailed" reasoning summary
+        reasoning_summary = "detailed" if "5.2" in self.model else self.reasoning_summary
+
+        # Build Codex request payload
+        payload = {
+            "model": self.model,
+            "instructions": system_instruction,
+            "input": [
+                {
+                    "type": "message",
+                    "role": msg.get("role", "user"),
+                    "content": msg.get("content", ""),
+                }
+                for msg in user_messages
+            ],
+            "tools": [],
+            "tool_choice": "auto",
+            "parallel_tool_calls": True,
+            "reasoning": {"summary": reasoning_summary},
+            "store": False,  # Codex uses stateless mode
+            "stream": True,  # SSE streaming
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": str(uuid.uuid4()),
+        }
+
+        headers = {
+            "Authorization": f"Bearer {self.access_token}",
+            "Content-Type": "application/json",
+            "OpenAI-Account-ID": self.account_id,
+            "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)",
+            "Origin": "https://chatgpt.com",
+        }
+
+        url = f"{self.base_url}/codex/responses"
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.post(url, json=payload, headers=headers, timeout=120.0)
+                response.raise_for_status()
+
+                # Parse SSE stream
+                content = await self._parse_sse_stream(response)
+
+                # Handle structured output
+                if response_format is not None:
+                    # Models may wrap JSON in markdown
+                    clean_content = content
+                    if "```json" in content:
+                        clean_content = content.split("```json")[1].split("```")[0].strip()
+                    elif "```" in content:
+                        clean_content = content.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_content)
+                    except json.JSONDecodeError as e:
+                        logger.warning(f"Codex JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = e
+                            continue
+                        raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=0,  # Codex doesn't report token counts in SSE
+                    output_tokens=0,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                # Estimate tokens for tracing
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(content) // 4
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=result if isinstance(result, str) else json.dumps(result),
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    duration=duration,
+                    finish_reason=None,
+                    error=None,
+                )
+
+                if return_usage:
+                    # Codex doesn't provide token counts, estimate based on content
+                    estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                    estimated_output = len(content) // 4
+                    token_usage = TokenUsage(
+                        input_tokens=estimated_input,
+                        output_tokens=estimated_output,
+                        total_tokens=estimated_input + estimated_output,
+                    )
+                    return result, token_usage
+
+                return result
+
+            except httpx.HTTPStatusError as e:
+                last_exception = e
+                status_code = e.response.status_code
+
+                # Fast fail on auth errors
+                if status_code in (401, 403):
+                    logger.error(f"Codex auth error (HTTP {status_code}): {e.response.text[:200]}")
+                    raise RuntimeError(
+                        "Codex authentication failed. Your OAuth token may have expired.\n"
+                        "Run 'codex auth login' to re-authenticate."
+                    ) from e
+
+                # Log the actual error message from the API
+                error_detail = e.response.text[:500] if hasattr(e.response, "text") else str(e)
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(
+                        f"Codex HTTP error {status_code} (attempt {attempt + 1}/{max_retries + 1}): {error_detail}"
+                    )
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(
+                        f"Codex HTTP error after {max_retries + 1} attempts: Status {status_code}, Detail: {error_detail}"
+                    )
+                    raise
+
+            except httpx.RequestError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Codex connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Codex connection error after {max_retries + 1} attempts: {e}")
+                    raise
+
+            except Exception as e:
+                logger.error(f"Unexpected Codex error: {type(e).__name__}: {e}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Codex call failed after all retries")
+
+    async def _parse_sse_stream(self, response: httpx.Response) -> str:
+        """
+        Parse Server-Sent Events (SSE) stream from Codex API.
+
+        Args:
+            response: HTTP response with SSE stream.
+
+        Returns:
+            Extracted text content from stream.
+        """
+        full_text = ""
+        event_type = None
+
+        async for line in response.aiter_lines():
+            if not line:
+                continue
+
+            # Track event type
+            if line.startswith("event: "):
+                event_type = line[7:]
+
+            # Parse data
+            elif line.startswith("data: "):
+                data_str = line[6:]
+                if data_str == "[DONE]":
+                    break
+
+                try:
+                    data = json.loads(data_str)
+
+                    # Extract content based on event type
+                    if event_type == "response.text.delta" and "delta" in data:
+                        full_text += data["delta"]
+                    elif event_type == "response.content_part.delta" and "delta" in data:
+                        full_text += data["delta"]
+                    # Check for item content
+                    elif "item" in data:
+                        item = data["item"]
+                        if "content" in item:
+                            content = item["content"]
+                            if isinstance(content, list):
+                                for part in content:
+                                    if isinstance(part, dict) and "text" in part:
+                                        full_text += part["text"]
+                            elif isinstance(content, str):
+                                full_text += content
+
+                except json.JSONDecodeError:
+                    # Skip malformed JSON events
+                    pass
+
+        return full_text
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make API call with tool calling support.
+
+        Parses Codex SSE stream to extract tool calls from response.output_item.done events.
+        Tools are converted from OpenAI format to Codex format (flat structure at top level).
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature.
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Prepare system instructions
+        system_instruction = ""
+        user_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction += ("\n\n" + content) if system_instruction else content
+            elif role == "tool":
+                # Handle tool results
+                user_messages.append(
+                    {
+                        "type": "message",
+                        "role": "user",
+                        "content": f"Tool result: {content}",
+                    }
+                )
+            else:
+                user_messages.append(
+                    {
+                        "type": "message",
+                        "role": role,
+                        "content": content,
+                    }
+                )
+
+        # Convert tools to Codex format
+        # Codex expects tools with type and name/description/parameters at top level
+        codex_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            codex_tools.append(
+                {
+                    "type": "function",
+                    "name": func.get("name", ""),
+                    "description": func.get("description", ""),
+                    "parameters": func.get("parameters", {}),
+                }
+            )
+
+        # gpt-5.2-codex only supports "detailed" reasoning summary
+        reasoning_summary = "detailed" if "5.2" in self.model else self.reasoning_summary
+
+        payload = {
+            "model": self.model,
+            "instructions": system_instruction,
+            "input": user_messages,
+            "tools": codex_tools,
+            "tool_choice": tool_choice,
+            "parallel_tool_calls": True,
+            "reasoning": {"summary": reasoning_summary},
+            "store": False,
+            "stream": True,
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": str(uuid.uuid4()),
+        }
+
+        headers = {
+            "Authorization": f"Bearer {self.access_token}",
+            "Content-Type": "application/json",
+            "OpenAI-Account-ID": self.account_id,
+            "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)",
+            "Origin": "https://chatgpt.com",
+        }
+
+        url = f"{self.base_url}/codex/responses"
+
+        # Debug logging for troubleshooting
+        logger.debug(f"Codex tool call request: url={url}, model={payload['model']}, tools={len(codex_tools)}")
+
+        try:
+            response = await self._client.post(url, json=payload, headers=headers, timeout=120.0)
+
+            # Log response details on error
+            if response.status_code != 200:
+                logger.error(f"Codex API error {response.status_code}: {response.text[:500]}")
+
+            response.raise_for_status()
+
+            # Parse SSE for tool calls and content
+            content, tool_calls = await self._parse_sse_tool_stream(response)
+
+            duration = time.time() - start_time
+            metrics = get_metrics_collector()
+            metrics.record_llm_call(
+                provider=self.provider,
+                model=self.model,
+                scope=scope,
+                duration=duration,
+                input_tokens=0,
+                output_tokens=0,
+                success=True,
+            )
+
+            # Record OpenTelemetry span
+            from hindsight_api.tracing import get_span_recorder
+
+            span_recorder = get_span_recorder()
+            # Convert LLMToolCall objects to dicts for span recording
+            tool_calls_dict = (
+                [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls] if tool_calls else None
+            )
+            span_recorder.record_llm_call(
+                provider=self.provider,
+                model=self.model,
+                scope=scope,
+                messages=messages,
+                response_content=content,
+                input_tokens=0,  # Codex doesn't provide token counts
+                output_tokens=0,
+                duration=duration,
+                finish_reason="tool_calls" if tool_calls else "stop",
+                error=None,
+                tool_calls=tool_calls_dict,
+            )
+
+            return LLMToolCallResult(
+                content=content,
+                tool_calls=tool_calls,
+                finish_reason="tool_calls" if tool_calls else "stop",
+                input_tokens=0,
+                output_tokens=0,
+            )
+
+        except Exception as e:
+            logger.error(f"Codex tool call error: {e}")
+            raise
+
+    async def _parse_sse_tool_stream(self, response: httpx.Response) -> tuple[str | None, list[LLMToolCall]]:
+        """
+        Parse SSE stream for tool calls and content.
+
+        Returns:
+            Tuple of (content, tool_calls).
+        """
+        content = ""
+        tool_calls: list[LLMToolCall] = []
+        event_type = None
+
+        async for line in response.aiter_lines():
+            if not line:
+                continue
+
+            if line.startswith("event: "):
+                event_type = line[7:]
+
+            elif line.startswith("data: "):
+                data_str = line[6:]
+                if data_str == "[DONE]":
+                    break
+
+                try:
+                    data = json.loads(data_str)
+
+                    # Extract text content
+                    if event_type == "response.text.delta" and "delta" in data:
+                        content += data["delta"]
+
+                    # Extract completed tool calls from response.output_item.done
+                    elif event_type == "response.output_item.done":
+                        item = data.get("item", {})
+                        if item.get("type") == "function_call" and item.get("status") == "completed":
+                            tool_name = item.get("name", "")
+                            arguments_str = item.get("arguments", "{}")
+                            call_id = item.get("call_id", "")
+
+                            try:
+                                arguments = json.loads(arguments_str)
+                            except json.JSONDecodeError:
+                                logger.warning(f"Failed to parse tool arguments: {arguments_str}")
+                                arguments = {}
+
+                            tool_calls.append(
+                                LLMToolCall(
+                                    id=call_id,
+                                    name=tool_name,
+                                    arguments=arguments,
+                                )
+                            )
+
+                except json.JSONDecodeError as e:
+                    logger.warning(f"Failed to parse SSE data: {e}, data_str: {data_str[:200]}")
+
+        return content if content else None, tool_calls
+
+    async def cleanup(self) -> None:
+        """Clean up HTTP client."""
+        await self._client.aclose()
@@ -0,0 +1,550 @@
+"""
+Google Gemini/VertexAI LLM provider.
+
+This provider supports both:
+1. Gemini API (api.generativeai.google.com) with API key authentication
+2. Vertex AI with service account or Application Default Credentials (ADC)
+"""
+
+import asyncio
+import json
+import logging
+import os
+import time
+from typing import Any
+
+from google import genai
+from google.genai import errors as genai_errors
+from google.genai import types as genai_types
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+# Vertex AI imports (optional)
+try:
+    import google.auth
+    from google.oauth2 import service_account
+
+    VERTEXAI_AVAILABLE = True
+except ImportError:
+    VERTEXAI_AVAILABLE = False
+
+
+class GeminiLLM(LLMInterface):
+    """
+    LLM provider for Google Gemini and Vertex AI.
+
+    Supports:
+    - Gemini API: provider="gemini", requires api_key
+    - Vertex AI: provider="vertexai", requires project_id and region, uses ADC or service account
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Gemini/VertexAI LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        self._client = None
+        self._is_vertexai = self.provider == "vertexai"
+
+        if self._is_vertexai:
+            self._init_vertexai(**kwargs)
+        else:
+            self._init_gemini()
+
+    def _init_gemini(self) -> None:
+        """Initialize Gemini API client."""
+        if not self.api_key:
+            raise ValueError("Gemini provider requires api_key")
+
+        self._client = genai.Client(api_key=self.api_key)
+        logger.info(f"Gemini API: model={self.model}")
+
+    def _init_vertexai(self, **kwargs: Any) -> None:
+        """Initialize Vertex AI client with project, region, and credentials."""
+        # Extract Vertex AI config from kwargs
+        project_id = kwargs.get("vertexai_project_id")
+        region = kwargs.get("vertexai_region", "us-central1")
+        service_account_key = kwargs.get("vertexai_service_account_key")
+        credentials = kwargs.get("vertexai_credentials")  # Pre-loaded credentials object
+
+        if not project_id:
+            raise ValueError(
+                "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID is required for Vertex AI provider. "
+                "Set it to your GCP project ID."
+            )
+
+        auth_method = "ADC"
+
+        # Use pre-loaded credentials if provided (passed from LLMProvider)
+        if credentials is not None:
+            auth_method = "service_account"
+        # Otherwise, load explicit service account credentials if path provided
+        elif service_account_key:
+            if not VERTEXAI_AVAILABLE:
+                raise ValueError(
+                    "Vertex AI service account auth requires 'google-auth' package. "
+                    "Install with: pip install google-auth"
+                )
+            credentials = service_account.Credentials.from_service_account_file(
+                service_account_key,
+                scopes=["https://www.googleapis.com/auth/cloud-platform"],
+            )
+            auth_method = "service_account"
+            logger.info(f"Vertex AI: Using service account key: {service_account_key}")
+
+        # Strip google/ prefix from model name — native SDK uses bare names
+        # e.g. "google/gemini-2.0-flash-lite-001" -> "gemini-2.0-flash-lite-001"
+        if self.model.startswith("google/"):
+            self.model = self.model[len("google/") :]
+
+        # Create Vertex AI client
+        client_kwargs: dict[str, Any] = {
+            "vertexai": True,
+            "project": project_id,
+            "location": region,
+        }
+        if credentials is not None:
+            client_kwargs["credentials"] = credentials
+
+        self._client = genai.Client(**client_kwargs)
+
+        logger.info(f"Vertex AI: project={project_id}, region={region}, model={self.model}, auth={auth_method}")
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Gemini/VertexAI provider is configured correctly.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            logger.info(f"Verifying {self.provider.upper()}: model={self.model}...")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=100,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"{self.provider.upper()} connection verified successfully")
+        except Exception as e:
+            raise RuntimeError(f"Failed to verify {self.provider.upper()} connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make a Gemini/VertexAI API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response (not supported by Gemini).
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported by Gemini).
+            return_usage: If True, return tuple (result, TokenUsage).
+
+        Returns:
+            If return_usage=False: Parsed response if response_format provided, else text.
+            If return_usage=True: Tuple of (result, TokenUsage).
+        """
+        start_time = time.time()
+
+        # Convert OpenAI-style messages to Gemini format
+        system_instruction = None
+        gemini_contents = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                if system_instruction:
+                    system_instruction += "\n\n" + content
+                else:
+                    system_instruction = content
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            if system_instruction:
+                system_instruction += schema_msg
+            else:
+                system_instruction = schema_msg
+
+        # Build generation config
+        config_kwargs: dict[str, Any] = {}
+        if system_instruction:
+            config_kwargs["system_instruction"] = system_instruction
+        if response_format is not None:
+            config_kwargs["response_mime_type"] = "application/json"
+            config_kwargs["response_schema"] = response_format
+        if temperature is not None:
+            config_kwargs["temperature"] = temperature
+
+        generation_config = genai_types.GenerateContentConfig(**config_kwargs) if config_kwargs else None
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=generation_config,
+                )
+
+                content = response.text
+
+                # Handle empty response
+                if content is None:
+                    block_reason = None
+                    if hasattr(response, "candidates") and response.candidates:
+                        candidate = response.candidates[0]
+                        if hasattr(candidate, "finish_reason"):
+                            block_reason = candidate.finish_reason
+
+                    if attempt < max_retries:
+                        logger.warning(f"Gemini returned empty response (reason: {block_reason}), retrying...")
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        raise RuntimeError(f"Gemini returned empty response after {max_retries + 1} attempts")
+
+                # Parse structured output if requested
+                if response_format is not None:
+                    json_data = json.loads(content)
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Extract token usage
+                input_tokens = 0
+                output_tokens = 0
+                if hasattr(response, "usage_metadata") and response.usage_metadata:
+                    usage = response.usage_metadata
+                    input_tokens = usage.prompt_token_count or 0
+                    output_tokens = usage.candidates_token_count or 0
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                finish_reason = None
+                if hasattr(response, "candidates") and response.candidates:
+                    if hasattr(response.candidates[0], "finish_reason"):
+                        finish_reason = str(response.candidates[0].finish_reason)
+                span_recorder = get_span_recorder()
+                from hindsight_api.tracing import _serialize_for_span
+
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0 and input_tokens > 0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=input_tokens + output_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except json.JSONDecodeError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    logger.warning("Gemini returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Gemini returned invalid JSON after {max_retries + 1} attempts")
+                    raise
+
+            except genai_errors.APIError as e:
+                # Fast fail on auth errors - these won't recover with retries
+                if e.code in (401, 403):
+                    logger.error(f"Gemini auth error (HTTP {e.code}), not retrying: {str(e)}")
+                    raise
+
+                # Retry on retryable errors (rate limits, server errors, client errors)
+                if e.code in (400, 429, 500, 502, 503, 504) or (e.code and e.code >= 500):
+                    last_exception = e
+                    if attempt < max_retries:
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                        await asyncio.sleep(backoff + jitter)
+                    else:
+                        logger.error(f"Gemini API error after {max_retries + 1} attempts: {str(e)}")
+                        raise
+                else:
+                    logger.error(f"Gemini API error: {type(e).__name__}: {str(e)}")
+                    raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Gemini call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Gemini call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make a Gemini/VertexAI API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens (not supported by Gemini).
+            temperature: Sampling temperature.
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools (Gemini uses "auto" only).
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Convert tools to Gemini format
+        gemini_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            gemini_tools.append(
+                genai_types.Tool(
+                    function_declarations=[
+                        genai_types.FunctionDeclaration(
+                            name=func.get("name", ""),
+                            description=func.get("description", ""),
+                            parameters=func.get("parameters"),
+                        )
+                    ]
+                )
+            )
+
+        # Convert messages
+        system_instruction = None
+        gemini_contents = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction = (system_instruction + "\n\n" + content) if system_instruction else content
+            elif role == "tool":
+                # Gemini uses function_response
+                gemini_contents.append(
+                    genai_types.Content(
+                        role="user",
+                        parts=[
+                            genai_types.Part(
+                                function_response=genai_types.FunctionResponse(
+                                    name=msg.get("name", ""),
+                                    response={"result": content},
+                                )
+                            )
+                        ],
+                    )
+                )
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+
+        config_kwargs: dict[str, Any] = {"tools": gemini_tools}
+        if system_instruction:
+            config_kwargs["system_instruction"] = system_instruction
+        if temperature is not None:
+            config_kwargs["temperature"] = temperature
+
+        config = genai_types.GenerateContentConfig(**config_kwargs)
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=config,
+                )
+
+                # Extract content and tool calls
+                content = None
+                tool_calls: list[LLMToolCall] = []
+
+                if response.candidates and response.candidates[0].content:
+                    parts = response.candidates[0].content.parts
+                    if parts:
+                        for part in parts:
+                            if hasattr(part, "text") and part.text:
+                                content = part.text
+                            if hasattr(part, "function_call") and part.function_call:
+                                fc = part.function_call
+                                tool_calls.append(
+                                    LLMToolCall(
+                                        id=f"gemini_{len(tool_calls)}",
+                                        name=fc.name,
+                                        arguments=dict(fc.args) if fc.args else {},
+                                    )
+                                )
+
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Extract token usage
+                input_tokens = 0
+                output_tokens = 0
+                if response.usage_metadata:
+                    input_tokens = response.usage_metadata.prompt_token_count or 0
+                    output_tokens = response.usage_metadata.candidates_token_count or 0
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except genai_errors.APIError as e:
+                # Fast fail on auth errors
+                if e.code in (401, 403):
+                    logger.error(f"Gemini auth error (HTTP {e.code}), not retrying: {str(e)}")
+                    raise
+
+                # Retry on retryable errors
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Gemini tool call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Gemini tool call failed")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close connections, etc.)."""
+        # Gemini client doesn't require explicit cleanup
+        pass
@@ -0,0 +1,301 @@
+"""
+Mock LLM provider for testing.
+
+This provider allows tests to record LLM calls and return configurable mock responses
+without making actual API calls to external LLM services.
+"""
+
+import logging
+from typing import Any
+
+from ..llm_interface import LLMInterface
+from ..response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+
+logger = logging.getLogger(__name__)
+
+
+class MockLLM(LLMInterface):
+    """
+    Mock LLM provider for testing.
+
+    This provider records all calls and returns configurable mock responses,
+    enabling tests to verify LLM interactions without making real API calls.
+
+    Example:
+        # Create mock provider
+        mock_llm = MockLLM(provider="mock", api_key="", base_url="", model="mock-model")
+
+        # Set mock response
+        mock_llm.set_mock_response({"answer": "test"})
+
+        # Make calls
+        result = await mock_llm.call(
+            messages=[{"role": "user", "content": "test"}],
+            response_format=MyResponseModel
+        )
+
+        # Verify calls
+        calls = mock_llm.get_mock_calls()
+        assert len(calls) == 1
+        assert calls[0]["scope"] == "memory"
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """
+        Initialize mock LLM provider.
+
+        Args:
+            provider: Provider name (should be "mock").
+            api_key: Not used for mock provider.
+            base_url: Not used for mock provider.
+            model: Model name for tracking.
+            reasoning_effort: Not used for mock provider.
+            **kwargs: Additional parameters (not used).
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Storage for test verification
+        self._mock_calls: list[dict] = []
+        self._mock_response: Any = None
+        self._mock_exception: Exception | None = None
+
+    async def verify_connection(self) -> None:
+        """
+        Verify mock provider (always succeeds).
+
+        Mock provider doesn't need connection verification since it doesn't
+        make real API calls.
+        """
+        logger.debug("Mock LLM: connection verification (always succeeds)")
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make a mock LLM API call.
+
+        Records the call for test verification and returns the configured mock response.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Not used in mock.
+            temperature: Not used in mock.
+            scope: Scope identifier for tracking.
+            max_retries: Not used in mock.
+            initial_backoff: Not used in mock.
+            max_backoff: Not used in mock.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Not used in mock.
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with mock token counts.
+        """
+        # Record the call for test verification
+        call_record = {
+            "provider": self.provider,
+            "model": self.model,
+            "messages": messages,
+            "response_format": response_format.__name__
+            if response_format and hasattr(response_format, "__name__")
+            else str(response_format),
+            "scope": scope,
+        }
+        self._mock_calls.append(call_record)
+        logger.debug(f"Mock LLM call recorded: scope={scope}, model={self.model}")
+
+        # Raise mock exception if configured
+        if self._mock_exception is not None:
+            raise self._mock_exception
+
+        # Record trace span (minimal for mock provider)
+        from hindsight_api.tracing import get_span_recorder
+
+        span_recorder = get_span_recorder()
+        span_recorder.record_llm_call(
+            provider=self.provider,
+            model=self.model,
+            scope=scope,
+            messages=messages,
+            response_content="mock response",
+            input_tokens=10,
+            output_tokens=5,
+            duration=0.001,  # Mock calls are instant
+            finish_reason="stop",
+            error=None,
+        )
+
+        # Return mock response
+        if self._mock_response is not None:
+            result = self._mock_response
+        elif response_format is not None:
+            # Try to create a minimal valid instance of the response format
+            try:
+                # For Pydantic models, try to create with minimal valid data
+                result = {"mock": True}
+            except Exception:
+                result = {"mock": True}
+        else:
+            result = "mock response"
+
+        if return_usage:
+            token_usage = TokenUsage(input_tokens=10, output_tokens=5, total_tokens=15)
+            return result, token_usage
+        return result
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make a mock LLM API call with tool/function calling support.
+
+        Records the call for test verification and returns the configured mock response.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Not used in mock.
+            temperature: Not used in mock.
+            scope: Scope identifier for tracking.
+            max_retries: Not used in mock.
+            initial_backoff: Not used in mock.
+            max_backoff: Not used in mock.
+            tool_choice: Not used in mock.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        # Record the call for test verification
+        call_record = {
+            "provider": self.provider,
+            "model": self.model,
+            "messages": messages,
+            "tools": [t.get("function", {}).get("name") for t in tools],
+            "scope": scope,
+        }
+        self._mock_calls.append(call_record)
+
+        # Raise mock exception if configured
+        if self._mock_exception is not None:
+            raise self._mock_exception
+
+        # Record OpenTelemetry span
+        from hindsight_api.tracing import get_span_recorder
+
+        span_recorder = get_span_recorder()
+
+        if self._mock_response is not None:
+            if isinstance(self._mock_response, LLMToolCallResult):
+                result = self._mock_response
+            elif isinstance(self._mock_response, list):
+                # Allow setting just tool calls as a list
+                result = LLMToolCallResult(
+                    tool_calls=[
+                        LLMToolCall(id=f"mock_{i}", name=tc["name"], arguments=tc.get("arguments", {}))
+                        for i, tc in enumerate(self._mock_response)
+                    ],
+                    finish_reason="tool_calls",
+                )
+            else:
+                result = LLMToolCallResult(content="mock response", finish_reason="stop")
+        else:
+            result = LLMToolCallResult(content="mock response", finish_reason="stop")
+
+        # Record span with mock values
+        # Convert LLMToolCall objects to dicts for span recording
+        tool_calls_dict = (
+            [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in result.tool_calls]
+            if result.tool_calls
+            else None
+        )
+        span_recorder.record_llm_call(
+            provider=self.provider,
+            model=self.model,
+            scope=scope,
+            messages=messages,
+            response_content=result.content,
+            input_tokens=10,  # Mock value
+            output_tokens=5,  # Mock value
+            duration=0.1,  # Mock value
+            finish_reason=result.finish_reason,
+            error=None,
+            tool_calls=tool_calls_dict,
+        )
+
+        return result
+
+    async def cleanup(self) -> None:
+        """Clean up resources (no-op for mock provider)."""
+        pass
+
+    def set_mock_response(self, response: Any) -> None:
+        """
+        Set the response to return from mock calls.
+
+        Args:
+            response: The response to return. Can be:
+                - A dict/Pydantic model for regular calls
+                - An LLMToolCallResult for tool calls
+                - A list of tool call dicts for tool calls
+                - Any other value to return as-is
+        """
+        self._mock_response = response
+
+    def set_mock_exception(self, exception: Exception) -> None:
+        """
+        Set an exception to raise from mock calls.
+
+        Args:
+            exception: The exception to raise on the next call.
+                      After raising, the exception is cleared.
+        """
+        self._mock_exception = exception
+
+    def get_mock_calls(self) -> list[dict]:
+        """
+        Get the list of recorded mock calls.
+
+        Returns:
+            List of call records, each containing:
+                - provider: Provider name
+                - model: Model name
+                - messages: Messages sent
+                - response_format/tools: Format or tools used
+                - scope: Call scope
+        """
+        return self._mock_calls
+
+    def clear_mock_calls(self) -> None:
+        """Clear the recorded mock calls and any set exception."""
+        self._mock_calls = []
+        self._mock_exception = None
@@ -0,0 +1,788 @@
+"""
+OpenAI-compatible LLM provider supporting OpenAI, Groq, Ollama, and LMStudio.
+
+This provider handles all OpenAI API-compatible models including:
+- OpenAI: GPT-4, GPT-4o, GPT-5, o1, o3 (reasoning models)
+- Groq: Fast inference with seed control and service tiers
+- Ollama: Local models with native streaming API support
+- LMStudio: Local models with OpenAI-compatible API
+
+Features:
+- Reasoning models with extended thinking (o1, o3, GPT-5 families)
+- Strict JSON schema enforcement (OpenAI)
+- Provider-specific parameters (Groq seed, service tier)
+- Native Ollama streaming for better structured output
+- Automatic token limit handling per model family
+"""
+
+import asyncio
+import json
+import logging
+import os
+import re
+import time
+from typing import Any
+
+import httpx
+from openai import APIConnectionError, APIStatusError, AsyncOpenAI, LengthFinishReasonError
+
+from hindsight_api.config import DEFAULT_LLM_TIMEOUT, ENV_LLM_TIMEOUT
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+# Seed applied to every Groq request for deterministic behavior
+DEFAULT_LLM_SEED = 4242
+
+
+class OpenAICompatibleLLM(LLMInterface):
+    """
+    LLM provider for OpenAI-compatible APIs.
+
+    Supports:
+    - OpenAI: Standard models (GPT-4, GPT-4o) and reasoning models (o1, o3, GPT-5)
+    - Groq: Fast inference with seed control and service tiers
+    - Ollama: Local models with native streaming API for better structured output
+    - LMStudio: Local models with OpenAI-compatible API
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        timeout: float | None = None,
+        groq_service_tier: str | None = None,
+        **kwargs: Any,
+    ):
+        """
+        Initialize OpenAI-compatible LLM provider.
+
+        Args:
+            provider: Provider name ("openai", "groq", "ollama", "lmstudio").
+            api_key: API key (optional for ollama/lmstudio).
+            base_url: Base URL for the API (uses defaults for groq/ollama/lmstudio if empty).
+            model: Model name.
+            reasoning_effort: Reasoning effort level for supported models ("low", "medium", "high").
+            timeout: Request timeout in seconds (uses env var or 300s default).
+            groq_service_tier: Groq service tier ("on_demand", "flex", "auto").
+            **kwargs: Additional provider-specific parameters.
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Validate provider
+        valid_providers = ["openai", "groq", "ollama", "lmstudio"]
+        if self.provider not in valid_providers:
+            raise ValueError(f"OpenAICompatibleLLM only supports: {', '.join(valid_providers)}. Got: {self.provider}")
+
+        # Set default base URLs
+        if not self.base_url:
+            if self.provider == "groq":
+                self.base_url = "https://api.groq.com/openai/v1"
+            elif self.provider == "ollama":
+                self.base_url = "http://localhost:11434/v1"
+            elif self.provider == "lmstudio":
+                self.base_url = "http://localhost:1234/v1"
+
+        # For ollama/lmstudio, use dummy key if not provided
+        if self.provider in ("ollama", "lmstudio") and not self.api_key:
+            self.api_key = "local"
+
+        # Validate API key for cloud providers
+        if self.provider in ("openai", "groq") and not self.api_key:
+            raise ValueError(f"API key is required for {self.provider}")
+
+        # Groq service tier configuration
+        self.groq_service_tier = groq_service_tier or os.getenv("HINDSIGHT_API_LLM_GROQ_SERVICE_TIER", "auto")
+
+        # Get timeout config
+        self.timeout = timeout or float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT)))
+
+        # Create OpenAI client
+        client_kwargs: dict[str, Any] = {"api_key": self.api_key, "max_retries": 0}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        if self.timeout:
+            client_kwargs["timeout"] = self.timeout
+
+        self._client = AsyncOpenAI(**client_kwargs)
+        logger.info(
+            f"OpenAI-compatible client initialized: provider={self.provider}, model={self.model}, "
+            f"base_url={self.base_url or 'default'}"
+        )
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            logger.info(f"Verifying connection: {self.provider}/{self.model}")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=100,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"Connection verified: {self.provider}/{self.model}")
+        except Exception as e:
+            raise RuntimeError(f"Connection verification failed for {self.provider}/{self.model}: {e}") from e
+
+    def _supports_reasoning_model(self) -> bool:
+        """Check if the current model is a reasoning model (o1, o3, GPT-5, DeepSeek)."""
+        model_lower = self.model.lower()
+        return any(x in model_lower for x in ["gpt-5", "o1", "o3", "deepseek"])
+
+    def _get_max_reasoning_tokens(self) -> int | None:
+        """Get max reasoning tokens for reasoning models."""
+        model_lower = self.model.lower()
+
+        # GPT-4 and GPT-4.1 models have different caps
+        if any(x in model_lower for x in ["gpt-4.1", "gpt-4-"]):
+            return 32000
+        elif "gpt-4o" in model_lower:
+            return 16384
+
+        return None
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (OpenAI only).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        # Handle Ollama with native API for structured output (better schema enforcement)
+        if self.provider == "ollama" and response_format is not None:
+            return await self._call_ollama_native(
+                messages=messages,
+                response_format=response_format,
+                max_completion_tokens=max_completion_tokens,
+                temperature=temperature,
+                max_retries=max_retries,
+                initial_backoff=initial_backoff,
+                max_backoff=max_backoff,
+                skip_validation=skip_validation,
+                scope=scope,
+                return_usage=return_usage,
+            )
+
+        start_time = time.time()
+
+        # Build call parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+        }
+
+        # Check if model supports reasoning parameter
+        is_reasoning_model = self._supports_reasoning_model()
+
+        # Apply model-specific token limits
+        if max_completion_tokens is not None:
+            max_tokens_cap = self._get_max_reasoning_tokens()
+            if max_tokens_cap and max_completion_tokens > max_tokens_cap:
+                max_completion_tokens = max_tokens_cap
+            # For reasoning models, enforce minimum to ensure space for reasoning + output
+            if is_reasoning_model and max_completion_tokens < 16000:
+                max_completion_tokens = 16000
+            call_params["max_completion_tokens"] = max_completion_tokens
+
+        # Temperature - reasoning models don't support custom temperature
+        if temperature is not None and not is_reasoning_model:
+            call_params["temperature"] = temperature
+
+        # Set reasoning_effort for reasoning models
+        if is_reasoning_model:
+            call_params["reasoning_effort"] = self.reasoning_effort
+
+        # Provider-specific parameters
+        if self.provider == "groq":
+            call_params["seed"] = DEFAULT_LLM_SEED
+            extra_body: dict[str, Any] = {}
+            # Add service_tier if configured
+            if self.groq_service_tier:
+                extra_body["service_tier"] = self.groq_service_tier
+            # Add reasoning parameters for reasoning models
+            if is_reasoning_model:
+                extra_body["include_reasoning"] = False
+            if extra_body:
+                call_params["extra_body"] = extra_body
+
+        # Prepare response format ONCE before retry loop
+        if response_format is not None:
+            schema = None
+            if hasattr(response_format, "model_json_schema"):
+                schema = response_format.model_json_schema()
+
+            if strict_schema and schema is not None:
+                # Use OpenAI's strict JSON schema enforcement
+                call_params["response_format"] = {
+                    "type": "json_schema",
+                    "json_schema": {
+                        "name": "response",
+                        "strict": True,
+                        "schema": schema,
+                    },
+                }
+            else:
+                # Soft enforcement: add schema to prompt and use json_object mode
+                if schema is not None:
+                    schema_msg = (
+                        f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+                    )
+
+                    if call_params["messages"] and call_params["messages"][0].get("role") == "system":
+                        first_msg = call_params["messages"][0]
+                        if isinstance(first_msg, dict) and isinstance(first_msg.get("content"), str):
+                            first_msg["content"] += schema_msg
+                    elif call_params["messages"]:
+                        first_msg = call_params["messages"][0]
+                        if isinstance(first_msg, dict) and isinstance(first_msg.get("content"), str):
+                            first_msg["content"] = schema_msg + "\n\n" + first_msg["content"]
+                if self.provider not in ("lmstudio", "ollama"):
+                    # LM Studio and Ollama don't support json_object response format reliably
+                    call_params["response_format"] = {"type": "json_object"}
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                if response_format is not None:
+                    response = await self._client.chat.completions.create(**call_params)
+
+                    content = response.choices[0].message.content
+
+                    # Strip reasoning model thinking tags
+                    # Supports: <think>, <thinking>, <reasoning>, |startthink|/|endthink|
+                    if content:
+                        original_len = len(content)
+                        content = re.sub(r"<think>.*?</think>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"<thinking>.*?</thinking>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"<reasoning>.*?</reasoning>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"\|startthink\|.*?\|endthink\|", "", content, flags=re.DOTALL)
+                        content = content.strip()
+                        if len(content) < original_len:
+                            logger.debug(f"Stripped {original_len - len(content)} chars of reasoning tokens")
+
+                    # For local models, they may wrap JSON in markdown code blocks
+                    if self.provider in ("lmstudio", "ollama"):
+                        clean_content = content
+                        if "```json" in content:
+                            clean_content = content.split("```json")[1].split("```")[0].strip()
+                        elif "```" in content:
+                            clean_content = content.split("```")[1].split("```")[0].strip()
+                        try:
+                            json_data = json.loads(clean_content)
+                        except json.JSONDecodeError:
+                            # Fallback to parsing raw content
+                            json_data = json.loads(content)
+                    else:
+                        # Log raw LLM response for debugging JSON parse issues
+                        try:
+                            json_data = json.loads(content)
+                        except json.JSONDecodeError as json_err:
+                            # Truncate content for logging
+                            content_preview = content[:500] if content else "<empty>"
+                            if content and len(content) > 700:
+                                content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                            logger.warning(
+                                f"JSON parse error from LLM response (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                                f"  Model: {self.provider}/{self.model}\n"
+                                f"  Content length: {len(content) if content else 0} chars\n"
+                                f"  Content preview: {content_preview!r}\n"
+                                f"  Finish reason: {response.choices[0].finish_reason if response.choices else 'unknown'}"
+                            )
+                            # Retry on JSON parse errors
+                            if attempt < max_retries:
+                                backoff = min(initial_backoff * (2**attempt), max_backoff)
+                                await asyncio.sleep(backoff)
+                                last_exception = json_err
+                                continue
+                            else:
+                                logger.error(f"JSON parse error after {max_retries + 1} attempts, giving up")
+                                raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    response = await self._client.chat.completions.create(**call_params)
+                    result = response.choices[0].message.content
+
+                # Record token usage metrics
+                duration = time.time() - start_time
+                usage = response.usage
+                input_tokens = usage.prompt_tokens or 0 if usage else 0
+                output_tokens = usage.completion_tokens or 0 if usage else 0
+                total_tokens = usage.total_tokens or 0 if usage else 0
+
+                # Record LLM metrics
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import _serialize_for_span, get_span_recorder
+
+                finish_reason = response.choices[0].finish_reason if response.choices else None
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0 and usage:
+                    ratio = max(1, output_tokens) / max(1, input_tokens)
+                    cached_tokens = 0
+                    if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
+                        cached_tokens = getattr(usage.prompt_tokens_details, "cached_tokens", 0) or 0
+                    cache_info = f", cached_tokens={cached_tokens}" if cached_tokens > 0 else ""
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"total_tokens={total_tokens}{cache_info}, time={duration:.3f}s, ratio out/in={ratio:.2f}"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=total_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except LengthFinishReasonError as e:
+                logger.warning(f"LLM output exceeded token limits: {str(e)}")
+                raise OutputTooLongError(
+                    "LLM output exceeded token limits. Input may need to be split into smaller chunks."
+                ) from e
+
+            except APIConnectionError as e:
+                last_exception = e
+                status_code = getattr(e, "status_code", None) or getattr(
+                    getattr(e, "response", None), "status_code", None
+                )
+                logger.warning(f"APIConnectionError (HTTP {status_code}), attempt {attempt + 1}: {str(e)[:200]}")
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Connection error after {max_retries + 1} attempts: {str(e)}")
+                    raise
+
+            except APIStatusError as e:
+                # Fast fail only on 401 (unauthorized) and 403 (forbidden)
+                if e.status_code in (401, 403):
+                    logger.error(f"Auth error (HTTP {e.status_code}), not retrying: {str(e)}")
+                    raise
+
+                # Handle tool_use_failed error - model outputted in tool call format
+                if e.status_code == 400 and response_format is not None:
+                    try:
+                        error_body = e.body if hasattr(e, "body") else {}
+                        if isinstance(error_body, dict):
+                            error_info: dict[str, Any] = error_body.get("error") or {}
+                            if error_info.get("code") == "tool_use_failed":
+                                failed_gen = error_info.get("failed_generation", "")
+                                if failed_gen:
+                                    # Parse tool call format and convert to expected format
+                                    tool_call = json.loads(failed_gen)
+                                    tool_name = tool_call.get("name", "")
+                                    tool_args = tool_call.get("arguments", {})
+                                    converted = {"actions": [{"tool": tool_name, **tool_args}]}
+                                    if skip_validation:
+                                        result = converted
+                                    else:
+                                        result = response_format.model_validate(converted)
+
+                                    # Record metrics
+                                    duration = time.time() - start_time
+                                    metrics = get_metrics_collector()
+                                    metrics.record_llm_call(
+                                        provider=self.provider,
+                                        model=self.model,
+                                        scope=scope,
+                                        duration=duration,
+                                        input_tokens=0,
+                                        output_tokens=0,
+                                        success=True,
+                                    )
+                                    if return_usage:
+                                        return result, TokenUsage(input_tokens=0, output_tokens=0, total_tokens=0)
+                                    return result
+                    except (json.JSONDecodeError, KeyError, TypeError):
+                        pass  # Failed to parse tool_use_failed, continue with normal retry
+
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                    sleep_time = backoff + jitter
+                    await asyncio.sleep(sleep_time)
+                else:
+                    logger.error(f"API error after {max_retries + 1} attempts: {str(e)}")
+                    raise
+
+            except Exception:
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("LLM call failed after all retries with no exception captured")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Build call parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+            "tools": tools,
+            "tool_choice": tool_choice,
+        }
+
+        if max_completion_tokens is not None:
+            call_params["max_completion_tokens"] = max_completion_tokens
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        # Provider-specific parameters
+        if self.provider == "groq":
+            call_params["seed"] = DEFAULT_LLM_SEED
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.chat.completions.create(**call_params)
+
+                message = response.choices[0].message
+                finish_reason = response.choices[0].finish_reason
+
+                # Extract tool calls if present
+                tool_calls: list[LLMToolCall] = []
+                if message.tool_calls:
+                    for tc in message.tool_calls:
+                        try:
+                            args = json.loads(tc.function.arguments) if tc.function.arguments else {}
+                        except json.JSONDecodeError:
+                            args = {"_raw": tc.function.arguments}
+                        tool_calls.append(LLMToolCall(id=tc.id, name=tc.function.name, arguments=args))
+
+                content = message.content
+
+                # Record metrics
+                duration = time.time() - start_time
+                usage = response.usage
+                input_tokens = usage.prompt_tokens or 0 if usage else 0
+                output_tokens = usage.completion_tokens or 0 if usage else 0
+
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except APIConnectionError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+            except APIStatusError as e:
+                if e.status_code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+            except Exception:
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Tool call failed after all retries")
+
+    async def _call_ollama_native(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any,
+        max_completion_tokens: int | None,
+        temperature: float | None,
+        max_retries: int,
+        initial_backoff: float,
+        max_backoff: float,
+        skip_validation: bool,
+        scope: str = "memory",
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Call Ollama using native API with JSON schema enforcement.
+
+        Ollama's native API supports passing a full JSON schema in the 'format' parameter,
+        which provides better structured output control than the OpenAI-compatible API.
+        """
+        start_time = time.time()
+
+        # Get the JSON schema from the Pydantic model
+        schema = response_format.model_json_schema() if hasattr(response_format, "model_json_schema") else None
+
+        # Build the base URL for Ollama's native API
+        # Default OpenAI-compatible URL is http://localhost:11434/v1
+        # Native API is at http://localhost:11434/api/chat
+        base_url = self.base_url or "http://localhost:11434/v1"
+        if base_url.endswith("/v1"):
+            native_url = base_url[:-3] + "/api/chat"
+        else:
+            native_url = base_url.rstrip("/") + "/api/chat"
+
+        # Build request payload
+        payload: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+            "stream": False,
+        }
+
+        # Add schema as format parameter for structured output
+        if schema:
+            payload["format"] = schema
+
+        # Add optional parameters with optimized defaults for Ollama
+        options: dict[str, Any] = {
+            "num_ctx": 16384,  # 16k context window for larger prompts
+            "num_batch": 512,  # Optimal batch size for prompt processing
+        }
+        if max_completion_tokens:
+            options["num_predict"] = max_completion_tokens
+        if temperature is not None:
+            options["temperature"] = temperature
+        payload["options"] = options
+
+        last_exception = None
+
+        async with httpx.AsyncClient(timeout=300.0) as client:
+            for attempt in range(max_retries + 1):
+                try:
+                    response = await client.post(native_url, json=payload)
+                    response.raise_for_status()
+
+                    result = response.json()
+                    content = result.get("message", {}).get("content", "")
+
+                    # Parse JSON response
+                    try:
+                        json_data = json.loads(content)
+                    except json.JSONDecodeError as json_err:
+                        content_preview = content[:500] if content else "<empty>"
+                        if content and len(content) > 700:
+                            content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                        logger.warning(
+                            f"Ollama JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                            f"  Model: ollama/{self.model}\n"
+                            f"  Content length: {len(content) if content else 0} chars\n"
+                            f"  Content preview: {content_preview!r}"
+                        )
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = json_err
+                            continue
+                        else:
+                            raise
+
+                    # Extract token usage from Ollama response
+                    duration = time.time() - start_time
+                    input_tokens = result.get("prompt_eval_count", 0) or 0
+                    output_tokens = result.get("eval_count", 0) or 0
+                    total_tokens = input_tokens + output_tokens
+
+                    # Record LLM metrics
+                    metrics = get_metrics_collector()
+                    metrics.record_llm_call(
+                        provider=self.provider,
+                        model=self.model,
+                        scope=scope,
+                        duration=duration,
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        success=True,
+                    )
+
+                    # Validate against Pydantic model or return raw JSON
+                    if skip_validation:
+                        validated_result = json_data
+                    else:
+                        validated_result = response_format.model_validate(json_data)
+
+                    if return_usage:
+                        token_usage = TokenUsage(
+                            input_tokens=input_tokens,
+                            output_tokens=output_tokens,
+                            total_tokens=total_tokens,
+                        )
+                        return validated_result, token_usage
+                    return validated_result
+
+                except httpx.HTTPStatusError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(
+                            f"Ollama HTTP error (attempt {attempt + 1}/{max_retries + 1}): {e.response.status_code}"
+                        )
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama HTTP error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except httpx.RequestError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(f"Ollama connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama connection error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except Exception as e:
+                    logger.error(f"Unexpected error during Ollama call: {type(e).__name__}: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Ollama call failed after all retries")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close OpenAI client connections)."""
+        if hasattr(self, "_client") and self._client:
+            await self._client.close()
@@ -4,17 +4,15 @@ 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. Learn new insights (create/update mental models)
-4. Expand memories (get chunk/document context)
+3. Expand memories (get chunk/document context)
 """

 from .agent import ReflectAgentResult, run_reflect_agent
-from .models import MentalModelInput, ReflectAction, ReflectActionBatch
+from .models import ReflectAction, ReflectActionBatch

 __all__ = [
    "run_reflect_agent",
    "ReflectAgentResult",
    "ReflectAction",
    "ReflectActionBatch",
-    "MentalModelInput",
 ]
@@ -7,54 +7,31 @@ from typing import Any, Literal
 from pydantic import BaseModel, Field


-class MentalModelObservation(BaseModel):
-    """An observation within a mental model with its supporting memories."""
+class ObservationSection(BaseModel):
+    """A section within an observation with its supporting memories."""

-    title: str = Field(description="Observation header (can be empty for intro)")
-    text: str = Field(description="Observation content - no headers, use lists/tables/bold")
-    memory_ids: list[str] = Field(default_factory=list, description="Memory IDs supporting this observation")
-
-
-class MentalModelInput(BaseModel):
-    """Input for the learn tool to create a mental model placeholder.
-
-    The agent only specifies name and description - the actual content/observations
-    are generated during refresh, similar to pinned models.
-    """
-
-    name: str = Field(description="Human-readable name for the mental model")
-    description: str = Field(description="What to track - used as prompt for content generation during refresh")
-    entity_id: str | None = Field(default=None, description="Optional link to existing entity ID")
-
-
-class AnswerSection(BaseModel):
-    """A section of the answer with its supporting evidence (DEPRECATED)."""
-
-    title: str = Field(description="Section header/title")
-    text: str = Field(description="Section content")
+    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")
-    model_ids: list[str] = Field(default_factory=list, description="Mental model IDs supporting this section")


 class ReflectAction(BaseModel):
    """Single action the reflect agent can take."""

-    tool: Literal["list_mental_models", "get_mental_model", "recall", "learn", "expand", "done"] = Field(
-        description="Tool to invoke: list_mental_models, get_mental_model, recall, learn, expand, or done"
+    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
-    model_id: str | None = Field(default=None, description="Mental model ID for get_mental_model")
+    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)")
-    mental_model: MentalModelInput | None = Field(default=None, description="Mental model to create/update for learn")
    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")
-    sections: list[AnswerSection] | None = Field(default=None, description="DEPRECATED: Use answer field instead")
-    observations: list[MentalModelObservation] | None = Field(
-        default=None, description="Observations for done action (when output_mode=observations)"
+    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: str | None = Field(default=None, description="Well-formatted markdown answer for done action")
    answer_memory_ids: list[str] | None = Field(
        default=None, description="Memory IDs supporting the answer", alias="memory_ids"
    )
@@ -73,7 +50,8 @@ class ReflectActionBatch(BaseModel):
 class ToolCall(BaseModel):
    """A single tool call made during reflect."""

-    tool: str = Field(description="Tool name: lookup, recall, learn, expand")
+    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")
@@ -85,6 +63,8 @@ class LLMCall(BaseModel):

    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):
@@ -92,7 +72,15 @@ class DirectiveInfo(BaseModel):

    id: str = Field(description="Directive mental model ID")
    name: str = Field(description="Directive name")
-    rules: list[str] = Field(default_factory=list, description="Directive rules/observations that were applied")
+    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):
@@ -104,11 +92,18 @@ class ReflectAgentResult(BaseModel):
    )
    iterations: int = Field(default=0, description="Number of iterations taken")
    tools_called: int = Field(default=0, description="Total number of tool calls made")
-    mental_models_created: list[str] = Field(default_factory=list, description="IDs of mental models created/updated")
    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_model_ids: list[str] = Field(default_factory=list, description="Validated model 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"
    )
@@ -184,65 +184,3 @@ def compute_trend(
        return Trend.WEAKENING
    else:
        return Trend.STABLE
-
-
-class CandidateObservation(BaseModel):
-    """A candidate observation generated during the seed phase.
-
-    Candidates are preliminary observations that need evidence validation
-    before becoming full observations.
-    """
-
-    content: str = Field(description="The proposed observation content")
-    seed_memory_ids: list[str] = Field(default_factory=list, description="Memory IDs that inspired this candidate")
-
-
-class CandidateWithEvidence(BaseModel):
-    """A candidate observation with gathered supporting and contradicting evidence."""
-
-    candidate: CandidateObservation
-    supporting_memories: list[dict] = Field(default_factory=list, description="Memories that support this observation")
-    contradicting_memories: list[dict] = Field(
-        default_factory=list, description="Memories that contradict this observation"
-    )
-
-
-class MentalModelSnapshot(BaseModel):
-    """A versioned snapshot of a mental model's observations.
-
-    Used for tracking changes over time and enabling diff views.
-    """
-
-    version: int = Field(description="Version number (1-indexed)")
-    observations: list[Observation] = Field(default_factory=list, description="Observations at this version")
-    created_at: datetime = Field(
-        default_factory=lambda: datetime.now(timezone.utc), description="When this version was created"
-    )
-    reflect_summary: str | None = Field(default=None, description="Summary of changes in this version")
-
-
-def verify_evidence_quotes(
-    observation: Observation,
-    memories: dict[str, str],
-) -> tuple[bool, list[str]]:
-    """Verify that all evidence quotes exist in the referenced memories.
-
-    Args:
-        observation: The observation to verify
-        memories: Dict mapping memory_id to memory content
-
-    Returns:
-        Tuple of (is_valid, list of error messages)
-    """
-    errors = []
-
-    for evidence in observation.evidence:
-        memory_content = memories.get(evidence.memory_id)
-        if memory_content is None:
-            errors.append(f"Memory {evidence.memory_id} not found")
-            continue
-
-        if evidence.quote not in memory_content:
-            errors.append(f"Quote not found in memory {evidence.memory_id}: '{evidence.quote[:50]}...'")
-
-    return len(errors) == 0, errors
@@ -1,5 +1,10 @@
 """
 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
@@ -11,7 +16,7 @@ def _extract_directive_rules(directives: list[dict[str, Any]]) -> list[str]:
    Extract directive rules as a list of strings.

    Args:
-        directives: List of directive mental models with observations
+        directives: List of directives with name and content

    Returns:
        List of directive rule strings
@@ -19,25 +24,34 @@ def _extract_directive_rules(directives: list[dict[str, Any]]) -> list[str]:
    rules = []
    for directive in directives:
        directive_name = directive.get("name", "")
-        observations = directive.get("observations", [])
-        if observations:
-            for obs in observations:
-                # Support both Pydantic Observation objects and dicts
-                if hasattr(obs, "title"):
-                    title = obs.title
-                    content = obs.content
-                else:
-                    title = obs.get("title", "")
-                    content = obs.get("content", "")
-                if title and content:
-                    rules.append(f"**{title}**: {content}")
-                elif content:
-                    rules.append(content)
-        elif directive_name:
-            # Fallback to description if no observations
-            desc = directive.get("description", "")
-            if desc:
-                rules.append(f"**{directive_name}**: {desc}")
+        # 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


@@ -111,27 +125,38 @@ 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.

-    This is a simplified prompt since tools are defined separately via the tools parameter.
+    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", "")

-    no_info_rule = (
-        "- Only say 'I don't have information' AFTER trying list_mental_models AND recall with no relevant results"
-    )
-
    parts = []

-    # Inject directives at the VERY START for maximum prominence
+    # Anti-hallucination rule at the very top
+    parts.extend(
+        [
+            "CRITICAL: You MUST ONLY use information from retrieved tool results. NEVER make up names, people, events, or entities.",
+            "",
+        ]
+    )
+
+    # Inject directives after anti-hallucination rule
    if directives:
        parts.append(build_directives_section(directives))

@@ -145,10 +170,9 @@ def build_system_prompt_for_tools(
    parts.extend(
        [
            "## CRITICAL RULES",
-            "- You must NEVER fabricate information that has no basis in retrieved data",
+            "- ONLY use information from tool results - no external knowledge or guessing",
            "- You SHOULD synthesize, infer, and reason from the retrieved memories",
-            "- You MUST call recall() before saying you don't have information",
-            no_info_rule,
+            "- 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",
@@ -156,7 +180,56 @@ def build_system_prompt_for_tools(
            "- 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",
            "",
-            "## Query Strategy (IMPORTANT)",
+            "## 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')",
@@ -164,44 +237,84 @@ def build_system_prompt_for_tools(
            "  1. recall('lessons') - find all lesson-related memories",
            "  2. recall('teaching sessions') - alternative phrasing",
            "  3. recall('student progress') - find student-related memories",
-            "  4. recall('topics taught') - find subject matter",
            "",
            "Think: What ENTITIES and CONCEPTS does this question involve? Search for each separately.",
-            "- Questions about patterns → search for the individual instances first",
-            "- Questions comparing things → search for each thing separately",
-            "- Questions about relationships → search for each party involved",
            "",
-            "## Workflow",
        ]
    )

-    # Answer mode: include mental model lookup in workflow
+    # 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(
        [
-            "1. Review the pre-fetched mental models for relevant synthesized knowledge",
-            "2. If relevant, call get_mental_model(model_id) for full observations",
-            "3. DECOMPOSE the question into component searches (see Query Strategy above)",
-            "   - Identify entities and concepts in the question",
-            "   - Search for each separately with targeted queries",
-            "4. Run multiple recall() calls - don't just echo the user's question",
-            "5. Use expand() if you need more context on specific memories",
-            "6. BEFORE answering: Check if any person/project/concept from the memories deserves a mental model - use learn() if so",
-            "7. When ready, call done() with your answer and supporting memory_ids",
            "",
-            "## When to Use learn() - IMPORTANT",
-            "ACTIVELY look for opportunities to use learn() when you discover:",
-            "- A person mentioned in 2+ memories who has no mental model yet",
-            "- A project or concept the user asks about that has no mental model",
-            "- A pattern or topic worth tracking for future questions",
-            "",
-            "DO NOT wait to be asked - proactively create models when you see the need.",
-            "Example: learn(name='Project Alpha', description='Track goals, status, and key decisions for Project Alpha')",
-            "",
-            "## Output Format: Plain Text Answer",
-            "Call done() with a plain text 'answer' field.",
-            "- Do NOT use markdown formatting",
+            "## Output Format: Well-Formatted Markdown Answer",
+            "Call done() with a well-formatted markdown 'answer' field.",
+            "- USE markdown formatting for structure (headers, lists, bold, italic, code blocks, tables, etc.)",
+            "- CRITICAL: Add blank lines before and after block elements (tables, code blocks, lists)",
+            "- Format for clarity and readability with proper spacing and hierarchy",
            "- NEVER include memory IDs, UUIDs, or 'Memory references' in the answer text",
-            "- Put memory IDs ONLY in the memory_ids array parameter, not in the answer",
+            "- Put IDs ONLY in the memory_ids/mental_model_ids/observation_ids arrays, not in the answer",
        ]
    )

@@ -295,9 +408,10 @@ def build_agent_prompt(
    else:
        parts.append(
            "\n## Instructions\n"
-            "Start by calling list_mental_models() to see available mental models - they contain pre-synthesized knowledge. "
-            "If a relevant model exists, use get_mental_model(model_id) to get its observations. "
-            "Then use recall(query) for specific details not covered by mental models."
+            "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)
@@ -359,404 +473,41 @@ def build_final_prompt(
    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."
+        "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."
+        "Only say 'I don't have information' if the retrieved data is truly unrelated to the question.\n\n"
+        "IMPORTANT: Output ONLY the final answer. Do NOT include meta-commentary like "
+        '"I\'ll search..." or "Let me analyze...". Do NOT explain your reasoning process. '
+        "Just provide the direct synthesized answer."
    )

    return "\n".join(parts)


-FINAL_SYSTEM_PROMPT = """You are a thoughtful assistant that synthesizes answers from retrieved memories.
+FINAL_SYSTEM_PROMPT = """CRITICAL: You MUST ONLY use information from retrieved tool results. NEVER make up names, people, events, or entities.
+
+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 use information from tool results - no external knowledge or guessing

 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."""

-
-# =============================================================================
-# 4-Phase Mental Model Reflect Prompts
-# =============================================================================
-
-SEED_PHASE_SYSTEM_PROMPT = """You are analyzing memories to discover NEW patterns and generate candidate observations.
-
-Your task is to identify potential observations (beliefs, preferences, patterns, behaviors) that could be part of a mental model about this person/topic.
-
-## Important: Avoid Redundancy
-If existing observations are provided, DO NOT generate candidates that are essentially the same.
-Focus on discovering NEW patterns not already covered by existing observations.
-
-## Rules
- Generate 5-15 candidate observations for NEW patterns only
- Each candidate should be specific and testable (can be supported or contradicted by evidence)
- Note which memory IDs inspired each candidate (these are seeds, not final evidence)
- Focus on patterns that appear MULTIPLE TIMES across many memories - the more the better
- The best candidates are ones you can find 10, 20, or even 50+ supporting memories for
- Skip patterns that are already covered by existing observations
-
-## Output Format
-Return a JSON array of candidate observations:
-```json
-{
-  "candidates": [
-    {
-      "content": "The specific observation/belief/pattern - be detailed and specific",
-      "seed_memory_ids": ["memory_id_1", "memory_id_2", "memory_id_3"]
-    }
-  ]
-}
-```
-
-Focus on patterns that appear multiple times or have strong signals. Don't generate obvious or trivial observations.
-Prefer candidates with MORE seed memories - they're more likely to be real patterns.
-Return an empty candidates array if no genuinely new patterns are found."""
-
-
-def build_seed_phase_prompt(
-    memories: list[dict],
-    topic: str | None = None,
-    existing_observations: list[dict] | None = None,
-) -> str:
-    """Build the user prompt for the seed phase.
-
-    Args:
-        memories: List of memories to analyze
-        topic: Optional topic focus for the mental model
-        existing_observations: Optional list of existing observations to avoid rediscovering
-    """
-    parts = []
-
-    if topic:
-        parts.append(f"## Topic Focus\n{topic}\n")
-
-    # Include existing observations so we don't rediscover them
-    if existing_observations:
-        parts.append("## Existing Observations (DO NOT regenerate these)")
-        parts.append("These patterns are already tracked. Focus on discovering NEW patterns:\n")
-        for i, obs in enumerate(existing_observations, 1):
-            title = obs.get("title", "")
-            content = obs.get("content", "")
-            parts.append(f"{i}. **{title}**: {content}\n")
-        parts.append("")
-
-    parts.append("## Memories to Analyze")
-    parts.append("Review these memories and identify patterns, preferences, beliefs, and behaviors:\n")
-
-    for mem in memories:
-        mem_id = mem.get("id", "unknown")
-        content = mem.get("content", mem.get("text", ""))
-        timestamp = mem.get("timestamp", mem.get("created_at", ""))
-        parts.append(f"[{mem_id}] ({timestamp}): {content}\n")
-
-    parts.append("\n## Instructions")
-    if existing_observations:
-        parts.append("Generate candidate observations for NEW patterns not already covered above.")
-        parts.append("If all patterns are already covered by existing observations, return an empty candidates array.")
-    else:
-        parts.append("Generate candidate observations based on patterns you see in these memories.")
-    parts.append("Look for: recurring themes, stated preferences, behavioral patterns, beliefs, values, goals.")
-
-    return "\n".join(parts)
-
-
-VALIDATE_PHASE_SYSTEM_PROMPT = """You are validating candidate observations against evidence.
-
-For each candidate, you have:
- Supporting memories (evidence FOR the observation)
- Contradicting memories (evidence AGAINST the observation)
-
-## Your Task
-1. Evaluate each candidate based on the evidence
-2. For valid candidates, extract EXACT QUOTES from supporting memories
-3. Discard candidates with insufficient or contradicting evidence
-4. Merge similar candidates into single, refined observations
-
-## Rules for Quotes
- Quotes must be EXACT text from the memory, not paraphrased
- Each quote should directly support the observation
- The MORE evidence quotes, the BETTER - don't limit yourself, include ALL relevant quotes (10, 20, 50+)
- Observations with only 1-2 quotes are weak and should be discarded unless the evidence is exceptionally strong
- Stronger observations have more supporting evidence - aim for comprehensive coverage
-
-## Output Format
-Return validated observations with evidence:
-```json
-{
-  "observations": [
-    {
-      "title": "Short descriptive title (3-8 words) - like a headline",
-      "content": "The full observation content - detailed explanation of the pattern/belief",
-      "evidence": [
-        {
-          "memory_id": "exact_memory_id",
-          "quote": "Exact quote from the memory text",
-          "relevance": "Brief explanation of how this supports the observation",
-          "timestamp": "2024-01-15T10:00:00Z"
-        }
-      ]
-    }
-  ],
-  "discarded": [
-    {
-      "content": "The discarded candidate",
-      "reason": "Why it was discarded (insufficient evidence, contradicted, etc.)"
-    }
-  ],
-  "merged": [
-    {
-      "from": ["candidate 1 content", "candidate 2 content"],
-      "into": "The merged observation content"
-    }
-  ]
-}
-```
-
-## Title Guidelines
- Title should be a SHORT label (like "Prefers morning meetings" or "Coffee enthusiast")
- NOT a truncated version of the content
- Think of it as a category/tag for the observation
-
-Be rigorous: only keep observations with clear, verifiable evidence from multiple memories."""
-
-
-def build_validate_phase_prompt(candidates_with_evidence: list[dict]) -> str:
-    """Build the user prompt for the validate phase."""
-    parts = ["## Candidates to Validate\n"]
-
-    for i, item in enumerate(candidates_with_evidence, 1):
-        candidate = item.get("candidate", {})
-        supporting = item.get("supporting_memories", [])
-        contradicting = item.get("contradicting_memories", [])
-
-        parts.append(f"### Candidate {i}: {candidate.get('content', '')}")
-
-        if supporting:
-            parts.append("\n**Supporting Evidence:**")
-            for mem in supporting:
-                mem_id = mem.get("id", "unknown")
-                content = mem.get("content", mem.get("text", ""))
-                timestamp = mem.get("timestamp", mem.get("created_at", ""))
-                parts.append(f"- [{mem_id}] ({timestamp}): {content}")
-
-        if contradicting:
-            parts.append("\n**Contradicting Evidence:**")
-            for mem in contradicting:
-                mem_id = mem.get("id", "unknown")
-                content = mem.get("content", mem.get("text", ""))
-                timestamp = mem.get("timestamp", mem.get("created_at", ""))
-                parts.append(f"- [{mem_id}] ({timestamp}): {content}")
-
-        if not supporting and not contradicting:
-            parts.append("\n*No additional evidence found*")
-
-        parts.append("")
-
-    parts.append("## Instructions")
-    parts.append("1. Evaluate each candidate based on its evidence")
-    parts.append("2. Keep candidates with strong supporting evidence")
-    parts.append("3. Discard candidates with no evidence or strong contradictions")
-    parts.append("4. Merge similar candidates")
-    parts.append("5. Extract EXACT quotes (copy-paste from memory text) for evidence")
-
-    return "\n".join(parts)
-
-
-COMPARE_PHASE_SYSTEM_PROMPT = """You are merging new observations with an existing mental model.
-
-You have:
- EXISTING observations (from the current mental model)
- NEW observations (from this reflect cycle)
-
-## Your Task
-Produce the final, complete mental model by:
-1. Keeping existing observations that are still valid
-2. Updating existing observations with new evidence (ADD new evidence to existing)
-3. Adding new observations that don't overlap with existing
-4. Removing existing observations that are contradicted by new evidence
-5. Merging overlapping observations
-
-## Rules
- The final model should have no contradictions
- Each observation must have evidence with exact quotes
- COMBINE evidence from both existing and new observations
- If an existing observation has new supporting evidence, ADD ALL the new evidence to it
- Include ALL relevant evidence - the more quotes the better (10, 20, 50+ is great)
- Observations with more evidence are more reliable - don't limit the number of quotes
-
-## Output Format
-Return the complete, final mental model:
-```json
-{
-  "observations": [
-    {
-      "title": "Short descriptive title (3-8 words)",
-      "content": "Full observation content - detailed explanation",
-      "evidence": [
-        {
-          "memory_id": "id",
-          "quote": "exact quote",
-          "relevance": "explanation",
-          "timestamp": "ISO timestamp"
-        }
-      ],
-      "created_at": "ISO timestamp of when observation was first created"
-    }
-  ],
-  "changes": {
-    "kept": ["Observation that was kept unchanged"],
-    "updated": [{"from": "old content", "to": "new content", "reason": "why"}],
-    "added": ["New observation that was added"],
-    "removed": [{"content": "removed observation", "reason": "why removed"}],
-    "merged": [{"from": ["obs1", "obs2"], "into": "merged observation"}]
-  }
-}
-```"""
-
-
-def build_compare_phase_prompt(
-    existing_observations: list[dict],
-    new_observations: list[dict],
-) -> str:
-    """Build the user prompt for the compare phase."""
-    parts = []
-
-    parts.append("## Existing Mental Model Observations")
-    if existing_observations:
-        for i, obs in enumerate(existing_observations, 1):
-            title = obs.get("title", "")
-            content = obs.get("content", obs.get("text", ""))
-            evidence = obs.get("evidence", [])
-            parts.append(f"\n### Existing {i}: {title}")
-            parts.append(f"Content: {content}")
-            if evidence:
-                parts.append(f"Evidence ({len(evidence)} items):")
-                for ev in evidence[:5]:  # Show max 5 evidence items
-                    parts.append(f'  - [{ev.get("memory_id", "?")}]: "{ev.get("quote", "")}"')
-                if len(evidence) > 5:
-                    parts.append(f"  ... and {len(evidence) - 5} more")
-    else:
-        parts.append("*No existing observations*")
-
-    parts.append("\n## New Observations from This Reflect")
-    if new_observations:
-        for i, obs in enumerate(new_observations, 1):
-            title = obs.get("title", "")
-            content = obs.get("content", "")
-            evidence = obs.get("evidence", [])
-            parts.append(f"\n### New {i}: {title}")
-            parts.append(f"Content: {content}")
-            if evidence:
-                parts.append(f"Evidence ({len(evidence)} items):")
-                for ev in evidence:
-                    parts.append(f'  - [{ev.get("memory_id", "?")}]: "{ev.get("quote", "")}"')
-    else:
-        parts.append("*No new observations*")
-
-    parts.append("\n## Instructions")
-    parts.append("Merge these into a coherent, non-contradictory mental model.")
-    parts.append("Preserve all valid evidence. Remove stale or contradicted observations.")
-
-    return "\n".join(parts)
-
-
-# =============================================================================
-# UPDATE EXISTING Phase Prompts (for diff-based refresh)
-# =============================================================================
-
-UPDATE_EXISTING_SYSTEM_PROMPT = """You are updating existing observations with newly found evidence.
-
-For each existing observation, you have been given:
- The original observation (title, content, existing evidence)
- Newly found supporting memories
- Newly found contradicting memories
-
-## Your Task
-1. Extract EXACT QUOTES from new supporting memories to add to the observation
-2. Flag observations with strong contradicting evidence for potential removal
-3. Keep existing evidence intact - only ADD new evidence
-
-## Rules for Quotes
- Quotes must be EXACT text from the memory, not paraphrased
- Each quote should directly support the observation
- Include ALL relevant quotes from the new memories
-
-## Output Format
-Return updated observations with new evidence:
-```json
-{
-  "updated_observations": [
-    {
-      "title": "Original title",
-      "content": "Original content",
-      "existing_evidence_count": 5,
-      "new_evidence": [
-        {
-          "memory_id": "exact_memory_id",
-          "quote": "Exact quote from the memory text",
-          "relevance": "Brief explanation of how this supports the observation",
-          "timestamp": "2024-01-15T10:00:00Z"
-        }
-      ],
-      "has_contradiction": false,
-      "contradiction_note": null
-    }
-  ]
-}
-```
-
-If an observation has strong contradicting evidence, set has_contradiction=true and explain in contradiction_note."""
-
-
-def build_update_existing_prompt(observations_with_evidence: list[dict]) -> str:
-    """Build the user prompt for the update existing phase.
-
-    Args:
-        observations_with_evidence: List of existing observations with new evidence found
-    """
-    parts = ["## Existing Observations to Update\n"]
-
-    for i, item in enumerate(observations_with_evidence, 1):
-        obs = item.get("observation", {})
-        supporting = item.get("supporting_memories", [])
-        contradicting = item.get("contradicting_memories", [])
-
-        title = obs.get("title", "")
-        content = obs.get("content", "")
-        existing_evidence = obs.get("evidence", [])
-
-        parts.append(f"### Observation {i}: {title}")
-        parts.append(f"Content: {content}")
-        parts.append(f"Existing evidence count: {len(existing_evidence)}")
-
-        if supporting:
-            parts.append("\n**New Supporting Memories:**")
-            for mem in supporting:
-                mem_id = mem.get("id", "unknown")
-                mem_content = mem.get("content", mem.get("text", ""))
-                timestamp = mem.get("timestamp", mem.get("created_at", ""))
-                parts.append(f"- [{mem_id}] ({timestamp}): {mem_content}")
-
-        if contradicting:
-            parts.append("\n**New Contradicting Memories:**")
-            for mem in contradicting:
-                mem_id = mem.get("id", "unknown")
-                mem_content = mem.get("content", mem.get("text", ""))
-                timestamp = mem.get("timestamp", mem.get("created_at", ""))
-                parts.append(f"- [{mem_id}] ({timestamp}): {mem_content}")
-
-        if not supporting and not contradicting:
-            parts.append("\n*No new evidence found*")
-
-        parts.append("")
-
-    parts.append("## Instructions")
-    parts.append("1. Extract EXACT quotes from new supporting memories")
-    parts.append("2. Flag observations with strong contradictions")
-    parts.append("3. Return the updated observations with new evidence added")
-
-    return "\n".join(parts)
+FORMATTING: Use proper markdown formatting in your answer:
+- Headers (##, ###) for sections
+- Lists (bullet or numbered) for enumerations
+- Bold/italic for emphasis
+- Tables with proper syntax (ensure blank line before and after)
+- Code blocks where appropriate
+- CRITICAL: Always add blank lines before and after block elements (tables, code blocks, lists)
+- Proper spacing between sections
+
+CRITICAL: Output ONLY the final synthesized answer. Do NOT include:
+- Meta-commentary about what you're doing ("I'll search...", "Let me analyze...")
+- Explanations of your reasoning process
+- Descriptions of your approach
+Just provide the direct answer with proper markdown formatting."""
@@ -1,16 +1,17 @@
 """
 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 re
 import uuid
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
 from typing import TYPE_CHECKING, Any

-from .models import MentalModelInput
-from .observations import Observation, ObservationEvidence, Trend
-
 if TYPE_CHECKING:
    from asyncpg import Connection

@@ -19,156 +20,214 @@ if TYPE_CHECKING:

 logger = logging.getLogger(__name__)

-
-def generate_model_id(name: str) -> str:
-    """Generate a stable ID from mental model name."""
-    # Normalize: lowercase, replace spaces/special chars with hyphens
-    normalized = re.sub(r"[^a-z0-9]+", "-", name.lower()).strip("-")
-    # Truncate to reasonable length
-    return normalized[:50]
+# Observation is considered stale if not updated in this many days
+STALE_THRESHOLD_DAYS = 7


-def _parse_observations(observations_raw: list) -> list[Observation]:
-    """Parse raw observation dicts into typed Observation models."""
-    observations: list[Observation] = []
-    for obs in observations_raw:
-        if not isinstance(obs, dict):
-            continue
-
-        try:
-            parsed = Observation(
-                title=obs.get("title", ""),
-                content=obs.get("content", ""),
-                evidence=[
-                    ObservationEvidence(
-                        memory_id=ev.get("memory_id", ""),
-                        quote=ev.get("quote", ""),
-                        relevance=ev.get("relevance", ""),
-                        timestamp=ev.get("timestamp"),
-                    )
-                    for ev in obs.get("evidence", [])
-                    if isinstance(ev, dict)
-                ],
-                created_at=obs.get("created_at"),
-            )
-            observations.append(parsed)
-        except Exception as e:
-            logger.warning(f"Failed to parse observation: {e}")
-            continue
-
-    return observations
-
-
-async def tool_lookup(
+async def tool_search_mental_models(
    conn: "Connection",
    bank_id: str,
-    model_id: str | None = None,
+    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]:
    """
-    List or get mental models.
+    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
-        model_id: Optional specific model ID to get (if None, lists all)
-        tags: Optional tags to filter models (when listing)
+        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 either a list of models or a single model's details
+        Dict with matching mental models including content and freshness info
    """
-    if model_id:
-        # Get specific mental model with full details including observations
-        row = await conn.fetchrow(
-            """
-            SELECT id, subtype, name, description, observations, entity_id, last_updated
-            FROM mental_models
-            WHERE id = $1 AND bank_id = $2
-            """,
-            model_id,
-            bank_id,
-        )
-        if row:
-            # Parse observations JSON
-            obs_data = row["observations"] or {"observations": []}
-            if isinstance(obs_data, str):
-                import json
+    from ..memory_engine import fq_table
+    from ..search.tags import build_tags_where_clause

-                obs_data = json.loads(obs_data)
-            observations_raw = obs_data.get("observations", []) if isinstance(obs_data, dict) else obs_data
+    # Build filters dynamically
+    filters = ""
+    params: list[Any] = [bank_id, str(query_embedding), max_results]
+    next_param = 4

-            # Parse observations into typed models
-            observations = _parse_observations(observations_raw)
+    # Use the centralized tag filtering logic
+    if tags:
+        tag_clause, tag_params, next_param = build_tags_where_clause(tags, param_offset=next_param, match=tags_match)
+        filters += f" {tag_clause}"
+        params.extend(tag_params)

-            return {
-                "found": True,
-                "model": {
-                    "id": row["id"],
-                    "subtype": row["subtype"],
-                    "name": row["name"],
-                    "description": row["description"],
-                    "observations": observations,
-                    "entity_id": str(row["entity_id"]) if row["entity_id"] else None,
-                    "last_updated": row["last_updated"].isoformat() if row["last_updated"] else None,
-                },
+    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 {"found": False, "model_id": model_id}
-    else:
-        # List mental models (compact: id, name, description only)
-        # Full observations are retrieved via get_mental_model(model_id)
-        # NOTE: Directives (subtype='directive') are excluded from listing -
-        # they are injected into the system prompt, not discoverable via tools
-        # Filter by tags if provided
-        if tags:
-            if tags_match == "all":
-                # All tags must match
-                rows = await conn.fetch(
-                    """
-                    SELECT id, subtype, name, description
-                    FROM mental_models
-                    WHERE bank_id = $1 AND tags @> $2::varchar[] AND subtype != 'directive'
-                    ORDER BY last_updated DESC NULLS LAST, created_at DESC
-                    """,
-                    bank_id,
-                    tags,
-                )
-            else:
-                # Any tag matches (OR) - default
-                rows = await conn.fetch(
-                    """
-                    SELECT id, subtype, name, description
-                    FROM mental_models
-                    WHERE bank_id = $1 AND tags && $2::varchar[] AND subtype != 'directive'
-                    ORDER BY last_updated DESC NULLS LAST, created_at DESC
-                    """,
-                    bank_id,
-                    tags,
-                )
-        else:
-            rows = await conn.fetch(
-                """
-                SELECT id, subtype, name, description
-                FROM mental_models
-                WHERE bank_id = $1 AND subtype != 'directive'
-                ORDER BY last_updated DESC NULLS LAST, created_at DESC
+        )
+
+    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[])
                """,
-                bank_id,
+                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 {
-            "count": len(rows),
-            "models": [
-                {
-                    "id": row["id"],
-                    "subtype": row["subtype"],
-                    "name": row["name"],
-                    "description": row["description"],
-                }
-                for row in rows
-            ],
-        }
+    # 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(
@@ -185,6 +244,9 @@ async def tool_recall(
    """
    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
@@ -202,13 +264,14 @@ async def tool_recall(
    result = await memory_engine.recall_async(
        bank_id=bank_id,
        query=query,
-        fact_type=["experience", "world"],  # Exclude opinions
+        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 = []
@@ -230,85 +293,6 @@ async def tool_recall(
    }


-async def tool_learn(
-    conn: "Connection",
-    bank_id: str,
-    input: MentalModelInput,
-    tags: list[str] | None = None,
-) -> dict[str, Any]:
-    """
-    Create a mental model placeholder with subtype='learned'.
-
-    The agent only specifies name and description - actual observations are generated
-    in the background via refresh, similar to pinned models.
-
-    Args:
-        conn: Database connection
-        bank_id: Bank identifier
-        input: Mental model input data (name, description, optional entity_id)
-        tags: Tags to apply to new mental models (from reflect context)
-
-    Returns:
-        Dict with created model info including model_id for background generation
-    """
-    model_id = generate_model_id(input.name)
-
-    # Parse entity_id if provided
-    entity_uuid = None
-    if input.entity_id:
-        try:
-            entity_uuid = uuid.UUID(input.entity_id)
-        except ValueError:
-            logger.warning(f"Invalid entity_id format: {input.entity_id}")
-
-    # Check if model exists
-    existing = await conn.fetchrow(
-        "SELECT id FROM mental_models WHERE id = $1 AND bank_id = $2",
-        model_id,
-        bank_id,
-    )
-
-    if existing:
-        # Update description only - observations will be regenerated
-        await conn.execute(
-            """
-            UPDATE mental_models SET
-                description = $3,
-                entity_id = $4
-            WHERE id = $1 AND bank_id = $2
-            """,
-            model_id,
-            bank_id,
-            input.description,
-            entity_uuid,
-        )
-        status = "updated"
-    else:
-        # Insert new model placeholder - observations will be generated in background
-        await conn.execute(
-            """
-            INSERT INTO mental_models (id, bank_id, subtype, name, description, observations, entity_id, tags, created_at)
-            VALUES ($1, $2, 'learned', $3, $4, '{}'::jsonb, $5, $6, NOW())
-            """,
-            model_id,
-            bank_id,
-            input.name,
-            input.description,
-            entity_uuid,
-            tags or [],
-        )
-        status = "created"
-
-    logger.info(f"[REFLECT] Mental model '{model_id}' {status} in bank {bank_id} - pending background generation")
-
-    return {
-        "status": status,
-        "model_id": model_id,
-        "name": input.name,
-        "pending_generation": True,
-    }
-
-
 async def tool_expand(
    conn: "Connection",
    bank_id: str,
@@ -327,6 +311,8 @@ async def tool_expand(
    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"}

@@ -344,9 +330,9 @@ async def tool_expand(

    # Batch fetch all memory units
    memories = await conn.fetch(
-        """
+        f"""
        SELECT id, text, chunk_id, document_id, fact_type, context
-        FROM memory_units
+        FROM {fq_table("memory_units")}
        WHERE id = ANY($1) AND bank_id = $2
        """,
        valid_uuids,
@@ -363,9 +349,9 @@ async def tool_expand(
    chunk_map: dict[str, Any] = {}
    if chunk_ids:
        chunks = await conn.fetch(
-            """
+            f"""
            SELECT chunk_id, chunk_text, chunk_index, document_id
-            FROM chunks
+            FROM {fq_table("chunks")}
            WHERE chunk_id = ANY($1)
            """,
            chunk_ids,
@@ -385,9 +371,9 @@ async def tool_expand(
    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 documents
+            FROM {fq_table("documents")}
            WHERE id = ANY($1) AND bank_id = $2
            """,
            all_doc_ids,
@@ -2,36 +2,70 @@
 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_LIST_MENTAL_MODELS = {
+
+TOOL_SEARCH_MENTAL_MODELS = {
    "type": "function",
    "function": {
-        "name": "list_mental_models",
-        "description": "List all available mental models - your synthesized knowledge about entities, concepts, and events. Returns an array of models with id, name, and description.",
+        "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": {},
-            "required": [],
+            "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_GET_MENTAL_MODEL = {
+TOOL_SEARCH_OBSERVATIONS = {
    "type": "function",
    "function": {
-        "name": "get_mental_model",
-        "description": "Get full details of a specific mental model including all observations and memory references.",
+        "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": {
-                "model_id": {
+                "reason": {
                    "type": "string",
-                    "description": "ID of the mental model (from list_mental_models results)",
+                    "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": ["model_id"],
+            "required": ["reason", "query"],
        },
    },
 }
@@ -40,10 +74,19 @@ TOOL_RECALL = {
    "type": "function",
    "function": {
        "name": "recall",
-        "description": "Search memories using semantic + temporal retrieval. Returns relevant memories from experience and world knowledge, each with an 'id' you can reference.",
+        "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",
@@ -53,29 +96,7 @@ TOOL_RECALL = {
                    "description": "Optional limit on result size (default 2048). Use higher values for broader searches.",
                },
            },
-            "required": ["query"],
-        },
-    },
-}
-
-TOOL_LEARN = {
-    "type": "function",
-    "function": {
-        "name": "learn",
-        "description": "Create a new mental model to track an important recurring topic. Use when you discover a person, project, concept, or pattern that appears frequently and would benefit from synthesized knowledge. The model content will be generated automatically.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "name": {
-                    "type": "string",
-                    "description": "Human-readable name (e.g., 'Project Alpha', 'John Smith', 'Product Strategy')",
-                },
-                "description": {
-                    "type": "string",
-                    "description": "What to track and synthesize (e.g., 'Track goals, milestones, blockers, and key decisions for Project Alpha')",
-                },
-            },
-            "required": ["name", "description"],
+            "required": ["reason", "query"],
        },
    },
 }
@@ -88,6 +109,10 @@ TOOL_EXPAND = {
        "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"},
@@ -99,7 +124,7 @@ TOOL_EXPAND = {
                    "description": "chunk: surrounding text chunk, document: full source document",
                },
            },
-            "required": ["memory_ids", "depth"],
+            "required": ["reason", "memory_ids", "depth"],
        },
    },
 }
@@ -114,18 +139,23 @@ TOOL_DONE_ANSWER = {
            "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.",
+                    "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
                },
                "memory_ids": {
                    "type": "array",
                    "items": {"type": "string"},
                    "description": "Array of memory IDs that support your answer (put IDs here, NOT in answer text)",
                },
-                "model_ids": {
+                "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"],
        },
@@ -143,8 +173,6 @@ def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
    Args:
        directive_rules: List of directive rule strings
    """
-    from typing import Any, cast
-
    # Build rules list for description
    rules_list = "\n".join(f"  {i + 1}. {rule}" for i, rule in enumerate(directive_rules))

@@ -162,18 +190,23 @@ def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
                "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.",
+                        "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
                    },
                    "memory_ids": {
                        "type": "array",
                        "items": {"type": "string"},
                        "description": "Array of memory IDs that support your answer (put IDs here, NOT in answer text)",
                    },
-                    "model_ids": {
+                    "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]...'",
@@ -185,29 +218,28 @@ def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
    }


-def get_reflect_tools(enable_learn: bool = True, directive_rules: list[str] | None = None) -> list[dict]:
+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:
-        enable_learn: Whether to include the learn tool
        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 = []
-
-    # Include mental model tools for lookup
-    tools.append(TOOL_LIST_MENTAL_MODELS)
-    tools.append(TOOL_GET_MENTAL_MODEL)
-    tools.append(TOOL_RECALL)
-
-    if enable_learn:
-        tools.append(TOOL_LEARN)
-
-    tools.append(TOOL_EXPAND)
+    tools = [
+        TOOL_SEARCH_MENTAL_MODELS,
+        TOOL_SEARCH_OBSERVATIONS,
+        TOOL_RECALL,
+        TOOL_EXPAND,
+    ]

    # Use directive-aware done tool if directives are present
    if directive_rules:
@@ -10,8 +10,8 @@ from typing import Any

 from pydantic import BaseModel, ConfigDict, Field

-# Valid fact types for recall operations (excludes 'observation' which is internal, and 'opinion' which is deprecated)
-VALID_RECALL_FACT_TYPES = frozenset(["world", "experience"])
+# Valid fact types for recall operations (excludes 'opinion' which is deprecated)
+VALID_RECALL_FACT_TYPES = frozenset(["world", "experience", "observation"])


 class LLMToolCall(BaseModel):
@@ -28,12 +28,15 @@ class LLMToolCallResult(BaseModel):
    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")
@@ -47,13 +50,13 @@ class LLMCallTrace(BaseModel):
    duration_ms: int = Field(description="Execution time in milliseconds")


-class MentalModelRef(BaseModel):
-    """Reference to a mental model accessed during reflect."""
+class ObservationRef(BaseModel):
+    """Reference to an observation accessed during reflect."""

-    id: str = Field(description="Mental model ID")
-    name: str = Field(description="Mental model name")
-    type: str = Field(description="Mental model type: entity, concept, event")
-    subtype: str = Field(description="Mental model subtype: structural, emergent, learned")
+    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)")

@@ -63,7 +66,7 @@ class DirectiveRef(BaseModel):

    id: str = Field(description="Directive mental model ID")
    name: str = Field(description="Directive name")
-    rules: list[str] = Field(default_factory=list, description="Directive rules/observations that were applied")
+    content: str = Field(description="Directive content")


 class TokenUsage(BaseModel):
@@ -166,6 +169,28 @@ 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.
@@ -229,8 +254,15 @@ class ReflectResult(BaseModel):
                    ],
                    "experience": [],
                    "opinion": [],
+                    "mental_models": [],
+                    "directives": [
+                        {
+                            "id": "directive-123",
+                            "name": "Response Style",
+                            "rules": ["Always be concise"],
+                        }
+                    ],
                },
-                "new_opinions": ["Machine learning has great potential in healthcare"],
                "structured_output": {"summary": "ML in healthcare", "confidence": 0.9},
                "usage": {"input_tokens": 1500, "output_tokens": 500, "total_tokens": 2000},
            }
@@ -238,10 +270,9 @@ class ReflectResult(BaseModel):
    )

    text: str = Field(description="The formulated answer text")
-    based_on: dict[str, list[MemoryFact]] = Field(
-        description="Facts used to formulate the answer, organized by type (world, experience, opinion)"
+    based_on: dict[str, Any] = Field(
+        description="Facts used to formulate the answer, organized by type (world, experience, mental_models, directives)"
    )
-    new_opinions: list[str] = Field(default_factory=list, description="List of newly formed opinions during reflection")
    structured_output: dict[str, Any] | None = Field(
        default=None,
        description="Structured output parsed according to the provided response schema. Only present when response_schema was provided.",
@@ -258,34 +289,12 @@ class ReflectResult(BaseModel):
        default_factory=list,
        description="Trace of LLM calls made during reflection. Only present when include.tool_calls is enabled.",
    )
-    mental_models: list[MentalModelRef] = Field(
-        default_factory=list,
-        description="Mental models accessed during reflection, including directives (subtype='directive').",
-    )
    directives_applied: list[DirectiveRef] = Field(
        default_factory=list,
        description="Directive mental models that were applied during this 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):
    """
    An observation about an entity.
@@ -57,21 +57,25 @@ def _infer_temporal_date(fact_text: str, event_date: datetime) -> str | None:
    return None


-def _sanitize_text(text: str) -> str:
+def _sanitize_text(text: str | None) -> str | None:
    """
-    Sanitize text by removing invalid Unicode surrogate characters.
+    Sanitize text by removing characters that break downstream systems.

-    Surrogate characters (U+D800 to U+DFFF) are used in UTF-16 encoding
-    but cannot be encoded in UTF-8. They can appear in Python strings
-    from improperly decoded data (e.g., from JavaScript or broken files).
+    Removes:
+    - Null bytes (\\x00): Invalid in PostgreSQL UTF-8 encoding
+    - Unicode surrogates (U+D800-U+DFFF): Invalid in UTF-8, break LLM APIs

-    This function removes unpaired surrogates to prevent UnicodeEncodeError
-    when the text is sent to the LLM API.
+    Surrogate characters are used in UTF-16 encoding but cannot be encoded
+    in UTF-8. They can appear in Python strings from improperly decoded data
+    (e.g., from JavaScript or broken files). Null bytes commonly appear in
+    OCR output, PDF extraction, or copy-paste from binary sources.
    """
+    if text is None:
+        return None
    if not text:
        return text
-    # Remove surrogate characters (U+D800 to U+DFFF) using regex
-    # These are invalid in UTF-8 and cause encoding errors
+    # Remove null bytes and surrogate characters
+    text = text.replace("\x00", "")
    return re.sub(r"[\ud800-\udfff]", "", text)


@@ -114,11 +118,8 @@ class CausalRelation(BaseModel):
    """Causal relationship from this fact to a previous fact (stored format)."""

    target_fact_index: int = Field(description="Index of the related fact in the facts array (0-based).")
-    relation_type: Literal["caused_by", "enabled_by", "prevented_by"] = Field(
-        description="How this fact relates to the target: "
-        "'caused_by' = this fact was caused by the target, "
-        "'enabled_by' = this fact was enabled by the target, "
-        "'prevented_by' = this fact was prevented by the target"
+    relation_type: Literal["caused_by"] = Field(
+        description="How this fact relates to the target: 'caused_by' = this fact was caused by the target"
    )
    strength: float = Field(
        description="Strength of relationship (0.0 to 1.0)",
@@ -141,11 +142,8 @@ class FactCausalRelation(BaseModel):
        "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", "enabled_by", "prevented_by"] = Field(
-        description="How this fact relates to the target fact: "
-        "'caused_by' = this fact was caused by the target fact, "
-        "'enabled_by' = this fact was enabled by the target fact, "
-        "'prevented_by' = this fact was blocked/prevented by the target fact"
+    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",
@@ -438,34 +436,15 @@ def _chunk_conversation(turns: list[dict], max_chars: int) -> list[str]:
 # FACT EXTRACTION PROMPTS
 # =============================================================================

-# Concise extraction prompt (default) - selective, high-quality facts
-CONCISE_FACT_EXTRACTION_PROMPT = """Extract SIGNIFICANT facts from text. Be SELECTIVE - only extract facts worth remembering long-term.
+# 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.

-LANGUAGE RULE (CRITICAL): Output facts in the EXACT SAME language as the input text. If input is Japanese, output Japanese. If input is Chinese, output Chinese. NEVER translate to English. Preserve original language completely.
+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.

 {fact_types_instruction}

-══════════════════════════════════════════════════════════════════════════
-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.
+{extraction_guidelines}

 ══════════════════════════════════════════════════════════════════════════
 FACT FORMAT - BE CONCISE
@@ -513,7 +492,33 @@ ENTITIES
 ══════════════════════════════════════════════════════════════════════════

 Include: people names, organizations, places, key objects, abstract concepts (career, friendship, etc.)
-Always include "user" when fact is about the user.
+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
@@ -537,7 +542,26 @@ Output: ONLY 2 facts (skip coffee preference - too trivial):
 QUALITY OVER QUANTITY
 ══════════════════════════════════════════════════════════════════════════

-Ask: "Would this be useful to recall in 6 months?" If no, skip it."""
+Ask: "Would this be useful to recall in 6 months?" If no, skip it.
+
+IMPORTANT: Sensory/emotional details and observations that provide meaningful context
+about experiences ARE important to remember, even if they seem small (e.g., how food
+tasted, how someone looked, how loud music was). Extract these if they characterize
+an experience or person."""
+
+# 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)
@@ -622,6 +646,7 @@ For EVENTS (fact_kind="event") - MUST SET BOTH occurred_start AND occurred_end:
 - Convert relative dates → absolute using Event Date as reference
 - If Event Date is "Saturday, March 15, 2020", then "yesterday" = Friday, March 14, 2020
 - Dates mentioned in text (e.g., "in March 2020") should use THAT year, not current year
+- CRITICAL: If the content mentions an absolute date (e.g., "March 15, 2024", "2024-03-15"), you MUST extract it and set occurred_start in ISO format
 - Always include the day name (Monday, Tuesday, etc.) in the 'when' field
 - Set occurred_start AND occurred_end to WHEN IT HAPPENED (not when mentioned)
 - For single-day/point events: set occurred_end = occurred_start (same timestamp)
@@ -662,7 +687,7 @@ CAUSAL RELATIONSHIPS
 ══════════════════════════════════════════════════════════════════════════

 Link facts with causal_relations (max 2 per fact). target_index must be < this fact's index.
-Types: "caused_by", "enabled_by", "prevented_by"
+Type: "caused_by" (this fact was caused by the target fact)

 Example: "Lost job → couldn't pay rent → moved apartment"
 - Fact 0: Lost job, causal_relations: null
@@ -678,7 +703,6 @@ async def _extract_facts_from_chunk(
    context: str,
    llm_config: "LLMConfig",
    agent_name: str = None,
-    extract_opinions: bool = False,
 ) -> tuple[list[dict[str, str]], TokenUsage]:
    """
    Extract facts from a single chunk (internal helper for parallel processing).
@@ -686,17 +710,15 @@ async def _extract_facts_from_chunk(
    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 ""
+    import logging

-    # Determine which fact types to extract based on the flag
+    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
-    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."
-        )
+    fact_types_instruction = "Extract ONLY 'world' and 'assistant' type facts."

    # Check config for extraction mode and causal link extraction
    config = get_config()
@@ -704,13 +726,27 @@ async def _extract_facts_from_chunk(
    extract_causal_links = config.retain_extract_causal_links

    # Select base prompt based on extraction mode
-    if extraction_mode == "verbose":
+    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
-
-    # Format the prompt with fact types instruction
-    prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
+        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
@@ -723,12 +759,6 @@ async def _extract_facts_from_chunk(
    else:
        response_schema = FactExtractionResponseNoCausal

-    import logging
-
-    from openai import BadRequestError
-
-    logger = logging.getLogger(__name__)
-
    # Retry logic for JSON validation errors
    max_retries = 2
    last_error = None
@@ -739,9 +769,12 @@ async def _extract_facts_from_chunk(

    # Build user message with metadata and chunk content in a clear format
    # Format event_date with day of week for better temporal reasoning
+    # Handle both datetime objects and ISO string formats (from deserialized async tasks)
+    from .orchestrator import parse_datetime_flexible
+
+    event_date = parse_datetime_flexible(event_date)
    event_date_formatted = event_date.strftime("%A, %B %d, %Y")  # e.g., "Monday, June 10, 2024"
    user_message = f"""Extract facts from the following text chunk.
-{memory_bank_context}

 Chunk: {chunk_index + 1}/{total_chunks}
 Event Date: {event_date_formatted} ({event_date.isoformat()})
@@ -753,12 +786,28 @@ Text:
    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(
                messages=[{"role": "system", "content": prompt}, {"role": "user", "content": user_message}],
                response_format=response_schema,
-                scope="memory_extract_facts",
+                scope="retain_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,
                skip_validation=True,  # Get raw JSON, we'll validate leniently
                return_usage=True,
            )
@@ -823,7 +872,8 @@ Text:

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

                # Convert "assistant" → "experience" for storage
                if fact_type == "assistant":
@@ -840,7 +890,10 @@ Text:
                    else:
                        # Default to 'world' if we can't determine
                        fact_type = "world"
-                        logger.warning(f"Fact {i}: defaulting to 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})"
+                        )

                # Get fact_kind for temporal handling (but don't store it)
                fact_kind = llm_fact.get("fact_kind", "conversation")
@@ -958,6 +1011,29 @@ Text:

        except BadRequestError as e:
            last_error = e
+            error_str = str(e).lower()
+
+            # Check if error is related to max_tokens/completion_tokens not being supported
+            if any(
+                keyword in error_str
+                for keyword in [
+                    "max_tokens",
+                    "max_completion_tokens",
+                    "maximum context",
+                    "token limit",
+                    "context length",
+                ]
+            ):
+                # Provide helpful error message with configuration suggestions
+                raise ValueError(
+                    f"Model does not support the required output token limit.\n\n"
+                    f"The model '{llm_config.model}' (provider: {llm_config.provider}) failed with: {e}\n\n"
+                    f"You have two options to fix this:\n"
+                    f"  1. Use a different model that supports at least {config.retain_max_completion_tokens} output tokens\n"
+                    f"  2. Decrease HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS to a value your model supports\n"
+                    f"     (current value: {config.retain_max_completion_tokens}, must be > RETAIN_CHUNK_SIZE={config.retain_chunk_size})"
+                ) from e
+
            if "json_validate_failed" in str(e):
                logger.warning(
                    f"          [1.3.{chunk_index + 1}] Attempt {attempt + 1}/{max_retries} failed with JSON validation error: {e}"
@@ -980,7 +1056,6 @@ async def _extract_facts_with_auto_split(
    context: str,
    llm_config: LLMConfig,
    agent_name: str = None,
-    extract_opinions: bool = False,
 ) -> tuple[list[dict[str, str]], TokenUsage]:
    """
    Extract facts from a chunk with automatic splitting if output exceeds token limits.
@@ -996,7 +1071,6 @@ 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)
@@ -1015,7 +1089,6 @@ 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
@@ -1060,7 +1133,6 @@ 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,
@@ -1070,7 +1142,6 @@ async def _extract_facts_with_auto_split(
                context=context,
                llm_config=llm_config,
                agent_name=agent_name,
-                extract_opinions=extract_opinions,
            ),
        ]

@@ -1094,7 +1165,6 @@ async def extract_facts_from_text(
    llm_config: LLMConfig,
    agent_name: str,
    context: str = "",
-    extract_opinions: bool = False,
 ) -> tuple[list[Fact], list[tuple[str, int]], TokenUsage]:
    """
    Extract semantic facts from conversational or narrative text using LLM.
@@ -1111,7 +1181,6 @@ 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:
@@ -1139,7 +1208,6 @@ 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)
    ]
@@ -1171,7 +1239,7 @@ SECONDS_PER_FACT = 10


 async def extract_facts_from_contents(
-    contents: list[RetainContent], llm_config, agent_name: str, extract_opinions: bool = False
+    contents: list[RetainContent], llm_config, agent_name: str
 ) -> tuple[list[ExtractedFactType], list[ChunkMetadata], TokenUsage]:
    """
    Extract facts from multiple content items in parallel.
@@ -1186,7 +1254,6 @@ 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)
@@ -1205,7 +1272,6 @@ 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)

@@ -1310,31 +1376,26 @@ def _convert_causal_relations(relations_from_llm, fact_start_idx: int) -> list[C

 def _add_temporal_offsets(facts: list[ExtractedFactType], contents: list[RetainContent]) -> None:
    """
-    Add time offsets to preserve fact ordering within each content.
+    Add time offsets to preserve fact ordering across all contents.

-    This allows retrieval to distinguish between facts that happened earlier vs later
-    in the same conversation, even when the base event_date is the same.
+    This allows retrieval to distinguish between facts from different documents/conversations
+    even when they have the same base event_date, and also between facts within the same
+    conversation.
+
+    Uses absolute position across all facts to ensure unique timestamps.

    Modifies facts in place.
    """
-    # Group facts by content_index
-    current_content_idx = 0
-    content_fact_start = 0
+    from .orchestrator import parse_datetime_flexible

    for i, fact in enumerate(facts):
-        if fact.content_index != current_content_idx:
-            # Moved to next content
-            current_content_idx = fact.content_index
-            content_fact_start = i
+        # Use absolute position across all facts to ensure uniqueness across different contents
+        offset = timedelta(seconds=i * SECONDS_PER_FACT)

-        # Calculate position within this content
-        fact_position = i - content_fact_start
-        offset = timedelta(seconds=fact_position * SECONDS_PER_FACT)
-
-        # Apply offset to all temporal fields
+        # Apply offset to all temporal fields (handle both datetime objects and ISO strings)
        if fact.occurred_start:
-            fact.occurred_start = fact.occurred_start + offset
+            fact.occurred_start = parse_datetime_flexible(fact.occurred_start) + offset
        if fact.occurred_end:
-            fact.occurred_end = fact.occurred_end + offset
+            fact.occurred_end = parse_datetime_flexible(fact.occurred_end) + offset
        if fact.mentioned_at:
-            fact.mentioned_at = fact.mentioned_at + offset
+            fact.mentioned_at = parse_datetime_flexible(fact.mentioned_at) + offset
@@ -8,6 +8,7 @@ import json
 import logging

 from ..memory_engine import fq_table
+from .fact_extraction import _sanitize_text
 from .types import ProcessedFact

 logger = logging.getLogger(__name__)
@@ -47,7 +48,7 @@ async def insert_facts_batch(
    tags_list = []

    for fact in facts:
-        fact_texts.append(fact.fact_text)
+        fact_texts.append(_sanitize_text(fact.fact_text))
        # Convert embedding to string for asyncpg vector type
        embeddings.append(str(fact.embedding))
        # event_date: Use occurred_start if available, otherwise use mentioned_at
@@ -56,7 +57,7 @@ async def insert_facts_batch(
        occurred_starts.append(fact.occurred_start)
        occurred_ends.append(fact.occurred_end)
        mentioned_ats.append(fact.mentioned_at)
-        contexts.append(fact.context)
+        contexts.append(_sanitize_text(fact.context))
        fact_types.append(fact.fact_type)
        # confidence_score is only for opinion facts
        confidence_scores.append(1.0 if fact.fact_type == "opinion" else None)
@@ -157,7 +158,8 @@ async def handle_document_tracking(
    """
    import hashlib

-    # Calculate content hash
+    # Sanitize and calculate content hash
+    combined_content = _sanitize_text(combined_content) or ""
    content_hash = hashlib.sha256(combined_content.encode()).hexdigest()

    # Always delete old document first if it exists (cascades to units and links)
@@ -754,17 +754,14 @@ 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: "causes", "caused_by", "enables", or "prevents"
+            - relation_type: "caused_by"
            - strength: Float in [0.0, 1.0] representing relationship strength

    Returns:
        Number of causal links created

-    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)
+    Causal link type:
+    - "caused_by": This fact was caused by the target fact
    """
    if not unit_ids or not causal_relations_per_fact:
        return 0
@@ -787,8 +784,8 @@ async def create_causal_links_batch(
                relation_type = relation["relation_type"]
                strength = relation.get("strength", 1.0)

-                # Validate relation_type - must match database constraint
-                valid_types = {"causes", "caused_by", "enables", "prevents"}
+                # Validate relation_type - only "caused_by" is supported (DB constraint)
+                valid_types = {"caused_by"}
                if relation_type not in valid_types:
                    logger.error(
                        f"Invalid relation_type '{relation_type}' (type: {type(relation_type).__name__}) "
@@ -8,6 +8,7 @@ import logging
 import time
 import uuid
 from datetime import UTC, datetime
+from typing import Any

 from ..db_utils import acquire_with_retry
 from . import bank_utils
@@ -18,6 +19,39 @@ def utcnow():
    return datetime.now(UTC)


+def parse_datetime_flexible(value: Any) -> datetime:
+    """
+    Parse a datetime value that could be either a datetime object or an ISO string.
+
+    This handles datetime values from both direct Python calls and deserialized JSON
+    (where datetime objects are serialized as ISO strings).
+
+    Args:
+        value: Either a datetime object or an ISO format string
+
+    Returns:
+        datetime object (timezone-aware)
+
+    Raises:
+        TypeError: If value is neither datetime nor string
+        ValueError: If string is not a valid ISO datetime
+    """
+    if isinstance(value, datetime):
+        # Ensure timezone-aware
+        if value.tzinfo is None:
+            return value.replace(tzinfo=UTC)
+        return value
+    elif isinstance(value, str):
+        # Parse ISO format string (handles both 'Z' and '+00:00' timezone formats)
+        dt = datetime.fromisoformat(value.replace("Z", "+00:00"))
+        # Ensure timezone-aware
+        if dt.tzinfo is None:
+            return dt.replace(tzinfo=UTC)
+        return dt
+    else:
+        raise TypeError(f"Expected datetime or string, got {type(value).__name__}")
+
+
 from ..response_models import TokenUsage
 from . import (
    chunk_storage,
@@ -89,10 +123,18 @@ async def retain_batch(
        # Merge item-level tags with document-level tags
        item_tags = item.get("tags", []) or []
        merged_tags = list(set(item_tags + (document_tags or [])))
+
+        # Handle event_date: parse flexibly (handles both datetime objects and ISO strings)
+        event_date_value = item.get("event_date")
+        if event_date_value:
+            event_date_value = parse_datetime_flexible(event_date_value)
+        else:
+            event_date_value = utcnow()
+
        content = RetainContent(
            content=item["content"],
            context=item.get("context", ""),
-            event_date=item.get("event_date") or utcnow(),
+            event_date=event_date_value,
            metadata=item.get("metadata", {}),
            entities=item.get("entities", []),
            tags=merged_tags,
@@ -101,11 +143,8 @@ async def retain_batch(

    # 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, extract_opinions
-    )
+    extracted_facts, chunks, usage = await fact_extraction.extract_facts_from_contents(contents, llm_config, agent_name)
    log_buffer.append(
        f"[1] Extract facts: {len(extracted_facts)} facts, {len(chunks)} chunks from {len(contents)} contents in {time.time() - step_start:.3f}s"
    )
@@ -119,6 +158,13 @@ async def retain_batch(
                # Handle document tracking even with no facts
                if document_id:
                    combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
+                    # Collect tags from all content items and merge with document_tags
+                    all_tags = set(document_tags or [])
+                    for item in contents_dicts:
+                        item_tags = item.get("tags", []) or []
+                        all_tags.update(item_tags)
+                    merged_tags = list(all_tags)
+
                    retain_params = {}
                    if contents_dicts:
                        first_item = contents_dicts[0]
@@ -133,7 +179,7 @@ async def retain_batch(
                        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
+                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params, merged_tags
                    )
                else:
                    # Check for per-item document_ids
@@ -147,6 +193,13 @@ async def retain_batch(

                    for doc_id, doc_contents in contents_by_doc.items():
                        combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
+                        # Collect tags from all content items for this document and merge with document_tags
+                        all_tags = set(document_tags or [])
+                        for _, item in doc_contents:
+                            item_tags = item.get("tags", []) or []
+                            all_tags.update(item_tags)
+                        merged_tags = list(all_tags)
+
                        retain_params = {}
                        if doc_contents:
                            first_item = doc_contents[0][1]
@@ -161,7 +214,7 @@ async def retain_batch(
                            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
+                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params, merged_tags
                        )

        total_time = time.time() - start_time
@@ -213,6 +266,13 @@ async def retain_batch(
                # Legacy: single document_id parameter
                combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
                retain_params = {}
+                # Collect tags from all content items and merge with document_tags
+                all_tags = set(document_tags or [])
+                for item in contents_dicts:
+                    item_tags = item.get("tags", []) or []
+                    all_tags.update(item_tags)
+                merged_tags = list(all_tags)
+
                if contents_dicts:
                    first_item = contents_dicts[0]
                    if first_item.get("context"):
@@ -227,7 +287,7 @@ 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, merged_tags
                )
                document_ids_added.append(document_id)
                doc_id_mapping[None] = document_id  # For backwards compatibility
@@ -255,6 +315,13 @@ async def retain_batch(
                            # Combine content for this document
                            combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])

+                            # Collect tags from all content items for this document and merge with document_tags
+                            all_tags = set(document_tags or [])
+                            for _, item in doc_contents:
+                                item_tags = item.get("tags", []) or []
+                                all_tags.update(item_tags)
+                            merged_tags = list(all_tags)
+
                            # Extract retain params from first content item
                            retain_params = {}
                            if doc_contents:
@@ -277,7 +344,7 @@ async def retain_batch(
                                combined_content,
                                is_first_batch,
                                retain_params,
-                                document_tags,
+                                merged_tags,
                            )
                            document_ids_added.append(actual_doc_id)

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

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

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

@@ -155,7 +155,6 @@ class LinkExpansionRetriever(GraphRetriever):
                all_seeds.extend(temporal_seeds)

            if not all_seeds:
-                logger.debug("[LinkExpansion] No seeds found, returning empty results")
                return [], timings

            seed_ids = list({s.id for s in all_seeds})
@@ -164,30 +163,102 @@ class LinkExpansionRetriever(GraphRetriever):
            # Run entity and causal expansion sequentially on same connection
            query_start = time.time()

-            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,
-            )
+            # 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"""
@@ -211,11 +282,69 @@ class LinkExpansionRetriever(GraphRetriever):
                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 = 2
-            timings.edge_count = len(entity_rows) + len(causal_rows)
+            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] = {}

@@ -230,6 +359,12 @@ class LinkExpansionRetriever(GraphRetriever):
            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]
@@ -1,134 +0,0 @@
-"""
-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_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
@@ -188,7 +188,7 @@ 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. CRITICAL: ONLY use the facts and information provided in the prompt - do not make up names, events, or information that weren't mentioned. If you don't have enough information to answer, say so. 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."


 async def reflect(
@@ -330,8 +330,8 @@ class SearchTracer:
                RetrievalResult(
                    rank=rank,
                    node_id=doc_id,
-                    text=data.get("text", ""),
-                    context=data.get("context", ""),
+                    text=data.get("text") or "",
+                    context=data.get("context") or "",
                    event_date=data.get("event_date"),
                    fact_type=data.get("fact_type") or fact_type,
                    score=score,
@@ -144,17 +144,21 @@ class BrokerTaskBackend(TaskBackend):
        self,
        pool_getter: Callable[[], "asyncpg.Pool"],
        schema: str | None = None,
+        schema_getter: Callable[[], str | None] | None = None,
    ):
        """
        Initialize the broker task backend.

        Args:
            pool_getter: Callable that returns the asyncpg connection pool
-            schema: Database schema for multi-tenant support (optional)
+            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.
        """
        super().__init__()
        self._pool_getter = pool_getter
        self._schema = schema
+        self._schema_getter = schema_getter

    async def initialize(self):
        """Initialize the backend."""
@@ -178,9 +182,19 @@ class BrokerTaskBackend(TaskBackend):
        operation_id = task_dict.get("operation_id")
        task_type = task_dict.get("type", "unknown")
        bank_id = task_dict.get("bank_id")
-        payload_json = json.dumps(task_dict)

-        table = fq_table("async_operations", self._schema)
+        # Custom encoder to handle datetime objects
+        from datetime import datetime
+
+        def datetime_encoder(obj):
+            if isinstance(obj, datetime):
+                return obj.isoformat()
+            raise TypeError(f"Object of type {type(obj).__name__} is not JSON serializable")
+
+        payload_json = json.dumps(task_dict, default=datetime_encoder)
+
+        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
@@ -231,7 +245,8 @@ class BrokerTaskBackend(TaskBackend):
        import asyncio

        pool = self._pool_getter()
-        table = fq_table("async_operations", self._schema)
+        schema = self._schema_getter() if self._schema_getter else self._schema
+        table = fq_table("async_operations", schema)

        start_time = asyncio.get_event_loop().time()
        while asyncio.get_event_loop().time() - start_time < timeout:
@@ -19,7 +19,6 @@ 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.
@@ -36,7 +35,6 @@ async def extract_facts(
        context: Context about the conversation/document
        llm_config: LLM configuration to use
        agent_name: Optional agent name to help identify agent-related facts
-        extract_opinions: If True, extract ONLY opinions. If False, extract world and agent facts (no opinions)

    Returns:
        Tuple of (facts, chunks) where:
@@ -55,7 +53,6 @@ async def extract_facts(
        context=context,
        llm_config=llm_config,
        agent_name=agent_name,
-        extract_opinions=extract_opinions,
    )

    if not facts:
@@ -65,129 +62,3 @@ 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_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
@@ -16,25 +16,34 @@ with the system (e.g., running migrations for tenant schemas).
 """

 from hindsight_api.extensions.base import Extension
-from hindsight_api.extensions.builtin import ApiKeyTenantExtension
+from hindsight_api.extensions.builtin import ApiKeyTenantExtension, SupabaseTenantExtension
 from hindsight_api.extensions.context import DefaultExtensionContext, ExtensionContext
 from hindsight_api.extensions.http import HttpExtension
 from hindsight_api.extensions.loader import load_extension
+from hindsight_api.extensions.mcp import MCPExtension
 from hindsight_api.extensions.operation_validator import (
+    # Consolidation operation
+    ConsolidateContext,
+    ConsolidateResult,
+    # Mental Model operations
+    MentalModelGetContext,
+    MentalModelGetResult,
+    MentalModelRefreshContext,
+    MentalModelRefreshResult,
+    # Core operations
    OperationValidationError,
    OperationValidatorExtension,
    RecallContext,
    RecallResult,
    ReflectContext,
    ReflectResultContext,
-    RefreshMentalModelContext,
-    RefreshMentalModelResult,
    RetainContext,
    RetainResult,
    ValidationResult,
 )
 from hindsight_api.extensions.tenant import (
    AuthenticationError,
+    Tenant,
    TenantContext,
    TenantExtension,
 )
@@ -49,22 +58,32 @@ __all__ = [
    "DefaultExtensionContext",
    # HTTP Extension
    "HttpExtension",
-    # Operation Validator
+    # MCP Extension
+    "MCPExtension",
+    # Operation Validator - Core
    "OperationValidationError",
    "OperationValidatorExtension",
    "RecallContext",
    "RecallResult",
    "ReflectContext",
    "ReflectResultContext",
-    "RefreshMentalModelContext",
-    "RefreshMentalModelResult",
    "RetainContext",
    "RetainResult",
    "ValidationResult",
+    # Operation Validator - Consolidation
+    "ConsolidateContext",
+    "ConsolidateResult",
+    # Operation Validator - Mental Model
+    "MentalModelGetContext",
+    "MentalModelGetResult",
+    "MentalModelRefreshContext",
+    "MentalModelRefreshResult",
    # Tenant/Auth
    "ApiKeyTenantExtension",
+    "SupabaseTenantExtension",
    "AuthenticationError",
    "RequestContext",
+    "Tenant",
    "TenantContext",
    "TenantExtension",
 ]
@@ -6,13 +6,17 @@ They can be used directly or serve as examples for custom implementations.

 Available built-in extensions:
    - ApiKeyTenantExtension: Simple API key validation with public schema
+    - SupabaseTenantExtension: Supabase JWT validation with per-user schema isolation

 Example usage:
    HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
+    HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.supabase_tenant:SupabaseTenantExtension
 """

+from hindsight_api.extensions.builtin.supabase_tenant import SupabaseTenantExtension
 from hindsight_api.extensions.builtin.tenant import ApiKeyTenantExtension

 __all__ = [
    "ApiKeyTenantExtension",
+    "SupabaseTenantExtension",
 ]
@@ -0,0 +1,433 @@
+"""
+Supabase Tenant Extension for Hindsight
+
+Validates Supabase JWTs and maps authenticated users to isolated memory banks.
+Each user gets their own PostgreSQL schema based on their Supabase user ID.
+
+This extension enables multi-tenant memory isolation for applications using
+Supabase Auth - each authenticated user's memories are stored in a separate
+schema, ensuring complete data isolation.
+
+Features:
+    - Local JWT Verification: Validates tokens locally using JWKS public keys
+      (no network call per request)
+    - Automatic Schema Isolation: Each user gets {prefix}_{user_id} schema
+    - Zero User Management: Leverages your existing Supabase Auth setup
+    - Production Ready: Includes health checks, timeouts, key rotation handling,
+      and error handling
+    - Built-in: Ships with Hindsight, no extra installation needed
+    - Legacy Support: Falls back to /auth/v1/user endpoint for HS256 projects
+
+JWT Verification Strategy:
+    By default, JWTs are verified locally using public keys from the Supabase
+    JWKS endpoint (/auth/v1/.well-known/jwks.json). This is the Supabase-recommended
+    approach: no network call per request, fast, and secure.
+
+    If JWKS keys are unavailable (e.g., legacy HS256 projects), the extension
+    falls back to calling /auth/v1/user per request for validation. This requires
+    the service_role key to be configured.
+
+Configuration via environment variables:
+    HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.supabase_tenant:SupabaseTenantExtension
+    HINDSIGHT_API_TENANT_SUPABASE_URL=https://your-project.supabase.co
+
+    # Optional - only required for legacy HS256 projects or health checks
+    HINDSIGHT_API_TENANT_SUPABASE_SERVICE_KEY=your-service-role-key
+
+    # Optional
+    HINDSIGHT_API_TENANT_SCHEMA_PREFIX=user  # Default: "user" (creates user_<uuid> schemas)
+
+Usage:
+    Clients pass their Supabase JWT in the Authorization header:
+
+    curl -H "Authorization: Bearer <supabase_jwt>" \\
+        https://your-hindsight-server/v1/default/banks/my-bank/memories/recall
+
+Author: BrighterBalance (https://brighterbalance.app)
+License: MIT
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+import time
+
+import httpx
+import jwt as pyjwt
+from jwt import PyJWK
+
+from hindsight_api.extensions.tenant import AuthenticationError, Tenant, TenantContext, TenantExtension
+from hindsight_api.models import RequestContext
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["SupabaseTenantExtension"]
+
+# Minimum expected JWT length (JWTs are typically 100+ characters)
+MIN_TOKEN_LENGTH = 20
+
+# Timeout for Supabase API calls
+REQUEST_TIMEOUT_SECONDS = 10.0
+
+# JWKS cache TTL — Supabase Edge caches JWKS for 10 minutes, so we match that
+JWKS_CACHE_TTL_SECONDS = 600
+
+# Minimum interval between JWKS refreshes to avoid hammering the endpoint
+JWKS_MIN_REFRESH_INTERVAL_SECONDS = 30
+
+# Algorithms supported by Supabase Auth for asymmetric JWT signing
+SUPPORTED_ALGORITHMS = ["RS256", "ES256"]
+
+# Supabase user IDs are UUIDs — validate before using in schema names
+_UUID_RE = re.compile(r"^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$", re.IGNORECASE)
+
+# Schema prefix must be a valid Postgres identifier component (letters, digits, underscores)
+_SCHEMA_PREFIX_RE = re.compile(r"^[a-zA-Z_][a-zA-Z0-9_]*$")
+
+
+class SupabaseTenantExtension(TenantExtension):
+    """
+    TenantExtension that validates Supabase JWTs for multi-tenant isolation.
+
+    Each authenticated user gets their own PostgreSQL schema, ensuring complete
+    memory isolation between users. The schema name is derived from the user's
+    Supabase user ID (the ``sub`` claim in the JWT).
+
+    JWT verification uses JWKS (local, no network call per request) when
+    asymmetric keys are configured in Supabase, and falls back to the
+    ``/auth/v1/user`` endpoint for legacy HS256 projects.
+
+    Example:
+        User with ID "a1b2c3d4-e5f6-7890-abcd-ef1234567890"
+        gets schema "user_a1b2c3d4_e5f6_7890_abcd_ef1234567890"
+    """
+
+    def __init__(self, config: dict[str, str]) -> None:
+        """
+        Initialize with configuration from environment variables.
+
+        Config keys are derived from HINDSIGHT_API_TENANT_* env vars:
+        - HINDSIGHT_API_TENANT_SUPABASE_URL -> config["supabase_url"] (required)
+        - HINDSIGHT_API_TENANT_SUPABASE_SERVICE_KEY -> config["supabase_service_key"] (optional)
+        - HINDSIGHT_API_TENANT_SCHEMA_PREFIX -> config["schema_prefix"] (optional)
+
+        Args:
+            config: Dictionary of configuration values from environment
+
+        Raises:
+            ValueError: If required configuration is missing
+        """
+        super().__init__(config)
+
+        self.supabase_url = (config.get("supabase_url") or "").rstrip("/")
+        self.supabase_service_key = config.get("supabase_service_key")
+        self.schema_prefix = config.get("schema_prefix", "user")
+
+        # Track initialized schemas to avoid redundant migrations
+        self._initialized_schemas: set[str] = set()
+
+        # Reusable HTTP client (created on startup)
+        self._http_client: httpx.AsyncClient | None = None
+
+        # JWKS state
+        self._jwks_keys: dict[str, PyJWK] = {}
+        self._jwks_last_fetched: float = 0
+        self._use_jwks: bool = False
+
+        if not self.supabase_url:
+            raise ValueError(
+                "HINDSIGHT_API_TENANT_SUPABASE_URL is required. "
+                "Set it to your Supabase project URL (e.g., https://xxx.supabase.co)"
+            )
+
+        if not _SCHEMA_PREFIX_RE.match(self.schema_prefix):
+            raise ValueError(
+                f"Invalid schema_prefix '{self.schema_prefix}'. "
+                "Must be a valid Postgres identifier (letters, digits, underscores, starting with a letter or underscore)."
+            )
+
+    # ------------------------------------------------------------------
+    # Lifecycle
+    # ------------------------------------------------------------------
+
+    async def on_startup(self) -> None:
+        """
+        Called when Hindsight starts.
+
+        Creates a reusable HTTP client, fetches JWKS for local JWT verification,
+        and optionally verifies connectivity to Supabase.
+        """
+        logger.info("Initializing Supabase tenant extension")
+        logger.info("Supabase URL: %s", self.supabase_url)
+        logger.info("Schema prefix: %s_", self.schema_prefix)
+
+        self._http_client = httpx.AsyncClient(timeout=REQUEST_TIMEOUT_SECONDS)
+
+        # Attempt to fetch JWKS for fast local JWT verification
+        await self._try_init_jwks()
+
+        # Optional health check using service key
+        if self.supabase_service_key:
+            await self._health_check()
+
+    async def on_shutdown(self) -> None:
+        """Called when Hindsight shuts down. Closes the HTTP client."""
+        logger.info("Shutting down Supabase tenant extension")
+        if self._http_client:
+            await self._http_client.aclose()
+            self._http_client = None
+
+    # ------------------------------------------------------------------
+    # JWKS management
+    # ------------------------------------------------------------------
+
+    async def _try_init_jwks(self) -> None:
+        """Fetch JWKS and decide verification mode (local JWKS vs legacy endpoint)."""
+        try:
+            await self._fetch_jwks()
+            if self._jwks_keys:
+                self._use_jwks = True
+                logger.info(
+                    "JWKS loaded — using local JWT verification with %d key(s)",
+                    len(self._jwks_keys),
+                )
+                return
+
+            # JWKS endpoint returned no keys — project likely uses legacy HS256
+            logger.warning(
+                "JWKS endpoint returned no signing keys. "
+                "Falling back to /auth/v1/user endpoint for JWT verification. "
+                "For better performance, enable asymmetric JWT signing in your "
+                "Supabase dashboard (Project Settings → Auth → JWT Algorithm)."
+            )
+        except Exception as e:
+            logger.warning(
+                "Could not fetch JWKS (%s). Falling back to /auth/v1/user endpoint for JWT verification.",
+                e,
+            )
+
+        # Legacy mode requires service key
+        if not self.supabase_service_key:
+            raise ValueError(
+                "HINDSIGHT_API_TENANT_SUPABASE_SERVICE_KEY is required when JWKS "
+                "is not available. Either enable asymmetric JWT signing in your "
+                "Supabase project or provide the service_role key."
+            )
+        self._use_jwks = False
+
+    async def _fetch_jwks(self) -> None:
+        """Fetch public signing keys from the Supabase JWKS endpoint."""
+        if self._http_client is None:
+            raise RuntimeError("HTTP client not initialized")
+
+        url = f"{self.supabase_url}/auth/v1/.well-known/jwks.json"
+        response = await self._http_client.get(url)
+        response.raise_for_status()
+
+        jwks_data = response.json()
+        keys: dict[str, PyJWK] = {}
+        for key_data in jwks_data.get("keys", []):
+            kid = key_data.get("kid")
+            if kid:
+                keys[kid] = PyJWK(key_data)
+
+        self._jwks_keys = keys
+        self._jwks_last_fetched = time.monotonic()
+
+    async def _get_signing_key(self, token: str) -> PyJWK:
+        """
+        Resolve the signing key for a token from the JWKS cache.
+
+        If the key ID (``kid``) is not in the cache, triggers one JWKS refresh
+        to handle key rotation before raising an error.
+        """
+        header = pyjwt.get_unverified_header(token)
+        kid = header.get("kid")
+        if not kid:
+            raise AuthenticationError("Token missing key ID (kid) header")
+
+        # Refresh cache if stale
+        now = time.monotonic()
+        if now - self._jwks_last_fetched > JWKS_CACHE_TTL_SECONDS:
+            logger.debug("JWKS cache expired, refreshing")
+            await self._fetch_jwks()
+
+        if kid in self._jwks_keys:
+            return self._jwks_keys[kid]
+
+        # Key not found — try one forced refresh to handle key rotation,
+        # but only if we haven't just refreshed
+        if now - self._jwks_last_fetched > JWKS_MIN_REFRESH_INTERVAL_SECONDS:
+            logger.info("Signing key %s not in cache, refreshing JWKS for possible key rotation", kid)
+            await self._fetch_jwks()
+            if kid in self._jwks_keys:
+                return self._jwks_keys[kid]
+
+        raise AuthenticationError("Unable to find signing key for token")
+
+    # ------------------------------------------------------------------
+    # Authentication
+    # ------------------------------------------------------------------
+
+    async def authenticate(self, context: RequestContext) -> TenantContext:
+        """
+        Validate a Supabase JWT and return tenant context.
+
+        Uses local JWKS verification when available (no network call per
+        request), falling back to the ``/auth/v1/user`` endpoint for legacy
+        HS256 projects.
+
+        Args:
+            context: Request context containing the API key (JWT)
+
+        Returns:
+            TenantContext with schema_name set to ``{prefix}_{user_uuid}``
+
+        Raises:
+            AuthenticationError: If token is missing, invalid, or expired
+        """
+        token = context.api_key
+
+        if not token:
+            raise AuthenticationError("Missing Authorization header. Expected: Bearer <supabase_jwt>")
+
+        if len(token) < MIN_TOKEN_LENGTH:
+            raise AuthenticationError("Invalid token format")
+
+        if self._http_client is None:
+            raise AuthenticationError("Extension not initialized")
+
+        # Verify the JWT and extract user ID
+        if self._use_jwks:
+            user_id = await self._verify_token_jwks(token)
+        else:
+            user_id = await self._verify_token_legacy(token)
+
+        # Validate user ID format before using in schema name
+        if not _UUID_RE.match(user_id):
+            raise AuthenticationError("Invalid user ID format in token")
+
+        # Build isolated schema name — hyphens to underscores for Postgres compatibility
+        safe_user_id = user_id.replace("-", "_")
+        schema_name = f"{self.schema_prefix}_{safe_user_id}"
+
+        # Initialize schema on first access
+        if schema_name not in self._initialized_schemas:
+            await self._initialize_schema(schema_name)
+
+        return TenantContext(schema_name=schema_name)
+
+    async def _verify_token_jwks(self, token: str) -> str:
+        """
+        Verify a JWT locally using cached JWKS public keys.
+
+        Validates signature, expiration, issuer, and audience. Returns the
+        user ID from the ``sub`` claim.
+
+        Raises:
+            AuthenticationError: If the token is invalid or expired.
+        """
+        try:
+            signing_key = await self._get_signing_key(token)
+            payload = pyjwt.decode(
+                token,
+                signing_key.key,
+                algorithms=SUPPORTED_ALGORITHMS,
+                audience="authenticated",
+                issuer=f"{self.supabase_url}/auth/v1",
+            )
+        except pyjwt.ExpiredSignatureError:
+            raise AuthenticationError("Token has expired")
+        except pyjwt.InvalidAudienceError:
+            raise AuthenticationError("Invalid token audience")
+        except pyjwt.InvalidIssuerError:
+            raise AuthenticationError("Invalid token issuer")
+        except pyjwt.DecodeError:
+            raise AuthenticationError("Invalid token")
+        except AuthenticationError:
+            raise
+        except Exception as e:
+            raise AuthenticationError(f"Token verification failed: {e!s}")
+
+        user_id = payload.get("sub")
+        if not user_id:
+            raise AuthenticationError("Token valid but missing subject (sub) claim")
+        return user_id
+
+    async def _verify_token_legacy(self, token: str) -> str:
+        """
+        Verify a JWT by calling the Supabase ``/auth/v1/user`` endpoint.
+
+        This is the fallback for projects using legacy HS256 JWT signing.
+        Adds a network round-trip per request.
+
+        Raises:
+            AuthenticationError: If the token is invalid or the request fails.
+        """
+        try:
+            response = await self._http_client.get(
+                f"{self.supabase_url}/auth/v1/user",
+                headers={
+                    "Authorization": f"Bearer {token}",
+                    "apikey": self.supabase_service_key,
+                },
+            )
+
+            if response.status_code == 401:
+                raise AuthenticationError("Invalid or expired token")
+
+            if response.status_code != 200:
+                raise AuthenticationError(f"Authentication failed: {response.status_code}")
+
+            user_data = response.json()
+            user_id = user_data.get("id")
+
+            if not user_id:
+                raise AuthenticationError("Token valid but no user ID found")
+
+            return user_id
+
+        except AuthenticationError:
+            raise
+        except httpx.TimeoutException:
+            raise AuthenticationError("Authentication timeout - please retry")
+        except httpx.RequestError as e:
+            raise AuthenticationError(f"Connection error: {e!s}")
+
+    # ------------------------------------------------------------------
+    # Schema management
+    # ------------------------------------------------------------------
+
+    async def _initialize_schema(self, schema_name: str) -> None:
+        """Run migrations for a new tenant schema and cache the result."""
+        logger.info("Initializing schema: %s", schema_name)
+        try:
+            await self.context.run_migration(schema_name)
+            self._initialized_schemas.add(schema_name)
+            logger.info("Schema ready: %s", schema_name)
+        except Exception as e:
+            logger.error("Schema initialization failed for %s: %s", schema_name, e)
+            raise AuthenticationError(f"Failed to initialize tenant: {e!s}")
+
+    async def list_tenants(self) -> list[Tenant]:
+        """Return all tenant schemas that have been initialized."""
+        return [Tenant(schema=schema) for schema in self._initialized_schemas]
+
+    # ------------------------------------------------------------------
+    # Health check
+    # ------------------------------------------------------------------
+
+    async def _health_check(self) -> None:
+        """Verify connectivity to Supabase using the auth health endpoint."""
+        try:
+            response = await self._http_client.get(
+                f"{self.supabase_url}/auth/v1/health",
+                headers={"apikey": self.supabase_service_key},
+            )
+            if response.status_code == 200:
+                logger.info("Supabase connection verified")
+            else:
+                logger.warning("Supabase health check returned %d", response.status_code)
+        except Exception as e:
+            logger.warning("Could not verify Supabase connection: %s", e)
@@ -1,20 +1,60 @@
 """Built-in tenant extension implementations."""

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


+class DefaultTenantExtension(TenantExtension):
+    """
+    Default single-tenant extension with no authentication.
+
+    This is the default extension used when no tenant extension is configured.
+    It provides single-tenant behavior using the configured schema from
+    HINDSIGHT_API_DATABASE_SCHEMA (defaults to 'public').
+
+    Features:
+    - No authentication required (passes all requests)
+    - Uses configured schema from environment
+    - Perfect for single-tenant deployments without auth
+
+    Configuration:
+        HINDSIGHT_API_DATABASE_SCHEMA=your-schema (optional, defaults to 'public')
+
+    This is automatically enabled by default. To use custom authentication,
+    configure a different tenant extension:
+        HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
+    """
+
+    def __init__(self, config: dict[str, str]):
+        super().__init__(config)
+        # Cache the schema at initialization for consistency
+        # Support explicit schema override via config, otherwise use environment
+        self._schema = config.get("schema", get_config().database_schema)
+
+    async def authenticate(self, context: RequestContext) -> TenantContext:
+        """Return configured schema without any authentication."""
+        return TenantContext(schema_name=self._schema)
+
+    async def list_tenants(self) -> list[Tenant]:
+        """Return configured schema for single-tenant setup."""
+        return [Tenant(schema=self._schema)]
+
+
 class ApiKeyTenantExtension(TenantExtension):
    """
    Built-in tenant extension that validates API key against an environment variable.

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

    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.
@@ -25,9 +65,26 @@ class ApiKeyTenantExtension(TenantExtension):
        self.expected_api_key = config.get("api_key")
        if not self.expected_api_key:
            raise ValueError("HINDSIGHT_API_TENANT_API_KEY is required when using ApiKeyTenantExtension")
+        # Allow disabling MCP auth for backwards compatibility
+        self.mcp_auth_disabled = config.get("mcp_auth_disabled", "").lower() in ("true", "1", "yes")

    async def authenticate(self, context: RequestContext) -> TenantContext:
-        """Validate API key and return public schema context."""
+        """Validate API key and return configured schema context."""
        if context.api_key != self.expected_api_key:
            raise AuthenticationError("Invalid API key")
-        return TenantContext(schema_name="public")
+        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)]
+
+    async def authenticate_mcp(self, context: RequestContext) -> TenantContext:
+        """
+        Authenticate MCP requests.
+
+        If mcp_auth_disabled is set, skip authentication for backwards compatibility.
+        Otherwise, delegate to authenticate().
+        """
+        if self.mcp_auth_disabled:
+            return TenantContext(schema_name=get_config().database_schema)
+        return await self.authenticate(context)
@@ -0,0 +1,42 @@
+"""MCP Extension for registering additional MCP tools.
+
+This extension allows external packages (like hindsight-cloud) to register
+additional MCP tools on the Hindsight MCP server.
+
+Example:
+    HINDSIGHT_API_MCP_EXTENSION=hindsight_cloud.extensions:CloudMCPExtension
+"""
+
+import logging
+from abc import abstractmethod
+
+from fastmcp import FastMCP
+
+from hindsight_api import MemoryEngine
+from hindsight_api.extensions.base import Extension
+
+logger = logging.getLogger(__name__)
+
+
+class MCPExtension(Extension):
+    """Base class for MCP extensions that register additional tools.
+
+    Subclass this to add MCP tools in extension packages.
+
+    Example:
+        class CloudMCPExtension(MCPExtension):
+            def register_tools(self, mcp: FastMCP, memory: MemoryEngine) -> None:
+                @mcp.tool()
+                async def my_custom_tool(query: str) -> str:
+                    return "result"
+    """
+
+    @abstractmethod
+    def register_tools(self, mcp: FastMCP, memory: MemoryEngine) -> None:
+        """Register additional MCP tools.
+
+        Args:
+            mcp: FastMCP server instance to register tools on
+            memory: MemoryEngine instance for accessing memory operations
+        """
+        pass
@@ -1,4 +1,4 @@
-"""Operation Validator Extension for validating retain/recall/reflect operations."""
+"""Operation Validator Extension for validating retain/recall/reflect/consolidate operations."""

 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
@@ -97,15 +97,16 @@ class ReflectContext:
    context: str | None = None


-@dataclass
-class RefreshMentalModelContext:
-    """Context for a refresh mental model operation validation (pre-operation).
+# =============================================================================
+# Consolidation Pre-operation Context
+# =============================================================================

-    Contains ALL user-provided parameters for the refresh mental model operation.
-    """
+
+@dataclass
+class ConsolidateContext:
+    """Context for a consolidation operation validation (pre-operation)."""

    bank_id: str
-    model_id: str
    request_context: "RequestContext"


@@ -131,6 +132,10 @@ class RetainResult:
    unit_ids: list[list[str]]  # List of unit IDs per content item
    success: bool = True
    error: str | None = None
+    # Actual LLM token usage (populated by engine when available)
+    llm_input_tokens: int | None = None
+    llm_output_tokens: int | None = None
+    llm_total_tokens: int | None = None


@dataclass
@@ -176,30 +181,79 @@ class ReflectResultContext:
    error: str | None = None


-@dataclass
-class RefreshMentalModelResult:
-    """Result context for post-refresh-mental-model hook.
+# =============================================================================
+# Consolidation Post-operation Context
+# =============================================================================

-    Contains the operation parameters and the result including token usage.
-    """
+
+@dataclass
+class ConsolidateResult:
+    """Result context for post-consolidation hook."""

    bank_id: str
-    model_id: str
    request_context: "RequestContext"
    # Result
-    model_name: str | None = None
-    observations_count: int = 0
-    input_tokens: int = 0
-    output_tokens: int = 0
-    total_tokens: int = 0
-    duration_ms: int = 0
+    processed: int = 0
+    created: int = 0
+    updated: int = 0
+    success: bool = True
+    error: str | None = None
+
+
+# =============================================================================
+# Mental Model Contexts
+# =============================================================================
+
+
+@dataclass
+class MentalModelGetContext:
+    """Context for a mental model GET operation validation (pre-operation)."""
+
+    bank_id: str
+    mental_model_id: str
+    request_context: "RequestContext"
+
+
+@dataclass
+class MentalModelRefreshContext:
+    """Context for a mental model refresh/create operation validation (pre-operation)."""
+
+    bank_id: str
+    mental_model_id: str | None  # None for create (not yet assigned)
+    request_context: "RequestContext"
+
+
+@dataclass
+class MentalModelGetResult:
+    """Result context for post-mental-model-GET hook."""
+
+    bank_id: str
+    mental_model_id: str
+    request_context: "RequestContext"
+    output_tokens: int  # tokens in the returned content
+    success: bool = True
+    error: str | None = None
+
+
+@dataclass
+class MentalModelRefreshResult:
+    """Result context for post-mental-model-refresh hook."""
+
+    bank_id: str
+    mental_model_id: str
+    request_context: "RequestContext"
+    query_tokens: int  # tokens in source_query
+    output_tokens: int  # tokens in generated content
+    context_tokens: int  # tokens in context (if any)
+    facts_used: int  # facts referenced in based_on
+    mental_models_used: int  # mental models referenced in based_on
    success: bool = True
    error: str | None = None


 class OperationValidatorExtension(Extension, ABC):
    """
-    Validates and hooks into retain/recall/reflect operations.
+    Validates and hooks into retain/recall/reflect/consolidate operations.

    This extension allows implementing custom logic such as:
    - Rate limiting (pre-operation)
@@ -218,9 +272,13 @@ class OperationValidatorExtension(Extension, ABC):
        -> config = {"max_requests": "100"}

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

    # =========================================================================
@@ -298,25 +356,6 @@ class OperationValidatorExtension(Extension, ABC):
        """
        ...

-    @abstractmethod
-    async def validate_refresh_mental_model(self, ctx: RefreshMentalModelContext) -> ValidationResult:
-        """
-        Validate a refresh mental model operation before execution.
-
-        Called before the refresh mental model operation is processed.
-        Return ValidationResult.reject() to prevent the operation from executing.
-
-        Args:
-            ctx: Context containing all user-provided parameters:
-                - bank_id: Bank identifier
-                - model_id: Mental model ID to refresh
-                - request_context: Request context with auth info
-
-        Returns:
-            ValidationResult indicating whether the operation is allowed.
-        """
-        ...
-
    # =========================================================================
    # Post-operation hooks (optional - override to implement)
    # =========================================================================
@@ -378,26 +417,120 @@ class OperationValidatorExtension(Extension, ABC):
        """
        pass

-    async def on_refresh_mental_model_complete(self, result: RefreshMentalModelResult) -> None:
-        """
-        Called after a refresh mental model operation completes (success or failure).
+    # =========================================================================
+    # Consolidation - Pre-operation validation hook (optional - override to implement)
+    # =========================================================================

-        Override this method to implement post-operation logic such as:
-        - Token usage tracking and billing
-        - Audit logging
-        - Metrics collection
+    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
-                - model_id: Mental model ID
+                - 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
+
+    # =========================================================================
+    # Mental Model - Pre-operation validation hook (optional - override to implement)
+    # =========================================================================
+
+    async def validate_mental_model_get(self, ctx: MentalModelGetContext) -> ValidationResult:
+        """
+        Validate a mental model GET operation before execution.
+
+        Override to implement custom validation logic for mental model retrieval.
+
+        Args:
+            ctx: Context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier
                - request_context: Request context with auth info
-                - model_name: Name of the mental model (if success)
-                - observations_count: Number of observations generated
-                - input_tokens: Number of input tokens used
-                - output_tokens: Number of output tokens used
-                - total_tokens: Total tokens used (input + output)
-                - duration_ms: Total operation duration in milliseconds
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        return ValidationResult.accept()
+
+    async def validate_mental_model_refresh(self, ctx: MentalModelRefreshContext) -> ValidationResult:
+        """
+        Validate a mental model refresh/create operation before execution.
+
+        Override to implement custom validation logic for mental model refresh.
+
+        Args:
+            ctx: Context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier (None for create)
+                - request_context: Request context with auth info
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        return ValidationResult.accept()
+
+    # =========================================================================
+    # Mental Model - Post-operation hooks (optional - override to implement)
+    # =========================================================================
+
+    async def on_mental_model_get_complete(self, result: MentalModelGetResult) -> None:
+        """
+        Called after a mental model GET operation completes (success or failure).
+
+        Override to implement post-operation logic such as tracking or audit logging.
+
+        Args:
+            result: Result context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier
+                - output_tokens: Token count of the returned content
+                - success: Whether the operation succeeded
+                - error: Error message (if failed)
+        """
+        pass
+
+    async def on_mental_model_refresh_complete(self, result: MentalModelRefreshResult) -> None:
+        """
+        Called after a mental model refresh operation completes (success or failure).
+
+        Override to implement post-operation logic such as tracking or audit logging.
+
+        Args:
+            result: Result context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier
+                - query_tokens: Tokens in source_query
+                - output_tokens: Tokens in generated content
+                - context_tokens: Tokens in context
+                - facts_used: Number of facts referenced
+                - mental_models_used: Number of mental models referenced
                - success: Whether the operation succeeded
                - error: Error message (if failed)
        """
@@ -28,6 +28,18 @@ 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.
@@ -61,3 +73,36 @@ 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")].
+        """
+        ...
+
+    async def authenticate_mcp(self, context: RequestContext) -> TenantContext:
+        """
+        Authenticate MCP requests.
+
+        By default, this calls authenticate(). Override this method to provide
+        different authentication behavior for MCP endpoints (e.g., to disable
+        auth for backwards compatibility with existing MCP servers).
+
+        Args:
+            context: The action context containing API key and other auth data.
+
+        Returns:
+            TenantContext with the schema_name for database operations.
+
+        Raises:
+            AuthenticationError: If authentication fails.
+        """
+        return await self.authenticate(context)
@@ -20,14 +20,13 @@ import warnings

 import uvicorn

-from . import MemoryEngine
+from . import MemoryEngine, __version__
 from .api import create_app
 from .banner import print_banner
 from .config import DEFAULT_WORKERS, ENV_WORKERS, HindsightConfig, get_config
 from .daemon import (
    DEFAULT_DAEMON_PORT,
    DEFAULT_IDLE_TIMEOUT,
-    DaemonLock,
    IdleTimeoutMiddleware,
    daemonize,
 )
@@ -136,30 +135,15 @@ def main():

    # Daemon mode handling
    if args.daemon:
-        # Use fixed daemon port
-        args.port = DEFAULT_DAEMON_PORT
+        # Use port from args (may be custom for profiles)
+        if args.port == config.port:  # No custom port specified
+            args.port = DEFAULT_DAEMON_PORT
        args.host = "127.0.0.1"  # Only bind to localhost for security

-        # Check if another daemon is already running
-        daemon_lock = DaemonLock()
-        if not daemon_lock.acquire():
-            print(f"Daemon already running (PID: {daemon_lock.get_pid()})", file=sys.stderr)
-            sys.exit(1)
-
        # Fork into background
+        # No lockfile needed - port binding prevents duplicate daemons
        daemonize()

-        # Re-acquire lock in child process
-        daemon_lock = DaemonLock()
-        if not daemon_lock.acquire():
-            sys.exit(1)
-
-        # Register cleanup to release lock
-        def release_lock():
-            daemon_lock.release()
-
-        atexit.register(release_lock)
-
    # Print banner (not in daemon mode)
    if not args.daemon:
        print()
@@ -170,32 +154,73 @@ 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_local_trust_remote_code=config.embeddings_local_trust_remote_code,
            embeddings_tei_url=config.embeddings_tei_url,
            embeddings_openai_base_url=config.embeddings_openai_base_url,
+            embeddings_cohere_api_key=config.embeddings_cohere_api_key,
+            embeddings_cohere_model=config.embeddings_cohere_model,
            embeddings_cohere_base_url=config.embeddings_cohere_base_url,
+            embeddings_litellm_api_base=config.embeddings_litellm_api_base,
+            embeddings_litellm_api_key=config.embeddings_litellm_api_key,
+            embeddings_litellm_model=config.embeddings_litellm_model,
            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_local_trust_remote_code=config.reranker_local_trust_remote_code,
            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_api_key=config.reranker_cohere_api_key,
+            reranker_cohere_model=config.reranker_cohere_model,
            reranker_cohere_base_url=config.reranker_cohere_base_url,
+            reranker_litellm_api_base=config.reranker_litellm_api_base,
+            reranker_litellm_api_key=config.reranker_litellm_api_key,
+            reranker_litellm_model=config.reranker_litellm_model,
            host=args.host,
            port=args.port,
            log_level=args.log_level,
@@ -205,13 +230,14 @@ def main():
            mpfp_top_k_neighbors=config.mpfp_top_k_neighbors,
            recall_max_concurrent=config.recall_max_concurrent,
            recall_connection_budget=config.recall_connection_budget,
-            observation_min_facts=config.observation_min_facts,
-            observation_top_entities=config.observation_top_entities,
            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_observations_async=config.retain_observations_async,
+            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,
@@ -223,10 +249,16 @@ def main():
            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,
+            worker_max_slots=config.worker_max_slots,
+            worker_consolidation_max_slots=config.worker_consolidation_max_slots,
            reflect_max_iterations=config.reflect_max_iterations,
            mental_model_refresh_concurrency=config.mental_model_refresh_concurrency,
+            otel_traces_enabled=config.otel_traces_enabled,
+            otel_exporter_otlp_endpoint=config.otel_exporter_otlp_endpoint,
+            otel_exporter_otlp_headers=config.otel_exporter_otlp_headers,
+            otel_service_name=config.otel_service_name,
+            otel_deployment_environment=config.otel_deployment_environment,
        )
    config.configure_logging()
    if not args.daemon:
@@ -331,11 +363,13 @@ def main():
            embeddings_provider=config.embeddings_provider,
            reranker_provider=config.reranker_provider,
            mcp_enabled=config.mcp_enabled,
+            version=__version__,
        )

    # 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():
@@ -346,12 +380,12 @@ def main():
                loop = asyncio.new_event_loop()
                asyncio.set_event_loop(loop)
                loop.run_until_complete(idle_middleware._check_idle())
-            except Exception:
-                pass
+            except Exception as e:
+                logging.error(f"Idle checker error: {e}", exc_info=True)

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

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


 if __name__ == "__main__":
@@ -44,7 +44,6 @@ 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,
@@ -53,6 +52,7 @@ 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
@@ -85,9 +85,6 @@ 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:
@@ -105,55 +102,17 @@ def create_local_mcp_server(bank_id: str, memory=None) -> FastMCP:

    mcp = FastMCP("hindsight")

-    @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
+    # 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
+    )

-        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": []}
+    register_mcp_tools(mcp, memory, config)

    return mcp

@@ -189,7 +189,7 @@ class MetricsCollectorBase:
        Args:
            provider: LLM provider name (openai, anthropic, gemini, groq, ollama, lmstudio)
            model: Model name
-            scope: Scope identifier (e.g., "memory", "reflect", "entity_observation")
+            scope: Scope identifier (e.g., "memory", "reflect", "consolidation")
            duration: Call duration in seconds
            input_tokens: Number of input/prompt tokens
            output_tokens: Number of output/completion tokens
@@ -321,7 +321,7 @@ class MetricsCollector(MetricsCollectorBase):
                pass

        Args:
-            operation: Operation name (retain, recall, reflect, entity_observation)
+            operation: Operation name (retain, recall, reflect, consolidation)
            bank_id: Memory bank ID
            source: Source of the operation (api, reflect, internal)
            budget: Optional budget level (low, mid, high)
@@ -371,7 +371,7 @@ class MetricsCollector(MetricsCollectorBase):
        Args:
            provider: LLM provider name (openai, anthropic, gemini, groq, ollama, lmstudio)
            model: Model name
-            scope: Scope identifier (e.g., "memory", "reflect", "entity_observation")
+            scope: Scope identifier (e.g., "memory", "reflect", "consolidation")
            duration: Call duration in seconds
            input_tokens: Number of input/prompt tokens
            output_tokens: Number of output/completion tokens
@@ -25,6 +25,8 @@ from alembic.config import Config
 from alembic.script.revision import ResolutionError
 from sqlalchemy import create_engine, text

+from .utils import mask_network_location
+
 logger = logging.getLogger(__name__)

 # Advisory lock ID for migrations (arbitrary unique number)
@@ -54,7 +56,7 @@ def _run_migrations_internal(database_url: str, script_location: str, schema: st
    """
    schema_name = schema or "public"
    logger.info(f"Running database migrations to head for schema '{schema_name}'...")
-    logger.info(f"Database URL: {database_url}")
+    logger.info(f"Database URL: {mask_network_location(database_url)}")
    logger.info(f"Script location: {script_location}")

    # Create Alembic configuration programmatically (no alembic.ini needed)
@@ -165,6 +167,81 @@ def run_migrations(
            logger.debug("Migration advisory lock acquired")

            try:
+                # Ensure pgvector extension is installed globally BEFORE schema migrations
+                # This is critical: the extension must exist database-wide before any schema
+                # migrations run, otherwise custom schemas won't have access to vector types
+                logger.debug("Checking pgvector extension availability...")
+
+                # First, check if extension already exists
+                ext_check = conn.execute(
+                    text(
+                        "SELECT extname, nspname FROM pg_extension e "
+                        "JOIN pg_namespace n ON e.extnamespace = n.oid "
+                        "WHERE extname = 'vector'"
+                    )
+                ).fetchone()
+
+                if ext_check:
+                    # Extension exists - check if in correct schema
+                    ext_schema = ext_check[1]
+                    if ext_schema == "public":
+                        logger.info("pgvector extension found in public schema - ready to use")
+                    else:
+                        # Extension in wrong schema - try to fix if we have permissions
+                        logger.warning(
+                            f"pgvector extension found in schema '{ext_schema}' instead of 'public'. "
+                            f"Attempting to relocate..."
+                        )
+                        try:
+                            conn.execute(text("DROP EXTENSION vector CASCADE"))
+                            conn.execute(text("SET search_path TO public"))
+                            conn.execute(text("CREATE EXTENSION vector"))
+                            conn.commit()
+                            logger.info("pgvector extension relocated to public schema")
+                        except Exception as e:
+                            # Failed to relocate - log but don't fail if extension exists somewhere
+                            logger.warning(
+                                f"Could not relocate pgvector extension to public schema: {e}. "
+                                f"Continuing with extension in '{ext_schema}' schema."
+                            )
+                            conn.rollback()
+                else:
+                    # Extension doesn't exist - try to install
+                    logger.info("pgvector extension not found, attempting to install...")
+                    try:
+                        conn.execute(text("SET search_path TO public"))
+                        conn.execute(text("CREATE EXTENSION vector"))
+                        conn.commit()
+                        logger.info("pgvector extension installed in public schema")
+                    except Exception as e:
+                        # Installation failed - this is only fatal if extension truly doesn't exist
+                        # Check one more time in case another process installed it
+                        conn.rollback()
+                        ext_recheck = conn.execute(
+                            text(
+                                "SELECT nspname FROM pg_extension e "
+                                "JOIN pg_namespace n ON e.extnamespace = n.oid "
+                                "WHERE extname = 'vector'"
+                            )
+                        ).fetchone()
+
+                        if ext_recheck:
+                            logger.warning(
+                                f"Could not install pgvector extension (permission denied?), "
+                                f"but extension exists in '{ext_recheck[0]}' schema. Continuing..."
+                            )
+                        else:
+                            # Extension truly doesn't exist and we can't install it
+                            logger.error(
+                                f"pgvector extension is not installed and cannot be installed: {e}. "
+                                f"Please ensure pgvector is installed by a database administrator. "
+                                f"See: https://github.com/pgvector/pgvector#installation"
+                            )
+                            raise RuntimeError(
+                                "pgvector extension is required but not installed. "
+                                "Please install it with: CREATE EXTENSION vector;"
+                            ) from e
+
                # Run migrations while holding the lock
                _run_migrations_internal(database_url, script_location, schema=schema)
            finally:
@@ -20,7 +20,8 @@ class RequestContext:
    api_key: str | None = None
    api_key_id: str | None = None  # UUID of the API key used for authentication
    tenant_id: str | None = None  # Tenant identifier (set by extension after auth)
-    internal: bool = False  # True for background/internal operations (not user-visible)
+    internal: bool = False  # True for background/internal operations (skips extension auth)
+    user_initiated: bool = False  # True for async operations that originated from a user request


 from pgvector.sqlalchemy import Vector
@@ -40,7 +40,7 @@ class EmbeddedPostgres:
            # Only set port if explicitly specified
            if self.port is not None:
                kwargs["port"] = self.port
-            self._pg0 = Pg0(**kwargs)  # type: ignore[invalid-argument-type] - dict kwargs
+            self._pg0 = Pg0(**kwargs)
        return self._pg0

    async def start(self, max_retries: int = 5, retry_delay: float = 4.0) -> str:
@@ -0,0 +1,480 @@
+"""
+OpenTelemetry distributed tracing instrumentation for Hindsight API.
+
+This module provides tracing for:
+- LLM API calls with full prompts/completions following GenAI semantic conventions
+- Token usage and model information
+- Error tracking and finish reasons
+
+Tracing is conditional and disabled by default. When enabled, traces are exported
+to Langfuse (or any OTLP-compatible backend) via OTLP HTTP protocol.
+"""
+
+import json
+import logging
+import time
+from typing import Any, Optional
+
+from opentelemetry import trace
+from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+from opentelemetry.sdk.resources import Resource
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.sdk.trace.export import BatchSpanProcessor
+from opentelemetry.trace import Status, StatusCode
+
+logger = logging.getLogger(__name__)
+
+
+def _serialize_for_span(obj: Any) -> str:
+    """Serialize an object for span recording, handling Pydantic models."""
+    if isinstance(obj, str):
+        return obj
+    if hasattr(obj, "model_dump_json"):
+        # Pydantic v2 model
+        return obj.model_dump_json()
+    if hasattr(obj, "json"):
+        # Pydantic v1 model
+        return obj.json()
+    if hasattr(obj, "model_dump"):
+        # Pydantic v2 model - convert to dict then json
+        return json.dumps(obj.model_dump())
+    if hasattr(obj, "dict"):
+        # Pydantic v1 model - convert to dict then json
+        return json.dumps(obj.dict())
+    # Fallback to json.dumps for dicts and other types
+    return json.dumps(obj)
+
+
+# No-op tracer for when tracing is disabled
+class NoOpTracer:
+    """No-op tracer that provides the same interface as OpenTelemetry Tracer but does nothing."""
+
+    def start_as_current_span(self, name: str, **kwargs):
+        """Return a no-op context manager that yields a NoOpSpan."""
+        from contextlib import contextmanager
+
+        @contextmanager
+        def noop_span_context():
+            yield NoOpSpan()
+
+        return noop_span_context()
+
+    def start_span(self, name: str, **kwargs):
+        """Return a no-op span."""
+        return NoOpSpan()
+
+
+class NoOpSpan:
+    """No-op span that provides the same interface as OpenTelemetry Span but does nothing."""
+
+    def set_attribute(self, key: str, value: Any) -> None:
+        """No-op."""
+        pass
+
+    def set_status(self, status: Any) -> None:
+        """No-op."""
+        pass
+
+    def record_exception(self, exception: Exception) -> None:
+        """No-op."""
+        pass
+
+    def add_event(self, name: str, attributes: dict | None = None) -> None:
+        """No-op."""
+        pass
+
+    def end(self, end_time: int | None = None) -> None:
+        """No-op."""
+        pass
+
+
+# Global tracer instance
+_tracer: trace.Tracer | NoOpTracer = NoOpTracer()
+_tracing_enabled: bool = False
+
+
+# GenAI semantic convention attribute names (based on v1.37 spec)
+class GenAIAttributes:
+    """GenAI semantic convention attribute names."""
+
+    # Operation and provider
+    OPERATION_NAME = "gen_ai.operation.name"
+    PROVIDER_NAME = "gen_ai.provider.name"
+
+    # Model information
+    REQUEST_MODEL = "gen_ai.request.model"
+    RESPONSE_MODEL = "gen_ai.response.model"
+
+    # Token usage
+    USAGE_INPUT_TOKENS = "gen_ai.usage.input_tokens"
+    USAGE_OUTPUT_TOKENS = "gen_ai.usage.output_tokens"
+
+    # Messages and prompts
+    SYSTEM_INSTRUCTIONS = "gen_ai.system_instructions"
+    INPUT_MESSAGES = "gen_ai.input.messages"
+    OUTPUT_MESSAGES = "gen_ai.output.messages"
+
+    # Response metadata
+    FINISH_REASONS = "gen_ai.response.finish_reasons"
+
+    # Error tracking
+    ERROR_TYPE = "error.type"
+
+
+# Provider name mapping (Hindsight internal -> GenAI semantic convention)
+PROVIDER_NAME_MAPPING = {
+    "openai": "openai",
+    "anthropic": "anthropic",
+    "gemini": "google",
+    "vertexai": "google",
+    "groq": "groq",
+    "ollama": "ollama",
+    "lmstudio": "lmstudio",
+    "openai-codex": "openai",
+    "claude-code": "anthropic",
+    "mock": "mock",
+}
+
+
+def initialize_tracing(
+    service_name: str,
+    endpoint: str,
+    headers: Optional[str] = None,
+    deployment_environment: str = "development",
+) -> None:
+    """
+    Initialize OpenTelemetry tracing with OTLP exporter.
+
+    Args:
+        service_name: Name of the service for resource attributes
+        endpoint: OTLP endpoint URL (e.g., https://cloud.langfuse.com/api/public/otel)
+        headers: Optional headers in format "key1=value1,key2=value2"
+        deployment_environment: Deployment environment (e.g., development, staging, production)
+    """
+    global _tracer, _tracing_enabled
+
+    # Create resource with service information
+    resource = Resource.create(
+        {
+            "service.name": service_name,
+            "service.version": "0.4.8",  # Could import from __version__
+            "deployment.environment.name": deployment_environment,
+        }
+    )
+
+    # Parse headers
+    headers_dict = {}
+    if headers:
+        for pair in headers.split(","):
+            if "=" in pair:
+                key, value = pair.split("=", 1)
+                headers_dict[key.strip()] = value.strip()
+
+    # Create OTLP HTTP exporter
+    # Note: Langfuse expects /v1/traces path appended to base endpoint
+    otlp_endpoint = endpoint if endpoint.endswith("/v1/traces") else f"{endpoint}/v1/traces"
+    otlp_exporter = OTLPSpanExporter(
+        endpoint=otlp_endpoint,
+        headers=headers_dict,
+    )
+
+    # Create tracer provider with batch processor
+    provider = TracerProvider(resource=resource)
+    provider.add_span_processor(BatchSpanProcessor(otlp_exporter))
+
+    # Set global tracer provider
+    trace.set_tracer_provider(provider)
+
+    # Get tracer for this application
+    _tracer = trace.get_tracer(__name__)
+    _tracing_enabled = True
+
+    logger.info(f"Tracing initialized: endpoint={otlp_endpoint}, service={service_name}")
+
+
+def get_tracer() -> trace.Tracer | NoOpTracer:
+    """
+    Get the global tracer instance.
+
+    Returns a no-op tracer if tracing is disabled, so callers don't need to check for None.
+    This improves code readability by allowing direct use without null checks.
+    """
+    return _tracer
+
+
+def create_operation_span(operation: str, bank_id: str | None = None):
+    """
+    Create a parent span for a Hindsight operation (retain, reflect, consolidation, etc.).
+
+    This creates the span hierarchy:
+    - hindsight.{operation} (parent)
+      - chat {model} (child LLM calls)
+
+    Args:
+        operation: Operation name (retain, reflect, consolidation, mental_model_refresh)
+        bank_id: Optional bank ID for context
+
+    Returns:
+        Span context manager
+    """
+    if not _tracing_enabled or _tracer is None:
+        # Return a no-op context manager
+        from contextlib import nullcontext
+
+        return nullcontext()
+
+    span_name = f"hindsight.{operation}"
+    span = _tracer.start_as_current_span(span_name)
+
+    # Add operation-specific attributes
+    if span and hasattr(span, "set_attribute"):
+        span.set_attribute("hindsight.operation", operation)
+        if bank_id:
+            span.set_attribute("hindsight.bank_id", bank_id)
+
+    return span
+
+
+def is_tracing_enabled() -> bool:
+    """Check if tracing is enabled."""
+    return _tracing_enabled
+
+
+# Maximum content length before truncation (to stay within span size limits)
+MAX_CONTENT_LENGTH = 100_000  # characters
+
+
+def _truncate_content(content: str) -> str:
+    """Truncate content if too large for span."""
+    if len(content) > MAX_CONTENT_LENGTH:
+        return content[:MAX_CONTENT_LENGTH] + f"\n\n[TRUNCATED: {len(content) - MAX_CONTENT_LENGTH} chars omitted]"
+    return content
+
+
+class LLMSpanRecorder:
+    """
+    Records OpenTelemetry spans for LLM calls following GenAI semantic conventions.
+    """
+
+    def __init__(self, tracer: trace.Tracer):
+        self.tracer = tracer
+
+    def record_llm_call(
+        self,
+        provider: str,
+        model: str,
+        scope: str,
+        messages: list[dict[str, str]],
+        response_content: Optional[str],
+        input_tokens: int,
+        output_tokens: int,
+        duration: float,
+        finish_reason: Optional[str] = None,
+        error: Optional[Exception] = None,
+        tool_calls: Optional[list[dict[str, Any]]] = None,
+    ) -> None:
+        """
+        Record a completed LLM call as a span with GenAI semantic conventions.
+
+        This creates a span AFTER the call completes, using timestamps to
+        set the correct start/end times. This approach works better with
+        the existing sync metrics recording pattern.
+
+        Args:
+            provider: Hindsight provider name
+            model: Model name
+            scope: Scope identifier (memory, reflect, consolidation, etc.)
+            messages: Input messages (chat history)
+            response_content: Response text from LLM
+            input_tokens: Input token count
+            output_tokens: Output token count
+            duration: Call duration in seconds
+            finish_reason: Reason the model stopped (stop, length, tool_calls, etc.)
+            error: Exception if call failed
+            tool_calls: List of tool calls made (for function calling)
+        """
+        try:
+            # Map provider name to GenAI semantic convention
+            genai_provider = PROVIDER_NAME_MAPPING.get(provider.lower(), provider.lower())
+
+            # Determine operation name based on scope/context
+            operation_name = "chat"  # Default for GenAI semantic conventions
+
+            # Create span name: "hindsight.{scope}" for consistency with parent spans
+            # Model info is available in span attributes (gen_ai.request.model)
+            if scope:
+                span_name = f"hindsight.{scope}"
+            else:
+                # Fallback to chat {model} if no scope provided
+                span_name = f"{operation_name} {model}"
+
+            # Calculate timestamps
+            end_time_ns = time.time_ns()
+            start_time_ns = end_time_ns - int(duration * 1_000_000_000)
+
+            # Create span with explicit timestamps
+            with self.tracer.start_as_current_span(
+                span_name,
+                start_time=start_time_ns,
+                end_on_exit=False,  # We'll set end time manually
+            ) as span:
+                # Set required attributes
+                span.set_attribute(GenAIAttributes.OPERATION_NAME, operation_name)
+                span.set_attribute(GenAIAttributes.PROVIDER_NAME, genai_provider)
+                span.set_attribute(GenAIAttributes.REQUEST_MODEL, model)
+                span.set_attribute(GenAIAttributes.RESPONSE_MODEL, model)
+                span.set_attribute(GenAIAttributes.USAGE_INPUT_TOKENS, input_tokens)
+                span.set_attribute(GenAIAttributes.USAGE_OUTPUT_TOKENS, output_tokens)
+
+                # Add custom attributes for Hindsight context
+                span.set_attribute("hindsight.scope", scope)
+                span.set_attribute("hindsight.provider.internal", provider)
+
+                # Add tool call information if present
+                if tool_calls:
+                    span.set_attribute("gen_ai.tool_calls.count", len(tool_calls))
+                    # Add tool names as comma-separated list
+                    tool_names = [tc.get("name", "") for tc in tool_calls]
+                    span.set_attribute("gen_ai.tool_calls.names", ",".join(tool_names))
+
+                # Format messages for GenAI conventions (as JSON)
+                input_messages_json = self._format_messages(messages)
+                output_messages_json = self._format_output(response_content, finish_reason)
+
+                # Extract system instructions if present
+                system_instructions = self._extract_system_instructions(messages)
+
+                # Add event with prompts/completions following v1.37 conventions
+                event_attrs = {}
+                if input_messages_json:
+                    event_attrs[GenAIAttributes.INPUT_MESSAGES] = input_messages_json
+                if output_messages_json:
+                    event_attrs[GenAIAttributes.OUTPUT_MESSAGES] = output_messages_json
+                if system_instructions:
+                    event_attrs[GenAIAttributes.SYSTEM_INSTRUCTIONS] = system_instructions
+                if finish_reason:
+                    event_attrs[GenAIAttributes.FINISH_REASONS] = json.dumps([finish_reason])
+
+                span.add_event(
+                    "gen_ai.client.inference.operation.details",
+                    attributes=event_attrs,
+                )
+
+                # Add individual tool call events with details
+                if tool_calls:
+                    for i, tc in enumerate(tool_calls):
+                        tool_event_attrs = {
+                            "tool.name": tc.get("name", ""),
+                            "tool.id": tc.get("id", ""),
+                            "tool.arguments": json.dumps(tc.get("arguments", {})),
+                        }
+                        span.add_event(f"gen_ai.tool_call.{i}", attributes=tool_event_attrs)
+
+                # Handle errors
+                if error:
+                    span.set_status(Status(StatusCode.ERROR, str(error)))
+                    span.set_attribute(GenAIAttributes.ERROR_TYPE, type(error).__name__)
+                    span.record_exception(error)
+                else:
+                    span.set_status(Status(StatusCode.OK))
+
+                # Set end time
+                span.end(end_time=end_time_ns)
+
+        except Exception as e:
+            # Don't let tracing errors break LLM calls
+            logger.error(f"Failed to record LLM span: {e}", exc_info=True)
+
+    def _format_messages(self, messages: list[dict[str, str]]) -> str:
+        """
+        Format messages into GenAI semantic convention format (JSON array).
+
+        Returns JSON string representation of message array.
+        """
+        try:
+            formatted = []
+            for msg in messages:
+                content = msg.get("content", "")
+                # Truncate if needed
+                if isinstance(content, str):
+                    content = _truncate_content(content)
+
+                formatted.append(
+                    {
+                        "role": msg.get("role", "user"),
+                        "content": content,
+                    }
+                )
+
+            return json.dumps(formatted)
+        except Exception as e:
+            logger.warning(f"Failed to format input messages: {e}")
+            return "[]"
+
+    def _format_output(
+        self,
+        content: Optional[str],
+        finish_reason: Optional[str],
+    ) -> str:
+        """Format output message into GenAI semantic convention format."""
+        try:
+            if content is None:
+                return "[]"
+
+            # Truncate if needed
+            if isinstance(content, str):
+                content = _truncate_content(content)
+
+            return json.dumps(
+                [
+                    {
+                        "role": "assistant",
+                        "content": content,
+                    }
+                ]
+            )
+        except Exception as e:
+            logger.warning(f"Failed to format output message: {e}")
+            return "[]"
+
+    def _extract_system_instructions(self, messages: list[dict[str, str]]) -> Optional[str]:
+        """Extract system instructions from messages if present."""
+        try:
+            for msg in messages:
+                if msg.get("role") == "system":
+                    content = msg.get("content", "")
+                    if isinstance(content, str):
+                        return _truncate_content(content)
+                    return str(content)
+        except Exception as e:
+            logger.warning(f"Failed to extract system instructions: {e}")
+        return None
+
+
+class NoOpLLMSpanRecorder:
+    """No-op span recorder for when tracing is disabled."""
+
+    def record_llm_call(self, **kwargs) -> None:
+        """No-op."""
+        pass
+
+
+# Global span recorder instance
+_span_recorder: Optional[LLMSpanRecorder] = None
+
+
+def get_span_recorder() -> LLMSpanRecorder | NoOpLLMSpanRecorder:
+    """Get the global span recorder (NoOp if tracing disabled)."""
+    if _span_recorder is None:
+        return NoOpLLMSpanRecorder()
+    return _span_recorder
+
+
+def create_span_recorder() -> LLMSpanRecorder:
+    """Create and set the global span recorder."""
+    global _span_recorder
+    tracer = get_tracer()
+    if tracer is None:
+        raise RuntimeError("Tracing not initialized. Call initialize_tracing() first.")
+    _span_recorder = LLMSpanRecorder(tracer)
+    return _span_recorder
@@ -0,0 +1,13 @@
+from urllib.parse import urlparse, urlunparse
+
+
+def mask_network_location(url):
+    if not url:
+        return url
+    parsed_url = urlparse(url)
+    masked_network_location = parsed_url.hostname or ""
+    if parsed_url.port:
+        masked_network_location += f":{parsed_url.port}"
+    if parsed_url.username or parsed_url.password:
+        masked_network_location = f"***:***@{masked_network_location}"
+    return urlunparse(parsed_url._replace(netloc=masked_network_location))
@@ -124,12 +124,6 @@ def main():
        default=config.worker_poll_interval_ms,
        help=f"Poll interval in milliseconds (default: {config.worker_poll_interval_ms}, env: HINDSIGHT_API_WORKER_POLL_INTERVAL_MS)",
    )
-    parser.add_argument(
-        "--batch-size",
-        type=int,
-        default=config.worker_batch_size,
-        help=f"Tasks to claim per poll (default: {config.worker_batch_size}, env: HINDSIGHT_API_WORKER_BATCH_SIZE)",
-    )
    parser.add_argument(
        "--max-retries",
        type=int,
@@ -168,8 +162,9 @@ def main():

    print(f"Starting Hindsight Worker: {args.worker_id}")
    print(f"  Poll interval: {args.poll_interval}ms")
-    print(f"  Batch size: {args.batch_size}")
    print(f"  Max retries: {args.max_retries}")
+    print(f"  Max slots: {config.worker_max_slots}")
+    print(f"  Consolidation max slots: {config.worker_consolidation_max_slots}")
    print(f"  HTTP server: {args.http_host}:{args.http_port}")
    print()

@@ -181,40 +176,83 @@ def main():
        nonlocal memory, poller
        import uvicorn

+        from ..extensions import OperationValidatorExtension, TenantExtension, load_extension
+
+        # Load tenant extension BEFORE creating MemoryEngine so it can
+        # set correct schema context during task execution. Without this,
+        # _authenticate_tenant sees no extension and resets schema to "public",
+        # causing worker writes to land in the wrong schema.
+        tenant_extension = load_extension("TENANT", TenantExtension)
+
+        # Load operation validator so workers can record usage metering
+        # for async operations (e.g. refresh_mental_model after consolidation)
+        operation_validator = load_extension("OPERATION_VALIDATOR", OperationValidatorExtension)
+        if operation_validator:
+            logger.info(f"Loaded operation validator: {operation_validator.__class__.__name__}")
+
        # Initialize MemoryEngine
        # Workers use SyncTaskBackend because they execute tasks directly,
        # they don't need to store tasks (they poll from DB)
        memory = MemoryEngine(
            run_migrations=False,  # Workers don't run migrations
            task_backend=SyncTaskBackend(),
+            tenant_extension=tenant_extension,
+            operation_validator=operation_validator,
        )

        await memory.initialize()

        print(f"Database connected: {config.database_url}")

-        # Create and start the poller
+        if tenant_extension:
+            print("Tenant extension loaded - schemas will be discovered dynamically on each poll")
+        else:
+            print(f"No tenant extension configured, using schema: {config.database_schema}")
+
+        # Create a single poller that handles all schemas dynamically
+        # Convert default schema to None for SQL compatibility (no schema prefix)
+        from hindsight_api.config import DEFAULT_DATABASE_SCHEMA
+
+        schema = None if config.database_schema == DEFAULT_DATABASE_SCHEMA else config.database_schema
        poller = WorkerPoller(
            pool=memory._pool,
            worker_id=args.worker_id,
            executor=memory.execute_task,
            poll_interval_ms=args.poll_interval,
-            batch_size=args.batch_size,
            max_retries=args.max_retries,
+            schema=schema,
+            tenant_extension=tenant_extension,
+            max_slots=config.worker_max_slots,
+            consolidation_max_slots=config.worker_consolidation_max_slots,
        )

        # Create the HTTP app for metrics/health
        app = create_worker_app(poller, memory)

-        # Setup signal handlers for graceful shutdown
+        # Setup signal handlers for graceful shutdown using asyncio
        shutdown_requested = asyncio.Event()
+        force_exit = False

-        def signal_handler(signum, frame):
-            print(f"\nReceived signal {signum}, initiating graceful shutdown...")
-            shutdown_requested.set()
+        loop = asyncio.get_event_loop()

-        signal.signal(signal.SIGINT, signal_handler)
-        signal.signal(signal.SIGTERM, signal_handler)
+        def signal_handler():
+            nonlocal force_exit
+            if shutdown_requested.is_set():
+                # Second signal = force exit
+                print("\nReceived second signal, forcing immediate exit...")
+                force_exit = True
+                # Restore default handler so third signal kills process
+                loop.remove_signal_handler(signal.SIGINT)
+                loop.remove_signal_handler(signal.SIGTERM)
+                sys.exit(1)
+            else:
+                print("\nReceived shutdown signal, initiating graceful shutdown...")
+                print("(Press Ctrl+C again to force immediate exit)")
+                shutdown_requested.set()
+
+        # Use asyncio's signal handlers which work properly with the event loop
+        loop.add_signal_handler(signal.SIGINT, signal_handler)
+        loop.add_signal_handler(signal.SIGTERM, signal_handler)

        # Create uvicorn config and server
        uvicorn_config = uvicorn.Config(
@@ -233,7 +271,10 @@ def main():
        print(f"Worker started. Metrics available at http://{args.http_host}:{args.http_port}/metrics")

        # Wait for shutdown signal
-        await shutdown_requested.wait()
+        try:
+            await shutdown_requested.wait()
+        except KeyboardInterrupt:
+            print("\nReceived interrupt, initiating graceful shutdown...")

        # Graceful shutdown
        print("Shutting down HTTP server...")
--- a/Show More
+++ b/Show More