fix(perf): fix CI install and remove fragile WorkerPoller kwarg

* feat(perf): add system performance test runner and CI workflow

Add `uv run perf-test` command that orchestrates retain throughput and
recall latency benchmarks using mock LLM + pg0 for deterministic,
LLM-independent baselines. Wraps existing recall_perf/retain_perf
building blocks without duplicating benchmark logic.

Also fixes _RRFReranker in recall_perf.py to include the cross_encoder
attribute now required by the engine's combined scoring path.

* feat(perf): add run-perf-test.sh script

* feat(perf): use run-perf-test.sh in CI, remove run-retain-perf.sh

Replace ad-hoc retain perf wrapper with the new system perf test
script in CI workflow and docs. The standalone retain_perf.py is still
available for ad-hoc document benchmarking.

* feat(perf): add suite input to workflow dispatch

.claude/skills/code-review/SKILL.md

@@ -18,15 +18,17 @@ Read and internalize these standards before writing code. The review steps below
 - Pydantic models for request/response
 - Ruff for linting (line-length 120)
 - No Python files at project root - maintain clean directory structure
-- **Never use multi-item tuple return values** - prefer dataclass or Pydantic model for structured returns
+- **Never use multi-item tuple return values** — not even for internal/private functions. Always use a dataclass or Pydantic model. No exceptions, no "it's just two values" shortcuts. If a function returns more than one value, define a named type for it.
 
 ### Type Safety with Pydantic Models
-**NEVER use raw `dict` types for structured data.** Always use Pydantic models:
+**NEVER use raw `dict` types for structured data** — this applies to all code, including internal helpers and private functions. If the dict has known keys, it must be a dataclass or Pydantic model:
 - Use Pydantic `BaseModel` for all data structures passed between functions
+- Use `@dataclass` for lightweight internal data containers when Pydantic validation isn't needed
 - Add `@field_validator` for type coercion (e.g., ensuring datetimes are timezone-aware)
 - Avoid `dict.get()` patterns - use typed model attributes instead
 - Parse external data (JSON, API responses) into Pydantic models at the boundary
 - This catches type errors at parse time, not deep in business logic
+- The only acceptable `dict` usage is for truly dynamic/unknown keys (e.g., arbitrary metadata, JSON blobs with no fixed schema)
 
 ```python
 # BAD - error-prone dict access
@@ -118,8 +120,8 @@ For each changed TypeScript file, check for:
 ### 5. Check type safety (Python)
 
 For each changed Python file, check for violations:
-- **No raw `dict` for structured data** — should use Pydantic models
-- **No multi-item tuple returns** — should use dataclass or Pydantic model
+- **No raw `dict` for structured data** — must use Pydantic model or dataclass, even for internal/private functions (only exception: truly dynamic/unknown keys)
+- **No multi-item tuple returns** — must use dataclass or Pydantic model, even for internal/private functions (no exceptions)
 - **Missing type hints** on function parameters and return types
 - **Missing `@field_validator`** for datetime fields that should be timezone-aware
 
@@ -147,7 +149,24 @@ For each non-trivial change:
 - **Changed approach** — does the comment include what was done before and why it changed?
 - **Stale comments** — do existing comments near the changed code still accurately describe the behavior?
 
-### 9. Review against other coding standards
+### 9. Check integration completeness
+
+If any files in `hindsight-integrations/` were added or changed, verify:
+- **Tests exist** — the integration must have tests that simulate/exercise the external framework (not just pure unit tests of helpers). Check for a `tests/` directory with meaningful test files.
+- **CI job exists** — check `.github/workflows/test.yml` for a corresponding `test-<name>-integration` job. If missing, flag it.
+- **Release process** — check that the integration name is in the `VALID_INTEGRATIONS` array in `scripts/release-integration.sh`. If missing, flag it.
+- **Code standards** — the integration code must follow all Python style rules (type hints, no raw dicts, no tuple returns, etc.).
+
+### 10. Check MCP tool registration completeness
+
+If any new MCP tools were added or existing tools renamed in `hindsight-api-slim/hindsight_api/mcp_tools.py`:
+- **`_ALL_TOOLS` set** in `mcp_tools.py` — must include the new tool name
+- **`tools_to_register` default set** in `register_mcp_tools()` in `mcp_tools.py` — must include the new tool name
+- **`_SINGLE_BANK_TOOLS` set** in `hindsight-api-slim/hindsight_api/api/mcp.py` — must include the new tool if it is bank-scoped (not a bank-management tool like `list_banks`/`create_bank`)
+- **`MCP_TOOL_GROUPS`** in `hindsight-control-plane/src/components/bank-config-view.tsx` — must include the new tool in the appropriate group for the UI tool selector
+- **Tool count assertions** in tests (e.g., `test_mcp_tools.py`) — must be updated to reflect the new count
+
+### 11. Review against other coding standards
 
 Check the diff for violations of the standards listed above:
 - Python files at project root (not allowed)
@@ -159,7 +178,7 @@ Check the diff for violations of the standards listed above:
 - Premature abstractions or speculative helpers
 - Backwards-compatibility hacks (unused vars, re-exports, "removed" comments)
 
-### 10. Report findings
+### 12. Report findings
 
 Present a clear summary organized by severity:
 
@@ -167,8 +186,10 @@ Present a clear summary organized by severity:
 - Unrelated commits on the branch
 - Lint failures
 - Missing type hints on public functions
-- Raw dict usage for structured data
+- Raw dict usage for structured data (including internal code)
+- Multi-item tuple returns (including internal code)
 - Missing tests for new endpoints
+- New integration missing tests, CI job, or release-integration.sh entry
 
 **Should fix** — issues that hurt code quality:
 - Dead code / unused imports missed by linter

.github/workflows/deploy-docs.yml

@@ -34,7 +34,7 @@ jobs:
         env:
           UMAMI_URL: https://analytics.hindsight.vectorize.io
           UMAMI_WEBSITE_ID: ${{ secrets.UMAMI_WEBSITE_ID }}
-      - uses: actions/upload-pages-artifact@v4
+      - uses: actions/upload-pages-artifact@v5
         with:
           path: hindsight-docs/build
   deploy:
@@ -44,5 +44,5 @@ jobs:
     runs-on: ubuntu-latest
     needs: build
     steps:
-      - uses: actions/deploy-pages@v4
+      - uses: actions/deploy-pages@v5
         id: deployment

.github/workflows/perf-test.yml

@@ -0,0 +1,93 @@
+name: Performance Tests
+
+on:
+  schedule:
+    # Run daily at 06:00 UTC
+    - cron: "0 6 * * *"
+  workflow_dispatch:
+    inputs:
+      scale:
+        description: "Test scale"
+        type: choice
+        options:
+          - tiny
+          - small
+          - medium
+          - large
+        default: small
+      suite:
+        description: "Suite to run (blank = all)"
+        type: choice
+        options:
+          - ""
+          - retain
+          - recall
+        default: ""
+      ref:
+        description: "Git ref to test (branch, tag, or SHA). Defaults to main."
+        type: string
+        default: ""
+
+concurrency:
+  group: perf-test
+  cancel-in-progress: true
+
+jobs:
+  perf-test:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v6
+      with:
+        ref: ${{ inputs.ref || github.ref }}
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v7
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v6
+      with:
+        python-version-file: ".python-version"
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v5
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api-slim/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-
+
+    - name: Pre-download models
+      working-directory: ./hindsight-api-slim
+      run: |
+        uv run --frozen --all-extras --index-strategy unsafe-best-match python -c "
+        from sentence_transformers import SentenceTransformer
+        print('Downloading embedding model...')
+        SentenceTransformer('BAAI/bge-small-en-v1.5')
+        print('Model downloaded successfully')
+        "
+
+    - name: Install hindsight-dev dependencies
+      run: |
+        cd hindsight-dev && uv sync --frozen --all-extras --index-strategy unsafe-best-match
+
+    - name: Run perf tests
+      run: |
+        SUITE_ARG=""
+        if [ -n "${{ inputs.suite }}" ]; then
+          SUITE_ARG="--suite ${{ inputs.suite }}"
+        fi
+        ./scripts/benchmarks/run-perf-test.sh \
+          --scale ${{ inputs.scale || 'small' }} \
+          $SUITE_ARG \
+          --output perf-results.json
+
+    - name: Upload perf results
+      if: always()
+      uses: actions/upload-artifact@v4
+      with:
+        name: perf-results-${{ github.sha }}
+        path: hindsight-dev/perf-results.json
+        retention-days: 90

.github/workflows/release-integration.yml

@@ -82,6 +82,15 @@ jobs:
         node-version: '22'
         registry-url: 'https://registry.npmjs.org'
 
+    # Guard: fail fast if the integration's lockfile resolves any dep from a
+    # monorepo workspace (link=true) or a relative file path. The release
+    # runner has no pre-built workspace `dist/` so `npm run build` would
+    # later fail at tsc with "Cannot find module". See:
+    # https://github.com/vectorize-io/hindsight/issues/… (0.6.0 openclaw retry)
+    - name: Check integration lockfile
+      if: steps.type.outputs.type == 'typescript'
+      run: ./scripts/check-integration-lockfiles.sh
+
     - name: Install dependencies
       if: steps.type.outputs.type == 'typescript'
       working-directory: ./hindsight-integrations/${{ steps.info.outputs.integration }}

.github/workflows/release.yml

@@ -150,6 +150,55 @@ jobs:
         path: hindsight-clients/typescript/*.tgz
         retention-days: 1
 
+  release-hindsight-all-npm:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v6
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v6
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+        cache: 'npm'
+        cache-dependency-path: package-lock.json
+
+    - name: Install dependencies
+      run: npm ci --workspace=hindsight-all-npm
+
+    - name: Build
+      run: npm run build --workspace=hindsight-all-npm
+
+    - name: Publish to npm
+      working-directory: ./hindsight-all-npm
+      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-all-npm
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v7
+      with:
+        name: hindsight-all-npm
+        path: hindsight-all-npm/*.tgz
+        retention-days: 1
+
   release-control-plane:
     runs-on: ubuntu-latest
     environment: npm
@@ -382,7 +431,7 @@ jobs:
     - uses: actions/checkout@v6
 
     - name: Install Helm
-      uses: azure/setup-helm@v4
+      uses: azure/setup-helm@v5
       with:
         version: 'latest'
 
@@ -407,7 +456,7 @@ 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-hindsight-all-npm, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
     permissions:
       contents: write
 
@@ -436,6 +485,12 @@ jobs:
         name: control-plane
         path: ./artifacts/control-plane
 
+    - name: Download hindsight-embed npm wrapper
+      uses: actions/download-artifact@v8
+      with:
+        name: hindsight-all-npm
+        path: ./artifacts/hindsight-all-npm
+
     - name: Download Rust CLI (Linux)
       uses: actions/download-artifact@v8
       with:
@@ -472,6 +527,8 @@ jobs:
         cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
         # TypeScript client
         cp artifacts/typescript-client/*.tgz release-assets/ || true
+        # hindsight-embed npm wrapper
+        cp artifacts/hindsight-all-npm/*.tgz release-assets/ || true
         # Control Plane
         cp artifacts/control-plane/*.tgz release-assets/ || true
         # Rust CLI binaries
@@ -483,7 +540,7 @@ jobs:
         ls -la release-assets/
 
     - name: Create GitHub Release
-      uses: softprops/action-gh-release@v2
+      uses: softprops/action-gh-release@v3
       with:
         files: release-assets/*
         generate_release_notes: true

.prettierrc.json

@@ -0,0 +1,7 @@
+{
+  "semi": true,
+  "singleQuote": false,
+  "tabWidth": 2,
+  "trailingComma": "es5",
+  "printWidth": 100
+}

CLAUDE.md

@@ -11,9 +11,15 @@ Hindsight is an agent memory system that provides long-term memory for AI agents
 
 ## Development Commands
 
+### Local Development (API + UI)
+```bash
+# Start both API server and control plane UI
+./scripts/dev/start.sh
+```
+
 ### API Server (Python/FastAPI)
 ```bash
-# Start API server (loads .env automatically)
+# Start API server only (loads .env automatically)
 ./scripts/dev/start-api.sh
 
 # Run all tests (parallelized with pytest-xdist)
@@ -62,8 +68,9 @@ cd hindsight-control-plane && npm run dev
 ./scripts/benchmarks/run-locomo.sh
 
 # Performance benchmarks
+./scripts/benchmarks/run-perf-test.sh                      # System perf (mock LLM + pg0)
+./scripts/benchmarks/run-perf-test.sh --scale tiny          # Quick smoke test
 ./scripts/benchmarks/run-consolidation.sh
-./scripts/benchmarks/run-retain-perf.sh --document <path>  # Requires API server running
 
 # Results viewer
 ./scripts/benchmarks/start-visualizer.sh  # View results at localhost:8001
@@ -73,17 +80,16 @@ cd hindsight-control-plane && npm run dev
 
 ### Monorepo Structure
 - **hindsight-api-slim/**: Core FastAPI server with memory engine (Python, uv)
-- **hindsight/**: Embedded Python bundle (hindsight-all package)
 - **hindsight-control-plane/**: Admin UI (Next.js, npm)
 - **hindsight-cli/**: CLI tool (Rust, cargo, uses progenitor for API client)
 - **hindsight-clients/**: Generated SDK clients (Python, TypeScript, Rust)
 - **hindsight-docs/**: Docusaurus documentation site
-- **hindsight-integrations/**: Framework integrations (LiteLLM, OpenAI)
+- **hindsight-integrations/**: Framework integrations (LiteLLM, CrewAI, LangGraph, Pydantic AI, AG2, Claude Code, etc.)
 - **hindsight-dev/**: Development tools and benchmarks
 
 ### Core Engine (hindsight-api-slim/hindsight_api/engine/)
-- `memory_engine.py`: Main orchestrator (~170KB) for retain/recall/reflect operations
-- `llm_wrapper.py`: LLM abstraction supporting OpenAI, Anthropic, Gemini, Groq, MiniMax, Ollama, LM Studio
+- `memory_engine.py`: Main orchestrator for retain/recall/reflect operations
+- `llm_wrapper.py`: LLM abstraction supporting OpenAI, Anthropic, Gemini, VertexAI, Groq, MiniMax, Ollama, LM Studio, LiteLLM, Claude Code
 - `embeddings.py`: Embedding generation (local sentence-transformers or TEI)
 - `cross_encoder.py`: Reranking (local or TEI)
 - `entity_resolver.py`: Entity extraction and normalization
@@ -96,13 +102,13 @@ cd hindsight-control-plane && npm run dev
 
 **search/**: Multi-strategy retrieval
 - `retrieval.py`: Main retrieval orchestrator
-- `graph_retrieval.py`: Entity/relationship graph traversal
-- `mpfp_retrieval.py`: Multi-Path Fact Propagation retrieval
+- `graph_retrieval.py`: Graph retrieval abstract base class
+- `link_expansion_retrieval.py`: Link expansion graph retrieval
 - `fusion.py`: Reciprocal rank fusion for combining results
 - `reranking.py`: Cross-encoder reranking
 
 ### API Layer (hindsight-api-slim/hindsight_api/api/)
-- `http.py`: FastAPI HTTP routers (~80KB) for all REST endpoints
+- `http.py`: FastAPI HTTP routers for all REST endpoints
 - `mcp.py`: Model Context Protocol server implementation
 
 Main operations:
@@ -174,6 +180,8 @@ Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
 
 **After completing any implementation work, run `/code-review`** to verify your changes against project standards (missing tests, dead code, type safety, etc.). Fix any "must fix" issues before considering the task done.
 
+**MANDATORY: Run `/code-review` before pushing code or creating a pull request.** Do not push or create a PR until all "must fix" issues are resolved.
+
 ### Memory Banks
 - Each bank is an isolated memory store (like a "brain" for one user/agent)
 - Banks have dispositions (skepticism, literalism, empathy traits 1-5) affecting reflect
@@ -204,6 +212,21 @@ When adding or modifying parameters in the dataplane API (hindsight-api), you mu
    - Update the client type definition in `lib/api.ts`
    - Update any UI components that need to use the new parameter
 
+### Adding New Integrations
+
+Every new integration in `hindsight-integrations/` must satisfy all of the following before it can be merged:
+
+1. **Tests are required** — tests must simulate or exercise the external system (mock the framework's interfaces and verify the integration actually calls Hindsight correctly). Pure unit tests of helper functions are not sufficient.
+2. **CI job** — add a test job in `.github/workflows/test.yml` following the existing pattern (e.g., `test-crewai-integration`). The job must build, install deps, and run `uv run pytest tests -v`. Also add the integration to `detect-changes` outputs so it only runs when its files change.
+3. **Release process** — add the integration name to the `VALID_INTEGRATIONS` array in `scripts/release-integration.sh` so it can be released via the standard release workflow.
+4. **Follow project code standards** — Python style, type safety, no raw dicts for structured data, no multi-item tuple returns (see `.claude/skills/code-review/SKILL.md`).
+
+If any of these are missing, the integration is incomplete and must not be pushed or merged.
+
+### Changelogs
+
+Never add "Unreleased" entries to changelogs (e.g. `hindsight-docs/src/pages/changelog/**`). Changelog entries are written by the release script (`./scripts/release-integration.sh`) when a version is actually cut. If a bug fix or feature needs documenting before release, describe it in the PR/commit — the release tooling will surface it in the published changelog section.
+
 ### Adding New API Configuration Flags
 
 Configuration follows a hierarchical system: **Global (env vars) → Tenant (via extension) → Bank (database)**.
@@ -216,17 +239,17 @@ Fields must be categorized as either **hierarchical** (can be overridden per-ten
    - Add `ENV_*` constant for the environment variable name (e.g., `ENV_MY_SETTING = "HINDSIGHT_API_MY_SETTING"`)
    - Add `DEFAULT_*` constant for the default value
    - Add field to `HindsightConfig` dataclass with type annotation
-   - **Mark as hierarchical or static** by adding to `_HIERARCHICAL_FIELDS` set (hierarchical) or leaving it out (static)
+   - **Mark as configurable** by adding to `_CONFIGURABLE_FIELDS` set if the field should be overridable per-tenant/bank via API
    - Add initialization in `from_env()` method
 
    ```python
-   # Hierarchical field (can be overridden per-bank)
-   _HIERARCHICAL_FIELDS = {
+   # Configurable field (can be overridden per-tenant/bank via API)
+   _CONFIGURABLE_FIELDS = {
        ...,
-       "my_setting",  # Add here for hierarchical
+       "my_setting",  # Add here for configurable
    }
 
-   # Static field - just don't add to _HIERARCHICAL_FIELDS
+   # Static field - just don't add to _CONFIGURABLE_FIELDS
    ```
 
 2. **main.py** (`hindsight-api-slim/hindsight_api/main.py`):

helm/hindsight/Chart.yaml

@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.4.21
-appVersion: "0.4.21"
+version: 0.5.4
+appVersion: "0.5.4"
 keywords:
   - ai
   - memory

helm/hindsight/templates/api-deployment.yaml

@@ -95,6 +95,27 @@ spec:
           {{- toYaml .Values.api.readinessProbe | nindent 10 }}
         resources:
           {{- toYaml .Values.api.resources | nindent 10 }}
+        {{- if or .Values.api.persistence.modelCache.enabled .Values.api.extraVolumeMounts }}
+        volumeMounts:
+        {{- if .Values.api.persistence.modelCache.enabled }}
+        - name: model-cache
+          mountPath: /home/hindsight/.cache
+        {{- end }}
+        {{- with .Values.api.extraVolumeMounts }}
+        {{- toYaml . | nindent 8 }}
+        {{- end }}
+        {{- end }}
+      {{- if or .Values.api.persistence.modelCache.enabled .Values.api.extraVolumes }}
+      volumes:
+      {{- if .Values.api.persistence.modelCache.enabled }}
+      - name: model-cache
+        persistentVolumeClaim:
+          claimName: {{ include "hindsight.fullname" . }}-api-model-cache
+      {{- end }}
+      {{- with .Values.api.extraVolumes }}
+      {{- toYaml . | nindent 6 }}
+      {{- end }}
+      {{- end }}
       {{- with .Values.nodeSelector }}
       nodeSelector:
         {{- toYaml . | nindent 8 }}

helm/hindsight/templates/api-model-cache-pvc.yaml

@@ -0,0 +1,21 @@
+{{- if and .Values.api.enabled .Values.api.persistence.modelCache.enabled }}
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: {{ include "hindsight.fullname" . }}-api-model-cache
+  labels:
+    {{- include "hindsight.api.labels" . | nindent 4 }}
+  {{- with .Values.api.persistence.modelCache.annotations }}
+  annotations:
+    {{- toYaml . | nindent 4 }}
+  {{- end }}
+spec:
+  accessModes:
+    {{- toYaml .Values.api.persistence.modelCache.accessModes | nindent 4 }}
+  {{- if .Values.api.persistence.modelCache.storageClass }}
+  storageClassName: {{ .Values.api.persistence.modelCache.storageClass }}
+  {{- end }}
+  resources:
+    requests:
+      storage: {{ .Values.api.persistence.modelCache.size }}
+{{- end }}

helm/hindsight/templates/worker-statefulset.yaml

@@ -95,6 +95,16 @@ spec:
           {{- toYaml .Values.worker.readinessProbe | nindent 10 }}
         resources:
           {{- toYaml .Values.worker.resources | nindent 10 }}
+        {{- if or .Values.worker.persistence.modelCache.enabled .Values.worker.extraVolumeMounts }}
+        volumeMounts:
+        {{- if .Values.worker.persistence.modelCache.enabled }}
+        - name: model-cache
+          mountPath: /home/hindsight/.cache
+        {{- end }}
+        {{- with .Values.worker.extraVolumeMounts }}
+        {{- toYaml . | nindent 8 }}
+        {{- end }}
+        {{- end }}
       {{- with .Values.nodeSelector }}
       nodeSelector:
         {{- toYaml . | nindent 8 }}
@@ -107,4 +117,26 @@ spec:
       tolerations:
         {{- toYaml . | nindent 8 }}
       {{- end }}
+      {{- with .Values.worker.extraVolumes }}
+      volumes:
+      {{- toYaml . | nindent 6 }}
+      {{- end }}
+  {{- if .Values.worker.persistence.modelCache.enabled }}
+  volumeClaimTemplates:
+  - metadata:
+      name: model-cache
+      {{- with .Values.worker.persistence.modelCache.annotations }}
+      annotations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+    spec:
+      accessModes:
+        {{- toYaml .Values.worker.persistence.modelCache.accessModes | nindent 8 }}
+      {{- if .Values.worker.persistence.modelCache.storageClass }}
+      storageClassName: {{ .Values.worker.persistence.modelCache.storageClass }}
+      {{- end }}
+      resources:
+        requests:
+          storage: {{ .Values.worker.persistence.modelCache.size }}
+  {{- end }}
 {{- end }}

helm/hindsight/values.yaml

@@ -67,6 +67,33 @@ api:
   # Pod affinity/anti-affinity (overrides global affinity for this component)
   # affinity: {}
 
+  # Persistent volume for local model cache (reranker, embeddings)
+  # Models are downloaded to /home/hindsight/.cache on first use.
+  # Without persistence, models are re-downloaded on every pod restart.
+  persistence:
+    modelCache:
+      enabled: false
+      size: 5Gi
+      storageClass: ""
+      accessModes:
+        - ReadWriteOnce
+      annotations: {}
+
+  # Extra volume mounts for the api container
+  # e.g.
+  # extraVolumeMounts:
+  #   - name: my-volume
+  #     mountPath: /mnt/my-volume
+  extraVolumeMounts: []
+
+  # Extra volumes for the api pod
+  # e.g.
+  # extraVolumes:
+  #   - name: my-volume
+  #     configMap:
+  #       name: my-configmap
+  extraVolumes: []
+
   # Environment variables
   env:
     #HINDSIGHT_API_LLM_PROVIDER: "groq"
@@ -140,6 +167,32 @@ worker:
   # Pod affinity/anti-affinity (overrides global affinity for this component)
   # affinity: {}
 
+  # Persistent volume for local model cache (reranker, embeddings)
+  # Uses volumeClaimTemplates since worker is a StatefulSet.
+  persistence:
+    modelCache:
+      enabled: false
+      size: 5Gi
+      storageClass: ""
+      accessModes:
+        - ReadWriteOnce
+      annotations: {}
+
+  # Extra volume mounts for the worker container
+  # e.g.
+  # extraVolumeMounts:
+  #   - name: my-volume
+  #     mountPath: /mnt/my-volume
+  extraVolumeMounts: []
+
+  # Extra volumes for the worker pod
+  # e.g.
+  # extraVolumes:
+  #   - name: my-volume
+  #     configMap:
+  #       name: my-configmap
+  extraVolumes: []
+
   # Secret environment variables (inherited from api.secrets if not specified)
   secrets: {}
 

hindsight-all-npm/.gitignore

@@ -0,0 +1,4 @@
+node_modules
+dist
+*.tgz
+.DS_Store

hindsight-all-npm/README.md

@@ -0,0 +1,80 @@
+# @vectorize-io/hindsight-all
+
+Node.js equivalent of the Python [`hindsight-all`](https://pypi.org/project/hindsight-all/) package — programmatic lifecycle manager for a local Hindsight daemon. Use this when you want to embed Hindsight in a Node application without hand-rolling subprocess management.
+
+This package deliberately does **not** ship an HTTP client. Once the daemon is running, talk to it with [`@vectorize-io/hindsight-client`](https://www.npmjs.com/package/@vectorize-io/hindsight-client) against `server.getBaseUrl()`. The two packages compose — one owns the daemon process, the other owns the HTTP API surface.
+
+## Requirements
+
+- **Node.js >= 22** — uses global `fetch` and `AbortSignal.timeout`.
+- **`uv` / `uvx`** on `PATH` — used to download and run the underlying `hindsight-embed` daemon on first use. Install via <https://docs.astral.sh/uv/>.
+
+## Install
+
+```bash
+npm install @vectorize-io/hindsight-all @vectorize-io/hindsight-client
+```
+
+## Example
+
+```ts
+import { HindsightServer, consoleLogger } from '@vectorize-io/hindsight-all';
+import { HindsightClient } from '@vectorize-io/hindsight-client';
+
+const server = new HindsightServer({
+  profile: 'my-app',
+  port: 9077,
+  env: {
+    HINDSIGHT_API_LLM_PROVIDER: 'anthropic',
+    HINDSIGHT_API_LLM_API_KEY: process.env.ANTHROPIC_API_KEY,
+    HINDSIGHT_API_LLM_MODEL: 'claude-sonnet-4-20250514',
+    HINDSIGHT_EMBED_DAEMON_IDLE_TIMEOUT: '0',
+  },
+  logger: consoleLogger,
+});
+
+await server.start();
+
+const client = new HindsightClient({ baseUrl: server.getBaseUrl() });
+
+await client.retain('user-123', 'User prefers dark mode and concise answers.', {
+  documentId: 'pref-2026-04-01',
+});
+
+const recall = await client.recall('user-123', 'what are the user preferences?');
+console.log(recall.results);
+
+await server.stop();
+```
+
+For a remote Hindsight API, skip `HindsightServer` entirely and just point `HindsightClient` at the remote URL.
+
+## Open config — forward-compatible with new daemon flags
+
+`HindsightServerOptions` is designed so every new environment variable or CLI flag in the underlying Hindsight daemon can be used without waiting for a wrapper release:
+
+- **`env`** accepts an arbitrary `Record<string, string>`. Every entry is exported into the daemon process and written into the profile config via `--env KEY=VALUE`.
+- **`extraProfileCreateArgs`** / **`extraDaemonStartArgs`** append raw args to the respective commands.
+
+## Development against a local checkout
+
+If you're hacking on the Python `hindsight-embed` package in the same monorepo, point the server at the local path — it'll use `uv run --directory <path>` instead of `uvx`:
+
+```ts
+new HindsightServer({
+  embedPackagePath: '/path/to/hindsight-embed',
+  // ...
+});
+```
+
+## API surface
+
+- `HindsightServer` — daemon lifecycle (`start`, `stop`, `checkHealth`, `getBaseUrl`, `getProfile`).
+- `Logger` interface plus `silentLogger` (default) and `consoleLogger` helpers.
+- `getEmbedCommand(opts)` — low-level helper that returns the `[cmd, ...args]` tuple used to invoke the underlying Python CLI.
+
+For memory operations (retain, recall, reflect, bank management, stats) use [`@vectorize-io/hindsight-client`](https://www.npmjs.com/package/@vectorize-io/hindsight-client).
+
+## License
+
+MIT

hindsight-all-npm/package.json

@@ -0,0 +1,57 @@
+{
+  "name": "@vectorize-io/hindsight-all",
+  "version": "0.5.4",
+  "description": "Node.js programmatic lifecycle manager for Hindsight — embeds a local hindsight daemon in a Node application. Pair with @vectorize-io/hindsight-client for memory operations.",
+  "main": "dist/index.js",
+  "types": "dist/index.d.ts",
+  "type": "module",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
+    }
+  },
+  "keywords": [
+    "hindsight",
+    "hindsight-all",
+    "memory",
+    "ai",
+    "agent",
+    "long-term-memory",
+    "llm",
+    "embedded-server"
+  ],
+  "author": "Vectorize <[email protected]>",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/vectorize-io/hindsight.git",
+    "directory": "hindsight-all-npm"
+  },
+  "files": [
+    "dist",
+    "README.md"
+  ],
+  "scripts": {
+    "build": "tsup",
+    "dev": "tsup --watch",
+    "clean": "rm -rf dist",
+    "test": "vitest run src",
+    "test:watch": "vitest src",
+    "prepublishOnly": "npm run clean && npm run build"
+  },
+  "devDependencies": {
+    "@types/node": "^22.0.0",
+    "tsup": "^8.5.1",
+    "typescript": "^5.7.0",
+    "vitest": "^4.1.2"
+  },
+  "engines": {
+    "node": ">=22"
+  },
+  "overrides": {
+    "rollup": "^4.59.0",
+    "picomatch": ">=2.3.2 <3.0.0 || >=4.0.4",
+    "vite": ">=8.0.5"
+  }
+}

hindsight-all-npm/src/command.test.ts

@@ -0,0 +1,32 @@
+import { describe, it, expect } from 'vitest';
+import { getEmbedCommand } from './command.js';
+
+describe('getEmbedCommand', () => {
+  it('defaults to uvx hindsight-embed@latest', () => {
+    expect(getEmbedCommand()).toEqual(['uvx', 'hindsight-embed@latest']);
+  });
+
+  it('honours an explicit version', () => {
+    expect(getEmbedCommand({ embedVersion: '0.5.0' })).toEqual(['uvx', '[email protected]']);
+  });
+
+  it('treats an empty version as latest', () => {
+    expect(getEmbedCommand({ embedVersion: '' })).toEqual(['uvx', 'hindsight-embed@latest']);
+  });
+
+  it('uses uv run --directory when a local path is given', () => {
+    expect(getEmbedCommand({ embedPackagePath: '/abs/path' })).toEqual([
+      'uv',
+      'run',
+      '--directory',
+      '/abs/path',
+      'hindsight-embed',
+    ]);
+  });
+
+  it('local path takes precedence over version', () => {
+    expect(
+      getEmbedCommand({ embedPackagePath: '/abs/path', embedVersion: '0.5.0' }),
+    ).toEqual(['uv', 'run', '--directory', '/abs/path', 'hindsight-embed']);
+  });
+});

hindsight-all-npm/src/command.ts

@@ -0,0 +1,25 @@
+/**
+ * Resolve the command that invokes the `hindsight-embed` Python CLI.
+ *
+ * - If `embedPackagePath` is set, runs the package from a local checkout via
+ *   `uv run --directory <path> hindsight-embed`. Used for in-repo development.
+ * - Otherwise runs it via `uvx hindsight-embed@<version>` so no global install
+ *   is required.
+ *
+ * Returns the argv as `[command, ...baseArgs]` suitable for `spawn()` /
+ * `execFile()` (never shell-interpolated).
+ */
+export interface EmbedCommandOptions {
+  /** Version spec passed to uvx (e.g. "latest", "0.5.0"). Default: "latest". */
+  embedVersion?: string;
+  /** Local checkout path. When set, overrides `embedVersion` and uses `uv run`. */
+  embedPackagePath?: string;
+}
+
+export function getEmbedCommand(opts: EmbedCommandOptions = {}): string[] {
+  if (opts.embedPackagePath) {
+    return ['uv', 'run', '--directory', opts.embedPackagePath, 'hindsight-embed'];
+  }
+  const version = opts.embedVersion && opts.embedVersion.length > 0 ? opts.embedVersion : 'latest';
+  return ['uvx', `hindsight-embed@${version}`];
+}

hindsight-all-npm/src/index.ts

@@ -0,0 +1,7 @@
+export { HindsightServer } from './server.js';
+export { getEmbedCommand } from './command.js';
+export { silentLogger, consoleLogger } from './logger.js';
+
+export type { Logger } from './logger.js';
+export type { EmbedCommandOptions } from './command.js';
+export type { HindsightServerOptions } from './types.js';

hindsight-all-npm/src/logger.ts

@@ -0,0 +1,29 @@
+/**
+ * Pluggable logger interface.
+ *
+ * This package does not own any logging infrastructure — consumers inject
+ * whatever they want (console, pino, openclaw's logger, a no-op). The default
+ * is silent so embedding this package never adds noise to an unrelated app.
+ */
+export interface Logger {
+  debug(msg: string): void;
+  info(msg: string): void;
+  warn(msg: string): void;
+  error(msg: string): void;
+}
+
+/** Logger that drops every call. Used when no logger is passed. */
+export const silentLogger: Logger = {
+  debug: () => {},
+  info: () => {},
+  warn: () => {},
+  error: () => {},
+};
+
+/** Logger that writes to the standard console. Handy for CLIs and tests. */
+export const consoleLogger: Logger = {
+  debug: (msg) => console.debug(msg),
+  info: (msg) => console.log(msg),
+  warn: (msg) => console.warn(msg),
+  error: (msg) => console.error(msg),
+};

hindsight-all-npm/src/server.test.ts

@@ -0,0 +1,35 @@
+import { describe, it, expect } from 'vitest';
+import { HindsightServer } from './server.js';
+
+describe('HindsightServer construction', () => {
+  it('defaults base URL to http://127.0.0.1:8888', () => {
+    const server = new HindsightServer();
+    expect(server.getBaseUrl()).toBe('http://127.0.0.1:8888');
+    expect(server.getProfile()).toBe('default');
+  });
+
+  it('honours custom profile, port, and host', () => {
+    const server = new HindsightServer({ profile: 'app', port: 9077, host: '0.0.0.0' });
+    expect(server.getProfile()).toBe('app');
+    expect(server.getBaseUrl()).toBe('http://0.0.0.0:9077');
+  });
+
+  it('accepts open env pass-through without complaining about unknown keys', () => {
+    const server = new HindsightServer({
+      env: {
+        HINDSIGHT_API_LLM_PROVIDER: 'openai',
+        HINDSIGHT_API_LLM_MODEL: 'gpt-4o-mini',
+        // A field that does not exist today — should still be accepted
+        HINDSIGHT_FUTURE_FLAG: 'enabled',
+      },
+    });
+    expect(server).toBeInstanceOf(HindsightServer);
+  });
+
+  it('exposes checkHealth that returns false when no daemon is running', async () => {
+    // Random high port that nothing is listening on.
+    const server = new HindsightServer({ port: 1, readyTimeoutMs: 100 });
+    const healthy = await server.checkHealth();
+    expect(healthy).toBe(false);
+  });
+});

hindsight-all-npm/src/server.ts

@@ -0,0 +1,322 @@
+import { spawn } from 'child_process';
+import { getEmbedCommand } from './command.js';
+import { silentLogger } from './logger.js';
+import type { Logger } from './logger.js';
+import type { HindsightServerOptions } from './types.js';
+
+const DEFAULT_PORT = 8888;
+const DEFAULT_HOST = '127.0.0.1';
+const DEFAULT_PROFILE = 'default';
+const DEFAULT_READY_TIMEOUT_MS = 30_000;
+const DEFAULT_READY_POLL_INTERVAL_MS = 1_000;
+
+/**
+ * Manages the lifecycle of a local Hindsight daemon from a Node.js process.
+ *
+ * On {@link start}, this class:
+ *   1. Resolves the `hindsight-embed` command (via `uvx` or a local `uv run`).
+ *   2. Runs `profile create <name> --merge --port <port> [--env K=V ...]`
+ *      with every entry in {@link HindsightServerOptions.env} forwarded as
+ *      an `--env` flag.
+ *   3. Runs `daemon --profile <name> start` and waits for the start command
+ *      to exit.
+ *   4. Polls `http://host:port/health` until it returns `200` or the
+ *      `readyTimeoutMs` budget is exhausted.
+ *
+ * On {@link stop}, it runs `daemon --profile <name> stop` and returns once
+ * the command exits (or after a short grace period).
+ *
+ * This is the Node.js equivalent of the Python `hindsight-all` package's
+ * `HindsightServer`: a thin programmatic lifecycle wrapper around the
+ * Hindsight daemon. It does NOT ship an HTTP client — once `start()`
+ * resolves, use `@vectorize-io/hindsight-client` against `getBaseUrl()` for
+ * retain / recall / reflect.
+ *
+ * The class is deliberately transparent about the daemon: new CLI flags or
+ * environment variables never require a code change here — callers can pass
+ * them via `env`, `extraProfileCreateArgs`, or `extraDaemonStartArgs`.
+ */
+export class HindsightServer {
+  private readonly profile: string;
+  private readonly port: number;
+  private readonly host: string;
+  private readonly baseUrl: string;
+  private readonly embedVersion: string | undefined;
+  private readonly embedPackagePath: string | undefined;
+  private readonly userEnv: Record<string, string | undefined>;
+  private readonly extraProfileCreateArgs: string[];
+  private readonly extraDaemonStartArgs: string[];
+  private readonly platformCpuWorkaround: boolean;
+  private readonly readyTimeoutMs: number;
+  private readonly readyPollIntervalMs: number;
+  private readonly logger: Logger;
+
+  constructor(opts: HindsightServerOptions = {}) {
+    this.profile = opts.profile ?? DEFAULT_PROFILE;
+    this.port = opts.port ?? DEFAULT_PORT;
+    this.host = opts.host ?? DEFAULT_HOST;
+    this.baseUrl = `http://${this.host}:${this.port}`;
+    this.embedVersion = opts.embedVersion;
+    this.embedPackagePath = opts.embedPackagePath;
+    this.userEnv = opts.env ?? {};
+    this.extraProfileCreateArgs = opts.extraProfileCreateArgs ?? [];
+    this.extraDaemonStartArgs = opts.extraDaemonStartArgs ?? [];
+    this.platformCpuWorkaround = opts.platformCpuWorkaround ?? (process.platform === 'darwin');
+    this.readyTimeoutMs = opts.readyTimeoutMs ?? DEFAULT_READY_TIMEOUT_MS;
+    this.readyPollIntervalMs = opts.readyPollIntervalMs ?? DEFAULT_READY_POLL_INTERVAL_MS;
+    this.logger = opts.logger ?? silentLogger;
+  }
+
+  /** The base URL the daemon listens on (`http://host:port`). */
+  getBaseUrl(): string {
+    return this.baseUrl;
+  }
+
+  /** The profile name this server operates on. */
+  getProfile(): string {
+    return this.profile;
+  }
+
+  /**
+   * Ensure the daemon is configured and running. Idempotent — the underlying
+   * `profile create --merge` and `daemon start` commands tolerate re-runs.
+   */
+  async start(): Promise<void> {
+    this.logger.info(`[hindsight] starting daemon for profile "${this.profile}"`);
+
+    const env = this.buildEnv();
+    await this.configureProfile(env);
+    await this.startDaemon(env);
+    await this.waitForReady();
+
+    this.logger.info(`[hindsight] daemon ready at ${this.baseUrl}`);
+  }
+
+  /** Stop the daemon. Never throws — logs and resolves even on failure. */
+  async stop(): Promise<void> {
+    this.logger.info(`[hindsight] stopping daemon for profile "${this.profile}"`);
+
+    const [cmd, ...baseArgs] = getEmbedCommand({
+      embedVersion: this.embedVersion,
+      embedPackagePath: this.embedPackagePath,
+    });
+    const args = [...baseArgs, 'daemon', '--profile', this.profile, 'stop'];
+
+    const child = spawn(cmd, args, { stdio: 'pipe' });
+    this.pipeOutput(child, 'daemon.stop');
+
+    await new Promise<void>((resolve) => {
+      const timeout = setTimeout(() => {
+        this.logger.warn(`[hindsight] daemon stop timed out after 5s`);
+        resolve();
+      }, 5_000);
+      child.on('exit', () => {
+        clearTimeout(timeout);
+        this.logger.info(`[hindsight] daemon stopped`);
+        resolve();
+      });
+      child.on('error', (err) => {
+        clearTimeout(timeout);
+        this.logger.warn(`[hindsight] error stopping daemon: ${err.message}`);
+        resolve();
+      });
+    });
+  }
+
+  /** Probe `/health` once with a short timeout. */
+  async checkHealth(): Promise<boolean> {
+    try {
+      const res = await fetch(`${this.baseUrl}/health`, {
+        signal: AbortSignal.timeout(2_000),
+      });
+      return res.ok;
+    } catch {
+      return false;
+    }
+  }
+
+  // -------------------------------------------------------------------------
+  // Internal
+  // -------------------------------------------------------------------------
+
+  /**
+   * Merge the process env, the caller-supplied `env`, and (on macOS) the
+   * embeddings CPU workaround. Caller-supplied values always win over the
+   * workaround; undefined values are dropped.
+   */
+  private buildEnv(): NodeJS.ProcessEnv {
+    const merged: NodeJS.ProcessEnv = { ...process.env };
+
+    if (this.platformCpuWorkaround && process.platform === 'darwin') {
+      merged['HINDSIGHT_API_EMBEDDINGS_LOCAL_FORCE_CPU'] = '1';
+      merged['HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU'] = '1';
+    }
+
+    for (const [key, value] of Object.entries(this.userEnv)) {
+      if (value !== undefined) {
+        merged[key] = value;
+      }
+    }
+
+    return merged;
+  }
+
+  /**
+   * Run `profile create <name> --merge --port <port> [--env K=V ...]`.
+   * Every entry in the merged env that was passed via {@link userEnv} (or
+   * auto-applied by the CPU workaround) is forwarded as `--env`.
+   */
+  private async configureProfile(env: NodeJS.ProcessEnv): Promise<void> {
+    this.logger.info(`[hindsight] configuring profile "${this.profile}"`);
+
+    const [cmd, ...baseArgs] = getEmbedCommand({
+      embedVersion: this.embedVersion,
+      embedPackagePath: this.embedPackagePath,
+    });
+    const createArgs = [
+      ...baseArgs,
+      'profile',
+      'create',
+      this.profile,
+      '--merge',
+      '--port',
+      String(this.port),
+    ];
+
+    // Forward every env var that the caller intended for the daemon as --env.
+    // We only forward keys the caller explicitly set (userEnv) plus the CPU
+    // workaround values — not the entire process.env, to avoid leaking random
+    // host state into profile config.
+    const envForProfile = this.collectProfileEnv(env);
+    for (const [key, value] of Object.entries(envForProfile)) {
+      createArgs.push('--env', `${key}=${value}`);
+    }
+
+    createArgs.push(...this.extraProfileCreateArgs);
+
+    await this.runCommand(cmd, createArgs, env, 'profile.create');
+  }
+
+  /** Collect only the env vars that should be written into the profile file. */
+  private collectProfileEnv(env: NodeJS.ProcessEnv): Record<string, string> {
+    const out: Record<string, string> = {};
+
+    // 1. User-supplied env — always forwarded.
+    for (const [key, value] of Object.entries(this.userEnv)) {
+      if (value !== undefined) {
+        out[key] = value;
+      }
+    }
+
+    // 2. CPU workaround — only if auto-applied and not already overridden.
+    if (this.platformCpuWorkaround && process.platform === 'darwin') {
+      const cpuKeys = [
+        'HINDSIGHT_API_EMBEDDINGS_LOCAL_FORCE_CPU',
+        'HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU',
+      ];
+      for (const key of cpuKeys) {
+        if (!(key in out) && env[key] !== undefined) {
+          out[key] = env[key] as string;
+        }
+      }
+    }
+
+    return out;
+  }
+
+  private async startDaemon(env: NodeJS.ProcessEnv): Promise<void> {
+    const [cmd, ...baseArgs] = getEmbedCommand({
+      embedVersion: this.embedVersion,
+      embedPackagePath: this.embedPackagePath,
+    });
+    const args = [
+      ...baseArgs,
+      'daemon',
+      '--profile',
+      this.profile,
+      'start',
+      ...this.extraDaemonStartArgs,
+    ];
+
+    await this.runCommand(cmd, args, env, 'daemon.start');
+  }
+
+  /**
+   * Spawn `cmd` with `args`, pipe its output through the logger, and resolve
+   * once it exits with code 0. Rejects on non-zero exit or spawn error.
+   */
+  private async runCommand(
+    cmd: string,
+    args: string[],
+    env: NodeJS.ProcessEnv,
+    label: string,
+  ): Promise<void> {
+    const child = spawn(cmd, args, { stdio: 'pipe', env });
+    let output = '';
+    child.stdout?.on('data', (data: Buffer) => {
+      const text = data.toString();
+      output += text;
+      for (const line of text.trimEnd().split('\n')) {
+        if (line) this.logger.info(`[hindsight:${label}] ${line}`);
+      }
+    });
+    child.stderr?.on('data', (data: Buffer) => {
+      const text = data.toString();
+      output += text;
+      for (const line of text.trimEnd().split('\n')) {
+        if (line) this.logger.warn(`[hindsight:${label}] ${line}`);
+      }
+    });
+
+    await new Promise<void>((resolve, reject) => {
+      child.on('exit', (code) => {
+        if (code === 0) {
+          resolve();
+        } else {
+          reject(new Error(`${label} failed with code ${code}: ${output.trim()}`));
+        }
+      });
+      child.on('error', (err) => {
+        reject(new Error(`${label} failed to spawn: ${err.message}`, { cause: err }));
+      });
+    });
+  }
+
+  /** Stream a spawned child's stdout/stderr through the logger without blocking. */
+  private pipeOutput(child: ReturnType<typeof spawn>, label: string): void {
+    child.stdout?.on('data', (data: Buffer) => {
+      for (const line of data.toString().trimEnd().split('\n')) {
+        if (line) this.logger.info(`[hindsight:${label}] ${line}`);
+      }
+    });
+    child.stderr?.on('data', (data: Buffer) => {
+      for (const line of data.toString().trimEnd().split('\n')) {
+        if (line) this.logger.warn(`[hindsight:${label}] ${line}`);
+      }
+    });
+  }
+
+  /** Poll `/health` until it succeeds or `readyTimeoutMs` elapses. */
+  private async waitForReady(): Promise<void> {
+    const deadline = Date.now() + this.readyTimeoutMs;
+    let attempt = 0;
+    while (Date.now() < deadline) {
+      attempt++;
+      try {
+        const res = await fetch(`${this.baseUrl}/health`, {
+          signal: AbortSignal.timeout(this.readyPollIntervalMs),
+        });
+        if (res.ok) {
+          this.logger.debug(`[hindsight] health check passed (attempt ${attempt})`);
+          return;
+        }
+      } catch {
+        // expected while the daemon is still booting
+      }
+      await new Promise((resolve) => setTimeout(resolve, this.readyPollIntervalMs));
+    }
+    throw new Error(
+      `Hindsight daemon did not become ready within ${this.readyTimeoutMs}ms at ${this.baseUrl}`,
+    );
+  }
+}

hindsight-all-npm/src/types.ts

@@ -0,0 +1,54 @@
+import type { Logger } from './logger.js';
+
+/**
+ * Options for {@link HindsightServer}.
+ *
+ * The server is intentionally thin and pass-through: anything configurable
+ * on the daemon side (env vars or CLI flags) can be set here without needing
+ * a new dedicated option. Use {@link env} for `HINDSIGHT_*` / `OPENAI_API_KEY` /
+ * custom provider settings, and the two `extra*` arrays to append raw CLI
+ * args to `profile create` or `daemon start`.
+ *
+ * For talking to the daemon after `start()`, use `@vectorize-io/hindsight-client`
+ * against `server.getBaseUrl()`. This package does not ship its own HTTP
+ * client.
+ */
+export interface HindsightServerOptions {
+  /** Profile name used for `--profile <name>` on every sub-command. Default: `"default"`. */
+  profile?: string;
+  /** TCP port the daemon listens on. Default: `8888`. */
+  port?: number;
+  /** Hostname the daemon binds to (for health checks). Default: `127.0.0.1`. */
+  host?: string;
+  /** Version of the underlying `hindsight-embed` PyPI package to run via `uvx`. Default: `"latest"`. */
+  embedVersion?: string;
+  /** Local path to a `hindsight-embed` checkout — takes precedence over `embedVersion`. */
+  embedPackagePath?: string;
+  /**
+   * Environment variables passed to the daemon process AND written into the
+   * profile via repeated `--env KEY=VALUE` flags. This is the preferred way
+   * to surface any `HINDSIGHT_API_*` / `HINDSIGHT_EMBED_*` setting — adding a
+   * new daemon env var never requires a wrapper update.
+   *
+   * Values of `undefined` are dropped (so you can spread conditionally).
+   */
+  env?: Record<string, string | undefined>;
+  /** Extra args appended verbatim to `hindsight-embed profile create <name> --merge ...`. */
+  extraProfileCreateArgs?: string[];
+  /** Extra args appended verbatim to `hindsight-embed daemon --profile <name> start ...`. */
+  extraDaemonStartArgs?: string[];
+  /**
+   * On macOS, automatically set
+   * `HINDSIGHT_API_EMBEDDINGS_LOCAL_FORCE_CPU=1` and
+   * `HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU=1` to avoid Metal/MPS crashes in
+   * daemon mode. Default: `true` on `darwin`, ignored elsewhere. Any value set
+   * explicitly in {@link env} wins over the auto-applied value.
+   */
+  platformCpuWorkaround?: boolean;
+  /** Max time (ms) to wait for `/health` to return 200. Default: `30_000`. */
+  readyTimeoutMs?: number;
+  /** Polling interval (ms) while waiting for `/health`. Default: `1_000`. */
+  readyPollIntervalMs?: number;
+  /** Optional pluggable logger. Default: silent. */
+  logger?: Logger;
+}

hindsight-all-npm/tsconfig.json

@@ -0,0 +1,18 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "ES2022",
+    "lib": ["ES2022"],
+    "moduleResolution": "node",
+    "declaration": true,
+    "outDir": "./dist",
+    "rootDir": "./src",
+    "strict": true,
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "forceConsistentCasingInFileNames": true,
+    "resolveJsonModule": true
+  },
+  "include": ["src/**/*"],
+  "exclude": ["node_modules", "dist", "src/**/*.test.ts"]
+}

hindsight-all-npm/tsup.config.ts

@@ -0,0 +1,11 @@
+import { defineConfig } from 'tsup';
+
+export default defineConfig({
+  entry: ['src/index.ts'],
+  format: ['esm'],
+  dts: true,
+  outDir: 'dist',
+  clean: true,
+  sourcemap: true,
+  bundle: true,
+});

hindsight-all-npm/vitest.config.ts

@@ -0,0 +1,8 @@
+import { defineConfig } from 'vitest/config';
+
+export default defineConfig({
+  test: {
+    include: ['src/**/*.test.ts'],
+    environment: 'node',
+  },
+});

hindsight-all-slim/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hindsight-all-slim"
-version = "0.4.21"
+version = "0.5.4"
 description = "Hindsight: Agent Memory That Works Like Human Memory - Slim All-in-One Bundle"
 readme = "README.md"
 requires-python = ">=3.11"

hindsight-all/hindsight/embedded.py

@@ -71,7 +71,7 @@ class HindsightEmbedded:
         llm_model: Model name to use
         llm_base_url: Optional custom base URL for LLM API
         database_url: Optional database URL override (default: profile-specific pg0)
-        idle_timeout: Seconds before daemon auto-exits when idle (default: 300)
+        idle_timeout: Seconds before daemon auto-exits when idle (default: 0, disabled)
         log_level: Daemon log level (default: "info")
         ui: Whether to start the control plane web UI alongside the daemon (default: False)
         ui_port: Port for the UI. Defaults to daemon_port + 10000.
@@ -86,7 +86,7 @@ class HindsightEmbedded:
         llm_model: str = "openai/gpt-oss-120b",
         llm_base_url: Optional[str] = None,
         database_url: Optional[str] = None,
-        idle_timeout: int = 300,
+        idle_timeout: int = 0,
         log_level: str = "info",
         ui: bool = False,
         ui_port: Optional[int] = None,
@@ -102,7 +102,7 @@ class HindsightEmbedded:
             llm_model: Model name to use
             llm_base_url: Optional custom base URL for LLM API
             database_url: Optional database URL override
-            idle_timeout: Seconds before daemon auto-exits when idle
+            idle_timeout: Seconds before daemon auto-exits when idle (0 = disabled)
             log_level: Daemon log level
             ui: Whether to start the control plane web UI alongside the daemon
             ui_port: Port for the UI (defaults to daemon_port + 10000)
@@ -142,14 +142,37 @@ class HindsightEmbedded:
         self._memories_api: Optional[MemoriesAPI] = None
 
     def _ensure_started(self):
-        """Ensure daemon is running (thread-safe)."""
+        """Ensure daemon is running (thread-safe), restarting if crashed."""
         if self._started and self._client is not None:
-            return
+            if self._manager.is_running(self.profile):
+                return
+            # Daemon crashed — reset state and fall through to restart
+            logger.warning(
+                "Daemon for profile '%s' is no longer responsive, restarting...",
+                self.profile,
+            )
+            try:
+                self._client.close()
+            except Exception:
+                logger.debug("Error closing stale client", exc_info=True)
+            self._client = None
+            self._started = False
 
         with self._lock:
             # Double-check after acquiring lock
             if self._started and self._client is not None:
-                return
+                if self._manager.is_running(self.profile):
+                    return
+                logger.warning(
+                    "Daemon for profile '%s' is no longer responsive (lock path), restarting...",
+                    self.profile,
+                )
+                try:
+                    self._client.close()
+                except Exception:
+                    logger.debug("Error closing stale client", exc_info=True)
+                self._client = None
+                self._started = False
 
             if self._closed:
                 raise RuntimeError(
@@ -190,12 +213,32 @@ class HindsightEmbedded:
         if self._closed:
             return
 
-        with self._lock:
+        acquired = self._lock.acquire(timeout=5.0)
+        if not acquired:
+            # Lock is held by another thread (e.g. _ensure_started).
+            # Mark closed to prevent new operations but skip shared-state
+            # teardown — the daemon's idle timeout handles the rest.
+            logger.warning(
+                "Cleanup lock acquisition timed out for profile '%s'; "
+                "marking closed, daemon will idle-stop on its own",
+                self.profile,
+            )
+            self._closed = True
+            return
+
+        try:
             if self._closed:
                 return
 
             if self._client is not None:
-                self._client.close()
+                try:
+                    self._client.close()
+                except Exception:
+                    logger.debug(
+                        "Error closing client for profile '%s'",
+                        self.profile,
+                        exc_info=True,
+                    )
                 self._client = None
 
             # Stop UI if it was started
@@ -209,6 +252,8 @@ class HindsightEmbedded:
                 self._manager.stop(self.profile)
 
             self._closed = True
+        finally:
+            self._lock.release()
 
     def close(self, stop_daemon: bool = False):
         """
@@ -231,23 +276,10 @@ class HindsightEmbedded:
         This allows HindsightEmbedded to expose all HindsightClient methods
         without manually wrapping each one.
         """
-        # Ensure server is started before proxying
+        # Ensure server is started (and restart if crashed) before proxying
         self._ensure_started()
 
-        # Get the attribute from the underlying client
-        attr = getattr(self._client, name)
-
-        # If it's a callable, wrap it to ensure server is started
-        # (shouldn't be needed since _ensure_started already called, but defensive)
-        if callable(attr):
-
-            def wrapper(*args, **kwargs):
-                self._ensure_started()
-                return attr(*args, **kwargs)
-
-            return wrapper
-
-        return attr
+        return getattr(self._client, name)
 
     def __enter__(self):
         """Context manager entry - ensures server is started."""
@@ -372,11 +404,8 @@ class HindsightEmbedded:
         """
         Get the underlying Hindsight client for direct access.
 
-        WARNING: Using this property directly means daemon restarts won't be
-        handled automatically. Prefer using the API namespaces (banks, mental_models,
-        directives, memories) or direct method calls on HindsightEmbedded instead.
-
-        Ensures daemon is started before returning the client.
+        Ensures daemon is started (and restarts it if it has crashed) before
+        returning the client.
 
         Returns:
             Hindsight: The underlying client instance
@@ -387,9 +416,8 @@ class HindsightEmbedded:
 
             embedded = HindsightEmbedded(profile="myapp", ...)
 
-            # Direct access (not recommended - daemon crashes won't be handled)
             client = embedded.client
-            banks = client.list_banks()  # If daemon crashes, this will fail
+            banks = client.list_banks()
             ```
         """
         self._ensure_started()
@@ -403,8 +431,13 @@ class HindsightEmbedded:
 
     @property
     def is_running(self) -> bool:
-        """Check if the client is initialized."""
-        return self._started and not self._closed and self._client is not None
+        """Check if the client is initialized and the daemon is responsive."""
+        return (
+            self._started
+            and not self._closed
+            and self._client is not None
+            and self._manager.is_running(self.profile)
+        )
 
     @property
     def ui_url(self) -> str:

hindsight-all/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "hindsight-all"
-version = "0.4.21"
+version = "0.5.4"
 description = "Hindsight: Agent Memory That Works Like Human Memory - All-in-One Bundle"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -20,6 +20,9 @@ hindsight-client = { workspace = true }
 hindsight-embed = { workspace = true }
 
 [project.optional-dependencies]
+local-llm = [
+    "hindsight-api-slim[local-llm]>=0.4.17",
+]
 test = [
     "pytest>=7.0.0",
     "pytest-asyncio>=0.21.0",

hindsight-all/tests/test_cleanup_timeout.py

@@ -0,0 +1,56 @@
+"""
+Unit test for _cleanup lock timeout behavior.
+
+Verifies that _cleanup completes even when the lock is held by another thread,
+instead of hanging indefinitely (fixes #952).
+"""
+
+import threading
+import time
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+def test_cleanup_completes_when_lock_held():
+    """
+    _cleanup should complete (best-effort) even when self._lock is held
+    by another thread, e.g. during a long _ensure_started call.
+    """
+    with patch.dict("sys.modules", {
+        "hindsight_client": MagicMock(),
+        "hindsight_embed": MagicMock(),
+        "hindsight.api_namespaces": MagicMock(),
+    }):
+        from hindsight.embedded import HindsightEmbedded
+
+        client = HindsightEmbedded.__new__(HindsightEmbedded)
+        client.profile = "test"
+        client._lock = threading.Lock()
+        client._closed = False
+        client._client = None
+        client._started = False
+        client._ui = False
+
+        # Simulate another thread holding the lock
+        client._lock.acquire()
+
+        cleanup_done = threading.Event()
+
+        def run_cleanup():
+            client._cleanup()
+            cleanup_done.set()
+
+        t = threading.Thread(target=run_cleanup)
+        t.start()
+
+        # Cleanup should complete within the timeout (5s) + margin
+        assert cleanup_done.wait(timeout=8.0), (
+            "_cleanup hung instead of timing out on lock acquisition"
+        )
+
+        # Release the lock from the simulating thread
+        client._lock.release()
+        t.join(timeout=1.0)
+
+        assert client._closed, "Client should be marked as closed after cleanup"

hindsight-all/tests/test_embedded.py

@@ -401,3 +401,42 @@ def test_embedded_ui_flag(llm_config):
 
     finally:
         client.close()
+
+
+def test_embedded_daemon_crash_recovery(llm_config):
+    """
+    Test that HindsightEmbedded recovers when the daemon crashes.
+
+    Simulates a crash by stopping the daemon, then verifies
+    that the next operation transparently restarts it.
+    """
+    profile = f"test_crash_{uuid.uuid4().hex[:8]}"
+    bank_id = f"bank_{uuid.uuid4().hex[:8]}"
+
+    client = HindsightEmbedded(profile=profile, log_level="info", **llm_config)
+
+    try:
+        # Start daemon and store a memory
+        result = client.retain(bank_id=bank_id, content="Before crash")
+        assert result.success, "Initial retain should succeed"
+        assert client.is_running, "Daemon should be running"
+
+        original_url = client.url
+
+        # Simulate daemon crash by stopping it
+        client._manager.stop(client.profile)
+        assert not client._manager.is_running(client.profile), (
+            "Daemon should be stopped after simulated crash"
+        )
+
+        # Next operation should transparently restart the daemon
+        result2 = client.retain(bank_id=bank_id, content="After crash recovery")
+        assert result2.success, "Retain after crash recovery should succeed"
+        assert client.is_running, "Daemon should be running again after recovery"
+
+        # Verify recall still works
+        recall_result = client.recall(bank_id=bank_id, query="crash")
+        assert isinstance(recall_result.results, list), "Recall should return results"
+
+    finally:
+        client.close()

hindsight-api-slim/hindsight_api/__init__.py

@@ -46,4 +46,4 @@ __all__ = [
     "RemoteTEICrossEncoder",
     "LLMConfig",
 ]
-__version__ = "0.4.21"
+__version__ = "0.5.4"

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

@@ -375,6 +375,140 @@ def decommission_worker(
         typer.echo(f"No tasks found for worker '{worker_id}'")
 
 
+async def _decommission_all_workers(db_url: str, schema: str = "public") -> list[dict[str, Any]]:
+    """Release all processing tasks from all workers, setting them back to pending status."""
+    is_pg0, instance_name, _ = parse_pg0_url(db_url)
+    if is_pg0:
+        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
+    resolved_url = await resolve_database_url(db_url)
+
+    conn = await asyncpg.connect(resolved_url)
+    try:
+        table = _fq_table("async_operations", schema)
+        rows = await conn.fetch(
+            f"""
+            UPDATE {table}
+            SET status = 'pending', worker_id = NULL, claimed_at = NULL, updated_at = now()
+            WHERE status = 'processing'
+            RETURNING operation_id, worker_id, operation_type
+            """,
+        )
+        return [dict(r) for r in rows]
+    finally:
+        await conn.close()
+
+
+@app.command(name="decommission-workers")
+def decommission_workers(
+    schema: str = typer.Option("public", "--schema", "-s", help="Database schema"),
+    yes: bool = typer.Option(False, "--yes", "-y", help="Skip confirmation prompt"),
+):
+    """Release all processing tasks from all workers (sets status back to pending).
+
+    Use this command to recover from situations where one or more workers have crashed
+    or been removed without graceful shutdown. All tasks currently in 'processing' status
+    will be released back to the queue regardless of which worker owns them.
+    """
+    config = HindsightConfig.from_env()
+
+    if not config.database_url:
+        typer.echo("Error: Database URL not configured.", err=True)
+        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
+        raise typer.Exit(1)
+
+    if not yes:
+        typer.confirm(
+            "This will release ALL processing tasks from ALL workers back to pending. Continue?",
+            abort=True,
+        )
+
+    typer.echo(f"Decommissioning all workers (schema: {schema})...")
+
+    released = asyncio.run(_decommission_all_workers(config.database_url, schema))
+
+    if released:
+        # Group by worker_id for summary
+        by_worker: dict[str, int] = {}
+        for row in released:
+            wid = row["worker_id"] or "unknown"
+            by_worker[wid] = by_worker.get(wid, 0) + 1
+
+        typer.echo(f"Released {len(released)} task(s):")
+        for wid, count in by_worker.items():
+            typer.echo(f"  {wid}: {count} task(s)")
+    else:
+        typer.echo("No processing tasks found")
+
+
+async def _worker_status(db_url: str, schema: str = "public") -> list[dict[str, Any]]:
+    """Get all processing tasks grouped by worker with their last update time."""
+    is_pg0, instance_name, _ = parse_pg0_url(db_url)
+    if is_pg0:
+        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
+    resolved_url = await resolve_database_url(db_url)
+
+    conn = await asyncpg.connect(resolved_url)
+    try:
+        table = _fq_table("async_operations", schema)
+        rows = await conn.fetch(
+            f"""
+            SELECT worker_id, operation_id, operation_type, bank_id,
+                   claimed_at, updated_at,
+                   now() - claimed_at AS running_for,
+                   now() - updated_at AS last_update_ago
+            FROM {table}
+            WHERE status = 'processing'
+            ORDER BY worker_id, claimed_at
+            """,
+        )
+        return [dict(r) for r in rows]
+    finally:
+        await conn.close()
+
+
+@app.command(name="worker-status")
+def worker_status(
+    schema: str = typer.Option("public", "--schema", "-s", help="Database schema"),
+):
+    """Show all currently processing tasks grouped by worker.
+
+    Displays each worker's active tasks with operation type, bank, how long
+    the task has been running, and when it was last updated. Useful for
+    identifying dead workers with orphaned tasks.
+    """
+    config = HindsightConfig.from_env()
+
+    if not config.database_url:
+        typer.echo("Error: Database URL not configured.", err=True)
+        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
+        raise typer.Exit(1)
+
+    rows = asyncio.run(_worker_status(config.database_url, schema))
+
+    if not rows:
+        typer.echo("No processing tasks found")
+        return
+
+    # Group by worker_id
+    by_worker: dict[str, list[dict[str, Any]]] = {}
+    for row in rows:
+        wid = row["worker_id"] or "unknown"
+        by_worker.setdefault(wid, []).append(row)
+
+    typer.echo(f"Processing tasks across {len(by_worker)} worker(s):\n")
+    for wid, tasks in by_worker.items():
+        typer.echo(f"Worker: {wid} ({len(tasks)} task(s))")
+        for task in tasks:
+            op_id = str(task["operation_id"])[:8]
+            running_for = task["running_for"]
+            last_update = task["last_update_ago"]
+            typer.echo(
+                f"  {op_id}  {task['operation_type']:<20s} bank={task['bank_id']}"
+                f"  running={running_for}  last_update={last_update} ago"
+            )
+        typer.echo("")
+
+
 def main():
     app()
 

hindsight-api-slim/hindsight_api/alembic/versions/2eee35aa3cfc_case_insensitive_entities_trgm_index.py

@@ -0,0 +1,45 @@
+"""Recreate entities trigram index on LOWER(canonical_name) for case-insensitive matching
+
+The previous GIN trigram index on canonical_name was case-sensitive, causing
+"Alice" and "alice" to have different trigram sets. This recreates it on
+LOWER(canonical_name) so the % operator matches case-insensitively.
+
+Revision ID: 2eee35aa3cfc
+Revises: d6e7f8a9b0c1
+Create Date: 2026-03-31
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "2eee35aa3cfc"
+down_revision: str | Sequence[str] | None = "d6e7f8a9b0c1"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    # Drop the old case-sensitive trigram index
+    op.execute("DROP INDEX IF EXISTS entities_canonical_name_trgm_idx")
+    # Create case-insensitive trigram index on LOWER(canonical_name)
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS entities_canonical_name_lower_trgm_idx "
+        f"ON {schema}entities USING GIN (LOWER(canonical_name) gin_trgm_ops)"
+    )
+
+
+def downgrade() -> None:
+    op.execute("DROP INDEX IF EXISTS entities_canonical_name_lower_trgm_idx")
+    schema = _get_schema_prefix()
+    # Restore original case-sensitive index
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS entities_canonical_name_trgm_idx "
+        f"ON {schema}entities USING GIN (canonical_name gin_trgm_ops)"
+    )

hindsight-api-slim/hindsight_api/alembic/versions/8c6fa6f7230b_merge_v053_divergent_heads.py

@@ -0,0 +1,40 @@
+"""Merge divergent migration heads for v0.5.3
+
+v0.5.3 shipped with two migration heads that were never unified:
+
+  * ``c4x5y6z7a8b9`` — delta-refresh chain
+    (``add_last_refreshed_source_query`` ->
+     ``add_structured_content_to_mental_models`` ->
+     ``backsweep_orphan_observations_v2``)
+
+  * ``h3i4j5k6l7m8`` — per-bank vector indexes / audit log chain
+    (the ``merge_heads_and_add_unit_entities_index`` subtree)
+
+Both fork from ``z1u2v3w4x5y6``. Upgrades from v0.5.2 still succeed — the
+walker applies the three c4x5 revisions and leaves the database stamped at
+both heads — but the result is a split DAG: ``alembic upgrade head``
+(singular) is ambiguous, and any future migration has to pick one head as
+its parent, orphaning the other.
+
+This revision linearises the DAG into a single head. It has no schema
+effect.
+
+Revision ID: 8c6fa6f7230b
+Revises: c4x5y6z7a8b9, h3i4j5k6l7m8
+Create Date: 2026-04-18
+"""
+
+from collections.abc import Sequence
+
+revision: str = "8c6fa6f7230b"
+down_revision: str | Sequence[str] | None = ("c4x5y6z7a8b9", "h3i4j5k6l7m8")
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    pass
+
+
+def downgrade() -> None:
+    pass

hindsight-api-slim/hindsight_api/alembic/versions/a2v3w4x5y6z7_add_last_refreshed_source_query.py

@@ -0,0 +1,38 @@
+"""Add last_refreshed_source_query column to mental_models
+
+Revision ID: a2v3w4x5y6z7
+Revises: z1u2v3w4x5y6
+Create Date: 2026-04-15
+
+Tracks the source_query that was used during the most recent refresh.
+Used by delta-mode refresh to detect when the query has changed: if it has,
+delta mode falls back to a full regeneration because the surgical-edit
+assumption (same topic, new facts) no longer holds.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "a2v3w4x5y6z7"
+down_revision: str | Sequence[str] | None = "z1u2v3w4x5y6"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD COLUMN IF NOT EXISTS last_refreshed_source_query TEXT
+    """)
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS last_refreshed_source_query")

hindsight-api-slim/hindsight_api/alembic/versions/a4b5c6d7e8f9_fix_per_bank_vector_index_type.py

@@ -0,0 +1,142 @@
+"""Fix per-bank vector indexes to match configured extension
+
+Revision ID: a4b5c6d7e8f9
+Revises: 2eee35aa3cfc
+Create Date: 2026-04-01
+
+Migration d5e6f7a8b9c0 hardcoded HNSW when creating per-bank partial vector
+indexes, ignoring HINDSIGHT_API_VECTOR_EXTENSION. Banks that existed when that
+migration ran got HNSW indexes even when pgvectorscale (DiskANN) or vchord
+was configured.
+
+This migration detects the mismatch and recreates the affected indexes with
+the correct type. Skipped entirely when the configured extension is pgvector
+(the default), since those indexes are already correct.
+"""
+
+import os
+from collections.abc import Sequence
+
+from alembic import context, op
+from sqlalchemy import text
+
+revision: str = "a4b5c6d7e8f9"
+down_revision: str | Sequence[str] | None = "2eee35aa3cfc"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+_FACT_TYPES: dict[str, str] = {
+    "world": "worl",
+    "experience": "expr",
+    "observation": "obsv",
+}
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def _target_index_type() -> str | None:
+    """Return the target index type, or None if pgvector (no fix needed)."""
+    ext = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
+    if ext == "pgvectorscale":
+        return "diskann"
+    elif ext == "vchord":
+        return "vchordrq"
+    return None
+
+
+def _vector_index_using_clause() -> str:
+    """Return the USING clause based on the configured vector extension."""
+    ext = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
+    if ext == "pgvectorscale":
+        return "USING diskann (embedding vector_cosine_ops) WITH (num_neighbors = 50)"
+    elif ext == "vchord":
+        return "USING vchordrq (embedding vector_l2_ops)"
+    else:
+        return "USING hnsw (embedding vector_cosine_ops)"
+
+
+def upgrade() -> None:
+    target = _target_index_type()
+    if target is None:
+        # pgvector — indexes are already HNSW, nothing to fix
+        return
+
+    bind = op.get_bind()
+    schema_name = context.config.get_main_option("target_schema")
+    schema = _get_schema_prefix()
+    table_ref = f'"{schema_name}".memory_units' if schema_name else "memory_units"
+    banks_ref = f'"{schema_name}".banks' if schema_name else "banks"
+    using_clause = _vector_index_using_clause()
+    pg_schema = schema_name or "public"
+
+    rows = bind.execute(text(f"SELECT bank_id, internal_id FROM {banks_ref}")).fetchall()  # noqa: S608
+    for row in rows:
+        bank_id = row[0]
+        internal_id = str(row[1]).replace("-", "")[:16]
+        escaped_bank_id = bank_id.replace("'", "''")
+        for ft, ft_short in _FACT_TYPES.items():
+            idx_name = f"idx_mu_emb_{ft_short}_{internal_id}"
+
+            # Check if this index exists and what type it is
+            idx_info = bind.execute(
+                text("SELECT indexdef FROM pg_indexes WHERE schemaname = :schema AND indexname = :idx"),
+                {"schema": pg_schema, "idx": idx_name},
+            ).fetchone()
+
+            if idx_info is None:
+                # Index doesn't exist — create it with the correct type
+                bind.execute(
+                    text(
+                        f"CREATE INDEX IF NOT EXISTS {idx_name} "
+                        f"ON {table_ref} {using_clause} "
+                        f"WHERE fact_type = '{ft}' AND bank_id = '{escaped_bank_id}'"
+                    )
+                )
+                continue
+
+            indexdef = idx_info[0].lower()
+            if target in indexdef:
+                # Already the correct type
+                continue
+
+            # Wrong type — drop and recreate
+            bind.execute(text(f"DROP INDEX IF EXISTS {schema}{idx_name}"))
+            bind.execute(
+                text(
+                    f"CREATE INDEX IF NOT EXISTS {idx_name} "
+                    f"ON {table_ref} {using_clause} "
+                    f"WHERE fact_type = '{ft}' AND bank_id = '{escaped_bank_id}'"
+                )
+            )
+
+
+def downgrade() -> None:
+    # Downgrade recreates indexes as HNSW (the original hardcoded behavior)
+    target = _target_index_type()
+    if target is None:
+        return
+
+    bind = op.get_bind()
+    schema_name = context.config.get_main_option("target_schema")
+    schema = _get_schema_prefix()
+    table_ref = f'"{schema_name}".memory_units' if schema_name else "memory_units"
+    banks_ref = f'"{schema_name}".banks' if schema_name else "banks"
+
+    rows = bind.execute(text(f"SELECT bank_id, internal_id FROM {banks_ref}")).fetchall()  # noqa: S608
+    for row in rows:
+        bank_id = row[0]
+        internal_id = str(row[1]).replace("-", "")[:16]
+        escaped_bank_id = bank_id.replace("'", "''")
+        for ft, ft_short in _FACT_TYPES.items():
+            idx_name = f"idx_mu_emb_{ft_short}_{internal_id}"
+            bind.execute(text(f"DROP INDEX IF EXISTS {schema}{idx_name}"))
+            bind.execute(
+                text(
+                    f"CREATE INDEX IF NOT EXISTS {idx_name} "
+                    f"ON {table_ref} USING hnsw (embedding vector_cosine_ops) "
+                    f"WHERE fact_type = '{ft}' AND bank_id = '{escaped_bank_id}'"
+                )
+            )

hindsight-api-slim/hindsight_api/alembic/versions/b3w4x5y6z7a8_add_structured_content_to_mental_models.py

@@ -0,0 +1,44 @@
+"""Add structured_content JSONB column to mental_models
+
+Revision ID: b3w4x5y6z7a8
+Revises: a2v3w4x5y6z7
+Create Date: 2026-04-16
+
+Stores the structured representation of a mental model document (sections,
+blocks). The plain ``content`` column remains the rendered markdown shown to
+users. ``structured_content`` is the source of truth for delta-mode refreshes:
+each refresh applies a list of typed operations to the structured doc, then
+re-renders to markdown — so unchanged sections come through byte-identical
+without an LLM round-trip.
+
+Nullable: existing markdown-only mental models continue to work in full mode;
+the column is populated lazily the first time a model is refreshed in delta
+mode.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "b3w4x5y6z7a8"
+down_revision: str | Sequence[str] | None = "a2v3w4x5y6z7"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD COLUMN IF NOT EXISTS structured_content JSONB
+    """)
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS structured_content")

hindsight-api-slim/hindsight_api/alembic/versions/c4x5y6z7a8b9_backsweep_orphan_observations_v2.py

@@ -0,0 +1,66 @@
+"""backsweep_orphan_observations_v2
+
+Re-run of Pass 2 from migration ``g7h8i9j0k1l2_backsweep_orphan_observations``
+to sweep observations that became orphaned between then and now.
+
+Why we need it again:
+  ``fact_storage.handle_document_tracking`` (the retain/upsert path) deleted
+  the existing document via the FK cascade — which removes the source
+  ``memory_units`` — but never invalidated the observations derived from
+  them.  Only the explicit ``MemoryEngine.delete_document`` API called
+  ``_delete_stale_observations_for_memories``.  Every document re-ingest
+  therefore left orphan observations whose ``source_memory_ids`` arrays
+  pointed at IDs that no longer existed in ``memory_units``.
+
+  ``handle_document_tracking`` now calls the same cleanup helper before the
+  cascade, so no new orphans will accumulate going forward.  This migration
+  cleans up the historical residue.
+
+Identical to Pass 2 of g7h8i9j0k1l2.  Pass 1 (memory_units whose bank is
+gone) is intentionally not re-run; that scenario has no fresh source.
+
+Revision ID: c4x5y6z7a8b9
+Revises: b3w4x5y6z7a8
+Create Date: 2026-04-16
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "c4x5y6z7a8b9"
+down_revision: str | Sequence[str] | None = "b3w4x5y6z7a8"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    mu = f"{schema}memory_units"
+
+    # Delete observations whose every source_memory_id refers to a now-deleted
+    # memory_unit (or the array is empty).  Observations with at least one
+    # surviving source are left alone — the consolidation engine will refresh
+    # their text on the next pass.
+    op.execute(
+        f"""
+        DELETE FROM {mu} orphan
+        WHERE orphan.fact_type = 'observation'
+          AND NOT EXISTS (
+              SELECT 1
+              FROM {mu} src
+              WHERE src.id = ANY(orphan.source_memory_ids)
+                AND src.bank_id = orphan.bank_id
+          )
+        """
+    )
+
+
+def downgrade() -> None:
+    # Deleted rows cannot be restored.
+    pass

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

@@ -1,4 +1,4 @@
-"""Add internal_id to banks and per-(bank, fact_type) partial HNSW indexes
+"""Add internal_id to banks and per-(bank, fact_type) partial vector indexes
 
 Revision ID: d5e6f7a8b9c0
 Revises: a3b4c5d6e7f8
@@ -6,25 +6,20 @@ Create Date: 2026-03-11
 
 This migration:
 1. Adds internal_id UUID column to banks (stable identifier for index naming)
-2. Drops the global HNSW index (competes with per-bank partial indexes)
-3. Creates per-(bank_id, fact_type) partial HNSW indexes for all existing banks
-   (new banks get indexes created at bank-creation time via bank_utils.create_bank_hnsw_indexes)
+2. Drops the global vector index (competes with per-bank partial indexes)
+3. Creates per-(bank_id, fact_type) partial vector indexes for all existing banks
+   using the configured vector extension (HNSW for pgvector, DiskANN for
+   pgvectorscale, vchordrq for vchord).
+   (new banks get indexes created at bank-creation time via bank_utils.create_bank_vector_indexes)
 
 Why per-(bank, fact_type) indexes:
 - fact_type-only partial indexes are never chosen by the planner when bank_id is in the WHERE
   clause, because the idx_memory_units_bank_id B-tree index always wins at planning time.
 - Per-(bank, fact_type) partial indexes have both predicates matching → planner selects them.
-- The global HNSW index competes for larger partitions (world, observation) and must be dropped.
-
-For large deployments, create indexes CONCURRENTLY before running this migration:
-    SELECT internal_id, bank_id FROM banks;
-    -- for each bank and each fact_type in (world, experience, observation):
-    CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_mu_emb_{ft}_{uid16}
-        ON memory_units USING hnsw (embedding vector_cosine_ops)
-        WHERE fact_type = '{ft}' AND bank_id = '{bank_id}';
-    DROP INDEX CONCURRENTLY IF EXISTS idx_memory_units_embedding;
+- The global vector index competes for larger partitions (world, observation) and must be dropped.
 """
 
+import os
 from collections.abc import Sequence
 
 from alembic import context, op
@@ -35,7 +30,7 @@ down_revision: str | Sequence[str] | None = "c3d4e5f6g7h8"
 branch_labels: str | Sequence[str] | None = None
 depends_on: str | Sequence[str] | None = None
 
-_HNSW_FACT_TYPES: dict[str, str] = {
+_FACT_TYPES: dict[str, str] = {
     "world": "worl",
     "experience": "expr",
     "observation": "obsv",
@@ -47,6 +42,17 @@ def _get_schema_prefix() -> str:
     return f'"{schema}".' if schema else ""
 
 
+def _vector_index_using_clause() -> str:
+    """Return the USING clause based on the configured vector extension."""
+    ext = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
+    if ext == "pgvectorscale":
+        return "USING diskann (embedding vector_cosine_ops) WITH (num_neighbors = 50)"
+    elif ext == "vchord":
+        return "USING vchordrq (embedding vector_l2_ops)"
+    else:
+        return "USING hnsw (embedding vector_cosine_ops)"
+
+
 def upgrade() -> None:
     schema = _get_schema_prefix()
 
@@ -56,33 +62,35 @@ def upgrade() -> None:
     )
     op.execute(f"ALTER TABLE {schema}banks ADD CONSTRAINT banks_internal_id_unique UNIQUE (internal_id)")
 
-    # 2. Drop any fact_type-only partial HNSW indexes that may exist from prior migrations
+    # 2. Drop any fact_type-only partial indexes that may exist from prior migrations
     #    (bank_id B-tree always wins over them when bank_id is in the WHERE clause)
     op.execute(f"DROP INDEX IF EXISTS {schema}idx_mu_emb_world")
     op.execute(f"DROP INDEX IF EXISTS {schema}idx_mu_emb_observation")
     op.execute(f"DROP INDEX IF EXISTS {schema}idx_mu_emb_experience")
 
-    # 4. Drop global HNSW index (competes with per-bank partial indexes)
+    # 4. Drop global vector index (competes with per-bank partial indexes)
     op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_embedding")
 
-    # 5. Create per-(bank, fact_type) partial HNSW indexes for all existing banks
+    # 5. Create per-(bank, fact_type) partial vector indexes for all existing banks
+    #    using the configured extension (HNSW / DiskANN / vchordrq)
     bind = op.get_bind()
     schema_name = context.config.get_main_option("target_schema")
     table_ref = f'"{schema_name}".memory_units' if schema_name else "memory_units"
     banks_ref = f'"{schema_name}".banks' if schema_name else "banks"
+    using_clause = _vector_index_using_clause()
 
     rows = bind.execute(text(f"SELECT bank_id, internal_id FROM {banks_ref}")).fetchall()  # noqa: S608
     for row in rows:
         bank_id = row[0]
         internal_id = str(row[1]).replace("-", "")[:16]
         escaped_bank_id = bank_id.replace("'", "''")
-        for ft, ft_short in _HNSW_FACT_TYPES.items():
+        for ft, ft_short in _FACT_TYPES.items():
             idx_name = f"idx_mu_emb_{ft_short}_{internal_id}"
             # Index name is schema-unqualified (indexes live in the schema of their table)
             bind.execute(
                 text(
                     f"CREATE INDEX IF NOT EXISTS {idx_name} "
-                    f"ON {table_ref} USING hnsw (embedding vector_cosine_ops) "
+                    f"ON {table_ref} {using_clause} "
                     f"WHERE fact_type = '{ft}' AND bank_id = '{escaped_bank_id}'"
                 )
             )

hindsight-api-slim/hindsight_api/alembic/versions/d6e7f8a9b0c1_drop_documents_metadata_column.py

@@ -0,0 +1,39 @@
+"""Drop unused metadata column from documents table
+
+Revision ID: d6e7f8a9b0c1
+Revises: c2d3e4f5g6h7, c5d6e7f8a9b0
+Create Date: 2026-03-30
+
+The metadata column on documents was always stored as an empty dict {}.
+Actual document metadata is stored inside retain_params.metadata.
+
+This migration was originally shipped in v0.4.22, then its file was deleted
+in v0.5.0 (and its revision ID accidentally reused by 2eee35aa3cfc).
+Restoring the file so that databases stamped at this revision can upgrade
+cleanly to v0.5.x+.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "d6e7f8a9b0c1"
+down_revision: str | Sequence[str] | None = ("c2d3e4f5g6h7", "c5d6e7f8a9b0")
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}documents DROP COLUMN IF EXISTS metadata")
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}documents ADD COLUMN IF NOT EXISTS metadata jsonb DEFAULT '{{}}'")

hindsight-api-slim/hindsight_api/alembic/versions/f1a2b3c4d5e6_add_memory_links_composite_index.py

@@ -5,7 +5,7 @@ Revises: e0a1b2c3d4e5
 Create Date: 2025-01-12
 
 Add composite index on memory_links (from_unit_id, link_type, weight DESC)
-to optimize MPFP graph traversal queries that need top-k edges per type.
+to optimize graph traversal queries that need top-k edges per type.
 """
 
 from collections.abc import Sequence
@@ -26,7 +26,7 @@ def _get_schema_prefix() -> str:
 
 
 def upgrade() -> None:
-    """Add composite index for efficient MPFP edge loading."""
+    """Add composite index for efficient graph retrieval edge loading."""
     schema = _get_schema_prefix()
     # Create composite index for efficient top-k per (from_node, link_type) queries
     # This enables LATERAL joins to use index-only scans with early termination

hindsight-api-slim/hindsight_api/alembic/versions/g2h3i4j5k6l7_remove_opinion_fact_type.py

@@ -0,0 +1,83 @@
+"""remove_opinion_fact_type
+
+Revision ID: g2h3i4j5k6l7
+Revises: f1a2b3c4d5e6
+Create Date: 2026-04-02
+
+Remove the deprecated 'opinion' fact type: drop opinion-specific indexes,
+update CHECK constraints, delete any remaining opinion rows, and drop the
+confidence_score column (was only used for opinions, always NULL otherwise).
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "g2h3i4j5k6l7"
+down_revision: str | Sequence[str] | None = "f1a2b3c4d5e6"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (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()
+
+    # 1. Delete any remaining opinion rows
+    op.execute(f"DELETE FROM {schema}memory_units WHERE fact_type = 'opinion'")
+
+    # 2. Drop opinion-specific indexes
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_opinion_confidence")
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_opinion_date")
+
+    # 3. Drop confidence_score constraints and column (only used for opinions, always NULL otherwise)
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS confidence_score_fact_type_check")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_confidence_score_check")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS confidence_score")
+
+    # 4. Replace fact_type CHECK constraint
+    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 "
+        f"CHECK (fact_type IN ('world', 'experience', 'observation'))"
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+
+    # Restore confidence_score column
+    op.execute(f"ALTER TABLE {schema}memory_units ADD COLUMN IF NOT EXISTS confidence_score float")
+    op.execute(
+        f"ALTER TABLE {schema}memory_units ADD CONSTRAINT memory_units_confidence_score_check "
+        f"CHECK (confidence_score IS NULL OR (confidence_score >= 0.0 AND confidence_score <= 1.0))"
+    )
+    op.execute(
+        f"ALTER TABLE {schema}memory_units ADD CONSTRAINT confidence_score_fact_type_check "
+        f"CHECK ((fact_type = 'opinion' AND confidence_score IS NOT NULL) OR "
+        f"(fact_type = 'observation') OR "
+        f"(fact_type NOT IN ('opinion', 'observation') AND confidence_score IS NULL))"
+    )
+
+    # Restore original fact_type CHECK constraint (with opinion)
+    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 "
+        f"CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation'))"
+    )
+
+    # Recreate opinion indexes
+    op.execute(
+        f"CREATE INDEX idx_memory_units_opinion_confidence ON {schema}memory_units "
+        f"(bank_id, confidence_score DESC) WHERE fact_type = 'opinion'"
+    )
+    op.execute(
+        f"CREATE INDEX idx_memory_units_opinion_date ON {schema}memory_units "
+        f"(bank_id, event_date DESC) WHERE fact_type = 'opinion'"
+    )

hindsight-api-slim/hindsight_api/alembic/versions/h3i4j5k6l7m8_merge_heads_and_add_unit_entities_index.py

@@ -0,0 +1,42 @@
+"""Merge 3 migration heads and add unit_entities composite index
+
+Revision ID: h3i4j5k6l7m8
+Revises: a4b5c6d7e8f9, g2h3i4j5k6l7
+Create Date: 2026-04-07
+
+Merges three unmerged migration heads into one, and adds a composite index
+(entity_id, unit_id) on unit_entities for index-only scans in the LATERAL
+entity expansion query.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "h3i4j5k6l7m8"
+down_revision: str | Sequence[str] | None = ("a4b5c6d7e8f9", "g2h3i4j5k6l7")
+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()
+    # Composite index enables index-only scans for entity_id -> unit_id lookups
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_unit_entities_entity_unit ON {schema}unit_entities (entity_id, unit_id)"
+    )
+    # Drop the now-redundant single-column index
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_unit_entities_entity")
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_unit_entities_entity_unit")
+    # Restore the single-column index
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_unit_entities_entity ON {schema}unit_entities (entity_id)")

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

@@ -97,6 +97,7 @@ def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
     _SINGLE_BANK_TOOLS: frozenset[str] = frozenset(
         {
             "retain",
+            "sync_retain",
             "recall",
             "reflect",
             "list_mental_models",
@@ -156,24 +157,65 @@ def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
     return mcp
 
 
+def _get_mcp_tools(mcp: FastMCP) -> dict:
+    """Get tool name→object mapping, compatible with FastMCP 2.x and 3.x."""
+    # FastMCP 2.x: _tool_manager._tools
+    if hasattr(mcp, "_tool_manager"):
+        return mcp._tool_manager._tools  # type: ignore[union-attr]
+    # FastMCP 3.x: _local_provider._components with "tool:" prefix
+    if hasattr(mcp, "_local_provider"):
+        return {
+            k.split(":")[1].split("@")[0]: v
+            for k, v in mcp._local_provider._components.items()  # type: ignore[union-attr]
+            if k.startswith("tool:")
+        }
+    msg = "Cannot locate tools on FastMCP instance"
+    raise AttributeError(msg)
+
+
 def _make_tools_tolerant(mcp: FastMCP) -> None:
-    """Wrap all tool run methods to strip unknown arguments before validation.
+    """Wrap all tool run methods to strip unknown arguments and coerce string-encoded JSON.
 
     LLMs frequently add extra fields like "explanation" or "reasoning" to tool calls.
     FastMCP's Pydantic TypeAdapter rejects these with "Unexpected keyword argument".
-    This wraps each tool's run() to filter arguments to only known parameters.
+
+    LLMs also frequently serialize list/dict arguments as JSON strings instead of native
+    types (e.g., tags='["a","b"]' instead of tags=["a","b"]). This auto-coerces them.
+
+    This wraps each tool's run() to apply both fixes before validation.
     """
     try:
-        for name, tool in mcp._tool_manager._tools.items():
+        tools = _get_mcp_tools(mcp)
+        for name, tool in tools.items():
             if hasattr(tool, "parameters") and tool.parameters:
-                allowed = set(tool.parameters.get("properties", {}).keys())
+                properties = tool.parameters.get("properties", {})
+                allowed = set(properties.keys())
+
+                # Build sets of parameter names that expect array or object types.
+                # Handles both direct types {"type": "array"} and anyOf/oneOf unions
+                # like {"anyOf": [{"type": "array", ...}, {"type": "null"}]}.
+                array_params: set[str] = set()
+                object_params: set[str] = set()
+                for param_name, param_schema in properties.items():
+                    _collect_coercible_types(param_schema, param_name, array_params, object_params)
+
                 original_run = tool.run
 
-                async def _tolerant_run(arguments, _allowed=allowed, _orig=original_run):
+                async def _tolerant_run(
+                    arguments,
+                    _allowed=allowed,
+                    _orig=original_run,
+                    _array_params=array_params,
+                    _object_params=object_params,
+                ):
                     extra_keys = set(arguments.keys()) - _allowed
                     if extra_keys:
                         logger.debug(f"Stripping unknown arguments from tool call: {extra_keys}")
                         arguments = {k: v for k, v in arguments.items() if k in _allowed}
+
+                    # Coerce string-encoded JSON for list/dict parameters
+                    arguments = _coerce_string_json(arguments, _array_params, _object_params)
+
                     return await _orig(arguments)
 
                 # FunctionTool is a Pydantic model with extra='forbid', so use
@@ -183,6 +225,59 @@ def _make_tools_tolerant(mcp: FastMCP) -> None:
         logger.warning(f"Could not make tools tolerant of extra arguments: {e}")
 
 
+def _collect_coercible_types(schema: dict, param_name: str, array_params: set[str], object_params: set[str]) -> None:
+    """Check a JSON Schema property and add param_name to array_params/object_params if applicable."""
+    # Direct type
+    schema_type = schema.get("type")
+    if schema_type == "array":
+        array_params.add(param_name)
+        return
+    if schema_type == "object":
+        object_params.add(param_name)
+        return
+
+    # anyOf / oneOf unions (e.g., list[str] | None → {"anyOf": [{"type": "array"}, {"type": "null"}]})
+    for variant in schema.get("anyOf", []) + schema.get("oneOf", []):
+        variant_type = variant.get("type")
+        if variant_type == "array":
+            array_params.add(param_name)
+            return
+        if variant_type == "object":
+            object_params.add(param_name)
+            return
+
+
+def _coerce_string_json(arguments: dict, array_params: set[str], object_params: set[str]) -> dict:
+    """Auto-coerce string-encoded JSON arrays/objects to native types.
+
+    LLM agents frequently serialize list and dict tool arguments as JSON strings.
+    This is backward-compatible: native arrays/objects pass through unchanged.
+    """
+    for param_name in array_params:
+        val = arguments.get(param_name)
+        if isinstance(val, str):
+            try:
+                parsed = json.loads(val)
+                if isinstance(parsed, list):
+                    arguments = {**arguments, param_name: parsed}
+                    logger.debug(f"Coerced string to list for parameter '{param_name}'")
+            except (json.JSONDecodeError, TypeError):
+                pass
+
+    for param_name in object_params:
+        val = arguments.get(param_name)
+        if isinstance(val, str):
+            try:
+                parsed = json.loads(val)
+                if isinstance(parsed, dict):
+                    arguments = {**arguments, param_name: parsed}
+                    logger.debug(f"Coerced string to dict for parameter '{param_name}'")
+            except (json.JSONDecodeError, TypeError):
+                pass
+
+    return arguments
+
+
 class MCPMiddleware:
     """ASGI middleware that intercepts MCP requests and routes to appropriate MCP server.
 

hindsight-api-slim/hindsight_api/config.py

@@ -131,10 +131,12 @@ ENV_LLM_MAX_BACKOFF = "HINDSIGHT_API_LLM_MAX_BACKOFF"
 ENV_LLM_TIMEOUT = "HINDSIGHT_API_LLM_TIMEOUT"
 ENV_LLM_GROQ_SERVICE_TIER = "HINDSIGHT_API_LLM_GROQ_SERVICE_TIER"
 ENV_LLM_OPENAI_SERVICE_TIER = "HINDSIGHT_API_LLM_OPENAI_SERVICE_TIER"
+ENV_LLM_EXTRA_BODY = "HINDSIGHT_API_LLM_EXTRA_BODY"
 
 # Defaults for service tiers
 DEFAULT_LLM_GROQ_SERVICE_TIER = "auto"  # "on_demand", "flex", or "auto"
 DEFAULT_LLM_OPENAI_SERVICE_TIER = None  # None (default) or "flex" (50% cheaper)
+DEFAULT_LLM_EXTRA_BODY = None  # None = no extra body params; JSON dict merged into OpenAI extra_body
 
 # Per-operation LLM configuration (optional, falls back to global LLM config)
 ENV_RETAIN_LLM_PROVIDER = "HINDSIGHT_API_RETAIN_LLM_PROVIDER"
@@ -175,6 +177,15 @@ 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_EMBEDDINGS_OPENAI_BATCH_SIZE = "HINDSIGHT_API_EMBEDDINGS_OPENAI_BATCH_SIZE"
+
+# Gemini/Vertex AI embeddings configuration
+ENV_EMBEDDINGS_GEMINI_API_KEY = "HINDSIGHT_API_EMBEDDINGS_GEMINI_API_KEY"
+ENV_EMBEDDINGS_GEMINI_MODEL = "HINDSIGHT_API_EMBEDDINGS_GEMINI_MODEL"
+ENV_EMBEDDINGS_GEMINI_OUTPUT_DIMENSIONALITY = "HINDSIGHT_API_EMBEDDINGS_GEMINI_OUTPUT_DIMENSIONALITY"
+ENV_EMBEDDINGS_VERTEXAI_PROJECT_ID = "HINDSIGHT_API_EMBEDDINGS_VERTEXAI_PROJECT_ID"
+ENV_EMBEDDINGS_VERTEXAI_REGION = "HINDSIGHT_API_EMBEDDINGS_VERTEXAI_REGION"
+ENV_EMBEDDINGS_VERTEXAI_SERVICE_ACCOUNT_KEY = "HINDSIGHT_API_EMBEDDINGS_VERTEXAI_SERVICE_ACCOUNT_KEY"
 
 # Cohere configuration (separate for embeddings and reranker)
 ENV_EMBEDDINGS_COHERE_API_KEY = "HINDSIGHT_API_EMBEDDINGS_COHERE_API_KEY"
@@ -184,6 +195,13 @@ 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"
 
+# OpenRouter configuration (embeddings and reranker)
+ENV_OPENROUTER_API_KEY = "HINDSIGHT_API_OPENROUTER_API_KEY"
+ENV_EMBEDDINGS_OPENROUTER_API_KEY = "HINDSIGHT_API_EMBEDDINGS_OPENROUTER_API_KEY"
+ENV_EMBEDDINGS_OPENROUTER_MODEL = "HINDSIGHT_API_EMBEDDINGS_OPENROUTER_MODEL"
+ENV_RERANKER_OPENROUTER_API_KEY = "HINDSIGHT_API_RERANKER_OPENROUTER_API_KEY"
+ENV_RERANKER_OPENROUTER_MODEL = "HINDSIGHT_API_RERANKER_OPENROUTER_MODEL"
+
 # Deprecated: Legacy shared Cohere API key (for backward compatibility)
 ENV_COHERE_API_KEY = "HINDSIGHT_API_COHERE_API_KEY"
 
@@ -200,6 +218,8 @@ ENV_RERANKER_LITELLM_MAX_TOKENS_PER_DOC = "HINDSIGHT_API_RERANKER_LITELLM_MAX_TO
 ENV_EMBEDDINGS_LITELLM_SDK_API_KEY = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_API_KEY"
 ENV_EMBEDDINGS_LITELLM_SDK_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_MODEL"
 ENV_EMBEDDINGS_LITELLM_SDK_API_BASE = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_API_BASE"
+ENV_EMBEDDINGS_LITELLM_SDK_OUTPUT_DIMENSIONS = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_OUTPUT_DIMENSIONS"
+ENV_EMBEDDINGS_LITELLM_SDK_ENCODING_FORMAT = "HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_ENCODING_FORMAT"
 ENV_RERANKER_LITELLM_SDK_API_KEY = "HINDSIGHT_API_RERANKER_LITELLM_SDK_API_KEY"
 ENV_RERANKER_LITELLM_SDK_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_SDK_MODEL"
 ENV_RERANKER_LITELLM_SDK_API_BASE = "HINDSIGHT_API_RERANKER_LITELLM_SDK_API_BASE"
@@ -219,6 +239,7 @@ ENV_RERANKER_LOCAL_BATCH_SIZE = "HINDSIGHT_API_RERANKER_LOCAL_BATCH_SIZE"
 ENV_RERANKER_TEI_URL = "HINDSIGHT_API_RERANKER_TEI_URL"
 ENV_RERANKER_TEI_BATCH_SIZE = "HINDSIGHT_API_RERANKER_TEI_BATCH_SIZE"
 ENV_RERANKER_TEI_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_TEI_MAX_CONCURRENT"
+ENV_RERANKER_TEI_HTTP_TIMEOUT = "HINDSIGHT_API_RERANKER_TEI_HTTP_TIMEOUT"
 ENV_RERANKER_MAX_CANDIDATES = "HINDSIGHT_API_RERANKER_MAX_CANDIDATES"
 ENV_RERANKER_FLASHRANK_MODEL = "HINDSIGHT_API_RERANKER_FLASHRANK_MODEL"
 ENV_RERANKER_FLASHRANK_CACHE_DIR = "HINDSIGHT_API_RERANKER_FLASHRANK_CACHE_DIR"
@@ -228,6 +249,16 @@ ENV_RERANKER_ZEROENTROPY_API_KEY = "HINDSIGHT_API_RERANKER_ZEROENTROPY_API_KEY"
 ENV_RERANKER_ZEROENTROPY_MODEL = "HINDSIGHT_API_RERANKER_ZEROENTROPY_MODEL"
 ENV_RERANKER_ZEROENTROPY_BASE_URL = "HINDSIGHT_API_RERANKER_ZEROENTROPY_BASE_URL"
 
+# SiliconFlow configuration (reranker only; Cohere-compatible /rerank endpoint)
+ENV_RERANKER_SILICONFLOW_API_KEY = "HINDSIGHT_API_RERANKER_SILICONFLOW_API_KEY"
+ENV_RERANKER_SILICONFLOW_MODEL = "HINDSIGHT_API_RERANKER_SILICONFLOW_MODEL"
+ENV_RERANKER_SILICONFLOW_BASE_URL = "HINDSIGHT_API_RERANKER_SILICONFLOW_BASE_URL"
+
+# Google Discovery Engine reranker configuration
+ENV_RERANKER_GOOGLE_MODEL = "HINDSIGHT_API_RERANKER_GOOGLE_MODEL"
+ENV_RERANKER_GOOGLE_PROJECT_ID = "HINDSIGHT_API_RERANKER_GOOGLE_PROJECT_ID"
+ENV_RERANKER_GOOGLE_SERVICE_ACCOUNT_KEY = "HINDSIGHT_API_RERANKER_GOOGLE_SERVICE_ACCOUNT_KEY"
+
 ENV_VECTOR_EXTENSION = "HINDSIGHT_API_VECTOR_EXTENSION"
 ENV_TEXT_SEARCH_EXTENSION = "HINDSIGHT_API_TEXT_SEARCH_EXTENSION"
 
@@ -236,17 +267,20 @@ ENV_PORT = "HINDSIGHT_API_PORT"
 ENV_BASE_PATH = "HINDSIGHT_API_BASE_PATH"
 ENV_LOG_LEVEL = "HINDSIGHT_API_LOG_LEVEL"
 ENV_LOG_FORMAT = "HINDSIGHT_API_LOG_FORMAT"
+ENV_LOG_JSON_FIELDS = "HINDSIGHT_API_LOG_JSON_FIELDS"
 ENV_WORKERS = "HINDSIGHT_API_WORKERS"
 ENV_MCP_ENABLED = "HINDSIGHT_API_MCP_ENABLED"
 ENV_MCP_ENABLED_TOOLS = "HINDSIGHT_API_MCP_ENABLED_TOOLS"
 ENV_MCP_STATELESS = "HINDSIGHT_API_MCP_STATELESS"
 ENV_ENABLE_BANK_CONFIG_API = "HINDSIGHT_API_ENABLE_BANK_CONFIG_API"
+ENV_DEFAULT_BANK_TEMPLATE = "HINDSIGHT_API_DEFAULT_BANK_TEMPLATE"
 ENV_GRAPH_RETRIEVER = "HINDSIGHT_API_GRAPH_RETRIEVER"
-ENV_MPFP_TOP_K_NEIGHBORS = "HINDSIGHT_API_MPFP_TOP_K_NEIGHBORS"
 ENV_RECALL_MAX_CONCURRENT = "HINDSIGHT_API_RECALL_MAX_CONCURRENT"
 ENV_RECALL_CONNECTION_BUDGET = "HINDSIGHT_API_RECALL_CONNECTION_BUDGET"
 ENV_RECALL_MAX_QUERY_TOKENS = "HINDSIGHT_API_RECALL_MAX_QUERY_TOKENS"
 ENV_MENTAL_MODEL_REFRESH_CONCURRENCY = "HINDSIGHT_API_MENTAL_MODEL_REFRESH_CONCURRENCY"
+ENV_LINK_EXPANSION_PER_ENTITY_LIMIT = "HINDSIGHT_API_LINK_EXPANSION_PER_ENTITY_LIMIT"
+ENV_LINK_EXPANSION_TIMEOUT = "HINDSIGHT_API_LINK_EXPANSION_TIMEOUT"
 
 # OpenTelemetry tracing configuration
 ENV_OTEL_TRACES_ENABLED = "HINDSIGHT_API_OTEL_TRACES_ENABLED"
@@ -254,6 +288,7 @@ 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"
+ENV_METRICS_INCLUDE_BANK_ID = "HINDSIGHT_API_METRICS_INCLUDE_BANK_ID"
 
 # Vertex AI configuration
 ENV_LLM_VERTEXAI_PROJECT_ID = "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID"
@@ -275,6 +310,7 @@ ENV_RETAIN_BATCH_TOKENS = "HINDSIGHT_API_RETAIN_BATCH_TOKENS"
 ENV_RETAIN_ENTITY_LOOKUP = "HINDSIGHT_API_RETAIN_ENTITY_LOOKUP"
 ENV_RETAIN_BATCH_ENABLED = "HINDSIGHT_API_RETAIN_BATCH_ENABLED"
 ENV_RETAIN_BATCH_POLL_INTERVAL_SECONDS = "HINDSIGHT_API_RETAIN_BATCH_POLL_INTERVAL_SECONDS"
+ENV_RETAIN_CHUNK_BATCH_SIZE = "HINDSIGHT_API_RETAIN_CHUNK_BATCH_SIZE"
 
 # File storage configuration
 ENV_FILE_STORAGE_TYPE = "HINDSIGHT_API_FILE_STORAGE_TYPE"
@@ -300,12 +336,14 @@ ENV_FILE_DELETE_AFTER_RETAIN = "HINDSIGHT_API_FILE_DELETE_AFTER_RETAIN"
 # Observations settings (consolidated knowledge from facts)
 ENV_ENABLE_OBSERVATIONS = "HINDSIGHT_API_ENABLE_OBSERVATIONS"
 ENV_CONSOLIDATION_BATCH_SIZE = "HINDSIGHT_API_CONSOLIDATION_BATCH_SIZE"
+ENV_CONSOLIDATION_MAX_MEMORIES_PER_ROUND = "HINDSIGHT_API_CONSOLIDATION_MAX_MEMORIES_PER_ROUND"
 ENV_CONSOLIDATION_LLM_BATCH_SIZE = "HINDSIGHT_API_CONSOLIDATION_LLM_BATCH_SIZE"
 ENV_CONSOLIDATION_MAX_TOKENS = "HINDSIGHT_API_CONSOLIDATION_MAX_TOKENS"
 ENV_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS = "HINDSIGHT_API_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS"
 ENV_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION = (
     "HINDSIGHT_API_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION"
 )
+ENV_CONSOLIDATION_MAX_ATTEMPTS = "HINDSIGHT_API_CONSOLIDATION_MAX_ATTEMPTS"
 ENV_OBSERVATIONS_MISSION = "HINDSIGHT_API_OBSERVATIONS_MISSION"
 ENV_MAX_OBSERVATIONS_PER_SCOPE = "HINDSIGHT_API_MAX_OBSERVATIONS_PER_SCOPE"
 ENV_ENABLE_OBSERVATION_HISTORY = "HINDSIGHT_API_ENABLE_OBSERVATION_HISTORY"
@@ -317,6 +355,14 @@ ENV_WEBHOOK_SECRET = "HINDSIGHT_API_WEBHOOK_SECRET"
 ENV_WEBHOOK_EVENT_TYPES = "HINDSIGHT_API_WEBHOOK_EVENT_TYPES"
 ENV_WEBHOOK_DELIVERY_POLL_INTERVAL_SECONDS = "HINDSIGHT_API_WEBHOOK_DELIVERY_POLL_INTERVAL_SECONDS"
 
+# Built-in llama.cpp configuration (for provider=llamacpp)
+ENV_LLAMACPP_MODEL_PATH = "HINDSIGHT_API_LLAMACPP_MODEL_PATH"
+ENV_LLAMACPP_GPU_LAYERS = "HINDSIGHT_API_LLAMACPP_GPU_LAYERS"
+ENV_LLAMACPP_CONTEXT_SIZE = "HINDSIGHT_API_LLAMACPP_CONTEXT_SIZE"
+ENV_LLAMACPP_CHAT_FORMAT = "HINDSIGHT_API_LLAMACPP_CHAT_FORMAT"
+ENV_LLAMACPP_NO_GRAMMAR = "HINDSIGHT_API_LLAMACPP_NO_GRAMMAR"
+ENV_LLAMACPP_EXTRA_ARGS = "HINDSIGHT_API_LLAMACPP_EXTRA_ARGS"
+
 # Optimization flags
 ENV_SKIP_LLM_VERIFICATION = "HINDSIGHT_API_SKIP_LLM_VERIFICATION"
 ENV_LAZY_RERANKER = "HINDSIGHT_API_LAZY_RERANKER"
@@ -337,7 +383,19 @@ ENV_WORKER_POLL_INTERVAL_MS = "HINDSIGHT_API_WORKER_POLL_INTERVAL_MS"
 ENV_WORKER_MAX_RETRIES = "HINDSIGHT_API_WORKER_MAX_RETRIES"
 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"
+
+# Per-operation-type slot reservations. Each entry maps an operation_type
+# (as stored in async_operations.operation_type) to its env var and default.
+# Adding a new operation type here is the ONLY change needed to make it
+# reservable via env var — config fields, from_env(), and the
+# worker_slot_reservations property all derive from this dict.
+WORKER_SLOT_RESERVATION_TYPES: dict[str, tuple[str, int]] = {
+    "consolidation": ("HINDSIGHT_API_WORKER_CONSOLIDATION_MAX_SLOTS", 2),
+    "retain": ("HINDSIGHT_API_WORKER_RETAIN_MAX_SLOTS", 0),
+    "file_convert_retain": ("HINDSIGHT_API_WORKER_FILE_CONVERT_RETAIN_MAX_SLOTS", 0),
+    "refresh_mental_model": ("HINDSIGHT_API_WORKER_REFRESH_MENTAL_MODEL_MAX_SLOTS", 0),
+}
+ENV_RETAIN_MAX_CONCURRENT = "HINDSIGHT_API_RETAIN_MAX_CONCURRENT"
 
 # Reflect agent settings
 ENV_REFLECT_MAX_ITERATIONS = "HINDSIGHT_API_REFLECT_MAX_ITERATIONS"
@@ -345,6 +403,20 @@ ENV_REFLECT_MAX_CONTEXT_TOKENS = "HINDSIGHT_API_REFLECT_MAX_CONTEXT_TOKENS"
 ENV_REFLECT_WALL_TIMEOUT = "HINDSIGHT_API_REFLECT_WALL_TIMEOUT"
 ENV_REFLECT_MISSION = "HINDSIGHT_API_REFLECT_MISSION"
 ENV_REFLECT_SOURCE_FACTS_MAX_TOKENS = "HINDSIGHT_API_REFLECT_SOURCE_FACTS_MAX_TOKENS"
+ENV_RECALL_INCLUDE_CHUNKS = "HINDSIGHT_API_RECALL_INCLUDE_CHUNKS"
+ENV_RECALL_MAX_TOKENS = "HINDSIGHT_API_RECALL_MAX_TOKENS"
+ENV_RECALL_CHUNKS_MAX_TOKENS = "HINDSIGHT_API_RECALL_CHUNKS_MAX_TOKENS"
+
+# Recall budget mapping (budget enum -> thinking_budget integer)
+ENV_RECALL_BUDGET_FUNCTION = "HINDSIGHT_API_RECALL_BUDGET_FUNCTION"
+ENV_RECALL_BUDGET_FIXED_LOW = "HINDSIGHT_API_RECALL_BUDGET_FIXED_LOW"
+ENV_RECALL_BUDGET_FIXED_MID = "HINDSIGHT_API_RECALL_BUDGET_FIXED_MID"
+ENV_RECALL_BUDGET_FIXED_HIGH = "HINDSIGHT_API_RECALL_BUDGET_FIXED_HIGH"
+ENV_RECALL_BUDGET_ADAPTIVE_LOW = "HINDSIGHT_API_RECALL_BUDGET_ADAPTIVE_LOW"
+ENV_RECALL_BUDGET_ADAPTIVE_MID = "HINDSIGHT_API_RECALL_BUDGET_ADAPTIVE_MID"
+ENV_RECALL_BUDGET_ADAPTIVE_HIGH = "HINDSIGHT_API_RECALL_BUDGET_ADAPTIVE_HIGH"
+ENV_RECALL_BUDGET_MIN = "HINDSIGHT_API_RECALL_BUDGET_MIN"
+ENV_RECALL_BUDGET_MAX = "HINDSIGHT_API_RECALL_BUDGET_MAX"
 
 # Audit log settings
 ENV_AUDIT_LOG_ENABLED = "HINDSIGHT_API_AUDIT_LOG_ENABLED"
@@ -369,6 +441,7 @@ PROVIDER_DEFAULT_MODELS = {
     "groq": "openai/gpt-oss-120b",
     "minimax": "MiniMax-M2.7",
     "ollama": "gemma3:12b",
+    "llamacpp": "gemma-4-e2b-it",
     "lmstudio": "local-model",
     "vertexai": "google/gemini-2.5-flash-lite",
     "openai-codex": "gpt-5.2-codex",
@@ -378,10 +451,18 @@ PROVIDER_DEFAULT_MODELS = {
     "litellm": "gpt-4o-mini",
     "bedrock": "us.amazon.nova-2-lite-v1:0",
     "volcano": "doubao-pro-32k",
+    "openrouter": "qwen/qwen3.5-9b",
 }
 DEFAULT_LLM_MODEL = "gpt-4o-mini"  # Fallback if provider not in table
+# Built-in llama.cpp defaults
+DEFAULT_LLAMACPP_GPU_LAYERS = -1  # -1 = offload all layers to GPU (Metal/CUDA)
+DEFAULT_LLAMACPP_CONTEXT_SIZE = 8192
+DEFAULT_LLAMACPP_CHAT_FORMAT = None  # None = auto-detect from GGUF metadata
+DEFAULT_LLAMACPP_NO_GRAMMAR = False  # True = disable JSON grammar enforcement (faster but less reliable)
+DEFAULT_LLAMACPP_EXTRA_ARGS = None  # Space-separated extra CLI args for llama.cpp server
+
 DEFAULT_LLM_MAX_CONCURRENT = 32
-DEFAULT_LLM_MAX_RETRIES = 10  # Max retry attempts for LLM API calls
+DEFAULT_LLM_MAX_RETRIES = 3  # 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
@@ -399,6 +480,9 @@ 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_EMBEDDINGS_OPENAI_BATCH_SIZE = 100
+DEFAULT_EMBEDDINGS_GEMINI_MODEL = "gemini-embedding-001"
+DEFAULT_EMBEDDINGS_GEMINI_OUTPUT_DIMENSIONALITY = 768
 DEFAULT_EMBEDDING_DIMENSION = 384
 
 DEFAULT_RERANKER_PROVIDER = "local"
@@ -413,6 +497,7 @@ DEFAULT_RERANKER_LOCAL_BUCKET_BATCHING = False  # Length-sorted bucket batching:
 DEFAULT_RERANKER_LOCAL_BATCH_SIZE = 32  # Batch size for local reranker predict() calls
 DEFAULT_RERANKER_TEI_BATCH_SIZE = 128
 DEFAULT_RERANKER_TEI_MAX_CONCURRENT = 8
+DEFAULT_RERANKER_TEI_HTTP_TIMEOUT = 30.0  # HTTP timeout for TEI reranker requests (seconds)
 DEFAULT_RERANKER_MAX_CANDIDATES = 300
 DEFAULT_RERANKER_FLASHRANK_MODEL = "ms-marco-MiniLM-L-12-v2"  # Best balance of speed and quality
 DEFAULT_RERANKER_FLASHRANK_CACHE_DIR = None  # Use default cache directory
@@ -420,8 +505,17 @@ DEFAULT_RERANKER_FLASHRANK_CACHE_DIR = None  # Use default cache directory
 DEFAULT_EMBEDDINGS_COHERE_MODEL = "embed-english-v3.0"
 DEFAULT_RERANKER_COHERE_MODEL = "rerank-english-v3.0"
 
+# OpenRouter defaults
+DEFAULT_EMBEDDINGS_OPENROUTER_MODEL = "perplexity/pplx-embed-v1-0.6b"
+DEFAULT_RERANKER_OPENROUTER_MODEL = "cohere/rerank-v3.5"
+
 DEFAULT_RERANKER_ZEROENTROPY_MODEL = "zerank-2"
 
+DEFAULT_RERANKER_SILICONFLOW_MODEL = "BAAI/bge-reranker-v2-m3"
+DEFAULT_RERANKER_SILICONFLOW_BASE_URL = "https://api.siliconflow.cn/v1"
+
+DEFAULT_RERANKER_GOOGLE_MODEL = "semantic-ranker-default-004"
+
 # Vector extension (pgvector, vchord, or pgvectorscale)
 DEFAULT_VECTOR_EXTENSION = "pgvector"  # Options: "pgvector", "vchord", "pgvectorscale"
 
@@ -436,6 +530,7 @@ DEFAULT_RERANKER_LITELLM_MAX_TOKENS_PER_DOC: int | None = None
 
 # LiteLLM SDK defaults
 DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL = "cohere/embed-english-v3.0"
+DEFAULT_EMBEDDINGS_LITELLM_SDK_ENCODING_FORMAT = "float"
 DEFAULT_RERANKER_LITELLM_SDK_MODEL = "cohere/rerank-english-v3.0"
 
 DEFAULT_HOST = "0.0.0.0"
@@ -448,12 +543,14 @@ DEFAULT_MCP_ENABLED = True
 DEFAULT_MCP_ENABLED_TOOLS: list[str] | None = None  # None = all tools enabled
 DEFAULT_MCP_STATELESS = False  # False = stateful (supports SSE/GET); True = stateless (POST-only)
 DEFAULT_ENABLE_BANK_CONFIG_API = True
-DEFAULT_GRAPH_RETRIEVER = "link_expansion"  # Options: "link_expansion", "mpfp", "bfs"
-DEFAULT_MPFP_TOP_K_NEIGHBORS = 20  # Fan-out limit per node in MPFP graph traversal
+DEFAULT_DEFAULT_BANK_TEMPLATE: dict | None = None  # BankTemplateManifest dict applied to newly-created banks
+DEFAULT_GRAPH_RETRIEVER = "link_expansion"
 DEFAULT_RECALL_MAX_CONCURRENT = 32  # Max concurrent recall operations per worker
 DEFAULT_RECALL_CONNECTION_BUDGET = 4  # Max concurrent DB connections per recall operation
 DEFAULT_RECALL_MAX_QUERY_TOKENS = 500  # Maximum tokens allowed in recall query
 DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY = 8  # Max concurrent mental model refreshes
+DEFAULT_LINK_EXPANSION_PER_ENTITY_LIMIT = 200  # Max target units per entity in graph expansion
+DEFAULT_LINK_EXPANSION_TIMEOUT = 10.0  # Timeout (seconds) for entity expansion query
 
 # Retain settings
 DEFAULT_RETAIN_MAX_COMPLETION_TOKENS = 64000  # Max tokens for fact extraction LLM call
@@ -465,6 +562,9 @@ DEFAULT_RETAIN_MISSION = None  # Declarative spec of what to retain (injected in
 DEFAULT_RETAIN_CUSTOM_INSTRUCTIONS = None  # Custom extraction guidelines (only used when mode="custom")
 DEFAULT_RETAIN_DEFAULT_STRATEGY = None  # Default strategy name (None = no strategy override)
 DEFAULT_RETAIN_STRATEGIES: dict | None = None  # Named retain strategies (dict of name → config overrides)
+DEFAULT_RETAIN_CHUNK_BATCH_SIZE = (
+    100  # Max chunks per streaming batch. Each chunk produces ~17 facts, so 100 chunks = ~1700 facts/batch.
+)
 DEFAULT_RETAIN_BATCH_TOKENS = 10_000  # ~40KB of text  # Max chars per sub-batch for async retain auto-splitting
 DEFAULT_RETAIN_ENTITY_LOOKUP = "trigram"  # "full" or "trigram"
 DEFAULT_RETAIN_BATCH_ENABLED = False  # Use LLM Batch API for fact extraction (only when async=True)
@@ -483,7 +583,11 @@ DEFAULT_FILE_DELETE_AFTER_RETAIN = True  # Delete file bytes after retain (saves
 DEFAULT_ENABLE_OBSERVATIONS = True  # Observations enabled by default
 DEFAULT_ENABLE_OBSERVATION_HISTORY = True  # Observation history tracking enabled by default
 DEFAULT_ENABLE_MENTAL_MODEL_HISTORY = True  # Mental model history tracking enabled by default
+DEFAULT_CONSOLIDATION_MAX_ATTEMPTS = 3  # Outer retry attempts for consolidation LLM batch calls
 DEFAULT_CONSOLIDATION_BATCH_SIZE = 50  # Memories to load per batch (internal memory optimization)
+DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND = (
+    100  # Max memories per consolidation round (0 = unlimited). Limits how long one bank holds a worker slot.
+)
 DEFAULT_CONSOLIDATION_LLM_BATCH_SIZE = 8  # Facts per LLM call (1 = no batching; >1 = batch mode)
 DEFAULT_CONSOLIDATION_MAX_TOKENS = 512  # Max tokens for recall when finding related observations
 DEFAULT_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS = (
@@ -511,13 +615,32 @@ DEFAULT_WORKER_POLL_INTERVAL_MS = 500  # Poll database every 500ms
 DEFAULT_WORKER_MAX_RETRIES = 3  # Max retries before marking task failed
 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
+DEFAULT_RETAIN_MAX_CONCURRENT = 4  # Max concurrent retain DB phases (HNSW reads + writes). Limits I/O contention.
 
 # Reflect agent settings
 DEFAULT_REFLECT_MAX_ITERATIONS = 10  # Max tool call iterations before forcing response
 DEFAULT_REFLECT_MAX_CONTEXT_TOKENS = 100_000  # Max accumulated context tokens before forcing final prompt
 DEFAULT_REFLECT_WALL_TIMEOUT = 300  # Wall-clock timeout in seconds for the entire reflect operation (5 minutes)
 DEFAULT_REFLECT_SOURCE_FACTS_MAX_TOKENS = -1  # Token budget for source facts in search_observations (-1 = disabled)
+DEFAULT_RECALL_INCLUDE_CHUNKS = True  # Whether internal recall (e.g. mental model refresh) returns raw chunks
+DEFAULT_RECALL_MAX_TOKENS = 2048  # Token budget for facts returned by internal recall
+DEFAULT_RECALL_CHUNKS_MAX_TOKENS = 1000  # Token budget for raw chunks returned by internal recall
+
+# Recall budget mapping
+# "fixed": thinking_budget = recall_budget_fixed_<level> (preserves legacy behavior)
+# "adaptive": thinking_budget = round(max_tokens * recall_budget_adaptive_<level>),
+#             clamped to [recall_budget_min, recall_budget_max]
+RECALL_BUDGET_FUNCTIONS = ("fixed", "adaptive")
+DEFAULT_RECALL_BUDGET_FUNCTION = "fixed"
+DEFAULT_RECALL_BUDGET_FIXED_LOW = 100
+DEFAULT_RECALL_BUDGET_FIXED_MID = 300
+DEFAULT_RECALL_BUDGET_FIXED_HIGH = 1000
+# Adaptive defaults chosen to roughly match fixed defaults at max_tokens=4096
+DEFAULT_RECALL_BUDGET_ADAPTIVE_LOW = 0.025
+DEFAULT_RECALL_BUDGET_ADAPTIVE_MID = 0.075
+DEFAULT_RECALL_BUDGET_ADAPTIVE_HIGH = 0.25
+DEFAULT_RECALL_BUDGET_MIN = 20  # Floor for the adaptive function
+DEFAULT_RECALL_BUDGET_MAX = 2000  # Ceiling for the adaptive function
 
 # Disposition defaults (None = not set, fall back to bank DB value or 3)
 DEFAULT_DISPOSITION_SKEPTICISM = None
@@ -528,6 +651,7 @@ DEFAULT_DISPOSITION_EMPATHY = None
 DEFAULT_OTEL_TRACES_ENABLED = False  # Disabled by default for backward compatibility
 DEFAULT_OTEL_SERVICE_NAME = "hindsight-api"
 DEFAULT_OTEL_DEPLOYMENT_ENVIRONMENT = "development"
+DEFAULT_METRICS_INCLUDE_BANK_ID = False  # Disabled by default to avoid high-cardinality OTel metric growth
 
 # Audit log defaults
 DEFAULT_AUDIT_LOG_ENABLED = False  # Disabled by default
@@ -579,6 +703,10 @@ class JsonFormatter(logging.Formatter):
         logging.CRITICAL: "CRITICAL",
     }
 
+    def __init__(self, allowed_fields: frozenset[str] | None = None):
+        super().__init__()
+        self._allowed_fields = allowed_fields
+
     def format(self, record: logging.LogRecord) -> str:
         log_entry = {
             "severity": self.SEVERITY_MAP.get(record.levelno, "DEFAULT"),
@@ -587,10 +715,20 @@ class JsonFormatter(logging.Formatter):
             "logger": record.name,
         }
 
+        # Lazy import to avoid circular dependency (engine imports from config).
+        from hindsight_api.engine.memory_engine import _current_schema
+
+        tenant = _current_schema.get()
+        if tenant:
+            log_entry["tenant"] = tenant
+
         # Add exception info if present
         if record.exc_info:
             log_entry["exception"] = self.formatException(record.exc_info)
 
+        if self._allowed_fields is not None:
+            log_entry = {k: v for k, v in log_entry.items() if k in self._allowed_fields}
+
         return json.dumps(log_entry)
 
 
@@ -599,6 +737,25 @@ def _parse_str_list(value: str) -> list[str]:
     return [v.strip() for v in value.split(",") if v.strip()]
 
 
+def _parse_positive_int(name: str, raw: str | None, default: int) -> int:
+    """
+    Parse an env var that must be a positive integer (>= 1).
+
+    Falls back to ``default`` when unset/empty. Raises ValueError on non-integer
+    or non-positive values so misconfiguration fails fast instead of triggering
+    infinite loops or zero-step range() calls downstream.
+    """
+    if raw is None or raw == "":
+        return default
+    try:
+        parsed = int(raw)
+    except ValueError as e:
+        raise ValueError(f"{name} must be an integer, got {raw!r}") from e
+    if parsed < 1:
+        raise ValueError(f"{name} must be >= 1, got {parsed}")
+    return parsed
+
+
 def _validate_extraction_mode(mode: str) -> str:
     """Validate and normalize extraction mode."""
     mode_lower = mode.lower()
@@ -611,11 +768,43 @@ def _validate_extraction_mode(mode: str) -> str:
     return mode_lower
 
 
+def _validate_recall_budget_function(function: str) -> str:
+    """Validate and normalize recall budget function."""
+    function_lower = function.lower()
+    if function_lower not in RECALL_BUDGET_FUNCTIONS:
+        logger.warning(
+            f"Invalid recall budget function '{function}', must be one of {RECALL_BUDGET_FUNCTIONS}. "
+            f"Defaulting to '{DEFAULT_RECALL_BUDGET_FUNCTION}'."
+        )
+        return DEFAULT_RECALL_BUDGET_FUNCTION
+    return function_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)
 
 
+def _parse_default_bank_template(raw: str | None) -> dict | None:
+    """
+    Parse HINDSIGHT_API_DEFAULT_BANK_TEMPLATE as JSON.
+
+    The env var holds a BankTemplateManifest (JSON object) applied verbatim to
+    every newly-created bank. Full Pydantic validation is deferred to bank
+    creation time (to avoid pulling API models into config.py), but we fail
+    fast here if the value is not valid JSON or not a JSON object.
+    """
+    if raw is None or raw.strip() == "":
+        return DEFAULT_DEFAULT_BANK_TEMPLATE
+    try:
+        parsed = json.loads(raw)
+    except json.JSONDecodeError as e:
+        raise ValueError(f"Invalid {ENV_DEFAULT_BANK_TEMPLATE}: expected a JSON object, got invalid JSON: {e}") from e
+    if not isinstance(parsed, dict):
+        raise ValueError(f"Invalid {ENV_DEFAULT_BANK_TEMPLATE}: expected a JSON object, got {type(parsed).__name__}")
+    return parsed
+
+
 @dataclass
 class HindsightConfig:
     """Configuration container for Hindsight API."""
@@ -639,6 +828,9 @@ class HindsightConfig:
     llm_timeout: float
     llm_groq_service_tier: str  # Groq: "on_demand", "flex", or "auto"
     llm_openai_service_tier: str | None  # OpenAI: None (default) or "flex" (50% cheaper)
+    llm_extra_body: (
+        dict | None
+    )  # Extra body params merged into OpenAI-compatible API calls (e.g. {"chat_template_kwargs": {"enable_thinking": true}})
 
     # Vertex AI configuration
     llm_vertexai_project_id: str | None
@@ -648,6 +840,14 @@ class HindsightConfig:
     # Gemini safety settings (None = use Gemini defaults; list of dicts with category/threshold)
     llm_gemini_safety_settings: list | None
 
+    # Built-in llama.cpp configuration (for provider=llamacpp)
+    llamacpp_model_path: str | None  # Path to GGUF file (None = auto-download default)
+    llamacpp_gpu_layers: int  # -1 = all layers on GPU, 0 = CPU only
+    llamacpp_context_size: int  # Context window size
+    llamacpp_chat_format: str | None  # Chat template format (None = auto-detect from GGUF)
+    llamacpp_no_grammar: bool  # Disable JSON grammar enforcement (faster, less reliable)
+    llamacpp_extra_args: str | None  # Space-separated extra CLI args for llama.cpp server
+
     # Per-operation LLM configuration (None = use default LLM config)
     retain_llm_provider: str | None
     retain_llm_api_key: str | None
@@ -689,12 +889,23 @@ class HindsightConfig:
     embeddings_cohere_api_key: str | None
     embeddings_cohere_model: str
     embeddings_cohere_base_url: str | None
+    embeddings_openrouter_api_key: str | None
+    embeddings_openrouter_model: str
     embeddings_litellm_api_base: str
     embeddings_litellm_api_key: str | None
     embeddings_litellm_model: str
     embeddings_litellm_sdk_api_key: str | None
     embeddings_litellm_sdk_model: str
     embeddings_litellm_sdk_api_base: str | None
+    embeddings_litellm_sdk_output_dimensions: int | None
+    embeddings_litellm_sdk_encoding_format: str | None
+    # Gemini/Vertex AI embeddings
+    embeddings_gemini_api_key: str | None
+    embeddings_gemini_model: str
+    embeddings_gemini_output_dimensionality: int | None
+    embeddings_vertexai_project_id: str | None
+    embeddings_vertexai_region: str | None
+    embeddings_vertexai_service_account_key: str | None
 
     # Reranker
     reranker_provider: str
@@ -708,10 +919,13 @@ class HindsightConfig:
     reranker_tei_url: str | None
     reranker_tei_batch_size: int
     reranker_tei_max_concurrent: int
+    reranker_tei_http_timeout: float
     reranker_max_candidates: int
     reranker_cohere_api_key: str | None
     reranker_cohere_model: str
     reranker_cohere_base_url: str | None
+    reranker_openrouter_api_key: str | None
+    reranker_openrouter_model: str
     reranker_litellm_api_base: str
     reranker_litellm_api_key: str | None
     reranker_litellm_model: str
@@ -722,6 +936,12 @@ class HindsightConfig:
     reranker_zeroentropy_api_key: str | None
     reranker_zeroentropy_model: str
     reranker_zeroentropy_base_url: str | None
+    reranker_siliconflow_api_key: str | None
+    reranker_siliconflow_model: str
+    reranker_siliconflow_base_url: str
+    reranker_google_model: str
+    reranker_google_project_id: str | None
+    reranker_google_service_account_key: str | None
 
     # Server
     host: str
@@ -729,18 +949,23 @@ class HindsightConfig:
     base_path: str
     log_level: str
     log_format: str
+    log_json_fields: list[str] | None  # None = all fields; explicit list = allowlist
     mcp_enabled: bool
     mcp_enabled_tools: list[str] | None  # None = all tools; explicit list = allowlist
     mcp_stateless: bool  # True = stateless HTTP (POST-only); False = stateful (supports GET/SSE)
     enable_bank_config_api: bool
+    # Default bank template (static, server-level only). When set, the manifest is applied
+    # to every newly-created bank, overriding the env/config defaults for any fields it sets.
+    default_bank_template: dict | None
 
     # Recall
     graph_retriever: str
-    mpfp_top_k_neighbors: int
     recall_max_concurrent: int
     recall_connection_budget: int
     recall_max_query_tokens: int
     mental_model_refresh_concurrency: int
+    link_expansion_per_entity_limit: int
+    link_expansion_timeout: float
 
     # Retain settings
     retain_max_completion_tokens: int
@@ -755,6 +980,7 @@ class HindsightConfig:
     retain_batch_enabled: bool
     retain_batch_poll_interval_seconds: int
     retain_entity_lookup: str  # "full" or "trigram"
+    retain_chunk_batch_size: int  # Max chunks per streaming batch (0 = disabled)
 
     # File storage (static - server-level only)
     file_storage_type: str  # "native" (PostgreSQL) or "s3" (S3-compatible)
@@ -782,10 +1008,12 @@ class HindsightConfig:
     enable_observation_history: bool
     enable_mental_model_history: bool
     consolidation_batch_size: int
+    consolidation_max_memories_per_round: int
     consolidation_llm_batch_size: int
     consolidation_max_tokens: int
     consolidation_source_facts_max_tokens: int
     consolidation_source_facts_max_tokens_per_observation: int
+    consolidation_max_attempts: int
     observations_mission: str | None
     max_observations_per_scope: int
 
@@ -800,6 +1028,25 @@ class HindsightConfig:
     reflect_mission: str | None
     reflect_source_facts_max_tokens: int
 
+    # Recall settings (used by internal recall, e.g. during mental model refresh)
+    recall_include_chunks: bool
+    recall_max_tokens: int
+    recall_chunks_max_tokens: int
+
+    # Recall budget mapping: how the Budget enum (LOW/MID/HIGH) maps to thinking_budget integer.
+    # function="fixed": use the recall_budget_fixed_* values directly (legacy behavior).
+    # function="adaptive": compute round(max_tokens * recall_budget_adaptive_*),
+    #                      clamped to [recall_budget_min, recall_budget_max].
+    recall_budget_function: str
+    recall_budget_fixed_low: int
+    recall_budget_fixed_mid: int
+    recall_budget_fixed_high: int
+    recall_budget_adaptive_low: float
+    recall_budget_adaptive_mid: float
+    recall_budget_adaptive_high: float
+    recall_budget_min: int
+    recall_budget_max: int
+
     # Disposition settings (hierarchical - can be overridden per bank; None = fall back to DB)
     disposition_skepticism: int | None
     disposition_literalism: int | None
@@ -825,7 +1072,8 @@ class HindsightConfig:
     worker_max_retries: int
     worker_http_port: int
     worker_max_slots: int
-    worker_consolidation_max_slots: int
+    worker_slot_reservations: dict[str, int]
+    retain_max_concurrent: int
 
     # Reflect agent settings
     reflect_max_iterations: int
@@ -838,6 +1086,7 @@ class HindsightConfig:
     otel_exporter_otlp_headers: str | None
     otel_service_name: str
     otel_deployment_environment: str
+    metrics_include_bank_id: bool
 
     # Audit log configuration (static - server-level only)
     audit_log_enabled: bool  # Master switch for audit logging
@@ -850,6 +1099,10 @@ class HindsightConfig:
     webhook_event_types: list[str]  # Event types to deliver globally
     webhook_delivery_poll_interval_seconds: int  # How often the delivery worker polls
 
+    # Defaulted fields (source-compatible additions — existing direct constructor callers keep working).
+    # Keep at the end of the dataclass; Python forbids non-default fields after default fields.
+    embeddings_openai_batch_size: int = DEFAULT_EMBEDDINGS_OPENAI_BATCH_SIZE
+
     # Class-level sets for configuration categorization
 
     # CREDENTIAL_FIELDS: Never exposed via API, never configurable per-tenant/bank
@@ -868,8 +1121,13 @@ class HindsightConfig:
         "reranker_tei_base_url",
         "reranker_cohere_base_url",
         "reranker_zeroentropy_base_url",
+        "reranker_siliconflow_base_url",
         # Service Account Keys
         "llm_vertexai_service_account_key",
+        "embeddings_vertexai_service_account_key",
+        "reranker_google_service_account_key",
+        # Embeddings API keys
+        "embeddings_gemini_api_key",
         # File storage credentials
         "file_storage_s3_access_key_id",
         "file_storage_s3_secret_access_key",
@@ -892,12 +1150,14 @@ class HindsightConfig:
         "retain_custom_instructions",
         "retain_default_strategy",
         "retain_strategies",
+        "retain_chunk_batch_size",
         # Entity labels (controlled vocabulary for entity classification)
         "entity_labels",
         "entities_allow_free_form",
         # Consolidation settings
         "enable_observations",
         "consolidation_llm_batch_size",
+        "consolidation_max_memories_per_round",
         "consolidation_source_facts_max_tokens",
         "consolidation_source_facts_max_tokens_per_observation",
         "observations_mission",
@@ -905,6 +1165,20 @@ class HindsightConfig:
         # Reflect settings
         "reflect_mission",
         "reflect_source_facts_max_tokens",
+        # Recall settings (used by internal recall, e.g. mental model refresh)
+        "recall_include_chunks",
+        "recall_max_tokens",
+        "recall_chunks_max_tokens",
+        # Recall budget mapping (Budget enum -> thinking_budget integer)
+        "recall_budget_function",
+        "recall_budget_fixed_low",
+        "recall_budget_fixed_mid",
+        "recall_budget_fixed_high",
+        "recall_budget_adaptive_low",
+        "recall_budget_adaptive_mid",
+        "recall_budget_adaptive_high",
+        "recall_budget_min",
+        "recall_budget_max",
         # Disposition settings
         "disposition_skepticism",
         "disposition_literalism",
@@ -1011,6 +1285,16 @@ class HindsightConfig:
                 f"provider: {self.retain_llm_provider or self.llm_provider})"
             )
 
+        # Validate that sum of per-operation slot reservations does not exceed max_slots
+        total_reserved = sum(self.worker_slot_reservations.values())
+        if total_reserved > self.worker_max_slots:
+            reservation_details = ", ".join(f"{k}={v}" for k, v in self.worker_slot_reservations.items() if v > 0)
+            raise ValueError(
+                f"Sum of per-operation slot reservations ({total_reserved}: {reservation_details}) "
+                f"exceeds worker_max_slots ({self.worker_max_slots}). "
+                f"Reduce reservations or increase HINDSIGHT_API_WORKER_MAX_SLOTS."
+            )
+
     @classmethod
     def from_env(cls) -> "HindsightConfig":
         """Create configuration from environment variables."""
@@ -1037,6 +1321,7 @@ class HindsightConfig:
             llm_timeout=float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT))),
             llm_groq_service_tier=os.getenv(ENV_LLM_GROQ_SERVICE_TIER, DEFAULT_LLM_GROQ_SERVICE_TIER),
             llm_openai_service_tier=os.getenv(ENV_LLM_OPENAI_SERVICE_TIER, DEFAULT_LLM_OPENAI_SERVICE_TIER),
+            llm_extra_body=json.loads(os.getenv(ENV_LLM_EXTRA_BODY, "null")),
             # 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),
@@ -1044,6 +1329,14 @@ class HindsightConfig:
             or DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY,
             # Gemini safety settings (JSON-encoded list of {category, threshold} dicts)
             llm_gemini_safety_settings=json.loads(os.getenv(ENV_LLM_GEMINI_SAFETY_SETTINGS, "null")),
+            # Built-in llama.cpp configuration
+            llamacpp_model_path=os.getenv(ENV_LLAMACPP_MODEL_PATH) or None,
+            llamacpp_gpu_layers=int(os.getenv(ENV_LLAMACPP_GPU_LAYERS, str(DEFAULT_LLAMACPP_GPU_LAYERS))),
+            llamacpp_context_size=int(os.getenv(ENV_LLAMACPP_CONTEXT_SIZE, str(DEFAULT_LLAMACPP_CONTEXT_SIZE))),
+            llamacpp_chat_format=os.getenv(ENV_LLAMACPP_CHAT_FORMAT) or DEFAULT_LLAMACPP_CHAT_FORMAT,
+            llamacpp_no_grammar=os.getenv(ENV_LLAMACPP_NO_GRAMMAR, str(DEFAULT_LLAMACPP_NO_GRAMMAR)).lower()
+            in ("true", "1"),
+            llamacpp_extra_args=os.getenv(ENV_LLAMACPP_EXTRA_ARGS) or DEFAULT_LLAMACPP_EXTRA_ARGS,
             # 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,
@@ -1128,10 +1421,20 @@ class HindsightConfig:
             in ("true", "1"),
             embeddings_tei_url=os.getenv(ENV_EMBEDDINGS_TEI_URL),
             embeddings_openai_base_url=os.getenv(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None,
+            embeddings_openai_batch_size=_parse_positive_int(
+                ENV_EMBEDDINGS_OPENAI_BATCH_SIZE,
+                os.getenv(ENV_EMBEDDINGS_OPENAI_BATCH_SIZE),
+                DEFAULT_EMBEDDINGS_OPENAI_BATCH_SIZE,
+            ),
             # 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,
+            # OpenRouter embeddings (with fallback to shared OpenRouter key, then LLM key)
+            embeddings_openrouter_api_key=os.getenv(ENV_EMBEDDINGS_OPENROUTER_API_KEY)
+            or os.getenv(ENV_OPENROUTER_API_KEY)
+            or os.getenv(ENV_LLM_API_KEY),
+            embeddings_openrouter_model=os.getenv(ENV_EMBEDDINGS_OPENROUTER_MODEL, DEFAULT_EMBEDDINGS_OPENROUTER_MODEL),
             # 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),
@@ -1143,6 +1446,26 @@ class HindsightConfig:
                 ENV_EMBEDDINGS_LITELLM_SDK_MODEL, DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL
             ),
             embeddings_litellm_sdk_api_base=os.getenv(ENV_EMBEDDINGS_LITELLM_SDK_API_BASE) or None,
+            embeddings_litellm_sdk_output_dimensions=int(v)
+            if (v := os.getenv(ENV_EMBEDDINGS_LITELLM_SDK_OUTPUT_DIMENSIONS))
+            else None,
+            embeddings_litellm_sdk_encoding_format=os.getenv(
+                ENV_EMBEDDINGS_LITELLM_SDK_ENCODING_FORMAT, DEFAULT_EMBEDDINGS_LITELLM_SDK_ENCODING_FORMAT
+            ),
+            # Gemini/Vertex AI embeddings (with fallback to LLM keys)
+            embeddings_gemini_api_key=os.getenv(ENV_EMBEDDINGS_GEMINI_API_KEY) or os.getenv(ENV_LLM_API_KEY),
+            embeddings_gemini_model=os.getenv(ENV_EMBEDDINGS_GEMINI_MODEL, DEFAULT_EMBEDDINGS_GEMINI_MODEL),
+            embeddings_gemini_output_dimensionality=int(
+                os.getenv(
+                    ENV_EMBEDDINGS_GEMINI_OUTPUT_DIMENSIONALITY,
+                    str(DEFAULT_EMBEDDINGS_GEMINI_OUTPUT_DIMENSIONALITY),
+                )
+            ),
+            embeddings_vertexai_project_id=os.getenv(ENV_EMBEDDINGS_VERTEXAI_PROJECT_ID)
+            or os.getenv(ENV_LLM_VERTEXAI_PROJECT_ID),
+            embeddings_vertexai_region=os.getenv(ENV_EMBEDDINGS_VERTEXAI_REGION) or os.getenv(ENV_LLM_VERTEXAI_REGION),
+            embeddings_vertexai_service_account_key=os.getenv(ENV_EMBEDDINGS_VERTEXAI_SERVICE_ACCOUNT_KEY)
+            or os.getenv(ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY),
             # Reranker
             reranker_provider=os.getenv(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER),
             reranker_local_model=os.getenv(ENV_RERANKER_LOCAL_MODEL, DEFAULT_RERANKER_LOCAL_MODEL),
@@ -1171,11 +1494,19 @@ class HindsightConfig:
             reranker_tei_max_concurrent=int(
                 os.getenv(ENV_RERANKER_TEI_MAX_CONCURRENT, str(DEFAULT_RERANKER_TEI_MAX_CONCURRENT))
             ),
+            reranker_tei_http_timeout=float(
+                os.getenv(ENV_RERANKER_TEI_HTTP_TIMEOUT, str(DEFAULT_RERANKER_TEI_HTTP_TIMEOUT))
+            ),
             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,
+            # OpenRouter reranker (with fallback to shared OpenRouter key, then LLM key)
+            reranker_openrouter_api_key=os.getenv(ENV_RERANKER_OPENROUTER_API_KEY)
+            or os.getenv(ENV_OPENROUTER_API_KEY)
+            or os.getenv(ENV_LLM_API_KEY),
+            reranker_openrouter_model=os.getenv(ENV_RERANKER_OPENROUTER_MODEL, DEFAULT_RERANKER_OPENROUTER_MODEL),
             # 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),
@@ -1192,12 +1523,25 @@ class HindsightConfig:
             reranker_zeroentropy_api_key=os.getenv(ENV_RERANKER_ZEROENTROPY_API_KEY),
             reranker_zeroentropy_model=os.getenv(ENV_RERANKER_ZEROENTROPY_MODEL, DEFAULT_RERANKER_ZEROENTROPY_MODEL),
             reranker_zeroentropy_base_url=os.getenv(ENV_RERANKER_ZEROENTROPY_BASE_URL) or None,
+            # SiliconFlow reranker (Cohere-compatible /rerank endpoint)
+            reranker_siliconflow_api_key=os.getenv(ENV_RERANKER_SILICONFLOW_API_KEY),
+            reranker_siliconflow_model=os.getenv(ENV_RERANKER_SILICONFLOW_MODEL, DEFAULT_RERANKER_SILICONFLOW_MODEL),
+            reranker_siliconflow_base_url=os.getenv(
+                ENV_RERANKER_SILICONFLOW_BASE_URL, DEFAULT_RERANKER_SILICONFLOW_BASE_URL
+            ),
+            # Google Discovery Engine reranker (with fallback to LLM Vertex AI keys)
+            reranker_google_model=os.getenv(ENV_RERANKER_GOOGLE_MODEL, DEFAULT_RERANKER_GOOGLE_MODEL),
+            reranker_google_project_id=os.getenv(ENV_RERANKER_GOOGLE_PROJECT_ID)
+            or os.getenv(ENV_LLM_VERTEXAI_PROJECT_ID),
+            reranker_google_service_account_key=os.getenv(ENV_RERANKER_GOOGLE_SERVICE_ACCOUNT_KEY)
+            or os.getenv(ENV_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY),
             # Server
             host=os.getenv(ENV_HOST, DEFAULT_HOST),
             port=int(os.getenv(ENV_PORT, DEFAULT_PORT)),
             base_path=os.getenv(ENV_BASE_PATH, DEFAULT_BASE_PATH),
             log_level=os.getenv(ENV_LOG_LEVEL, DEFAULT_LOG_LEVEL),
             log_format=os.getenv(ENV_LOG_FORMAT, DEFAULT_LOG_FORMAT).lower(),
+            log_json_fields=_parse_str_list(os.getenv(ENV_LOG_JSON_FIELDS, "")) or None,
             mcp_enabled=os.getenv(ENV_MCP_ENABLED, str(DEFAULT_MCP_ENABLED)).lower() == "true",
             mcp_enabled_tools=[t.strip() for t in os.getenv(ENV_MCP_ENABLED_TOOLS).split(",") if t.strip()]
             if os.getenv(ENV_MCP_ENABLED_TOOLS)
@@ -1205,9 +1549,9 @@ class HindsightConfig:
             mcp_stateless=os.getenv(ENV_MCP_STATELESS, str(DEFAULT_MCP_STATELESS)).lower() == "true",
             enable_bank_config_api=os.getenv(ENV_ENABLE_BANK_CONFIG_API, str(DEFAULT_ENABLE_BANK_CONFIG_API)).lower()
             == "true",
+            default_bank_template=_parse_default_bank_template(os.getenv(ENV_DEFAULT_BANK_TEMPLATE)),
             # Recall
             graph_retriever=os.getenv(ENV_GRAPH_RETRIEVER, DEFAULT_GRAPH_RETRIEVER),
-            mpfp_top_k_neighbors=int(os.getenv(ENV_MPFP_TOP_K_NEIGHBORS, str(DEFAULT_MPFP_TOP_K_NEIGHBORS))),
             recall_max_concurrent=int(os.getenv(ENV_RECALL_MAX_CONCURRENT, str(DEFAULT_RECALL_MAX_CONCURRENT))),
             recall_connection_budget=int(
                 os.getenv(ENV_RECALL_CONNECTION_BUDGET, str(DEFAULT_RECALL_CONNECTION_BUDGET))
@@ -1216,6 +1560,10 @@ class HindsightConfig:
             mental_model_refresh_concurrency=int(
                 os.getenv(ENV_MENTAL_MODEL_REFRESH_CONCURRENCY, str(DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY))
             ),
+            link_expansion_per_entity_limit=int(
+                os.getenv(ENV_LINK_EXPANSION_PER_ENTITY_LIMIT, str(DEFAULT_LINK_EXPANSION_PER_ENTITY_LIMIT))
+            ),
+            link_expansion_timeout=float(os.getenv(ENV_LINK_EXPANSION_TIMEOUT, str(DEFAULT_LINK_EXPANSION_TIMEOUT))),
             # Optimization flags
             skip_llm_verification=os.getenv(ENV_SKIP_LLM_VERIFICATION, "false").lower() == "true",
             lazy_reranker=os.getenv(ENV_LAZY_RERANKER, "false").lower() == "true",
@@ -1242,6 +1590,7 @@ class HindsightConfig:
             retain_batch_poll_interval_seconds=int(
                 os.getenv(ENV_RETAIN_BATCH_POLL_INTERVAL_SECONDS, str(DEFAULT_RETAIN_BATCH_POLL_INTERVAL_SECONDS))
             ),
+            retain_chunk_batch_size=int(os.getenv(ENV_RETAIN_CHUNK_BATCH_SIZE, str(DEFAULT_RETAIN_CHUNK_BATCH_SIZE))),
             # File storage
             file_storage_type=os.getenv(ENV_FILE_STORAGE_TYPE, DEFAULT_FILE_STORAGE_TYPE),
             file_storage_s3_bucket=os.getenv(ENV_FILE_STORAGE_S3_BUCKET) or None,
@@ -1285,6 +1634,12 @@ class HindsightConfig:
             consolidation_batch_size=int(
                 os.getenv(ENV_CONSOLIDATION_BATCH_SIZE, str(DEFAULT_CONSOLIDATION_BATCH_SIZE))
             ),
+            consolidation_max_memories_per_round=int(
+                os.getenv(
+                    ENV_CONSOLIDATION_MAX_MEMORIES_PER_ROUND,
+                    str(DEFAULT_CONSOLIDATION_MAX_MEMORIES_PER_ROUND),
+                )
+            ),
             consolidation_llm_batch_size=int(
                 os.getenv(ENV_CONSOLIDATION_LLM_BATCH_SIZE, str(DEFAULT_CONSOLIDATION_LLM_BATCH_SIZE))
             ),
@@ -1300,6 +1655,9 @@ class HindsightConfig:
                     str(DEFAULT_CONSOLIDATION_SOURCE_FACTS_MAX_TOKENS_PER_OBSERVATION),
                 )
             ),
+            consolidation_max_attempts=int(
+                os.getenv(ENV_CONSOLIDATION_MAX_ATTEMPTS, str(DEFAULT_CONSOLIDATION_MAX_ATTEMPTS))
+            ),
             observations_mission=os.getenv(ENV_OBSERVATIONS_MISSION) or DEFAULT_OBSERVATIONS_MISSION,
             max_observations_per_scope=int(
                 os.getenv(ENV_MAX_OBSERVATIONS_PER_SCOPE, str(DEFAULT_MAX_OBSERVATIONS_PER_SCOPE))
@@ -1320,9 +1678,12 @@ class HindsightConfig:
             worker_max_retries=int(os.getenv(ENV_WORKER_MAX_RETRIES, str(DEFAULT_WORKER_MAX_RETRIES))),
             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))
-            ),
+            worker_slot_reservations={
+                op_type: int(os.getenv(env_var, str(default)))
+                for op_type, (env_var, default) in WORKER_SLOT_RESERVATION_TYPES.items()
+                if int(os.getenv(env_var, str(default))) > 0
+            },
+            retain_max_concurrent=int(os.getenv(ENV_RETAIN_MAX_CONCURRENT, str(DEFAULT_RETAIN_MAX_CONCURRENT))),
             # Reflect agent settings
             reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
             reflect_max_context_tokens=int(
@@ -1333,6 +1694,31 @@ class HindsightConfig:
             reflect_source_facts_max_tokens=int(
                 os.getenv(ENV_REFLECT_SOURCE_FACTS_MAX_TOKENS, str(DEFAULT_REFLECT_SOURCE_FACTS_MAX_TOKENS))
             ),
+            recall_include_chunks=os.getenv(ENV_RECALL_INCLUDE_CHUNKS, str(DEFAULT_RECALL_INCLUDE_CHUNKS)).lower()
+            in ("true", "1", "yes"),
+            recall_max_tokens=int(os.getenv(ENV_RECALL_MAX_TOKENS, str(DEFAULT_RECALL_MAX_TOKENS))),
+            recall_chunks_max_tokens=int(
+                os.getenv(ENV_RECALL_CHUNKS_MAX_TOKENS, str(DEFAULT_RECALL_CHUNKS_MAX_TOKENS))
+            ),
+            recall_budget_function=_validate_recall_budget_function(
+                os.getenv(ENV_RECALL_BUDGET_FUNCTION, DEFAULT_RECALL_BUDGET_FUNCTION)
+            ),
+            recall_budget_fixed_low=int(os.getenv(ENV_RECALL_BUDGET_FIXED_LOW, str(DEFAULT_RECALL_BUDGET_FIXED_LOW))),
+            recall_budget_fixed_mid=int(os.getenv(ENV_RECALL_BUDGET_FIXED_MID, str(DEFAULT_RECALL_BUDGET_FIXED_MID))),
+            recall_budget_fixed_high=int(
+                os.getenv(ENV_RECALL_BUDGET_FIXED_HIGH, str(DEFAULT_RECALL_BUDGET_FIXED_HIGH))
+            ),
+            recall_budget_adaptive_low=float(
+                os.getenv(ENV_RECALL_BUDGET_ADAPTIVE_LOW, str(DEFAULT_RECALL_BUDGET_ADAPTIVE_LOW))
+            ),
+            recall_budget_adaptive_mid=float(
+                os.getenv(ENV_RECALL_BUDGET_ADAPTIVE_MID, str(DEFAULT_RECALL_BUDGET_ADAPTIVE_MID))
+            ),
+            recall_budget_adaptive_high=float(
+                os.getenv(ENV_RECALL_BUDGET_ADAPTIVE_HIGH, str(DEFAULT_RECALL_BUDGET_ADAPTIVE_HIGH))
+            ),
+            recall_budget_min=int(os.getenv(ENV_RECALL_BUDGET_MIN, str(DEFAULT_RECALL_BUDGET_MIN))),
+            recall_budget_max=int(os.getenv(ENV_RECALL_BUDGET_MAX, str(DEFAULT_RECALL_BUDGET_MAX))),
             # Disposition settings (None = fall back to DB value)
             disposition_skepticism=int(os.getenv(ENV_DISPOSITION_SKEPTICISM))
             if os.getenv(ENV_DISPOSITION_SKEPTICISM)
@@ -1350,6 +1736,8 @@ class HindsightConfig:
             otel_exporter_otlp_headers=os.getenv(ENV_OTEL_EXPORTER_OTLP_HEADERS) or None,
             otel_service_name=os.getenv(ENV_OTEL_SERVICE_NAME, DEFAULT_OTEL_SERVICE_NAME),
             otel_deployment_environment=os.getenv(ENV_OTEL_DEPLOYMENT_ENVIRONMENT, DEFAULT_OTEL_DEPLOYMENT_ENVIRONMENT),
+            metrics_include_bank_id=os.getenv(ENV_METRICS_INCLUDE_BANK_ID, str(DEFAULT_METRICS_INCLUDE_BANK_ID)).lower()
+            in ("true", "1", "yes"),
             # Audit log configuration (static, server-level only)
             audit_log_enabled=os.getenv(ENV_AUDIT_LOG_ENABLED, str(DEFAULT_AUDIT_LOG_ENABLED)).lower() == "true",
             audit_log_actions=[
@@ -1421,7 +1809,8 @@ class HindsightConfig:
         handler.setLevel(self.get_python_log_level())
 
         if self.log_format == "json":
-            handler.setFormatter(JsonFormatter())
+            allowed = frozenset(self.log_json_fields) if self.log_json_fields else None
+            handler.setFormatter(JsonFormatter(allowed_fields=allowed))
         else:
             handler.setFormatter(logging.Formatter("%(asctime)s - %(levelname)s - %(name)s - %(message)s"))
 

hindsight-api-slim/hindsight_api/config_resolver.py

@@ -15,7 +15,12 @@ from typing import Any
 
 import asyncpg
 
-from hindsight_api.config import HindsightConfig, _get_raw_config, normalize_config_dict
+from hindsight_api.config import (
+    RECALL_BUDGET_FUNCTIONS,
+    HindsightConfig,
+    _get_raw_config,
+    normalize_config_dict,
+)
 from hindsight_api.engine.memory_engine import fq_table
 from hindsight_api.extensions.tenant import TenantExtension
 from hindsight_api.models import RequestContext
@@ -239,6 +244,15 @@ class ConfigResolver:
                 logger.warning(f"Failed to check permissions for bank {bank_id}: {e}")
                 # Continue without permission check (fail open for backward compatibility)
 
+        # Validate entity_labels structure
+        if "entity_labels" in normalized_updates and normalized_updates["entity_labels"] is not None:
+            from .engine.retain.entity_labels import parse_entity_labels
+
+            try:
+                parse_entity_labels(normalized_updates["entity_labels"])
+            except Exception as e:
+                raise ValueError(f"Invalid entity_labels format: {e}")
+
         # Validate retain_strategies: reject empty string keys
         if "retain_strategies" in normalized_updates and normalized_updates["retain_strategies"]:
             empty_keys = [k for k in normalized_updates["retain_strategies"] if not str(k).strip()]
@@ -247,6 +261,9 @@ class ConfigResolver:
                     "Strategy names must not be empty strings. Remove entries with empty names before saving."
                 )
 
+        # Validate recall budget fields
+        _validate_recall_budget_updates(normalized_updates)
+
         # Merge with existing config (JSONB || operator)
         async with self.pool.acquire() as conn:
             await conn.execute(
@@ -283,6 +300,53 @@ class ConfigResolver:
         logger.info(f"Reset bank config for {bank_id} to defaults")
 
 
+_RECALL_BUDGET_FIXED_KEYS = (
+    "recall_budget_fixed_low",
+    "recall_budget_fixed_mid",
+    "recall_budget_fixed_high",
+)
+_RECALL_BUDGET_ADAPTIVE_KEYS = (
+    "recall_budget_adaptive_low",
+    "recall_budget_adaptive_mid",
+    "recall_budget_adaptive_high",
+)
+
+
+def _validate_recall_budget_updates(updates: dict[str, Any]) -> None:
+    """Validate recall budget config updates. Raises ValueError on invalid input."""
+    if "recall_budget_function" in updates:
+        function = updates["recall_budget_function"]
+        if not isinstance(function, str) or function.lower() not in RECALL_BUDGET_FUNCTIONS:
+            raise ValueError(
+                f"recall_budget_function must be one of {sorted(RECALL_BUDGET_FUNCTIONS)}, got {function!r}"
+            )
+
+    for key in _RECALL_BUDGET_FIXED_KEYS:
+        if key in updates:
+            value = updates[key]
+            if not isinstance(value, int) or isinstance(value, bool) or value < 1:
+                raise ValueError(f"{key} must be a positive integer, got {value!r}")
+
+    for key in _RECALL_BUDGET_ADAPTIVE_KEYS:
+        if key in updates:
+            value = updates[key]
+            if isinstance(value, bool) or not isinstance(value, (int, float)) or value <= 0:
+                raise ValueError(f"{key} must be a positive number, got {value!r}")
+
+    for key in ("recall_budget_min", "recall_budget_max"):
+        if key in updates:
+            value = updates[key]
+            if not isinstance(value, int) or isinstance(value, bool) or value < 1:
+                raise ValueError(f"{key} must be a positive integer, got {value!r}")
+
+    if "recall_budget_min" in updates and "recall_budget_max" in updates:
+        if updates["recall_budget_min"] > updates["recall_budget_max"]:
+            raise ValueError(
+                f"recall_budget_min ({updates['recall_budget_min']}) must be <= "
+                f"recall_budget_max ({updates['recall_budget_max']})"
+            )
+
+
 def apply_strategy(config: HindsightConfig, strategy_name: str) -> HindsightConfig:
     """
     Apply a named retain strategy's overrides on top of a resolved config.

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

@@ -42,6 +42,34 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+async def _filter_live_source_memories(
+    conn: "Connection",
+    bank_id: str,
+    source_memory_ids: list[uuid.UUID],
+) -> list[uuid.UUID]:
+    """Return only the source memory ids that still exist in the bank.
+
+    Uses FOR SHARE to block concurrent deletes from removing a row between the
+    check and the subsequent insert/update. Combined with the delete path running
+    its stale-observation sweep *after* deleting the source row, this closes the
+    race window where consolidation would otherwise produce an orphan observation.
+    """
+    if not source_memory_ids:
+        return []
+    rows = await conn.fetch(
+        f"""
+        SELECT id
+        FROM {fq_table("memory_units")}
+        WHERE id = ANY($1::uuid[]) AND bank_id = $2
+        FOR SHARE
+        """,
+        source_memory_ids,
+        bank_id,
+    )
+    live = {row["id"] for row in rows}
+    return [mid for mid in source_memory_ids if mid in live]
+
+
 class _CreateAction(BaseModel):
     text: str
     source_fact_ids: list[str]  # memory UUIDs from the NEW FACTS list
@@ -219,6 +247,7 @@ async def run_consolidation_job(
 
     perf = ConsolidationPerfLog(bank_id)
     max_memories_per_batch = config.consolidation_batch_size
+    max_memories_per_round = config.consolidation_max_memories_per_round
     llm_batch_size = max(1, config.consolidation_llm_batch_size)
 
     # Check if consolidation is enabled
@@ -281,8 +310,17 @@ async def run_consolidation_job(
     # Track all unique tags from consolidated memories for mental model refresh filtering
     consolidated_tags: set[str] = set()
 
+    round_limit_enabled = max_memories_per_round > 0
+    round_remaining = max_memories_per_round if round_limit_enabled else float("inf")
+    hit_round_limit = False
+
     llm_batch_num = 0
     while True:
+        # Cap fetch size by remaining round budget
+        fetch_limit = (
+            min(max_memories_per_batch, int(round_remaining)) if round_limit_enabled else max_memories_per_batch
+        )
+
         # Fetch next batch of unconsolidated memories
         async with pool.acquire() as conn:
             t0 = time.time()
@@ -299,7 +337,7 @@ async def run_consolidation_job(
                 LIMIT $2
                 """,
                 bank_id,
-                max_memories_per_batch,
+                fetch_limit,
             )
             perf.record_timing("fetch_memories", time.time() - t0)
 
@@ -524,6 +562,25 @@ async def run_consolidation_job(
                 f" | avg={llm_batch_time / len(llm_batch):.3f}s/memory"
             )
 
+        # Update round budget after processing this DB fetch batch
+        if round_limit_enabled:
+            round_remaining -= len(memories)
+            if round_remaining <= 0:
+                hit_round_limit = True
+                break
+
+    # Re-submit consolidation if we hit the round limit and there's likely more work
+    if hit_round_limit:
+        remaining = total_count - stats["memories_processed"]
+        logger.info(
+            f"[CONSOLIDATION] bank={bank_id} hit round limit of {max_memories_per_round} memories,"
+            f" ~{remaining} remaining. Re-queuing consolidation."
+        )
+        try:
+            await memory_engine.submit_async_consolidation(bank_id=bank_id, request_context=request_context)
+        except Exception as e:
+            logger.warning(f"[CONSOLIDATION] bank={bank_id} failed to re-queue consolidation: {e}")
+
     # Build summary
     perf.log(
         f"[3] Results: {stats['memories_processed']} memories -> "
@@ -552,16 +609,21 @@ async def run_consolidation_job(
     if timing_parts:
         perf.log(f"[4] Timing breakdown: {', '.join(timing_parts)}")
 
-    # Trigger mental model refreshes for models with refresh_after_consolidation=true
-    # SECURITY: Only refresh mental models with matching tags (or all if no tags were consolidated)
-    mental_models_refreshed = await _trigger_mental_model_refreshes(
-        memory_engine=memory_engine,
-        bank_id=bank_id,
-        request_context=request_context,
-        consolidated_tags=list(consolidated_tags) if consolidated_tags else None,
-        perf=perf,
-    )
-    stats["mental_models_refreshed"] = mental_models_refreshed
+    # Trigger mental model refreshes only on the final round (when all memories are processed).
+    # If we hit the round limit and re-queued, skip MM refresh — the next round will handle it.
+    if hit_round_limit:
+        stats["mental_models_refreshed"] = 0
+        logger.info(f"[CONSOLIDATION] bank={bank_id} skipping mental model refresh (round limit hit, re-queued)")
+    else:
+        # SECURITY: Only refresh mental models with matching tags (or all if no tags were consolidated)
+        mental_models_refreshed = await _trigger_mental_model_refreshes(
+            memory_engine=memory_engine,
+            bank_id=bank_id,
+            request_context=request_context,
+            consolidated_tags=list(consolidated_tags) if consolidated_tags else None,
+            perf=perf,
+        )
+        stats["mental_models_refreshed"] = mental_models_refreshed
 
     perf.flush()
 
@@ -593,17 +655,15 @@ async def _trigger_mental_model_refreshes(
     """
     pool = memory_engine._pool
 
-    # Find mental models with refresh_after_consolidation=true
-    # SECURITY: Control which mental models get refreshed based on tags
+    # Find mental models with refresh_after_consolidation=true that are actually stale.
+    # The tag filter on the SELECT enforces the security boundary (never look outside the
+    # relevant tag scope); compute_mental_model_is_stale then verifies that new memories
+    # in the MM's scope really were ingested since its last refresh.
     async with pool.acquire() as conn:
         if consolidated_tags:
-            # Tagged memories were consolidated - refresh:
-            # 1. Mental models with overlapping tags (security boundary)
-            # 2. Untagged mental models (they're "global" and available to all contexts)
-            # DO NOT refresh mental models with different tags
-            rows = await conn.fetch(
+            candidates = await conn.fetch(
                 f"""
-                SELECT id, name, tags
+                SELECT id, name, tags, last_refreshed_at, trigger
                 FROM {fq_table("mental_models")}
                 WHERE bank_id = $1
                   AND (trigger->>'refresh_after_consolidation')::boolean = true
@@ -616,11 +676,9 @@ async def _trigger_mental_model_refreshes(
                 consolidated_tags,
             )
         else:
-            # Untagged memories were consolidated - only refresh untagged mental models
-            # SECURITY: Tagged mental models are NOT refreshed when untagged memories are consolidated
-            rows = await conn.fetch(
+            candidates = await conn.fetch(
                 f"""
-                SELECT id, name, tags
+                SELECT id, name, tags, last_refreshed_at, trigger
                 FROM {fq_table("mental_models")}
                 WHERE bank_id = $1
                   AND (trigger->>'refresh_after_consolidation')::boolean = true
@@ -629,6 +687,11 @@ async def _trigger_mental_model_refreshes(
                 bank_id,
             )
 
+        rows = []
+        for candidate in candidates:
+            if await memory_engine.compute_mental_model_is_stale(conn, bank_id, candidate):
+                rows.append(candidate)
+
     if not rows:
         return 0
 
@@ -889,6 +952,15 @@ async def _execute_update_action(
         logger.debug(f"Update skipped: observation {observation_id} not found in recall results")
         return
 
+    live_source_memory_ids = await _filter_live_source_memories(conn, bank_id, source_memory_ids)
+    if not live_source_memory_ids:
+        logger.debug(
+            f"Update skipped: all {len(source_memory_ids)} source memories for observation "
+            f"{observation_id} were deleted concurrently"
+        )
+        return
+    source_memory_ids = live_source_memory_ids
+
     from ...config import get_config
 
     history_entry = {
@@ -1131,14 +1203,16 @@ async def _consolidate_batch_with_llm(
     memories: list[dict[str, Any]],
     union_observations: "list[MemoryFact]",
     union_source_facts: "dict[str, MemoryFact]",
-    config: Any = None,
+    config: Any,
     remaining_observation_slots: int | None = None,
     max_observations_per_scope: int = -1,
 ) -> _BatchLLMResult:
     """Single LLM call for a batch of facts against a pooled set of observations."""
+    if config is None:
+        raise ValueError("config is required for _consolidate_batch_with_llm")
     if union_observations:
         obs_list = _build_observations_for_llm(union_observations, union_source_facts)
-        observations_text = json.dumps(obs_list, indent=2)
+        observations_text = json.dumps(obs_list, indent=2, ensure_ascii=False)
     else:
         observations_text = "[]"
 
@@ -1172,8 +1246,7 @@ async def _consolidate_batch_with_llm(
                 f"(out of {max_observations_per_scope}). Prefer UPDATE over CREATE when possible."
             )
 
-    observations_mission = config.observations_mission if config is not None else None
-    prompt_template = build_batch_consolidation_prompt(observations_mission, observation_capacity_note)
+    prompt_template = build_batch_consolidation_prompt(config.observations_mission, observation_capacity_note)
     prompt = prompt_template.format(
         facts_text=facts_lines,
         observations_text=observations_text,
@@ -1182,15 +1255,29 @@ async def _consolidate_batch_with_llm(
     # Use a constrained response model when observation limit is active
     response_model = _build_response_model(max_creates=remaining_observation_slots)
 
-    max_attempts = 3
+    max_attempts = config.consolidation_max_attempts
+    inner_max_retries = config.consolidation_llm_max_retries
     last_exc: Exception | None = None
+    # Pre-compute a stable identifier set for the batch so failure logs name the
+    # exact memories whose consolidation is failing — without this, an opaque
+    # "LLM batch call failed" line gives operators no way to find the offending
+    # input until adaptive bisection narrows the batch down to a single memory.
+    memory_ids = [str(m.get("id")) for m in memories]
+    if len(memory_ids) <= 5:
+        ids_label = ", ".join(memory_ids)
+    else:
+        ids_label = f"{', '.join(memory_ids[:3])}, ... +{len(memory_ids) - 3} more"
+    batch_label = f"{len(memory_ids)} memories [{ids_label}]"
     for attempt in range(1, max_attempts + 1):
         try:
-            response: _ConsolidationBatchResponse = await llm_config.call(
-                messages=[{"role": "user", "content": prompt}],
-                response_format=response_model,
-                scope="consolidation",
-            )
+            call_kwargs: dict[str, Any] = {
+                "messages": [{"role": "user", "content": prompt}],
+                "response_format": response_model,
+                "scope": "consolidation",
+            }
+            if inner_max_retries is not None:
+                call_kwargs["max_retries"] = inner_max_retries
+            response: _ConsolidationBatchResponse = await llm_config.call(**call_kwargs)
             # Defensive truncation: some LLM providers may not enforce JSON schema max_length
             creates = response.creates
             if remaining_observation_slots is not None and remaining_observation_slots >= 0:
@@ -1209,10 +1296,13 @@ async def _consolidate_batch_with_llm(
             )
         except Exception as exc:
             last_exc = exc
-            logger.warning(f"[CONSOLIDATION] LLM batch call failed (attempt {attempt}/{max_attempts}): {exc}")
+            logger.warning(
+                f"[CONSOLIDATION] LLM batch call failed (attempt {attempt}/{max_attempts}) for {batch_label}: {exc}"
+            )
 
     logger.error(
-        f"[CONSOLIDATION] LLM batch call failed after {max_attempts} attempts, skipping batch. Last error: {last_exc}"
+        f"[CONSOLIDATION] LLM batch call failed after {max_attempts} attempts for {batch_label}, "
+        f"skipping batch. Last error: {last_exc}"
     )
     return _BatchLLMResult(obs_count=len(union_observations), prompt_chars=len(prompt), failed=True)
 
@@ -1231,6 +1321,12 @@ async def _create_observation_directly(
     perf: ConsolidationPerfLog | None = None,
 ) -> dict[str, Any]:
     """Create an observation from one or more source memories with pre-processed text."""
+    live_source_memory_ids = await _filter_live_source_memories(conn, bank_id, source_memory_ids)
+    if not live_source_memory_ids:
+        logger.debug(f"Create skipped: all {len(source_memory_ids)} source memories were deleted concurrently")
+        return {"action": "skipped", "reason": "sources_deleted"}
+    source_memory_ids = live_source_memory_ids
+
     # Generate embedding for the observation (convert to string for pgvector)
     t0 = time.time()
     embeddings = await embedding_utils.generate_embeddings_batch(memory_engine.embeddings, [observation_text])

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

@@ -5,10 +5,24 @@ _DEFAULT_MISSION = "Track every detail: names, numbers, dates, places, and relat
 
 # Processing rules — always present regardless of mission
 _PROCESSING_RULES = """Processing rules (always apply):
-- REDUNDANT: same info worded differently → UPDATE the existing observation.
-- CONTRADICTION/UPDATE: capture both states with temporal markers ("used to X, now Y").
-- RESOLVE REFERENCES: when a new fact provides a concrete value resolving a vague placeholder in an existing observation (e.g. "home country", "hometown", "birthplace", "native language", "her ex", "that city"), UPDATE the observation to embed the resolved value explicitly. Example: new fact says "grandma in Sweden" + existing observation says "moved from her home country" → update to "home country is Sweden".
-- NEVER merge observations about different people or unrelated topics."""
+
+1. ONE OBSERVATION PER DISTINCT FACET: each observation tracks exactly one specific facet — a count ("has 3 items"), a named entity ("has a dog named Rex"), a relationship ("works at Google"), etc. Never merge different facets into one observation.
+
+2. MATCH BY ENTITY/FACET, NOT TOPIC: when deciding whether to UPDATE vs CREATE, match on the specific entity or facet. "Sold item X" updates only the X observation. "Now has 5 items" updates only the count observation. Do not update observations about different entities just because they share a general topic.
+
+3. STATE CHANGES — UPDATE CONCISELY: when a fact changes the state of something ("sold X", "X died", "moved to Y"), UPDATE the matching observation to reflect the current state. Include dates when available. Keep it concise — only information about THAT specific facet. Example: "User owned a dog named Rex who died on March 15, 2025". Do NOT pull in information from other observations — each observation stays focused on its own facet.
+
+4. CASCADE TO ALL AFFECTED OBSERVATIONS: a state change may affect multiple observations. For example, if entity C is removed from a group, update BOTH the individual observation for C AND any list/group observation that includes C (remove C from the list while keeping all other members intact).
+
+5. NO COMPUTATION: you do not have the full picture — never calculate, derive, or adjust numeric values. If the user says "I have 2 dogs" and then "I have a dog named Rex", do NOT update the count to 3 — you don't know if Rex is one of the 2 or a new one. If the user says "I sold X", do NOT decrement a count. Only update a count when the user explicitly states a new count. Synthesize and consolidate what was stated, but never do arithmetic or logical deductions.
+
+6. SAME FACET → UPDATE, NOT CREATE: a new count supersedes the old count — UPDATE the existing count observation, don't create a second one. If there's an existing observation for the same specific facet, always UPDATE it rather than creating a duplicate.
+
+7. PRESERVE HISTORY: observations that record significant events (sold, died, moved, changed) are important history — never DELETE them. Only delete an observation when it is restated identically or truly meaningless. Be very conservative with deletes.
+
+8. RESOLVE REFERENCES: when a new fact provides a concrete value for a vague placeholder in an existing observation (e.g., "home country" → "Sweden"), UPDATE to embed the resolved value.
+
+9. NEVER merge observations about different people or unrelated topics."""
 
 # Data section — format placeholders {facts_text} and {observations_text} are substituted at call time
 _BATCH_DATA_SECTION = """
@@ -26,8 +40,8 @@ Each observation includes:
 - source_memories: array of supporting facts with their text and dates
 
 Compare the facts against existing observations:
-- Same topic as an existing observation → UPDATE it (observation_id + source_fact_ids)
-- New topic with durable knowledge → CREATE a new observation (source_fact_ids)
+- Same facet as an existing observation → UPDATE it (observation_id + source_fact_ids)
+- New facet with durable knowledge → CREATE a new observation (source_fact_ids)
 - Cross-reference facts within the batch: a later fact may resolve a vague reference in an earlier one
 - Purely ephemeral facts → omit them unless the MISSION above explicitly targets such data (e.g. timestamped events, session state, screen content)"""
 

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

@@ -20,6 +20,7 @@ from ..config import (
     DEFAULT_RERANKER_COHERE_MODEL,
     DEFAULT_RERANKER_FLASHRANK_CACHE_DIR,
     DEFAULT_RERANKER_FLASHRANK_MODEL,
+    DEFAULT_RERANKER_GOOGLE_MODEL,
     DEFAULT_RERANKER_LITELLM_MAX_TOKENS_PER_DOC,
     DEFAULT_RERANKER_LITELLM_MODEL,
     DEFAULT_RERANKER_LITELLM_SDK_MODEL,
@@ -29,20 +30,26 @@ from ..config import (
     DEFAULT_RERANKER_LOCAL_MODEL,
     DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE,
     DEFAULT_RERANKER_PROVIDER,
+    DEFAULT_RERANKER_SILICONFLOW_BASE_URL,
+    DEFAULT_RERANKER_SILICONFLOW_MODEL,
     DEFAULT_RERANKER_TEI_BATCH_SIZE,
+    DEFAULT_RERANKER_TEI_HTTP_TIMEOUT,
     DEFAULT_RERANKER_TEI_MAX_CONCURRENT,
     DEFAULT_RERANKER_ZEROENTROPY_MODEL,
     ENV_RERANKER_COHERE_API_KEY,
     ENV_RERANKER_COHERE_MODEL,
     ENV_RERANKER_FLASHRANK_CACHE_DIR,
     ENV_RERANKER_FLASHRANK_MODEL,
+    ENV_RERANKER_GOOGLE_PROJECT_ID,
     ENV_RERANKER_LITELLM_SDK_API_KEY,
     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_SILICONFLOW_API_KEY,
     ENV_RERANKER_TEI_BATCH_SIZE,
+    ENV_RERANKER_TEI_HTTP_TIMEOUT,
     ENV_RERANKER_TEI_MAX_CONCURRENT,
     ENV_RERANKER_TEI_URL,
     ENV_RERANKER_ZEROENTROPY_API_KEY,
@@ -516,6 +523,84 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
         return await self._predict_async(pairs)
 
 
+class _CohereCompatibleRerankClient:
+    """
+    Internal HTTP client for Cohere-compatible /rerank endpoints.
+
+    Shared by all providers that speak the Cohere rerank wire format —
+    {model, query, documents[, top_n]} request and
+    {results: [{index, relevance_score}, ...]} response. This covers
+    SiliconFlow, ZeroEntropy, Jina, Voyage, BGE self-hosted, and Cohere
+    itself when reached via a custom base_url (e.g. Azure AI Foundry).
+
+    Not a CrossEncoderModel — providers compose it and expose their own
+    provider_name / initialization logging.
+    """
+
+    def __init__(
+        self,
+        api_key: str,
+        model: str,
+        rerank_url: str,
+        timeout: float = 60.0,
+        include_top_n: bool = True,
+    ):
+        self.api_key = api_key
+        self.model = model
+        self.rerank_url = rerank_url
+        self.timeout = timeout
+        self.include_top_n = include_top_n
+        self._async_client: httpx.AsyncClient | None = None
+
+    async def initialize(self) -> None:
+        if self._async_client is not None:
+            return
+        self._async_client = httpx.AsyncClient(
+            timeout=self.timeout,
+            headers={
+                "Authorization": f"Bearer {self.api_key}",
+                "Content-Type": "application/json",
+            },
+        )
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        if self._async_client is None:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        if not pairs:
+            return []
+
+        query_groups: dict[str, list[tuple[int, str]]] = {}
+        for idx, (query, text) in enumerate(pairs):
+            query_groups.setdefault(query, []).append((idx, text))
+
+        all_scores = [0.0] * len(pairs)
+
+        for query, indexed_texts in query_groups.items():
+            texts = [text for _, text in indexed_texts]
+            indices = [idx for idx, _ in indexed_texts]
+
+            body: dict[str, object] = {
+                "model": self.model,
+                "query": query,
+                "documents": texts,
+                "return_documents": False,
+            }
+            if self.include_top_n:
+                body["top_n"] = len(texts)
+
+            response = await self._async_client.post(self.rerank_url, json=body)
+            response.raise_for_status()
+            result = response.json()
+
+            for item in result.get("results", []):
+                original_idx = item["index"]
+                score = item["relevance_score"]
+                all_scores[indices[original_idx]] = score
+
+        return all_scores
+
+
 class CohereCrossEncoder(CrossEncoderModel):
     """
     Cohere cross-encoder implementation using the Cohere Rerank API.
@@ -544,6 +629,20 @@ class CohereCrossEncoder(CrossEncoderModel):
         self.base_url = base_url
         self.timeout = timeout
         self._client = None
+        # Used when base_url is set (Azure AI Foundry and other Cohere-compatible hosts).
+        # Azure endpoints already include the full invoke path, so rerank_url == base_url
+        # and top_n is omitted to match the existing Azure contract.
+        self._http_client: _CohereCompatibleRerankClient | None = (
+            _CohereCompatibleRerankClient(
+                api_key=api_key,
+                model=model,
+                rerank_url=base_url,
+                timeout=timeout,
+                include_top_n=False,
+            )
+            if base_url
+            else None
+        )
 
     @property
     def provider_name(self) -> str:
@@ -551,23 +650,24 @@ class CohereCrossEncoder(CrossEncoderModel):
 
     async def initialize(self) -> None:
         """Initialize the Cohere client."""
-        if self._client is not None:
+        if self._client is not None or (self._http_client and self._http_client._async_client):
             return
 
-        try:
-            import cohere
-        except ImportError:
-            raise ImportError("cohere is required for CohereCrossEncoder. Install it with: pip install cohere")
-
         base_url_msg = f" at {self.base_url}" if self.base_url else ""
         logger.info(f"Reranker: initializing Cohere provider with model {self.model}{base_url_msg}")
 
-        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
-        client_kwargs = {"api_key": self.api_key, "timeout": self.timeout}
-        if self.base_url:
-            client_kwargs["base_url"] = self.base_url
-        self._client = cohere.Client(**client_kwargs)
-        logger.info("Reranker: Cohere provider initialized")
+        if self._http_client is not None:
+            await self._http_client.initialize()
+            logger.info("Reranker: Cohere provider initialized (Cohere-compatible HTTP endpoint)")
+        else:
+            # For native Cohere API, use the official SDK
+            try:
+                import cohere
+            except ImportError:
+                raise ImportError("cohere is required for CohereCrossEncoder. Install it with: pip install cohere")
+
+            self._client = cohere.Client(api_key=self.api_key, timeout=self.timeout)
+            logger.info("Reranker: Cohere provider initialized")
 
     async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
         """
@@ -579,25 +679,24 @@ class CohereCrossEncoder(CrossEncoderModel):
         Returns:
             List of relevance scores
         """
-        if self._client is None:
+        if self._client is None and self._http_client is None:
             raise RuntimeError("Reranker not initialized. Call initialize() first.")
 
         if not pairs:
             return []
 
-        # Run sync Cohere API calls in thread pool
-        loop = asyncio.get_event_loop()
-        return await loop.run_in_executor(None, self._predict_sync, pairs)
+        if self._http_client is not None:
+            return await self._http_client.predict(pairs)
 
-    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """Synchronous predict implementation for Cohere API."""
-        # Group pairs by query for efficient batching
-        # Cohere rerank expects one query with multiple documents
+        # Run sync Cohere SDK calls in thread pool
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(None, self._predict_sync_sdk, pairs)
+
+    def _predict_sync_sdk(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """Synchronous predict using the native Cohere SDK."""
         query_groups: dict[str, list[tuple[int, str]]] = {}
         for idx, (query, text) in enumerate(pairs):
-            if query not in query_groups:
-                query_groups[query] = []
-            query_groups[query].append((idx, text))
+            query_groups.setdefault(query, []).append((idx, text))
 
         all_scores = [0.0] * len(pairs)
 
@@ -612,7 +711,6 @@ class CohereCrossEncoder(CrossEncoderModel):
                 return_documents=False,
             )
 
-            # Map scores back to original positions
             for result in response.results:
                 original_idx = result.index
                 score = result.relevance_score
@@ -639,89 +737,70 @@ class ZeroEntropyCrossEncoder(CrossEncoderModel):
         base_url: str | None = None,
         timeout: float = 60.0,
     ):
-        """
-        Initialize ZeroEntropy cross-encoder client.
-
-        Args:
-            api_key: ZeroEntropy API key
-            model: ZeroEntropy rerank model name (default: zerank-2)
-            base_url: Custom base URL for ZeroEntropy-compatible API (e.g., mock server or proxy)
-            timeout: Request timeout in seconds (default: 60.0)
-        """
-        self.api_key = api_key
         self.model = model
         self.base_url = base_url.rstrip("/") if base_url else self.DEFAULT_BASE_URL
-        self.rerank_url = f"{self.base_url}{self.RERANK_PATH}"
-        self.timeout = timeout
-        self._async_client: httpx.AsyncClient | None = None
+        self._client = _CohereCompatibleRerankClient(
+            api_key=api_key,
+            model=model,
+            rerank_url=f"{self.base_url}{self.RERANK_PATH}",
+            timeout=timeout,
+        )
 
     @property
     def provider_name(self) -> str:
         return "zeroentropy"
 
     async def initialize(self) -> None:
-        """Initialize the async HTTP client."""
-        if self._async_client is not None:
+        if self._client._async_client is not None:
             return
-
         logger.info(f"Reranker: initializing ZeroEntropy provider with model {self.model}")
-        self._async_client = httpx.AsyncClient(
-            timeout=self.timeout,
-            headers={
-                "Authorization": f"Bearer {self.api_key}",
-                "Content-Type": "application/json",
-            },
-        )
+        await self._client.initialize()
         logger.info("Reranker: ZeroEntropy provider initialized")
 
     async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
-        """
-        Score query-document pairs using the ZeroEntropy Rerank API.
+        return await self._client.predict(pairs)
 
-        Args:
-            pairs: List of (query, document) tuples to score
 
-        Returns:
-            List of relevance scores
-        """
-        if self._async_client is None:
-            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+class SiliconFlowCrossEncoder(CrossEncoderModel):
+    """
+    SiliconFlow cross-encoder implementation.
 
-        if not pairs:
-            return []
+    SiliconFlow (https://siliconflow.cn) exposes a Cohere-compatible /rerank
+    endpoint. Shares the HTTP client with ZeroEntropy/Cohere-custom-endpoint
+    via _CohereCompatibleRerankClient.
+    """
 
-        # Group pairs by query for efficient batching
-        query_groups: dict[str, list[tuple[int, str]]] = {}
-        for idx, (query, text) in enumerate(pairs):
-            if query not in query_groups:
-                query_groups[query] = []
-            query_groups[query].append((idx, text))
+    RERANK_PATH = "/rerank"
 
-        all_scores = [0.0] * len(pairs)
+    def __init__(
+        self,
+        api_key: str,
+        model: str = DEFAULT_RERANKER_SILICONFLOW_MODEL,
+        base_url: str = DEFAULT_RERANKER_SILICONFLOW_BASE_URL,
+        timeout: float = 60.0,
+    ):
+        self.model = model
+        self.base_url = base_url.rstrip("/")
+        self._client = _CohereCompatibleRerankClient(
+            api_key=api_key,
+            model=model,
+            rerank_url=f"{self.base_url}{self.RERANK_PATH}",
+            timeout=timeout,
+        )
 
-        for query, indexed_texts in query_groups.items():
-            texts = [text for _, text in indexed_texts]
-            indices = [idx for idx, _ in indexed_texts]
+    @property
+    def provider_name(self) -> str:
+        return "siliconflow"
 
-            response = await self._async_client.post(
-                self.rerank_url,
-                json={
-                    "model": self.model,
-                    "query": query,
-                    "documents": texts,
-                    "top_n": len(texts),
-                },
-            )
-            response.raise_for_status()
-            result = response.json()
+    async def initialize(self) -> None:
+        if self._client._async_client is not None:
+            return
+        logger.info(f"Reranker: initializing SiliconFlow provider at {self.base_url} with model {self.model}")
+        await self._client.initialize()
+        logger.info("Reranker: SiliconFlow provider initialized")
 
-            # Map scores back to original positions
-            for item in result.get("results", []):
-                original_idx = item["index"]
-                score = item["relevance_score"]
-                all_scores[indices[original_idx]] = score
-
-        return all_scores
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        return await self._client.predict(pairs)
 
 
 class RRFPassthroughCrossEncoder(CrossEncoderModel):
@@ -1173,14 +1252,31 @@ class JinaMLXCrossEncoder(CrossEncoderModel):
         if self._reranker is not None:
             return
 
+        # Pre-warm transformers.AutoTokenizer to fully populate the transformers
+        # namespace before mlx_lm imports it. transformers 5.x uses _LazyModule,
+        # which has an unguarded window where `from transformers import AutoTokenizer`
+        # raises ImportError if another thread is concurrently initializing the
+        # namespace (e.g. embeddings init in an executor thread).
+        # See: https://github.com/vectorize-io/hindsight/issues/994
+        import transformers
+
+        _ = transformers.AutoTokenizer
+
         try:
             import mlx.core  # noqa: F401
             import mlx_lm  # noqa: F401
-        except ImportError:
+        except ImportError as exc:
+            # Only swallow "package not installed" errors. Anything else (e.g. a
+            # transitive import failure inside mlx_lm) must surface verbatim so
+            # the real cause is debuggable instead of being masked by a generic
+            # "install mlx" message.
+            msg = str(exc)
+            if "mlx" not in msg and "mlx_lm" not in msg:
+                raise
             raise ImportError(
                 "mlx and mlx-lm are required for JinaMLXCrossEncoder. "
                 "Install with: pip install mlx>=0.31.0 mlx-lm>=0.31.1 safetensors>=0.6.2"
-            )
+            ) from exc
 
         loop = asyncio.get_event_loop()
         await loop.run_in_executor(None, self._load_model)
@@ -1188,6 +1284,7 @@ class JinaMLXCrossEncoder(CrossEncoderModel):
     def _load_model(self) -> None:
         """Download (if needed) and load the MLX reranker. Runs in a thread."""
         import os
+        import threading
 
         from huggingface_hub import snapshot_download
 
@@ -1203,6 +1300,10 @@ class JinaMLXCrossEncoder(CrossEncoderModel):
             model_path=model_path,
             projector_path=os.path.join(model_path, "projector.safetensors"),
         )
+        # MLX Metal GPU ops are not thread-safe — concurrent calls to
+        # Device::end_encoding() crash with SIGSEGV (NULL deref).
+        # Serialize all reranker inference through this lock.
+        self._mlx_lock = threading.Lock()
         logger.info("Reranker: jina-mlx provider initialized")
 
     def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
@@ -1216,13 +1317,14 @@ class JinaMLXCrossEncoder(CrossEncoderModel):
 
         all_scores = [0.0] * len(pairs)
 
-        for query, indexed_docs in query_groups.items():
-            docs = [doc for _, doc in indexed_docs]
-            indices = [idx for idx, _ in indexed_docs]
-            results = self._reranker.rerank(query, docs)
-            for result in results:
-                original_idx = result["index"]
-                all_scores[indices[original_idx]] = result["relevance_score"]
+        with self._mlx_lock:
+            for query, indexed_docs in query_groups.items():
+                docs = [doc for _, doc in indexed_docs]
+                indices = [idx for idx, _ in indexed_docs]
+                results = self._reranker.rerank(query, docs)
+                for result in results:
+                    original_idx = result["index"]
+                    all_scores[indices[original_idx]] = result["relevance_score"]
 
         return all_scores
 
@@ -1234,6 +1336,164 @@ class JinaMLXCrossEncoder(CrossEncoderModel):
         return await loop.run_in_executor(None, self._predict_sync, pairs)
 
 
+class GoogleCrossEncoder(CrossEncoderModel):
+    """
+    Google Discovery Engine cross-encoder using the Ranking REST API.
+
+    Uses httpx + google-auth for lightweight REST calls (no gRPC/protobuf).
+    Supports ADC (Application Default Credentials) or service account key file.
+
+    Available models:
+    - semantic-ranker-default-004: Best quality, 1024 tokens/record (recommended)
+    - semantic-ranker-fast-004: Lower latency, 1024 tokens/record
+
+    Max 200 records per API request. Location is always "global".
+    """
+
+    MAX_RECORDS_PER_REQUEST = 200
+    API_BASE = "https://discoveryengine.googleapis.com/v1"
+    SCOPES = ["https://www.googleapis.com/auth/cloud-platform"]
+
+    def __init__(
+        self,
+        project_id: str,
+        model: str = DEFAULT_RERANKER_GOOGLE_MODEL,
+        service_account_key: str | None = None,
+        location: str = "global",
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize Google Discovery Engine cross-encoder.
+
+        Args:
+            project_id: Google Cloud project ID
+            model: Ranking model name (default: semantic-ranker-default-004)
+            service_account_key: Path to service account JSON key file.
+                                If None, uses Application Default Credentials (ADC).
+            location: API location (default: "global")
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        self.project_id = project_id
+        self.model = model
+        self.service_account_key = service_account_key
+        self.location = location
+        self.timeout = timeout
+        self._credentials = None
+        self._client: httpx.Client | None = None
+        self._rank_url: str | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "google"
+
+    def _get_auth_headers(self) -> dict[str, str]:
+        """Get Authorization header with a fresh access token."""
+        import google.auth.transport.requests
+
+        if not self._credentials.valid:
+            self._credentials.refresh(google.auth.transport.requests.Request())
+        return {"Authorization": f"Bearer {self._credentials.token}"}
+
+    async def initialize(self) -> None:
+        """Initialize credentials and HTTP client."""
+        if self._client is not None:
+            return
+
+        auth_method = "ADC" if not self.service_account_key else "service_account"
+        logger.info(
+            f"Reranker: initializing Google Discovery Engine provider "
+            f"(project={self.project_id}, model={self.model}, auth={auth_method})"
+        )
+        if self.service_account_key:
+            try:
+                from google.oauth2 import service_account
+            except ImportError:
+                raise ImportError(
+                    "google-auth is required for GoogleCrossEncoder. Install it with: pip install google-auth"
+                )
+            self._credentials = service_account.Credentials.from_service_account_file(
+                self.service_account_key,
+                scopes=self.SCOPES,
+            )
+        else:
+            try:
+                import google.auth
+            except ImportError:
+                raise ImportError(
+                    "google-auth is required for GoogleCrossEncoder. Install it with: pip install google-auth"
+                )
+            self._credentials, _ = google.auth.default(scopes=self.SCOPES)
+
+        ranking_config = f"projects/{self.project_id}/locations/{self.location}/rankingConfigs/default_ranking_config"
+        self._rank_url = f"{self.API_BASE}/{ranking_config}:rank"
+        self._client = httpx.Client(timeout=self.timeout)
+
+        logger.info("Reranker: Google Discovery Engine provider initialized")
+
+    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """Synchronous predict via REST API."""
+        if not pairs:
+            return []
+
+        # Group pairs by query
+        query_groups: dict[str, list[tuple[int, str]]] = {}
+        for idx, (query, text) in enumerate(pairs):
+            if query not in query_groups:
+                query_groups[query] = []
+            query_groups[query].append((idx, text))
+
+        all_scores = [0.0] * len(pairs)
+
+        for query, indexed_texts in query_groups.items():
+            texts = [text for _, text in indexed_texts]
+            indices = [idx for idx, _ in indexed_texts]
+
+            # Process in batches of MAX_RECORDS_PER_REQUEST
+            for batch_start in range(0, len(texts), self.MAX_RECORDS_PER_REQUEST):
+                batch_texts = texts[batch_start : batch_start + self.MAX_RECORDS_PER_REQUEST]
+                batch_indices = indices[batch_start : batch_start + self.MAX_RECORDS_PER_REQUEST]
+
+                records = [{"id": str(i), "content": text} for i, text in enumerate(batch_texts)]
+
+                response = self._client.post(
+                    self._rank_url,
+                    headers=self._get_auth_headers(),
+                    json={
+                        "model": self.model,
+                        "query": query,
+                        "records": records,
+                        "topN": len(records),
+                    },
+                )
+                response.raise_for_status()
+                result = response.json()
+
+                for record in result.get("records", []):
+                    local_idx = int(record["id"])
+                    all_scores[batch_indices[local_idx]] = record["score"]
+
+        return all_scores
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Score query-document pairs using Google Discovery Engine Ranking API.
+
+        Args:
+            pairs: List of (query, document) tuples to score
+
+        Returns:
+            List of relevance scores (0-1, higher = more relevant)
+        """
+        if self._client is None:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        if not pairs:
+            return []
+
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(None, self._predict_sync, pairs)
+
+
 def create_cross_encoder_from_env() -> CrossEncoderModel:
     """
     Create a CrossEncoderModel instance based on configuration.
@@ -1254,6 +1514,7 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
             raise ValueError(f"{ENV_RERANKER_TEI_URL} is required when {ENV_RERANKER_PROVIDER} is 'tei'")
         return RemoteTEICrossEncoder(
             base_url=url,
+            timeout=config.reranker_tei_http_timeout,
             batch_size=config.reranker_tei_batch_size,
             max_concurrent=config.reranker_tei_max_concurrent,
         )
@@ -1276,6 +1537,18 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
             model=config.reranker_cohere_model,
             base_url=config.reranker_cohere_base_url,
         )
+    elif provider == "openrouter":
+        api_key = config.reranker_openrouter_api_key
+        if not api_key:
+            raise ValueError(
+                "HINDSIGHT_API_RERANKER_OPENROUTER_API_KEY, HINDSIGHT_API_OPENROUTER_API_KEY, "
+                f"or HINDSIGHT_API_LLM_API_KEY is required when {ENV_RERANKER_PROVIDER} is 'openrouter'"
+            )
+        return CohereCrossEncoder(
+            api_key=api_key,
+            model=config.reranker_openrouter_model,
+            base_url="https://openrouter.ai/api/v1/rerank",
+        )
     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)
@@ -1309,11 +1582,34 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
             api_key=api_key,
             model=config.reranker_zeroentropy_model,
         )
+    elif provider == "siliconflow":
+        api_key = config.reranker_siliconflow_api_key
+        if not api_key:
+            raise ValueError(
+                f"{ENV_RERANKER_SILICONFLOW_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'siliconflow'"
+            )
+        return SiliconFlowCrossEncoder(
+            api_key=api_key,
+            model=config.reranker_siliconflow_model,
+            base_url=config.reranker_siliconflow_base_url,
+        )
+    elif provider == "google":
+        project_id = config.reranker_google_project_id
+        if not project_id:
+            raise ValueError(
+                f"{ENV_RERANKER_GOOGLE_PROJECT_ID} (or HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID) "
+                f"is required when {ENV_RERANKER_PROVIDER} is 'google'"
+            )
+        return GoogleCrossEncoder(
+            project_id=project_id,
+            model=config.reranker_google_model,
+            service_account_key=config.reranker_google_service_account_key,
+        )
     elif provider == "rrf":
         return RRFPassthroughCrossEncoder()
     elif provider == "jina-mlx":
         return JinaMLXCrossEncoder()
     else:
         raise ValueError(
-            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'zeroentropy', 'flashrank', 'litellm', 'litellm-sdk', 'rrf', 'jina-mlx'"
+            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'zeroentropy', 'siliconflow', 'google', 'flashrank', 'litellm', 'litellm-sdk', 'rrf', 'jina-mlx'"
         )

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

@@ -19,6 +19,7 @@ import httpx
 
 from ..config import (
     DEFAULT_EMBEDDINGS_COHERE_MODEL,
+    DEFAULT_EMBEDDINGS_GEMINI_MODEL,
     DEFAULT_EMBEDDINGS_LITELLM_MODEL,
     DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL,
     DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU,
@@ -28,6 +29,7 @@ from ..config import (
     DEFAULT_EMBEDDINGS_PROVIDER,
     DEFAULT_LITELLM_API_BASE,
     ENV_EMBEDDINGS_COHERE_API_KEY,
+    ENV_EMBEDDINGS_GEMINI_API_KEY,
     ENV_EMBEDDINGS_LITELLM_SDK_API_KEY,
     ENV_EMBEDDINGS_LOCAL_FORCE_CPU,
     ENV_EMBEDDINGS_LOCAL_MODEL,
@@ -752,8 +754,10 @@ class LiteLLMSDKEmbeddings(Embeddings):
         api_key: str,
         model: str = DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL,
         api_base: str | None = None,
+        output_dimensions: int | None = None,
         batch_size: int = 100,
         timeout: float = 60.0,
+        encoding_format: str | None = "float",
     ):
         """
         Initialize LiteLLM SDK embeddings client.
@@ -762,14 +766,19 @@ class LiteLLMSDKEmbeddings(Embeddings):
             api_key: API key for the embedding provider
             model: Model name with provider prefix (e.g., "cohere/embed-english-v3.0")
             api_base: Custom base URL for API (optional)
+            output_dimensions: Optional output embedding dimensions (provider-dependent)
             batch_size: Maximum batch size for embedding requests (default: 100)
             timeout: Request timeout in seconds (default: 60.0)
+            encoding_format: Encoding format for embeddings (default: "float").
+                Set to None or empty string to omit (needed for Voyage AI, Gemini).
         """
         self.api_key = api_key
         self.model = model
         self.api_base = api_base
+        self.output_dimensions = output_dimensions
         self.batch_size = batch_size
         self.timeout = timeout
+        self.encoding_format = encoding_format or None
         self._litellm = None  # Will be set during initialization
         self._dimension: int | None = None
 
@@ -805,10 +814,13 @@ class LiteLLMSDKEmbeddings(Embeddings):
                 "model": self.model,
                 "input": ["test"],
                 "api_key": self.api_key,
-                "encoding_format": "float",
             }
+            if self.encoding_format:
+                embed_kwargs["encoding_format"] = self.encoding_format
             if self.api_base:
                 embed_kwargs["api_base"] = self.api_base
+            if self.output_dimensions is not None:
+                embed_kwargs["dimensions"] = self.output_dimensions
 
             # Use async embedding method (standard in litellm)
             response = await self._litellm.aembedding(**embed_kwargs)
@@ -852,10 +864,13 @@ class LiteLLMSDKEmbeddings(Embeddings):
                     "model": self.model,
                     "input": batch,
                     "api_key": self.api_key,
-                    "encoding_format": "float",
                 }
+                if self.encoding_format:
+                    embed_kwargs["encoding_format"] = self.encoding_format
                 if self.api_base:
                     embed_kwargs["api_base"] = self.api_base
+                if self.output_dimensions is not None:
+                    embed_kwargs["dimensions"] = self.output_dimensions
 
                 # Use sync embedding (litellm doesn't have async in thread-safe way)
                 response = self._litellm.embedding(**embed_kwargs)
@@ -877,6 +892,179 @@ class LiteLLMSDKEmbeddings(Embeddings):
         return all_embeddings
 
 
+class GeminiEmbeddings(Embeddings):
+    """
+    Google embeddings via the google.genai SDK.
+
+    Supports both:
+    1. Gemini API (api.generativeai.google.com) with API key authentication
+    2. Vertex AI with service account or Application Default Credentials (ADC)
+
+    Uses the embed_content API: client.models.embed_content(model, contents)
+    """
+
+    def __init__(
+        self,
+        model: str = DEFAULT_EMBEDDINGS_GEMINI_MODEL,
+        api_key: str | None = None,
+        vertexai_project_id: str | None = None,
+        vertexai_region: str | None = None,
+        vertexai_service_account_key: str | None = None,
+        output_dimensionality: int | None = None,
+        batch_size: int = 100,
+    ):
+        self.model = model
+        self.api_key = api_key
+        self.vertexai_project_id = vertexai_project_id
+        self.vertexai_region = vertexai_region or "us-central1"
+        self.vertexai_service_account_key = vertexai_service_account_key
+        self.output_dimensionality = output_dimensionality
+        self.batch_size = batch_size
+        self._client = None
+        self._dimension: int | None = None
+        self._is_vertexai = vertexai_project_id is not None
+        self._embed_config = None  # EmbedContentConfig, built during initialize()
+
+    @property
+    def provider_name(self) -> str:
+        return "google"
+
+    @property
+    def dimension(self) -> int:
+        if self._dimension is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+        return self._dimension
+
+    async def initialize(self) -> None:
+        """Initialize the Google genai client and detect embedding dimension."""
+        if self._client is not None:
+            return
+
+        from google import genai
+        from google.genai import types as genai_types
+
+        if self._is_vertexai:
+            self._init_vertexai(genai)
+        else:
+            self._init_gemini(genai)
+
+        # Build EmbedContentConfig if output_dimensionality is set
+        if self.output_dimensionality is not None:
+            self._embed_config = genai_types.EmbedContentConfig(
+                output_dimensionality=self.output_dimensionality,
+            )
+
+        # Detect dimension via a test embedding (respects output_dimensionality)
+        embed_kwargs = {"model": self.model, "contents": ["test"]}
+        if self._embed_config is not None:
+            embed_kwargs["config"] = self._embed_config
+
+        result = self._client.models.embed_content(**embed_kwargs)  # type: ignore[union-attr]
+        if result.embeddings and len(result.embeddings) > 0:
+            self._dimension = len(result.embeddings[0].values)
+
+        auth_mode = "vertex_ai" if self._is_vertexai else "api_key"
+        logger.info(
+            f"Embeddings: google provider initialized (auth: {auth_mode}, model: {self.model}, dim: {self._dimension})"
+        )
+
+    def _init_gemini(self, genai) -> None:
+        """Initialize Gemini API client with API key."""
+        if not self.api_key:
+            raise ValueError("Gemini embeddings provider requires an API key")
+
+        self._client = genai.Client(api_key=self.api_key)
+        logger.info(f"Embeddings: initializing Gemini provider with model {self.model}")
+
+    def _init_vertexai(self, genai) -> None:
+        """Initialize Vertex AI client with project, region, and credentials."""
+        if not self.vertexai_project_id:
+            raise ValueError(
+                "HINDSIGHT_API_EMBEDDINGS_VERTEXAI_PROJECT_ID (or HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID) "
+                "is required for Vertex AI embeddings provider."
+            )
+
+        auth_method = "ADC"
+        credentials = None
+
+        if self.vertexai_service_account_key:
+            try:
+                from google.oauth2 import service_account
+            except ImportError:
+                raise ImportError(
+                    "Vertex AI service account auth requires 'google-auth' package. "
+                    "Install with: pip install google-auth"
+                )
+            credentials = service_account.Credentials.from_service_account_file(
+                self.vertexai_service_account_key,
+                scopes=["https://www.googleapis.com/auth/cloud-platform"],
+            )
+            auth_method = "service_account"
+            logger.info(f"Embeddings: Vertex AI using service account key: {self.vertexai_service_account_key}")
+
+        # Strip google/ prefix from model name — native SDK uses bare names
+        if self.model.startswith("google/"):
+            self.model = self.model[len("google/") :]
+
+        client_kwargs = {
+            "vertexai": True,
+            "project": self.vertexai_project_id,
+            "location": self.vertexai_region,
+        }
+        if credentials is not None:
+            client_kwargs["credentials"] = credentials
+
+        self._client = genai.Client(**client_kwargs)
+        logger.info(
+            f"Embeddings: initializing Vertex AI provider "
+            f"(project={self.vertexai_project_id}, region={self.vertexai_region}, "
+            f"model={self.model}, auth={auth_method})"
+        )
+
+    def encode(self, texts: list[str]) -> list[list[float]]:
+        """
+        Generate embeddings using the Google genai SDK.
+
+        Args:
+            texts: List of text strings to encode
+
+        Returns:
+            List of embedding vectors
+        """
+        if self._client is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+
+        if not texts:
+            return []
+
+        all_embeddings = []
+
+        # Process in batches
+        for i in range(0, len(texts), self.batch_size):
+            batch = texts[i : i + self.batch_size]
+
+            embed_kwargs = {"model": self.model, "contents": batch}
+            if self._embed_config is not None:
+                embed_kwargs["config"] = self._embed_config
+
+            result = self._client.models.embed_content(**embed_kwargs)
+
+            all_embeddings.extend([emb.values for emb in result.embeddings])
+
+        # L2-normalize when output_dimensionality is set — Gemini only returns
+        # normalized vectors at full 3072 dims; truncated dims need re-normalization
+        # for accurate cosine similarity.
+        if self.output_dimensionality is not None:
+            import numpy as np
+
+            arr = np.array(all_embeddings)
+            norms = np.linalg.norm(arr, axis=1, keepdims=True)
+            norms[norms == 0] = 1
+            all_embeddings = (arr / norms).tolist()
+
+        return all_embeddings
+
+
 def create_embeddings_from_env() -> Embeddings:
     """
     Create an Embeddings instance based on configuration.
@@ -912,7 +1100,25 @@ def create_embeddings_from_env() -> Embeddings:
             )
         model = os.environ.get(ENV_EMBEDDINGS_OPENAI_MODEL, DEFAULT_EMBEDDINGS_OPENAI_MODEL)
         base_url = os.environ.get(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None
-        return OpenAIEmbeddings(api_key=api_key, model=model, base_url=base_url)
+        return OpenAIEmbeddings(
+            api_key=api_key,
+            model=model,
+            base_url=base_url,
+            batch_size=config.embeddings_openai_batch_size,
+        )
+    elif provider == "openrouter":
+        api_key = config.embeddings_openrouter_api_key
+        if not api_key:
+            raise ValueError(
+                "HINDSIGHT_API_EMBEDDINGS_OPENROUTER_API_KEY, HINDSIGHT_API_OPENROUTER_API_KEY, "
+                f"or {ENV_LLM_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'openrouter'"
+            )
+        return OpenAIEmbeddings(
+            api_key=api_key,
+            model=config.embeddings_openrouter_model,
+            base_url="https://openrouter.ai/api/v1",
+            batch_size=config.embeddings_openai_batch_size,
+        )
     elif provider == "cohere":
         api_key = config.embeddings_cohere_api_key
         if not api_key:
@@ -938,9 +1144,30 @@ def create_embeddings_from_env() -> Embeddings:
             api_key=api_key,
             model=config.embeddings_litellm_sdk_model,
             api_base=config.embeddings_litellm_sdk_api_base,
+            output_dimensions=config.embeddings_litellm_sdk_output_dimensions,
+            encoding_format=config.embeddings_litellm_sdk_encoding_format,
+        )
+    elif provider == "google":
+        vertexai_project_id = config.embeddings_vertexai_project_id
+        if vertexai_project_id:
+            api_key = None  # Vertex AI uses ADC or service account
+        else:
+            api_key = config.embeddings_gemini_api_key
+            if not api_key:
+                raise ValueError(
+                    f"{ENV_EMBEDDINGS_GEMINI_API_KEY} or {ENV_LLM_API_KEY} is required "
+                    f"when {ENV_EMBEDDINGS_PROVIDER} is 'google' (set VERTEXAI_PROJECT_ID for Vertex AI auth instead)"
+                )
+        return GeminiEmbeddings(
+            model=config.embeddings_gemini_model,
+            api_key=api_key,
+            vertexai_project_id=vertexai_project_id,
+            vertexai_region=config.embeddings_vertexai_region,
+            vertexai_service_account_key=config.embeddings_vertexai_service_account_key,
+            output_dimensionality=config.embeddings_gemini_output_dimensionality,
         )
     else:
         raise ValueError(
             f"Unknown embeddings provider: {provider}. "
-            f"Supported: 'local', 'tei', 'openai', 'cohere', 'litellm', 'litellm-sdk'"
+            f"Supported: 'local', 'tei', 'openai', 'cohere', 'google', 'litellm', 'litellm-sdk'"
         )

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

@@ -317,8 +317,13 @@ class EntityResolver:
         entity_texts = list(set(e["text"] for e in entities_data))
 
         # Fetch candidates for all unique entity texts in a single batched query.
-        # The trigram % operator uses the GIN index; the substring conditions cover
-        # exact prefix/suffix matches that trigrams might miss at low similarity.
+        # Uses the GIN trigram index on LOWER(canonical_name) for case-insensitive
+        # similarity lookup. Previous version also had LIKE '%...' substring fallbacks,
+        # but those forced full sequential scans of the entities table and caused
+        # TimeoutErrors on banks with 10k+ entities. Lowering the similarity threshold
+        # to 0.15 (from default 0.3) catches most substring relationships while
+        # staying fully index-based.
+        await conn.execute("SET pg_trgm.similarity_threshold = 0.15")
         rows = await conn.fetch(
             f"""
             SELECT DISTINCT ON (e.id)
@@ -327,16 +332,13 @@ class EntityResolver:
             FROM unnest($2::text[]) AS q(query_text)
             JOIN {fq_table("entities")} e ON (
                 e.bank_id = $1
-                AND (
-                    e.canonical_name % q.query_text
-                    OR LOWER(e.canonical_name) LIKE '%' || LOWER(q.query_text) || '%'
-                    OR LOWER(q.query_text) LIKE '%' || LOWER(e.canonical_name) || '%'
-                )
+                AND LOWER(e.canonical_name) % LOWER(q.query_text)
             )
             """,
             bank_id,
             entity_texts,
         )
+        await conn.execute("RESET pg_trgm.similarity_threshold")
 
         # Group candidates by query_text
         all_candidates: dict[str, list] = {t: [] for t in entity_texts}
@@ -808,14 +810,19 @@ class EntityResolver:
             return await self._link_units_to_entities_batch_impl(conn, unit_entity_pairs)
 
     async def _link_units_to_entities_batch_impl(self, conn, unit_entity_pairs: list[tuple[str, str]]):
-        # Batch insert all unit-entity links
-        await conn.executemany(
+        # Sorted bulk insert to prevent deadlocks from inconsistent lock ordering
+        # across concurrent transactions on the unit_entities unique index.
+        sorted_pairs = sorted(unit_entity_pairs)
+        unit_ids = [p[0] for p in sorted_pairs]
+        entity_ids = [p[1] for p in sorted_pairs]
+        await conn.execute(
             f"""
             INSERT INTO {fq_table("unit_entities")} (unit_id, entity_id)
-            VALUES ($1, $2)
+            SELECT u, e FROM unnest($1::uuid[], $2::uuid[]) AS t(u, e)
             ON CONFLICT DO NOTHING
             """,
-            unit_entity_pairs,
+            unit_ids,
+            entity_ids,
         )
 
         # Build map of unit -> entities for co-occurrence calculation

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

@@ -240,6 +240,7 @@ class MemoryEngineInterface(ABC):
         bank_id: str,
         *,
         fact_type: str | None = None,
+        delete_bank_profile: bool = True,
         request_context: "RequestContext",
     ) -> dict[str, int]:
         """
@@ -248,6 +249,8 @@ class MemoryEngineInterface(ABC):
         Args:
             bank_id: The memory bank ID.
             fact_type: If specified, only delete memories of this type.
+            delete_bank_profile: If True, also delete the bank profile row itself.
+                If False, only delete memories/entities/documents but preserve the bank.
             request_context: Request context for authentication.
 
         Returns:

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

@@ -122,6 +122,7 @@ _PROVIDERS_WITHOUT_API_KEY = frozenset(
     {
         "ollama",
         "lmstudio",
+        "llamacpp",
         "openai-codex",
         "claude-code",
         "mock",
@@ -146,6 +147,7 @@ def create_llm_provider(
     reasoning_effort: str,
     groq_service_tier: str | None = None,
     openai_service_tier: str | None = None,
+    extra_body: dict[str, Any] | None = None,
     vertexai_project_id: str | None = None,
     vertexai_region: str | None = None,
     vertexai_credentials: Any = None,
@@ -162,6 +164,7 @@ def create_llm_provider(
         reasoning_effort: Reasoning effort level for supported providers.
         groq_service_tier: Groq service tier (for Groq provider) - "on_demand", "flex", or "auto".
         openai_service_tier: OpenAI service tier (for OpenAI provider) - None (default) or "flex" (50% cheaper).
+        extra_body: Extra body params merged into OpenAI-compatible API calls.
         vertexai_project_id: Vertex AI project ID (for VertexAI provider).
         vertexai_region: Vertex AI region (for VertexAI provider).
         vertexai_credentials: Vertex AI credentials object (for VertexAI provider).
@@ -176,6 +179,7 @@ def create_llm_provider(
         CodexLLM,
         GeminiLLM,
         LiteLLMLLM,
+        LlamaCppLLM,
         MockLLM,
         NoneLLM,
         OpenAICompatibleLLM,
@@ -261,7 +265,25 @@ def create_llm_provider(
             reasoning_effort=reasoning_effort,
         )
 
-    elif provider_lower in ("openai", "groq", "ollama", "lmstudio", "minimax", "volcano"):
+    elif provider_lower == "llamacpp":
+        from ..config import get_config
+
+        config = get_config()
+        return LlamaCppLLM(
+            provider=provider,
+            api_key=api_key,
+            base_url=base_url,
+            model=model,
+            reasoning_effort=reasoning_effort,
+            model_path=config.llamacpp_model_path,
+            gpu_layers=config.llamacpp_gpu_layers,
+            context_size=config.llamacpp_context_size,
+            chat_format=config.llamacpp_chat_format,
+            no_grammar=config.llamacpp_no_grammar,
+            extra_args=config.llamacpp_extra_args,
+        )
+
+    elif provider_lower in ("openai", "groq", "ollama", "lmstudio", "minimax", "volcano", "openrouter"):
         return OpenAICompatibleLLM(
             provider=provider,
             api_key=api_key,
@@ -270,6 +292,7 @@ def create_llm_provider(
             reasoning_effort=reasoning_effort,
             groq_service_tier=groq_service_tier,
             openai_service_tier=openai_service_tier,
+            extra_body=extra_body,
         )
 
     else:
@@ -293,6 +316,7 @@ class LLMProvider:
         groq_service_tier: str | None = None,
         openai_service_tier: str | None = None,
         gemini_safety_settings: list | None = None,
+        extra_body: dict[str, Any] | None = None,
     ):
         """
         Initialize LLM provider.
@@ -306,6 +330,7 @@ class LLMProvider:
             groq_service_tier: Groq service tier ("on_demand", "flex", "auto") - from config.
             openai_service_tier: OpenAI service tier (None or "flex") - from config.
             gemini_safety_settings: Safety settings for Gemini/VertexAI providers.
+            extra_body: Extra body params merged into OpenAI-compatible API calls.
         """
         self.provider = provider.lower()
         self.api_key = api_key
@@ -317,6 +342,8 @@ class LLMProvider:
         self.openai_service_tier = openai_service_tier
         # Gemini safety settings (instance default; can be overridden per-request via context var)
         self.gemini_safety_settings = gemini_safety_settings
+        # Extra body params for OpenAI-compatible providers (e.g. chat_template_kwargs)
+        self.extra_body = extra_body
 
         # Validate provider
         valid_providers = [
@@ -326,6 +353,7 @@ class LLMProvider:
             "gemini",
             "anthropic",
             "lmstudio",
+            "llamacpp",
             "vertexai",
             "openai-codex",
             "claude-code",
@@ -335,6 +363,7 @@ class LLMProvider:
             "litellm",
             "bedrock",
             "volcano",
+            "openrouter",
         ]
         if self.provider not in valid_providers:
             raise ValueError(f"Invalid LLM provider: {self.provider}. Must be one of: {', '.join(valid_providers)}")
@@ -349,6 +378,8 @@ class LLMProvider:
                 self.base_url = "http://localhost:1234/v1"
             elif self.provider == "minimax":
                 self.base_url = "https://api.minimax.io/v1"
+            elif self.provider == "openrouter":
+                self.base_url = "https://openrouter.ai/api/v1"
 
         # Prepare Vertex AI config (if applicable)
         vertexai_project_id = None
@@ -413,6 +444,7 @@ class LLMProvider:
             reasoning_effort=self.reasoning_effort,
             groq_service_tier=self.groq_service_tier,
             openai_service_tier=self.openai_service_tier,
+            extra_body=self.extra_body,
             vertexai_project_id=vertexai_project_id,
             vertexai_region=vertexai_region,
             vertexai_credentials=vertexai_credentials,
@@ -504,6 +536,15 @@ class LLMProvider:
             OutputTooLongError: If output exceeds token limits.
             Exception: Re-raises API errors after retries exhausted.
         """
+        # Stage breadcrumb so the worker log shows which LLM call a task is
+        # currently inside; the stage_age field then reveals long JSON-schema
+        # retry loops (e.g. a small model that can't satisfy strict_schema).
+        # No-op outside a worker context.
+        from ..worker.stage import set_stage
+
+        structured = "+structured" if response_format is not None else ""
+        set_stage(f"llm.{self.provider}.{scope}{structured}")
+
         async with _global_llm_semaphore:
             # Delegate to provider implementation
             result = await self._provider_impl.call(
@@ -560,6 +601,10 @@ class LLMProvider:
         Returns:
             LLMToolCallResult with content and/or tool_calls.
         """
+        from ..worker.stage import set_stage
+
+        set_stage(f"llm.{self.provider}.{scope}+tools")
+
         async with _global_llm_semaphore:
             # Delegate to provider implementation
             result = await self._provider_impl.call_with_tools(
@@ -703,69 +748,45 @@ class LLMProvider:
         return ConfiguredLLMProvider(self, config.llm_gemini_safety_settings)
 
     async def cleanup(self) -> None:
-        """Clean up resources."""
-        pass
+        """Clean up resources (e.g. stop llamacpp subprocess)."""
+        if self._provider_impl:
+            await self._provider_impl.cleanup()
 
     @classmethod
-    def for_memory(cls) -> "LLMProvider":
-        """Create provider for memory operations from environment variables."""
-        provider = os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq")
-        api_key = os.getenv("HINDSIGHT_API_LLM_API_KEY", "")
+    def from_env(cls) -> "LLMProvider":
+        """Create provider from environment variables using config.py constants."""
+        from ..config import (
+            DEFAULT_LLM_MODEL,
+            DEFAULT_LLM_PROVIDER,
+            ENV_LLM_API_KEY,
+            ENV_LLM_BASE_URL,
+            ENV_LLM_EXTRA_BODY,
+            ENV_LLM_MODEL,
+            ENV_LLM_PROVIDER,
+        )
+
+        provider = os.getenv(ENV_LLM_PROVIDER, DEFAULT_LLM_PROVIDER)
+        api_key = os.getenv(ENV_LLM_API_KEY, "")
 
-        # API key not needed for openai-codex (uses OAuth), claude-code (uses Keychain OAuth),
-        # ollama (local), vertexai (uses GCP service account credentials),
-        # or litellm (uses provider-specific auth, e.g. AWS credentials for Bedrock)
         if not api_key and not requires_api_key(provider):
             pass  # Provider handles its own auth
         elif not api_key:
             raise ValueError(
-                "HINDSIGHT_API_LLM_API_KEY environment variable is required (unless using openai-codex, claude-code, or litellm)"
+                f"{ENV_LLM_API_KEY} environment variable is required (unless using openai-codex, claude-code, or litellm)"
             )
 
-        base_url = os.getenv("HINDSIGHT_API_LLM_BASE_URL", "")
-        model = os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b")
+        base_url = os.getenv(ENV_LLM_BASE_URL, "")
+        model = os.getenv(ENV_LLM_MODEL, DEFAULT_LLM_MODEL)
+        extra_body = json.loads(os.getenv(ENV_LLM_EXTRA_BODY, "null"))
 
-        return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="low")
-
-    @classmethod
-    def for_answer_generation(cls) -> "LLMProvider":
-        """Create provider for answer generation. Falls back to memory config if not set."""
-        provider = os.getenv("HINDSIGHT_API_ANSWER_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
-        api_key = os.getenv("HINDSIGHT_API_ANSWER_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY", ""))
-
-        # API key not needed for providers with their own auth mechanisms
-        if not api_key and not requires_api_key(provider):
-            pass  # Provider handles its own auth
-        elif not api_key:
-            raise ValueError(
-                "HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_ANSWER_LLM_API_KEY environment variable is required "
-                "(unless using openai-codex, claude-code, or litellm)"
-            )
-
-        base_url = os.getenv("HINDSIGHT_API_ANSWER_LLM_BASE_URL", os.getenv("HINDSIGHT_API_LLM_BASE_URL", ""))
-        model = os.getenv("HINDSIGHT_API_ANSWER_LLM_MODEL", os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b"))
-
-        return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="high")
-
-    @classmethod
-    def for_judge(cls) -> "LLMProvider":
-        """Create provider for judge/evaluator operations. Falls back to memory config if not set."""
-        provider = os.getenv("HINDSIGHT_API_JUDGE_LLM_PROVIDER", os.getenv("HINDSIGHT_API_LLM_PROVIDER", "groq"))
-        api_key = os.getenv("HINDSIGHT_API_JUDGE_LLM_API_KEY", os.getenv("HINDSIGHT_API_LLM_API_KEY", ""))
-
-        # API key not needed for providers with their own auth mechanisms
-        if not api_key and not requires_api_key(provider):
-            pass  # Provider handles its own auth
-        elif not api_key:
-            raise ValueError(
-                "HINDSIGHT_API_LLM_API_KEY or HINDSIGHT_API_JUDGE_LLM_API_KEY environment variable is required "
-                "(unless using openai-codex, claude-code, or litellm)"
-            )
-
-        base_url = os.getenv("HINDSIGHT_API_JUDGE_LLM_BASE_URL", os.getenv("HINDSIGHT_API_LLM_BASE_URL", ""))
-        model = os.getenv("HINDSIGHT_API_JUDGE_LLM_MODEL", os.getenv("HINDSIGHT_API_LLM_MODEL", "openai/gpt-oss-120b"))
-
-        return cls(provider=provider, api_key=api_key, base_url=base_url, model=model, reasoning_effort="high")
+        return cls(
+            provider=provider,
+            api_key=api_key,
+            base_url=base_url,
+            model=model,
+            reasoning_effort="low",
+            extra_body=extra_body,
+        )
 
 
 class ConfiguredLLMProvider:

hindsight-api-slim/hindsight_api/engine/providers/__init__.py

@@ -9,6 +9,7 @@ from .claude_code_llm import ClaudeCodeLLM
 from .codex_llm import CodexLLM
 from .gemini_llm import GeminiLLM
 from .litellm_llm import LiteLLMLLM
+from .llamacpp_llm import LlamaCppLLM
 from .mock_llm import MockLLM
 from .none_llm import NoneLLM
 from .openai_compatible_llm import OpenAICompatibleLLM
@@ -18,6 +19,7 @@ __all__ = [
     "ClaudeCodeLLM",
     "CodexLLM",
     "GeminiLLM",
+    "LlamaCppLLM",
     "LiteLLMLLM",
     "MockLLM",
     "NoneLLM",

hindsight-api-slim/hindsight_api/engine/providers/anthropic_llm.py

@@ -153,7 +153,7 @@ class AnthropicLLM(LLMInterface):
         # 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)}"
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2, ensure_ascii=False)}"
             if system_prompt:
                 system_prompt += schema_msg
             else:

hindsight-api-slim/hindsight_api/engine/providers/claude_code_llm.py

@@ -171,7 +171,7 @@ class ClaudeCodeLLM(LLMInterface):
         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"
+                f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2, ensure_ascii=False)}\n\n"
                 "Respond with ONLY the JSON, no markdown formatting."
             )
             user_content += schema_instruction
@@ -455,9 +455,14 @@ class ClaudeCodeLLM(LLMInterface):
         # else: tool_choice == "auto" or unspecified - use default behavior (no changes needed)
 
         # Configure SDK options with MCP server
+        # tools=[] disables built-in CLI tools (Read, Write, Bash, ToolSearch, etc.)
+        # Without this, Claude Code CLI defers MCP tools when too many built-in tools
+        # are loaded, forcing Claude to use ToolSearch first — which wastes the max_turns
+        # budget and prevents direct MCP tool calls.
         options = ClaudeAgentOptions(
             system_prompt=system_prompt if system_prompt else None,
-            max_turns=1,  # Single-turn for API-style interactions
+            tools=[],  # Disable built-in tools so MCP tools load eagerly
+            max_turns=2,  # Allow tool call + tool result round-trip
             mcp_servers=mcp_servers_config,
             allowed_tools=allowed_tool_names,
         )

hindsight-api-slim/hindsight_api/engine/providers/codex_llm.py

@@ -205,7 +205,7 @@ class CodexLLM(LLMInterface):
         # 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)}"
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2, ensure_ascii=False)}"
             system_instruction += schema_msg
 
         # gpt-5.2-codex only supports "detailed" reasoning summary

hindsight-api-slim/hindsight_api/engine/providers/gemini_llm.py

@@ -23,6 +23,7 @@ from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
 from hindsight_api.engine.llm_wrapper import parse_llm_json
 from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
 from hindsight_api.metrics import get_metrics_collector
+from hindsight_api.worker.stage import set_stage
 
 logger = logging.getLogger(__name__)
 
@@ -174,7 +175,7 @@ class GeminiLLM(LLMInterface):
         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).
+            max_completion_tokens: Maximum tokens in response (mapped to Gemini's max_output_tokens).
             temperature: Sampling temperature (0.0-2.0).
             scope: Scope identifier for tracking.
             max_retries: Maximum retry attempts.
@@ -211,7 +212,7 @@ class GeminiLLM(LLMInterface):
         # 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)}"
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2, ensure_ascii=False)}"
             if system_instruction:
                 system_instruction += schema_msg
             else:
@@ -226,6 +227,11 @@ class GeminiLLM(LLMInterface):
             config_kwargs["response_schema"] = response_format
         if temperature is not None:
             config_kwargs["temperature"] = temperature
+        # Gemini's equivalent of OpenAI-style max_completion_tokens is max_output_tokens.
+        # Without it the model can produce arbitrarily long responses, ignoring the
+        # caller's intended cap (e.g. mental_models max_tokens during refresh).
+        if max_completion_tokens is not None:
+            config_kwargs["max_output_tokens"] = max_completion_tokens
 
         # Apply safety settings: context var (per-request bank override) takes precedence over instance default
         effective_safety_settings = _safety_settings_ctx.get()
@@ -242,6 +248,8 @@ class GeminiLLM(LLMInterface):
         last_exception = None
 
         for attempt in range(max_retries + 1):
+            if attempt > 0:
+                set_stage(f"llm.gemini.{scope}.attempt={attempt + 1}/{max_retries + 1}")
             try:
                 response = await asyncio.wait_for(
                     self._client.aio.models.generate_content(
@@ -398,7 +406,7 @@ class GeminiLLM(LLMInterface):
         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).
+            max_completion_tokens: Maximum tokens (mapped to Gemini's max_output_tokens).
             temperature: Sampling temperature.
             scope: Scope identifier for tracking.
             max_retries: Maximum retry attempts.
@@ -490,6 +498,10 @@ class GeminiLLM(LLMInterface):
             config_kwargs["system_instruction"] = system_instruction
         if temperature is not None:
             config_kwargs["temperature"] = temperature
+        # See note in `call`: Gemini's max_output_tokens is the equivalent of
+        # OpenAI-style max_completion_tokens.
+        if max_completion_tokens is not None:
+            config_kwargs["max_output_tokens"] = max_completion_tokens
 
         # Map OpenAI-style tool_choice to Gemini FunctionCallingConfig
         if tool_choice == "required":
@@ -527,6 +539,8 @@ class GeminiLLM(LLMInterface):
 
         last_exception = None
         for attempt in range(max_retries + 1):
+            if attempt > 0:
+                set_stage(f"llm.gemini.tools.attempt={attempt + 1}/{max_retries + 1}")
             try:
                 response = await asyncio.wait_for(
                     self._client.aio.models.generate_content(

hindsight-api-slim/hindsight_api/engine/providers/litellm_llm.py

@@ -21,6 +21,7 @@ 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
+from hindsight_api.worker.stage import set_stage
 
 logger = logging.getLogger(__name__)
 
@@ -141,6 +142,8 @@ class LiteLLMLLM(LLMInterface):
         last_exception = None
 
         for attempt in range(max_retries + 1):
+            if attempt > 0:
+                set_stage(f"llm.litellm.{scope}.attempt={attempt + 1}/{max_retries + 1}")
             try:
                 response = await self._litellm.acompletion(**call_kwargs)
 
@@ -283,6 +286,8 @@ class LiteLLMLLM(LLMInterface):
 
         last_exception = None
         for attempt in range(max_retries + 1):
+            if attempt > 0:
+                set_stage(f"llm.litellm.tools.attempt={attempt + 1}/{max_retries + 1}")
             try:
                 response = await self._litellm.acompletion(**call_kwargs)
 

hindsight-api-slim/hindsight_api/engine/providers/llamacpp_llm.py

@@ -0,0 +1,428 @@
+"""
+Built-in llama.cpp LLM provider for fully offline operation.
+
+Manages a llama-cpp-python server as a subprocess, downloads GGUF models
+from HuggingFace on first use, and delegates inference to the OpenAI-compatible API.
+
+Usage:
+    HINDSIGHT_API_LLM_PROVIDER=llamacpp
+    HINDSIGHT_API_LLAMACPP_MODEL_PATH=~/.hindsight/models/gemma-4-E2B-it-Q4_K_M.gguf
+    HINDSIGHT_API_LLAMACPP_GPU_LAYERS=-1  # -1 = all layers on GPU
+    HINDSIGHT_API_LLAMACPP_CONTEXT_SIZE=8192
+"""
+
+import asyncio
+import logging
+import os
+import signal
+import socket
+import subprocess
+import sys
+import time
+from pathlib import Path
+from typing import Any
+
+from hindsight_api.engine.llm_interface import LLMInterface
+from hindsight_api.engine.response_models import LLMToolCallResult
+
+logger = logging.getLogger(__name__)
+
+# Default GGUF model for offline mode
+DEFAULT_LLAMACPP_HF_REPO = "bartowski/google_gemma-4-E2B-it-GGUF"
+DEFAULT_LLAMACPP_HF_FILENAME = "google_gemma-4-E2B-it-Q4_K_M.gguf"
+DEFAULT_LLAMACPP_MODEL_ALIAS = "gemma-4-e2b-it"
+
+MODELS_DIR = Path.home() / ".hindsight" / "models"
+
+# Singleton server instance — shared across all LlamaCppLLM instances
+# (retain, reflect, consolidation each create their own LLMProvider,
+# but they should all share one llama.cpp server process)
+_shared_server: "LlamaCppServer | None" = None
+_shared_server_lock = asyncio.Lock()
+
+
+def _find_free_port() -> int:
+    """Find a free TCP port on localhost."""
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(("127.0.0.1", 0))
+        return s.getsockname()[1]
+
+
+def _download_default_model() -> Path:
+    """Download the default GGUF model from HuggingFace if not already cached.
+
+    Returns:
+        Path to the downloaded GGUF file.
+    """
+    try:
+        from huggingface_hub import hf_hub_download
+    except ImportError:
+        raise ImportError(
+            "huggingface-hub is required for automatic model download. "
+            "Install with: pip install 'hindsight-api-slim[local-llm]'"
+        )
+
+    MODELS_DIR.mkdir(parents=True, exist_ok=True)
+    target = MODELS_DIR / DEFAULT_LLAMACPP_HF_FILENAME
+
+    if target.exists():
+        logger.info(f"Using cached model: {target}")
+        return target
+
+    logger.info(
+        f"Downloading {DEFAULT_LLAMACPP_HF_FILENAME} from {DEFAULT_LLAMACPP_HF_REPO} (~3.5 GB, first run only)..."
+    )
+
+    downloaded = hf_hub_download(
+        repo_id=DEFAULT_LLAMACPP_HF_REPO,
+        filename=DEFAULT_LLAMACPP_HF_FILENAME,
+        local_dir=str(MODELS_DIR),
+    )
+
+    logger.info(f"Model downloaded: {downloaded}")
+    return Path(downloaded)
+
+
+def _resolve_model_path(model_path: str | None) -> Path:
+    """Resolve the model path, downloading the default if needed.
+
+    Args:
+        model_path: Explicit path to a GGUF file, or None to use the default.
+
+    Returns:
+        Resolved Path to the GGUF file.
+    """
+    if model_path:
+        p = Path(model_path).expanduser()
+        if not p.exists():
+            raise FileNotFoundError(
+                f"GGUF model not found: {p}\n"
+                f"Set HINDSIGHT_API_LLAMACPP_MODEL_PATH to a valid .gguf file, "
+                f"or remove the setting to auto-download the default model."
+            )
+        return p
+
+    return _download_default_model()
+
+
+class LlamaCppServer:
+    """Manages a llama-cpp-python OpenAI-compatible server as a subprocess."""
+
+    def __init__(
+        self,
+        model_path: Path,
+        port: int,
+        gpu_layers: int = -1,
+        context_size: int = 8192,
+        chat_format: str | None = None,
+        extra_args: str | None = None,
+    ):
+        self.model_path = model_path
+        self.port = port
+        self.gpu_layers = gpu_layers
+        self.context_size = context_size
+        self.chat_format = chat_format
+        self.extra_args = extra_args
+        self._process: subprocess.Popen | None = None
+
+    @property
+    def base_url(self) -> str:
+        return f"http://127.0.0.1:{self.port}/v1"
+
+    async def start(self) -> None:
+        """Start the llama.cpp server subprocess."""
+        cmd = [
+            sys.executable,
+            "-m",
+            "llama_cpp.server",
+            "--model",
+            str(self.model_path),
+            "--host",
+            "127.0.0.1",
+            "--port",
+            str(self.port),
+            "--n_gpu_layers",
+            str(self.gpu_layers),
+            "--n_ctx",
+            str(self.context_size),
+            "--flash_attn",
+            "true",
+            "--n_batch",
+            "2048",
+            # Prompt cache: reuse KV cache for repeated system prompts
+            "--cache",
+            "true",
+        ]
+        # Only pass chat_format if explicitly set (most GGUF models have it embedded)
+        if self.chat_format:
+            cmd.extend(["--chat_format", self.chat_format])
+        # User-provided extra args (e.g. "--type_k 1 --type_v 1 --n_threads 8")
+        if self.extra_args:
+            cmd.extend(self.extra_args.split())
+
+        logger.info(f"Starting llama.cpp server: {' '.join(cmd)}")
+
+        # Write stderr to a log file to avoid pipe buffer deadlock
+        # (llama.cpp outputs a lot of model metadata on stderr during loading)
+        self._log_path = MODELS_DIR / "llamacpp_server.log"
+        self._log_file = open(self._log_path, "w")
+
+        self._process = subprocess.Popen(
+            cmd,
+            stdout=subprocess.DEVNULL,
+            stderr=self._log_file,
+            # Ensure the subprocess is killed when the parent exits
+            preexec_fn=os.setsid if hasattr(os, "setsid") else None,
+        )
+
+        # Wait for the server to be ready
+        await self._wait_for_ready()
+
+    async def _wait_for_ready(self, timeout: float = 120.0) -> None:
+        """Wait for the llama.cpp server to accept connections."""
+        import httpx
+
+        start = time.monotonic()
+        url = f"http://127.0.0.1:{self.port}/v1/models"
+        last_log = start
+
+        while time.monotonic() - start < timeout:
+            # Check if process died
+            if self._process and self._process.poll() is not None:
+                stderr = ""
+                try:
+                    stderr = self._log_path.read_text()[-2000:]
+                except Exception:
+                    pass
+                raise RuntimeError(f"llama.cpp server exited with code {self._process.returncode}.\nstderr: {stderr}")
+
+            try:
+                async with httpx.AsyncClient() as client:
+                    resp = await client.get(url, timeout=5.0)
+                    if resp.status_code == 200:
+                        logger.info(f"llama.cpp server ready on port {self.port}")
+                        return
+            except (httpx.ConnectError, httpx.TimeoutException, httpx.ConnectTimeout):
+                pass
+
+            # Log progress every 15s
+            now = time.monotonic()
+            if now - last_log > 15:
+                elapsed = int(now - start)
+                logger.info(f"Waiting for llama.cpp server to load model... ({elapsed}s)")
+                last_log = now
+
+            await asyncio.sleep(1.0)
+
+        # Timeout — read the log to help debug
+        stderr = ""
+        try:
+            stderr = self._log_path.read_text()[-2000:]
+        except Exception:
+            pass
+        raise TimeoutError(
+            f"llama.cpp server did not become ready within {timeout}s.\n"
+            f"Check model compatibility and available memory.\n"
+            f"Server log: {stderr}"
+        )
+
+    async def stop(self) -> None:
+        """Stop the llama.cpp server subprocess."""
+        if self._process is None:
+            return
+
+        logger.info("Stopping llama.cpp server...")
+        try:
+            # Send SIGTERM to the process group
+            if hasattr(os, "killpg"):
+                os.killpg(os.getpgid(self._process.pid), signal.SIGTERM)
+            else:
+                self._process.terminate()
+
+            # Wait up to 10s for graceful shutdown
+            try:
+                self._process.wait(timeout=10)
+            except subprocess.TimeoutExpired:
+                if hasattr(os, "killpg"):
+                    os.killpg(os.getpgid(self._process.pid), signal.SIGKILL)
+                else:
+                    self._process.kill()
+                self._process.wait(timeout=5)
+        except (ProcessLookupError, OSError):
+            pass  # Process already exited
+        finally:
+            self._process = None
+            if hasattr(self, "_log_file") and self._log_file:
+                self._log_file.close()
+                self._log_file = None
+            logger.info("llama.cpp server stopped")
+
+
+class LlamaCppLLM(LLMInterface):
+    """
+    Built-in llama.cpp provider.
+
+    Manages a llama-cpp-python server subprocess and delegates to OpenAICompatibleLLM
+    for actual inference calls. Handles model downloading and server lifecycle.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        model_path: str | None = None,
+        gpu_layers: int = -1,
+        context_size: int = 8192,
+        chat_format: str | None = None,
+        no_grammar: bool = False,
+        extra_args: str | None = None,
+        **kwargs: Any,
+    ):
+        super().__init__(
+            provider=provider,
+            api_key=api_key or "llamacpp",
+            base_url=base_url or "",
+            model=model or DEFAULT_LLAMACPP_MODEL_ALIAS,
+            reasoning_effort=reasoning_effort,
+        )
+        self._model_path_str = model_path
+        self._gpu_layers = gpu_layers
+        self._context_size = context_size
+        self._chat_format = chat_format
+        self._no_grammar = no_grammar
+        self._extra_args = extra_args
+        self._server: LlamaCppServer | None = None
+        self._delegate: Any = None  # OpenAICompatibleLLM, created after server starts
+        self._initialized = False
+
+    async def _ensure_initialized(self) -> None:
+        """Lazy initialization: download model + start shared server on first use."""
+        if self._initialized:
+            return
+
+        global _shared_server
+
+        from .openai_compatible_llm import OpenAICompatibleLLM
+
+        async with _shared_server_lock:
+            if _shared_server is None:
+                # Resolve and potentially download the model
+                model_path = _resolve_model_path(self._model_path_str)
+                logger.info(f"Using GGUF model: {model_path}")
+
+                # Start the shared llama.cpp server
+                port = _find_free_port()
+                _shared_server = LlamaCppServer(
+                    model_path=model_path,
+                    port=port,
+                    gpu_layers=self._gpu_layers,
+                    context_size=self._context_size,
+                    chat_format=self._chat_format,
+                    extra_args=self._extra_args,
+                )
+                await _shared_server.start()
+
+        self._server = _shared_server
+
+        # Create the delegate that talks to the shared server's OpenAI-compatible API
+        if self._no_grammar:
+            logger.info("Grammar enforcement disabled (HINDSIGHT_API_LLAMACPP_NO_GRAMMAR=true)")
+        self._delegate = OpenAICompatibleLLM(
+            provider="llamacpp",
+            api_key="llamacpp",
+            base_url=self._server.base_url,
+            model=self.model,
+            reasoning_effort=self.reasoning_effort,
+        )
+
+        self._initialized = True
+
+    async def verify_connection(self) -> None:
+        """Verify the llama.cpp server is running and can generate text."""
+        await self._ensure_initialized()
+        # Make a simple test call to verify the model can actually generate
+        await self._delegate.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("llama.cpp LLM verification passed")
+
+    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:
+        """Delegate call to the OpenAI-compatible API."""
+        await self._ensure_initialized()
+        return await self._delegate.call(
+            messages=messages,
+            response_format=response_format,
+            max_completion_tokens=max_completion_tokens,
+            temperature=temperature,
+            scope=scope,
+            max_retries=max_retries,
+            initial_backoff=initial_backoff,
+            max_backoff=max_backoff,
+            skip_validation=skip_validation,
+            strict_schema=strict_schema,
+            return_usage=return_usage,
+        )
+
+    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:
+        """Delegate tool calls to the OpenAI-compatible API."""
+        await self._ensure_initialized()
+        return await self._delegate.call_with_tools(
+            messages=messages,
+            tools=tools,
+            max_completion_tokens=max_completion_tokens,
+            temperature=temperature,
+            scope=scope,
+            max_retries=max_retries,
+            initial_backoff=initial_backoff,
+            max_backoff=max_backoff,
+            tool_choice=tool_choice,
+        )
+
+    async def cleanup(self) -> None:
+        """Stop the shared llama.cpp server."""
+        global _shared_server
+
+        if self._delegate:
+            await self._delegate.cleanup()
+            self._delegate = None
+
+        # Stop the shared server (only the first cleanup call actually stops it)
+        async with _shared_server_lock:
+            if _shared_server is not None:
+                await _shared_server.stop()
+                _shared_server = None
+
+        self._server = None
+        self._initialized = False

hindsight-api-slim/hindsight_api/engine/providers/openai_compatible_llm.py

@@ -33,6 +33,7 @@ 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
+from hindsight_api.worker.stage import set_stage
 
 logger = logging.getLogger(__name__)
 
@@ -59,6 +60,32 @@ def _strip_code_fences(content: str) -> str:
         return content
 
 
+def _summarize_status_error(e: APIStatusError, body_max: int = 400) -> str:
+    """Render an APIStatusError with status code + truncated response body.
+
+    Without this, retry loops only log "API error after N attempts" with the
+    bare exception message — losing the provider's actual error payload, which
+    is the only thing that explains *why* a request failed (rate limit reason,
+    invalid tool schema, model overloaded, etc.).
+    """
+    body: Any = getattr(e, "body", None)
+    if body is None:
+        try:
+            body = e.response.text
+        except Exception:
+            body = None
+    if isinstance(body, (dict, list)):
+        try:
+            body_str = json.dumps(body, default=str, ensure_ascii=False)
+        except Exception:
+            body_str = str(body)
+    else:
+        body_str = str(body or "").strip()
+    if len(body_str) > body_max:
+        body_str = body_str[:body_max] + "...TRUNCATED"
+    return f"HTTP {e.status_code}: {body_str or '<no body>'}"
+
+
 class OpenAICompatibleLLM(LLMInterface):
     """
     LLM provider for OpenAI-compatible APIs.
@@ -80,6 +107,7 @@ class OpenAICompatibleLLM(LLMInterface):
         reasoning_effort: str = "low",
         timeout: float | None = None,
         groq_service_tier: str | None = None,
+        extra_body: dict[str, Any] | None = None,
         **kwargs: Any,
     ):
         """
@@ -93,12 +121,13 @@ class OpenAICompatibleLLM(LLMInterface):
             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").
+            extra_body: Extra body params merged into every API call.
             **kwargs: Additional provider-specific parameters.
         """
         super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
 
         # Validate provider
-        valid_providers = ["openai", "groq", "ollama", "lmstudio", "minimax", "volcano"]
+        valid_providers = ["openai", "groq", "ollama", "lmstudio", "llamacpp", "minimax", "volcano", "openrouter"]
         if self.provider not in valid_providers:
             raise ValueError(f"OpenAICompatibleLLM only supports: {', '.join(valid_providers)}. Got: {self.provider}")
 
@@ -112,18 +141,22 @@ class OpenAICompatibleLLM(LLMInterface):
                 self.base_url = "http://localhost:1234/v1"
             elif self.provider == "minimax":
                 self.base_url = "https://api.minimax.io/v1"
+            elif self.provider == "openrouter":
+                self.base_url = "https://openrouter.ai/api/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", "minimax") and not self.api_key:
+        if self.provider in ("openai", "groq", "minimax", "openrouter") and not self.api_key:
             raise ValueError(f"API key is required for {self.provider}")
 
         # Service tier configuration (from config, not env vars)
         self.groq_service_tier = groq_service_tier
         self.openai_service_tier = kwargs.get("openai_service_tier")
+        # User-configured extra body params (merged into every API call)
+        self._config_extra_body = extra_body or {}
 
         # Get timeout config
         self.timeout = timeout or float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT)))
@@ -187,6 +220,36 @@ class OpenAICompatibleLLM(LLMInterface):
 
         return None
 
+    def _max_tokens_param_name(self) -> str:
+        """Return the correct parameter name for limiting response tokens.
+
+        Native OpenAI, Azure OpenAI, Groq, and llamacpp accept 'max_completion_tokens'.
+        Mistral and other OpenAI-compatible endpoints that haven't adopted the newer
+        parameter name require 'max_tokens', so when the openai provider is configured
+        with a non-Azure custom base_url we fall back to the widely-supported
+        'max_tokens'.
+
+        Reasoning models (GPT-5, o1, o3) only accept 'max_completion_tokens' and reject
+        'max_tokens' outright, so they always use the new parameter name regardless of
+        base_url.
+        """
+        # Reasoning models (GPT-5, o1, o3, ...) only accept max_completion_tokens.
+        # Azure OpenAI + GPT-5 is the canonical example: issue #978.
+        if self._supports_reasoning_model():
+            return "max_completion_tokens"
+        # Native OpenAI (no custom base URL), Groq, and llamacpp use max_completion_tokens
+        if self.provider in ("groq", "llamacpp"):
+            return "max_completion_tokens"
+        if self.provider == "openai" and not self.base_url:
+            return "max_completion_tokens"
+        # Azure OpenAI is fully OpenAI-API-compatible — detect it by hostname so users
+        # can keep provider=openai + an Azure base_url (the documented setup).
+        if self.provider == "openai" and self.base_url and ".openai.azure.com" in self.base_url:
+            return "max_completion_tokens"
+        # openai with custom base_url, ollama, lmstudio, minimax, volcano —
+        # use the widely-supported max_tokens
+        return "max_tokens"
+
     async def call(
         self,
         messages: list[dict[str, str]],
@@ -259,9 +322,7 @@ class OpenAICompatibleLLM(LLMInterface):
             # 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
+            call_params[self._max_tokens_param_name()] = max_completion_tokens
         if temperature is not None and not is_reasoning_model:
             # MiniMax requires temperature in (0.0, 1.0] — clamp accordingly
             if self.provider == "minimax":
@@ -273,17 +334,17 @@ class OpenAICompatibleLLM(LLMInterface):
             call_params["reasoning_effort"] = self.reasoning_effort
 
         # Provider-specific parameters
+        extra_body: dict[str, Any] = {**self._config_extra_body}
         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
+        if extra_body:
+            call_params["extra_body"] = extra_body
 
         # Prepare response format ONCE before retry loop
         if response_format is not None:
@@ -304,9 +365,7 @@ class OpenAICompatibleLLM(LLMInterface):
             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)}"
-                    )
+                    schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2, ensure_ascii=False)}"
 
                     if call_params["messages"] and call_params["messages"][0].get("role") == "system":
                         first_msg = call_params["messages"][0]
@@ -316,13 +375,23 @@ class OpenAICompatibleLLM(LLMInterface):
                         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", "volcano"):
-                    # LM Studio, Ollama and Volcano don't support json_object response format reliably
+                # Providers that skip json_object grammar enforcement
+                skip_grammar = self.provider in ("lmstudio", "ollama", "volcano")
+                if self.provider == "llamacpp":
+                    from hindsight_api.config import get_config
+
+                    skip_grammar = get_config().llamacpp_no_grammar
+                if not skip_grammar:
                     call_params["response_format"] = {"type": "json_object"}
 
         last_exception = None
 
         for attempt in range(max_retries + 1):
+            # Surface attempt count in worker stage so JSON-schema retry loops
+            # are visible from logs (small models on strict structured output
+            # often loop here). Cheap no-op outside worker context.
+            if attempt > 0:
+                set_stage(f"llm.{self.provider}.{scope}.attempt={attempt + 1}/{max_retries + 1}")
             try:
                 if response_format is not None:
                     response = await self._client.chat.completions.create(**call_params)
@@ -505,12 +574,19 @@ class OpenAICompatibleLLM(LLMInterface):
 
                 last_exception = e
                 if attempt < max_retries:
+                    logger.warning(
+                        f"APIStatusError ({self.provider}/{self.model}, scope={scope}, "
+                        f"attempt {attempt + 1}/{max_retries + 1}): {_summarize_status_error(e)}"
+                    )
                     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)}")
+                    logger.error(
+                        f"API error after {max_retries + 1} attempts ({self.provider}/{self.model}, "
+                        f"scope={scope}): {_summarize_status_error(e)}"
+                    )
                     raise
 
             except Exception:
@@ -573,7 +649,7 @@ class OpenAICompatibleLLM(LLMInterface):
         }
 
         if max_completion_tokens is not None:
-            call_params["max_completion_tokens"] = max_completion_tokens
+            call_params[self._max_tokens_param_name()] = max_completion_tokens
         if temperature is not None:
             # MiniMax requires temperature in (0.0, 1.0] — clamp accordingly
             if self.provider == "minimax":
@@ -581,12 +657,17 @@ class OpenAICompatibleLLM(LLMInterface):
             call_params["temperature"] = temperature
 
         # Provider-specific parameters
+        extra_body: dict[str, Any] = {**self._config_extra_body}
         if self.provider == "groq":
             call_params["seed"] = DEFAULT_LLM_SEED
+        if extra_body:
+            call_params["extra_body"] = extra_body
 
         last_exception = None
 
         for attempt in range(max_retries + 1):
+            if attempt > 0:
+                set_stage(f"llm.{self.provider}.tools.attempt={attempt + 1}/{max_retries + 1}")
             try:
                 response = await self._client.chat.completions.create(**call_params)
 
@@ -656,18 +737,41 @@ class OpenAICompatibleLLM(LLMInterface):
 
             except APIConnectionError as e:
                 last_exception = e
+                status_code = getattr(e, "status_code", None) or getattr(
+                    getattr(e, "response", None), "status_code", None
+                )
                 if attempt < max_retries:
+                    logger.warning(
+                        f"APIConnectionError in tool call ({self.provider}/{self.model}, scope={scope}, "
+                        f"attempt {attempt + 1}/{max_retries + 1}, HTTP {status_code}): {str(e)[:200]}"
+                    )
                     await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
                     continue
+                logger.error(
+                    f"Connection error in tool call after {max_retries + 1} attempts "
+                    f"({self.provider}/{self.model}, scope={scope}): {str(e)}"
+                )
                 raise
 
             except APIStatusError as e:
                 if e.status_code in (401, 403):
+                    logger.error(
+                        f"Auth error in tool call (HTTP {e.status_code}, {self.provider}/{self.model}), "
+                        f"not retrying: {_summarize_status_error(e)}"
+                    )
                     raise
                 last_exception = e
                 if attempt < max_retries:
+                    logger.warning(
+                        f"APIStatusError in tool call ({self.provider}/{self.model}, scope={scope}, "
+                        f"attempt {attempt + 1}/{max_retries + 1}): {_summarize_status_error(e)}"
+                    )
                     await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
                     continue
+                logger.error(
+                    f"API error in tool call after {max_retries + 1} attempts "
+                    f"({self.provider}/{self.model}, scope={scope}): {_summarize_status_error(e)}"
+                )
                 raise
 
             except Exception:
@@ -715,6 +819,7 @@ class OpenAICompatibleLLM(LLMInterface):
             "model": self.model,
             "messages": messages,
             "stream": False,
+            "think": False,  # Disable thinking for reasoning models (qwen3.5, etc.)
         }
 
         # Add schema as format parameter for structured output
@@ -736,6 +841,8 @@ class OpenAICompatibleLLM(LLMInterface):
 
         async with httpx.AsyncClient(timeout=300.0) as client:
             for attempt in range(max_retries + 1):
+                if attempt > 0:
+                    set_stage(f"llm.ollama_native.{scope}.attempt={attempt + 1}/{max_retries + 1}")
                 try:
                     response = await client.post(native_url, json=payload)
                     response.raise_for_status()
@@ -867,7 +974,7 @@ class OpenAICompatibleLLM(LLMInterface):
         logger.info(f"Submitting batch with {len(requests)} requests to {self.provider}")
 
         # Format requests as JSONL
-        jsonl_content = "\n".join(json.dumps(req) for req in requests)
+        jsonl_content = "\n".join(json.dumps(req, ensure_ascii=False) for req in requests)
 
         # Upload file to provider (wrap in BytesIO with filename)
         file_bytes = io.BytesIO(jsonl_content.encode("utf-8"))

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

@@ -137,7 +137,21 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
             "RETURN_AS_TIMEZONE_AWARE": False,
         }
 
-        results = self._search_dates(query, settings=settings)
+        # Wrap dateparser in a defensive try/except. dateparser has been
+        # observed to crash with internal errors (e.g., IndexError from
+        # locale.translate_search) on certain query inputs. A parser bug
+        # should not bring down the whole search/consolidation pipeline —
+        # treat any failure as "no temporal constraint found" so the caller
+        # can fall back to non-temporal retrieval.
+        try:
+            results = self._search_dates(query, settings=settings)
+        except Exception as e:
+            logger.warning(
+                "dateparser raised %s on query (treating as no temporal constraint): %s",
+                type(e).__name__,
+                e,
+            )
+            return QueryAnalysis(temporal_constraint=None)
 
         if not results:
             return QueryAnalysis(temporal_constraint=None)

hindsight-api-slim/hindsight_api/engine/reflect/agent.py

@@ -17,7 +17,12 @@ from typing import TYPE_CHECKING, Any, Awaitable, Callable
 import tiktoken
 
 from .models import DirectiveInfo, LLMCall, ReflectAgentResult, TokenUsageSummary, ToolCall
-from .prompts import FINAL_SYSTEM_PROMPT, _extract_directive_rules, build_final_prompt, build_system_prompt_for_tools
+from .prompts import (
+    _extract_directive_rules,
+    build_final_prompt,
+    build_final_system_prompt,
+    build_system_prompt_for_tools,
+)
 from .tools_schema import get_reflect_tools
 
 
@@ -186,7 +191,7 @@ async def _generate_structured_output(
         DynamicModel = create_model("StructuredResponse", **fields)
 
         # Include the full schema in the prompt for better LLM guidance
-        schema_str = json.dumps(response_schema, indent=2)
+        schema_str = json.dumps(response_schema, indent=2, ensure_ascii=False)
 
         # Build field descriptions for the prompt
         field_descriptions = []
@@ -446,7 +451,7 @@ async def run_reflect_agent(
             llm_start = time.time()
             response, usage = await llm_config.call(
                 messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "system", "content": build_final_system_prompt(bank_profile.get("mission"))},
                     {"role": "user", "content": prompt},
                 ],
                 scope="reflect",
@@ -503,7 +508,7 @@ async def run_reflect_agent(
             llm_start = time.time()
             response, usage = await llm_config.call(
                 messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "system", "content": build_final_system_prompt(bank_profile.get("mission"))},
                     {"role": "user", "content": prompt},
                 ],
                 scope="reflect",
@@ -606,7 +611,7 @@ async def run_reflect_agent(
             llm_start = time.time()
             response, usage = await llm_config.call(
                 messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "system", "content": build_final_system_prompt(bank_profile.get("mission"))},
                     {"role": "user", "content": prompt},
                 ],
                 scope="reflect",
@@ -649,9 +654,57 @@ async def run_reflect_agent(
 
         # No tool calls - LLM wants to respond with text
         if not result.tool_calls:
-            if result.content:
+            # When directives are present but no evidence has been gathered,
+            # the LLM tends to echo directive content verbatim as its answer.
+            # Fall through to the final-prompt path which doesn't include
+            # directives and handles "no data" gracefully.
+            has_gathered_evidence = (
+                bool(available_memory_ids) or bool(available_mental_model_ids) or bool(available_observation_ids)
+            )
+            directive_leak_risk = directives and not has_gathered_evidence
+            if result.content and not directive_leak_risk:
                 answer = _clean_answer_text(result.content.strip())
 
+                # The call_with_tools call above is intentionally uncapped so the
+                # LLM has headroom to emit tool-call JSON plus any intermediate
+                # reasoning. But when the LLM short-circuits and returns text
+                # directly, that text becomes the user-visible final answer and
+                # must respect max_tokens like the forced-final paths do. If it
+                # overshoots, run one extra capped call to rewrite it within
+                # the cap.
+                if max_tokens is not None and len(_TIKTOKEN_ENCODING.encode(answer)) > max_tokens:
+                    rewrite_start = time.time()
+                    rewritten, rewrite_usage = await llm_config.call(
+                        messages=[
+                            {
+                                "role": "system",
+                                "content": (
+                                    "Rewrite the user's text so it fits within the requested token "
+                                    "budget. Preserve the key facts and structure; drop lower-priority "
+                                    "detail. Respond with the rewritten text only, no preamble."
+                                ),
+                            },
+                            {
+                                "role": "user",
+                                "content": f"Target budget: {max_tokens} tokens.\n\nText to rewrite:\n{answer}",
+                            },
+                        ],
+                        scope="reflect",
+                        max_completion_tokens=max_tokens,
+                        return_usage=True,
+                    )
+                    total_input_tokens += rewrite_usage.input_tokens
+                    total_output_tokens += rewrite_usage.output_tokens
+                    llm_trace.append(
+                        {
+                            "scope": "final_rewrite",
+                            "duration_ms": int((time.time() - rewrite_start) * 1000),
+                            "input_tokens": rewrite_usage.input_tokens,
+                            "output_tokens": rewrite_usage.output_tokens,
+                        }
+                    )
+                    answer = _clean_answer_text(rewritten.strip())
+
                 # Generate structured output if schema provided
                 structured_output = None
                 if response_schema and answer:
@@ -679,7 +732,7 @@ async def run_reflect_agent(
             llm_start = time.time()
             response, usage = await llm_config.call(
                 messages=[
-                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "system", "content": build_final_system_prompt(bank_profile.get("mission"))},
                     {"role": "user", "content": prompt},
                 ],
                 scope="reflect",
@@ -743,7 +796,8 @@ async def run_reflect_agent(
                         "content": json.dumps(
                             {
                                 "error": "You must search for information first. Use search_mental_models(), search_observations(), or recall() before providing your final answer."
-                            }
+                            },
+                            ensure_ascii=False,
                         ),
                     }
                 )
@@ -805,7 +859,8 @@ async def run_reflect_agent(
                         "content": json.dumps(
                             {
                                 "error": f"Tool '{_normalize_tool_name(tc.name)}' is not available. Use only the tools provided to you."
-                            }
+                            },
+                            ensure_ascii=False,
                         ),
                     }
                 )
@@ -876,7 +931,7 @@ async def run_reflect_agent(
                         "role": "tool",
                         "tool_call_id": tc.id,
                         "name": tc.name,  # Required by Gemini
-                        "content": json.dumps(output, default=str),
+                        "content": json.dumps(output, default=str, ensure_ascii=False),
                     }
                 )
 
@@ -899,7 +954,7 @@ async def run_reflect_agent(
                 )
 
                 try:
-                    output_chars = len(json.dumps(output))
+                    output_chars = len(json.dumps(output, ensure_ascii=False))
                 except (TypeError, ValueError):
                     output_chars = len(str(output))
 
@@ -936,7 +991,7 @@ def _tool_call_to_dict(tc: "LLMToolCall") -> dict[str, Any]:
         "type": "function",
         "function": {
             "name": tc.name,
-            "arguments": json.dumps(tc.arguments),
+            "arguments": json.dumps(tc.arguments, ensure_ascii=False),
         },
     }
     if tc.thought_signature is not None:
@@ -1034,7 +1089,7 @@ async def _execute_tool_with_timing(
         # Set attributes
         span.set_attribute("hindsight.tool.name", normalized_name)
         span.set_attribute("hindsight.tool.id", tc.id)
-        span.set_attribute("hindsight.tool.arguments", json.dumps(tc.arguments))
+        span.set_attribute("hindsight.tool.arguments", json.dumps(tc.arguments, ensure_ascii=False))
 
         try:
             result = await _execute_tool(

hindsight-api-slim/hindsight_api/engine/reflect/delta_ops.py

@@ -0,0 +1,307 @@
+"""Delta operations for structured mental models.
+
+The LLM's job during a delta refresh is to emit a list of these operations,
+each targeting an existing section (by id) or referencing a position relative
+to one.  ``apply_operations`` validates and applies each op in turn against a
+copy of the document; invalid ops (unknown ``section_id``, out-of-range
+``block_index``, malformed payloads) are dropped with a debug-friendly reason.
+
+Sections and blocks not mentioned by any op are physically copied through
+unchanged — there is no LLM-mediated re-emission of unchanged text, so prose
+drift is structurally impossible.
+
+Why operations and not "output the new structured doc":
+- "Output the new doc" still asks the LLM to *generate* every section's
+  blocks, including ones it didn't intend to modify, which gives it the same
+  opportunity to drift.
+- Operations make the no-change case mechanical: zero ops → identical doc.
+- Operations are auditable: each refresh produces a log of exactly what
+  changed, useful for debugging the LLM's behaviour and explaining diffs.
+
+Failure modes are by design conservative: an operation list that fails to
+parse against the Pydantic schema, or an LLM that returns invalid ops, results
+in zero changes — the document stays as-is. The structure can only get better
+or stay the same per refresh, never get worse.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Annotated, Any, Literal, Union
+
+from pydantic import BaseModel, ConfigDict, Field
+
+from .structured_doc import (
+    Block,
+    Section,
+    StructuredDocument,
+    make_unique_id,
+    slugify_heading,
+)
+
+logger = logging.getLogger(__name__)
+
+
+# Op payloads ---------------------------------------------------------------
+
+
+class _OpBase(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+
+class AppendBlockOp(_OpBase):
+    """Add a new block at the end of an existing section."""
+
+    op: Literal["append_block"] = "append_block"
+    section_id: str
+    block: Block
+
+
+class InsertBlockOp(_OpBase):
+    """Insert a new block at ``index`` in an existing section.
+
+    ``index`` may equal ``len(section.blocks)`` (append) but not be greater.
+    """
+
+    op: Literal["insert_block"] = "insert_block"
+    section_id: str
+    index: int = Field(ge=0)
+    block: Block
+
+
+class ReplaceBlockOp(_OpBase):
+    """Replace the block at ``index`` of an existing section."""
+
+    op: Literal["replace_block"] = "replace_block"
+    section_id: str
+    index: int = Field(ge=0)
+    block: Block
+
+
+class RemoveBlockOp(_OpBase):
+    """Remove the block at ``index`` of an existing section."""
+
+    op: Literal["remove_block"] = "remove_block"
+    section_id: str
+    index: int = Field(ge=0)
+
+
+class AddSectionOp(_OpBase):
+    """Add a brand-new section.
+
+    ``after_section_id`` is optional; when omitted the new section is appended
+    at the end. ``new_id`` is optional; when omitted we slugify the heading
+    and disambiguate against existing IDs.
+    """
+
+    op: Literal["add_section"] = "add_section"
+    heading: str
+    level: int = Field(default=2, ge=1, le=6)
+    blocks: list[Block] = Field(default_factory=list)
+    after_section_id: str | None = None
+    new_id: str | None = None
+
+
+class RemoveSectionOp(_OpBase):
+    """Remove an entire section by id."""
+
+    op: Literal["remove_section"] = "remove_section"
+    section_id: str
+
+
+class ReplaceSectionBlocksOp(_OpBase):
+    """Replace all blocks of a section in one go.
+
+    Used when most of a section's contents are stale and rebuilding it as a
+    unit is clearer than emitting many block-level ops. The section's heading
+    and id are preserved.
+    """
+
+    op: Literal["replace_section_blocks"] = "replace_section_blocks"
+    section_id: str
+    blocks: list[Block] = Field(default_factory=list)
+
+
+class RenameSectionOp(_OpBase):
+    """Rename a section's heading. The id is unchanged so future ops still resolve."""
+
+    op: Literal["rename_section"] = "rename_section"
+    section_id: str
+    new_heading: str
+
+
+Operation = Annotated[
+    Union[
+        AppendBlockOp,
+        InsertBlockOp,
+        ReplaceBlockOp,
+        RemoveBlockOp,
+        AddSectionOp,
+        RemoveSectionOp,
+        ReplaceSectionBlocksOp,
+        RenameSectionOp,
+    ],
+    Field(discriminator="op"),
+]
+
+
+class DeltaOperationList(BaseModel):
+    """Container for the operations produced by an LLM delta call."""
+
+    model_config = ConfigDict(extra="forbid")
+    operations: list[Operation] = Field(default_factory=list)
+
+
+# Application ---------------------------------------------------------------
+
+
+class AppliedDelta(BaseModel):
+    """Outcome of applying a list of operations to a document."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    document: StructuredDocument
+    applied: list[dict[str, Any]] = Field(default_factory=list)
+    skipped: list[dict[str, Any]] = Field(default_factory=list)
+
+    @property
+    def changed(self) -> bool:
+        return len(self.applied) > 0
+
+
+def _op_summary(op: Operation) -> dict[str, Any]:
+    """Compact dict suitable for the audit trail."""
+    data = op.model_dump()
+    return {k: v for k, v in data.items() if k != "block" and k != "blocks"} | {
+        "op": data["op"],
+    }
+
+
+def apply_operations(
+    doc: StructuredDocument,
+    operations: list[Operation],
+) -> AppliedDelta:
+    """Apply a list of operations to a document, returning a new document.
+
+    The original document is never mutated. Invalid operations (unknown
+    section, out-of-range index, name collision when adding a section) are
+    skipped and recorded in ``skipped`` with a ``reason`` string.
+    """
+    new_doc = doc.model_copy(deep=True)
+    applied: list[dict[str, Any]] = []
+    skipped: list[dict[str, Any]] = []
+
+    def skip(op: Operation, reason: str) -> None:
+        entry = _op_summary(op)
+        entry["reason"] = reason
+        skipped.append(entry)
+        logger.debug(f"[STRUCTURED_DELTA] skipping op {entry}")
+
+    for op in operations:
+        if isinstance(op, AppendBlockOp):
+            section = new_doc.section_by_id(op.section_id)
+            if section is None:
+                skip(op, f"unknown section_id: {op.section_id}")
+                continue
+            section.blocks.append(op.block)
+            applied.append(_op_summary(op))
+            continue
+
+        if isinstance(op, InsertBlockOp):
+            section = new_doc.section_by_id(op.section_id)
+            if section is None:
+                skip(op, f"unknown section_id: {op.section_id}")
+                continue
+            if op.index > len(section.blocks):
+                skip(
+                    op,
+                    f"index out of range: {op.index} > {len(section.blocks)}",
+                )
+                continue
+            section.blocks.insert(op.index, op.block)
+            applied.append(_op_summary(op))
+            continue
+
+        if isinstance(op, ReplaceBlockOp):
+            section = new_doc.section_by_id(op.section_id)
+            if section is None:
+                skip(op, f"unknown section_id: {op.section_id}")
+                continue
+            if op.index >= len(section.blocks):
+                skip(
+                    op,
+                    f"index out of range: {op.index} >= {len(section.blocks)}",
+                )
+                continue
+            section.blocks[op.index] = op.block
+            applied.append(_op_summary(op))
+            continue
+
+        if isinstance(op, RemoveBlockOp):
+            section = new_doc.section_by_id(op.section_id)
+            if section is None:
+                skip(op, f"unknown section_id: {op.section_id}")
+                continue
+            if op.index >= len(section.blocks):
+                skip(
+                    op,
+                    f"index out of range: {op.index} >= {len(section.blocks)}",
+                )
+                continue
+            section.blocks.pop(op.index)
+            applied.append(_op_summary(op))
+            continue
+
+        if isinstance(op, AddSectionOp):
+            existing_ids = {s.id for s in new_doc.sections}
+            base_id = op.new_id or slugify_heading(op.heading)
+            section_id = make_unique_id(base_id, existing_ids)
+            new_section = Section(
+                id=section_id,
+                heading=op.heading,
+                level=op.level,
+                blocks=list(op.blocks),
+            )
+            if op.after_section_id is None:
+                new_doc.sections.append(new_section)
+            else:
+                idx = new_doc.section_index(op.after_section_id)
+                if idx is None:
+                    skip(op, f"unknown after_section_id: {op.after_section_id}")
+                    continue
+                new_doc.sections.insert(idx + 1, new_section)
+            entry = _op_summary(op)
+            entry["assigned_id"] = section_id
+            applied.append(entry)
+            continue
+
+        if isinstance(op, RemoveSectionOp):
+            idx = new_doc.section_index(op.section_id)
+            if idx is None:
+                skip(op, f"unknown section_id: {op.section_id}")
+                continue
+            new_doc.sections.pop(idx)
+            applied.append(_op_summary(op))
+            continue
+
+        if isinstance(op, ReplaceSectionBlocksOp):
+            section = new_doc.section_by_id(op.section_id)
+            if section is None:
+                skip(op, f"unknown section_id: {op.section_id}")
+                continue
+            section.blocks = list(op.blocks)
+            applied.append(_op_summary(op))
+            continue
+
+        if isinstance(op, RenameSectionOp):
+            section = new_doc.section_by_id(op.section_id)
+            if section is None:
+                skip(op, f"unknown section_id: {op.section_id}")
+                continue
+            section.heading = op.new_heading
+            applied.append(_op_summary(op))
+            continue
+
+        skip(op, f"unhandled op type: {type(op).__name__}")  # pragma: no cover
+
+    return AppliedDelta(document=new_doc, applied=applied, skipped=skipped)

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

@@ -18,6 +18,9 @@ _TIKTOKEN_ENCODING = tiktoken.get_encoding("cl100k_base")
 # The remainder covers the system prompt, question, bank context, and output tokens.
 _FINAL_PROMPT_CONTEXT_FRACTION = 0.8
 
+_DEFAULT_ROLE = "You are a reflection agent that answers questions by reasoning over retrieved memories."
+_DEFAULT_FINAL_ROLE = "You are a thoughtful assistant that synthesizes answers from retrieved memories."
+
 
 def _extract_directive_rules(directives: list[dict[str, Any]]) -> list[str]:
     """Extract directive rules as a list of strings."""
@@ -133,7 +136,9 @@ def build_system_prompt_for_tools(
 
     parts.extend(
         [
-            "You are a reflection agent that answers questions by reasoning over retrieved memories.",
+            mission.strip() if mission else _DEFAULT_ROLE,
+            "",
+            "Answer the user's question by reasoning over retrieved memories.",
             "",
         ]
     )
@@ -369,7 +374,7 @@ def build_agent_prompt(
             output = entry["output"]
             # Format as proper JSON for LLM readability
             try:
-                output_str = json.dumps(output, indent=2, default=str)
+                output_str = json.dumps(output, indent=2, default=str, ensure_ascii=False)
             except (TypeError, ValueError):
                 output_str = str(output)
             parts.append(f"\n### Call {i}: {tool}\n```json\n{output_str}\n```")
@@ -444,7 +449,7 @@ def build_final_prompt(
             tool = entry["tool"]
             output = entry["output"]
             try:
-                output_str = json.dumps(output, indent=2, default=str)
+                output_str = json.dumps(output, indent=2, default=str, ensure_ascii=False)
             except (TypeError, ValueError):
                 output_str = str(output)
             block = f"\n### From {tool}:\n```json\n{output_str}\n```"
@@ -479,9 +484,9 @@ def build_final_prompt(
     return "\n".join(parts)
 
 
-FINAL_SYSTEM_PROMPT = """CRITICAL: You MUST ONLY use information from retrieved tool results. NEVER make up names, people, events, or entities.
+_FINAL_SYSTEM_PROMPT_BASE = """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.
+{role_section}
 
 Your approach:
 - Reason over the retrieved memories to answer the question
@@ -508,3 +513,213 @@ CRITICAL: Output ONLY the final synthesized answer. Do NOT include:
 Just provide the direct answer with proper markdown formatting.
 
 CRITICAL: This is a NON-CONVERSATIONAL system. NEVER ask follow-up questions, offer to search again, suggest alternatives, or end with anything like "Would you like me to..." or "Let me know if...". The user cannot reply. Your answer must be complete and self-contained."""
+
+
+def build_final_system_prompt(mission: str | None = None) -> str:
+    """Build the final synthesis system prompt, using mission as role when set."""
+    role_section = mission.strip() if mission else _DEFAULT_FINAL_ROLE
+    return _FINAL_SYSTEM_PROMPT_BASE.format(role_section=role_section)
+
+
+# Backward-compatible constant for non-identity missions
+FINAL_SYSTEM_PROMPT = build_final_system_prompt()
+
+
+STRUCTURED_DELTA_SYSTEM_PROMPT = """You are integrating *new information* into an existing structured document.
+
+You will be given:
+1. TOPIC — the question this document answers. Content that does not help
+   answer this question is OFF-TOPIC and should be removed.
+2. CURRENT DOCUMENT (JSON) — the existing structured mental model. Each section
+   has a stable ``id``, a ``heading``, a ``level`` (1..6), and an ordered list
+   of ``blocks``. Blocks are typed: ``paragraph``, ``bullet_list``,
+   ``ordered_list``, or ``code``.
+3. NEW INFORMATION SYNTHESIS (markdown) — a synthesis showing how the new facts
+   relate to the document's topic. Use it to understand context and relevance,
+   but do NOT copy its formatting or wording wholesale.
+4. SUPPORTING FACTS — observations and facts created since the last refresh.
+   These are genuinely new — they were NOT available when the current document
+   was written.
+
+Your task: output a JSON object ``{"operations": [...]}``. Applied to CURRENT
+DOCUMENT, the operations must produce a document that best answers the TOPIC
+by integrating the new facts.
+
+RULES
+- These facts are NEW since the last refresh. The existing document already
+  captures all prior information from earlier refreshes. Your job is to
+  integrate the new facts into the existing document.
+- **Preserve existing content**: The current document was built from prior facts
+  that you cannot see. Do NOT remove or replace existing sections just because
+  the new facts do not reference them. Only remove content when the new facts
+  explicitly contradict or supersede it.
+- **Merge overlapping topics**: When new facts cover topics that overlap with
+  existing sections, merge the new information INTO the existing section
+  rather than creating duplicates. When new facts provide more specific or
+  authoritative guidance on a topic already covered generically, update the
+  existing content to reflect the more specific guidance.
+- **Preserve examples**: Concrete examples, before/after pairs, sample sentences,
+  and illustrative ✅/❌ comparisons are MORE valuable than abstract rules.
+  When facts contain examples, include them. Never drop an example to make
+  room for an abstract restatement of the same point.
+- Operations target sections by ``section_id`` (use the ``id`` field of the
+  section in CURRENT DOCUMENT, NOT the heading). Block operations target
+  blocks by ``index`` (0-based, against the section's current block list).
+- **Add** new content with ``append_block``, ``insert_block``, or ``add_section``
+  when facts introduce information not yet covered. Prefer extending an
+  existing section over creating a new one.
+- **Update** existing content with ``replace_block`` or ``replace_section_blocks``
+  when new facts provide corrections, updates, or more specific information
+  about topics already in the document.
+- **Remove** content with ``remove_block`` or ``remove_section`` ONLY when
+  the new facts explicitly contradict or supersede it.
+- NEVER emit operations whose only effect is to reword unchanged content.
+- NEVER emit operations to "normalize" formatting (numbered → bulleted, casing
+  changes, paragraph → list, etc).
+- Every operation MUST be justifiable by a specific fact in SUPPORTING FACTS.
+- Output ``{"operations": []}`` only if the new facts are already reflected
+  in the document (e.g., from a concurrent update).
+
+ALLOWED OPERATIONS (each line shows the JSON shape)
+- ``{"op": "append_block", "section_id": "...", "block": {...}}``
+- ``{"op": "insert_block", "section_id": "...", "index": N, "block": {...}}``
+- ``{"op": "replace_block", "section_id": "...", "index": N, "block": {...}}``
+- ``{"op": "remove_block", "section_id": "...", "index": N}``
+- ``{"op": "add_section", "heading": "...", "level": 2, "blocks": [...], "after_section_id": "..."}``
+- ``{"op": "remove_section", "section_id": "..."}``
+- ``{"op": "replace_section_blocks", "section_id": "...", "blocks": [...]}``
+- ``{"op": "rename_section", "section_id": "...", "new_heading": "..."}``
+
+Block shapes
+- ``{"type": "paragraph", "text": "..."}``
+- ``{"type": "bullet_list", "items": ["...", "..."]}``
+- ``{"type": "ordered_list", "items": ["...", "..."]}``
+- ``{"type": "code", "language": "json", "text": "..."}``
+
+OUTPUT FORMAT
+Return ONLY a single JSON object on its own, with no prose before or after,
+no markdown code fences, no commentary. The object must have exactly one
+top-level key, ``operations``, whose value is an array of operation objects
+(empty array when nothing changes).
+
+Examples
+- No changes needed → ``{"operations": []}``
+- Add one bullet to an existing "Members" section →
+  ``{"operations": [{"op": "append_block", "section_id": "members",
+  "block": {"type": "bullet_list", "items": ["Carol — junior engineer"]}}]}``
+- Replace a paragraph that has been corrected by new facts →
+  ``{"operations": [{"op": "replace_block", "section_id": "overview",
+  "index": 0, "block": {"type": "paragraph", "text": "Updated summary."}}]}``
+- Remove an obsolete block →
+  ``{"operations": [{"op": "remove_block", "section_id": "status", "index": 2}]}``"""
+
+
+def build_structured_delta_prompt(
+    *,
+    current_document_json: str,
+    candidate_markdown: str,
+    supporting_facts: list[dict[str, Any]],
+    source_query: str,
+    max_output_tokens: int | None = None,
+) -> str:
+    """Build the user prompt for a structured-delta mental model refresh.
+
+    The LLM's job is to emit operations against ``current_document_json``;
+    the surrounding ``candidate_markdown`` and ``supporting_facts`` are
+    references for *what new information exists*, not templates to mimic.
+
+    ``max_output_tokens`` is surfaced in the prompt so the model can keep its
+    op list within the provider's response cap. The actual cap is enforced by
+    the caller; this is just an advisory anchor — without it the model often
+    returns op lists whose JSON gets truncated mid-string.
+    """
+    fact_lines: list[str] = []
+    for f in supporting_facts:
+        fid = f.get("id", "")
+        text = (f.get("text") or "").strip().replace("\n", " ")
+        ftype = f.get("type", "")
+        fact_lines.append(f"- [{ftype}:{fid}] {text}")
+    facts_block = "\n".join(fact_lines) if fact_lines else "(no supporting facts retrieved)"
+
+    budget_hint = ""
+    if max_output_tokens is not None:
+        budget_hint = (
+            f"\n\n## Output budget\n"
+            f"Your JSON response must fit within ~{max_output_tokens} tokens. If you "
+            "would need more than this to express every change, prefer the highest-"
+            "leverage edits first (a few ``replace_section_blocks`` ops over many "
+            "block-level ops) so the response always parses as valid JSON."
+        )
+
+    return (
+        f"## Topic\n{source_query}\n\n"
+        f"## CURRENT DOCUMENT (apply ops to this; reference section ids as listed)\n"
+        f"```json\n{current_document_json}\n```\n\n"
+        f"## NEW INFORMATION SYNTHESIS (context for how new facts relate to the topic)\n"
+        f"```markdown\n{candidate_markdown}\n```\n\n"
+        f"## SUPPORTING FACTS (new since last refresh — integrate these)\n{facts_block}"
+        f"{budget_hint}\n\n"
+        "## Task\n"
+        "Output a JSON object matching the operations schema. Integrate the new "
+        "supporting facts into CURRENT DOCUMENT. Add, update, or remove content "
+        "as needed. Preserve unchanged sections and blocks by not mentioning them."
+    )
+
+
+DELTA_SYSTEM_PROMPT = """You are performing a surgical delta update to an existing mental model document.
+
+You will be given:
+1. CURRENT DOCUMENT: the existing mental model content (markdown).
+2. CANDIDATE UPDATE: a freshly generated synthesis based on the latest retrieved memories.
+3. SUPPORTING FACTS: the observations and facts that support the CANDIDATE UPDATE.
+
+Your task: produce an updated version of the CURRENT DOCUMENT that reflects the new reality, with the MINIMUM possible changes.
+
+ABSOLUTE RULES:
+- Preserve unchanged content BYTE-FOR-BYTE. If a sentence, heading, bullet, code block, or section is still accurate according to the CANDIDATE UPDATE and SUPPORTING FACTS, copy it verbatim — same wording, same punctuation, same whitespace, same markdown structure.
+- Do NOT reformat, rephrase, or re-style content that is still accurate. No "light edits for clarity", no reordering for flow, no synonym swaps.
+- Remove content that is contradicted by the CANDIDATE UPDATE or SUPPORTING FACTS (stale content).
+- Add new content ONLY when the SUPPORTING FACTS contain information not already in the CURRENT DOCUMENT.
+- When adding new content, prefer appending to an existing relevant section. Creating a new section is acceptable when the new information does not fit any existing section.
+- When creating a new section, match the heading style, tone, and formatting conventions used in the CURRENT DOCUMENT.
+- Every assertion in your output MUST be grounded in either (a) the CURRENT DOCUMENT (preserved) or (b) the SUPPORTING FACTS. Never introduce outside knowledge.
+- If nothing in the SUPPORTING FACTS contradicts or extends the CURRENT DOCUMENT, return the CURRENT DOCUMENT UNCHANGED, character for character.
+
+OUTPUT FORMAT:
+- Output ONLY the updated markdown document. No preamble, no explanation, no diff markers, no commentary.
+- Do not wrap the output in code fences unless the CURRENT DOCUMENT itself was entirely a code fence."""
+
+
+def build_delta_prompt(
+    *,
+    current_content: str,
+    candidate_content: str,
+    supporting_facts: list[dict[str, Any]],
+    source_query: str,
+) -> str:
+    """Build the user prompt for a delta-mode mental model refresh.
+
+    Args:
+        current_content: The existing mental model content (to preserve as much as possible).
+        candidate_content: Fresh synthesis from the reflect agent reflecting new reality.
+        supporting_facts: Flat list of fact dicts (id, text, type) supporting the candidate.
+        source_query: The mental model's source query, for topical framing.
+    """
+    fact_lines: list[str] = []
+    for f in supporting_facts:
+        fid = f.get("id", "")
+        text = (f.get("text") or "").strip().replace("\n", " ")
+        ftype = f.get("type", "")
+        fact_lines.append(f"- [{ftype}:{fid}] {text}")
+    facts_block = "\n".join(fact_lines) if fact_lines else "(no supporting facts retrieved)"
+
+    return (
+        f"## Topic\n{source_query}\n\n"
+        f"## CURRENT DOCUMENT\n```markdown\n{current_content}\n```\n\n"
+        f"## CANDIDATE UPDATE\n```markdown\n{candidate_content}\n```\n\n"
+        f"## SUPPORTING FACTS\n{facts_block}\n\n"
+        "## Task\n"
+        "Produce the updated mental model document by applying the minimum necessary changes "
+        "to CURRENT DOCUMENT so that it reflects CANDIDATE UPDATE and SUPPORTING FACTS. "
+        "Preserve unchanged content byte-for-byte. Output only the final markdown."
+    )

hindsight-api-slim/hindsight_api/engine/reflect/structured_doc.py

@@ -0,0 +1,301 @@
+"""Structured representation of a mental model document.
+
+Why this exists
+---------------
+Storing mental models as raw markdown forces every refresh to round-trip prose
+through an LLM, which then drifts on stylistic details (numbered vs bulleted
+lists, casing, separator lines, paraphrasing) even when instructed to preserve
+content byte-for-byte. The intrinsic mechanism of an LLM is to *generate* the
+next token from a gestalt of the input — not to copy tokens verbatim — so any
+"preserve unchanged content" instruction is fundamentally a soft constraint.
+
+The fix is to give the LLM no opportunity to drift on unchanged content. We
+keep an authoritative structured representation of the document; the markdown
+shown to users is a deterministic render of that structure. Delta refreshes
+emit *operations* against the structure (see ``delta_ops.py``); sections and
+blocks not mentioned by any operation are physically untouched.
+
+Schema (v1)
+-----------
+A document is an ordered list of ``Section``s.  Each section has:
+- ``id``    : stable slug derived from ``heading`` (used as the operation
+              target across refreshes; surviving renames is a separate
+              concern handled by an explicit ``rename`` op).
+- ``heading``: the markdown heading text (without the ``#`` prefix).
+- ``level`` : 1 (``#``) … 6 (``######``).  Default 2.
+- ``blocks``: ordered list of typed blocks — paragraph, bullet_list,
+              ordered_list, code.
+
+The schema is intentionally narrow: it covers what real mental-model documents
+actually contain (the kind a coding agent writes for itself or a user writes as
+a "skill" doc).  Tables, images, and raw HTML are out of scope until needed.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Annotated, Literal, Union
+
+from pydantic import BaseModel, ConfigDict, Field
+
+# Blocks ---------------------------------------------------------------------
+
+
+class ParagraphBlock(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    type: Literal["paragraph"] = "paragraph"
+    text: str
+
+
+class BulletListBlock(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    type: Literal["bullet_list"] = "bullet_list"
+    items: list[str] = Field(default_factory=list)
+
+
+class OrderedListBlock(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    type: Literal["ordered_list"] = "ordered_list"
+    items: list[str] = Field(default_factory=list)
+
+
+class CodeBlock(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    type: Literal["code"] = "code"
+    language: str = ""
+    text: str
+
+
+Block = Annotated[
+    Union[ParagraphBlock, BulletListBlock, OrderedListBlock, CodeBlock],
+    Field(discriminator="type"),
+]
+
+
+# Section / Document ---------------------------------------------------------
+
+
+class Section(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    id: str
+    heading: str
+    level: int = Field(default=2, ge=1, le=6)
+    blocks: list[Block] = Field(default_factory=list)
+
+
+class StructuredDocument(BaseModel):
+    """Top-level structured representation of a mental model."""
+
+    model_config = ConfigDict(extra="forbid")
+    version: Literal[1] = 1
+    sections: list[Section] = Field(default_factory=list)
+
+    def section_by_id(self, section_id: str) -> Section | None:
+        for s in self.sections:
+            if s.id == section_id:
+                return s
+        return None
+
+    def section_index(self, section_id: str) -> int | None:
+        for i, s in enumerate(self.sections):
+            if s.id == section_id:
+                return i
+        return None
+
+
+# Slug helpers ---------------------------------------------------------------
+
+_SLUG_RX = re.compile(r"[^a-z0-9]+")
+
+
+def slugify_heading(heading: str) -> str:
+    """Stable, deterministic slug from a heading.
+
+    "Stop Conditions" -> "stop-conditions"
+    "Inputs and Context" -> "inputs-and-context"
+    """
+    slug = _SLUG_RX.sub("-", heading.strip().lower()).strip("-")
+    return slug or "section"
+
+
+def make_unique_id(base: str, existing: set[str]) -> str:
+    """Disambiguate by appending -2, -3, … if the slug is already in use."""
+    if base not in existing:
+        return base
+    i = 2
+    while f"{base}-{i}" in existing:
+        i += 1
+    return f"{base}-{i}"
+
+
+# Renderer -------------------------------------------------------------------
+
+
+def render_block(block: Block) -> str:
+    """Render a single block to markdown. No trailing newline."""
+    if isinstance(block, ParagraphBlock):
+        return block.text.rstrip()
+    if isinstance(block, BulletListBlock):
+        return "\n".join(f"- {item.rstrip()}" for item in block.items)
+    if isinstance(block, OrderedListBlock):
+        return "\n".join(f"{i + 1}. {item.rstrip()}" for i, item in enumerate(block.items))
+    if isinstance(block, CodeBlock):
+        fence_lang = block.language or ""
+        return f"```{fence_lang}\n{block.text}\n```"
+    raise TypeError(f"Unknown block type: {type(block)!r}")
+
+
+def render_section(section: Section) -> str:
+    """Render a section: heading + blank line + blocks separated by blank lines."""
+    parts = ["#" * section.level + " " + section.heading.strip()]
+    for block in section.blocks:
+        parts.append("")  # blank line before each block
+        parts.append(render_block(block))
+    return "\n".join(parts)
+
+
+def render_document(doc: StructuredDocument) -> str:
+    """Render the whole document. Sections separated by a single blank line.
+
+    The output is byte-stable: same structured input always produces the same
+    markdown, modulo the inherent ordering of sections/blocks/items.
+    """
+    if not doc.sections:
+        return ""
+    return "\n\n".join(render_section(s) for s in doc.sections) + "\n"
+
+
+# Parser ---------------------------------------------------------------------
+#
+# The parser is intentionally lenient: it accepts the markdown produced by
+# our own renderer (round-trip-safe) and the markdown an LLM tends to produce
+# for mental-model documents.  It is *not* a general CommonMark parser — it
+# does not need to be.  When it cannot classify a block it falls back to a
+# paragraph so that no content is silently dropped.
+
+_HEADING_RX = re.compile(r"^(#{1,6})\s+(.+?)\s*$")
+_BULLET_RX = re.compile(r"^\s*[-*+]\s+(.*)$")
+_ORDERED_RX = re.compile(r"^\s*\d+[.)]\s+(.*)$")
+_FENCE_RX = re.compile(r"^```([A-Za-z0-9_+-]*)\s*$")
+
+
+def _strip_separators(lines: list[str]) -> list[str]:
+    """Drop horizontal-rule lines (`---`, `***`) used as section separators.
+
+    Our renderer never emits these, but LLM output frequently includes them
+    between sections; treating them as blank lines avoids parsing them as
+    paragraphs.
+    """
+    return ["" if re.fullmatch(r"\s*([-*_])\1{2,}\s*", line) else line for line in lines]
+
+
+def _split_blocks(lines: list[str]) -> list[list[str]]:
+    """Group consecutive non-blank lines into block chunks."""
+    chunks: list[list[str]] = []
+    current: list[str] = []
+    in_fence = False
+    for line in lines:
+        if _FENCE_RX.match(line):
+            current.append(line)
+            in_fence = not in_fence
+            continue
+        if in_fence:
+            current.append(line)
+            continue
+        if line.strip() == "":
+            if current:
+                chunks.append(current)
+                current = []
+        else:
+            current.append(line)
+    if current:
+        chunks.append(current)
+    return chunks
+
+
+def _parse_block(chunk: list[str]) -> Block:
+    """Parse a single non-empty chunk into a block."""
+    if chunk and _FENCE_RX.match(chunk[0]):
+        m = _FENCE_RX.match(chunk[0])
+        lang = m.group(1) if m else ""
+        body_lines = chunk[1:]
+        if body_lines and _FENCE_RX.match(body_lines[-1]):
+            body_lines = body_lines[:-1]
+        return CodeBlock(language=lang, text="\n".join(body_lines))
+
+    if all(_BULLET_RX.match(line) for line in chunk):
+        items = []
+        for line in chunk:
+            m = _BULLET_RX.match(line)
+            assert m is not None
+            items.append(m.group(1).strip())
+        return BulletListBlock(items=items)
+
+    if all(_ORDERED_RX.match(line) for line in chunk):
+        items = []
+        for line in chunk:
+            m = _ORDERED_RX.match(line)
+            assert m is not None
+            items.append(m.group(1).strip())
+        return OrderedListBlock(items=items)
+
+    return ParagraphBlock(text=" ".join(line.strip() for line in chunk).strip())
+
+
+def parse_markdown(markdown: str) -> StructuredDocument:
+    """Best-effort parse of a markdown document into the structured schema.
+
+    Sections are introduced by ATX headings (``#``..``######``).  Anything
+    before the first heading is wrapped into an implicit "Overview" section
+    so we never silently drop user content.  Section IDs are unique slugs of
+    their headings.
+    """
+    raw_lines = (markdown or "").splitlines()
+    lines = _strip_separators(raw_lines)
+
+    sections: list[Section] = []
+    used_ids: set[str] = set()
+    pending: list[str] = []
+    current: Section | None = None
+
+    def flush_pending_into(section: Section) -> None:
+        if not pending:
+            return
+        for chunk in _split_blocks(pending):
+            section.blocks.append(_parse_block(chunk))
+        pending.clear()
+
+    for line in lines:
+        m = _HEADING_RX.match(line)
+        if m:
+            if current is not None:
+                flush_pending_into(current)
+                sections.append(current)
+            elif pending:
+                # Content before the first heading: wrap in implicit section.
+                base = "overview"
+                section_id = make_unique_id(base, used_ids)
+                used_ids.add(section_id)
+                implicit = Section(id=section_id, heading="Overview", level=2)
+                flush_pending_into(implicit)
+                sections.append(implicit)
+            level = len(m.group(1))
+            heading = m.group(2).strip()
+            section_id = make_unique_id(slugify_heading(heading), used_ids)
+            used_ids.add(section_id)
+            current = Section(id=section_id, heading=heading, level=level)
+        else:
+            pending.append(line)
+
+    if current is not None:
+        flush_pending_into(current)
+        sections.append(current)
+    elif pending:
+        base = "overview"
+        section_id = make_unique_id(base, used_ids)
+        used_ids.add(section_id)
+        implicit = Section(id=section_id, heading="Overview", level=2)
+        flush_pending_into(implicit)
+        sections.append(implicit)
+
+    return StructuredDocument(sections=sections)

hindsight-api-slim/hindsight_api/engine/reflect/tools.py

@@ -9,6 +9,7 @@ Implements hierarchical retrieval:
 
 import logging
 import uuid
+from dataclasses import replace
 from datetime import datetime, timezone
 from typing import TYPE_CHECKING, Any
 
@@ -22,6 +23,7 @@ logger = logging.getLogger(__name__)
 
 
 async def tool_search_mental_models(
+    memory_engine: "MemoryEngine",
     conn: "Connection",
     bank_id: str,
     query: str,
@@ -31,7 +33,6 @@ async def tool_search_mental_models(
     tags_match: str = "any",
     tag_groups: "list | None" = None,
     exclude_ids: list[str] | None = None,
-    pending_consolidation: int = 0,
 ) -> dict[str, Any]:
     """
     Search user-curated mental models by semantic similarity.
@@ -81,7 +82,7 @@ async def tool_search_mental_models(
         f"""
         SELECT
             id, name, content,
-            tags, created_at, last_refreshed_at,
+            tags, created_at, last_refreshed_at, trigger,
             1 - (embedding <=> $2::vector) as relevance
         FROM {fq_table("mental_models")}
         WHERE bank_id = $1 AND embedding IS NOT NULL {filters}
@@ -98,10 +99,9 @@ async def tool_search_mental_models(
         if last_refreshed_at and last_refreshed_at.tzinfo is None:
             last_refreshed_at = last_refreshed_at.replace(tzinfo=timezone.utc)
 
-        # A mental model is stale when there are memories that haven't been consolidated yet —
-        # the same signal used for observations staleness.
-        is_stale = pending_consolidation > 0
-        staleness_reason = f"{pending_consolidation} memories pending consolidation" if is_stale else None
+        # Per-MM staleness: new in-scope memories since last refresh (includes pending).
+        is_stale = await memory_engine.compute_mental_model_is_stale(conn, bank_id, row)
+        staleness_reason = "new in-scope memories ingested since last refresh" if is_stale else None
 
         mental_models.append(
             {
@@ -135,6 +135,8 @@ async def tool_search_observations(
     last_consolidated_at: datetime | None = None,
     pending_consolidation: int = 0,
     source_facts_max_tokens: int = -1,
+    created_after: datetime | None = None,
+    created_before: datetime | None = None,
 ) -> dict[str, Any]:
     """
     Search consolidated observations using recall.
@@ -162,17 +164,24 @@ async def tool_search_observations(
     if include_source_facts and source_facts_max_tokens > 0:
         recall_kwargs["max_source_facts_tokens"] = source_facts_max_tokens
 
+    # Use an internal request context so this recall is not billed as a
+    # user-facing operation. The reflect caller is already billed for the
+    # overall reflect operation; double-billing the sub-recalls would
+    # overcharge the customer.
+    internal_ctx = replace(request_context, internal=True)
     result = await memory_engine.recall_async(
         bank_id=bank_id,
         query=query,
         fact_type=["observation"],
         max_tokens=max_tokens,
         enable_trace=False,
-        request_context=request_context,
+        request_context=internal_ctx,
         tags=tags,
         tags_match=tags_match,
         tag_groups=tag_groups,
         include_source_facts=include_source_facts,
+        created_after=created_after,
+        created_before=created_before,
         _connection_budget=1,
         _quiet=True,
         **recall_kwargs,
@@ -208,6 +217,9 @@ async def tool_recall(
     connection_budget: int = 1,
     max_chunk_tokens: int = 1000,
     fact_types: list[str] | None = None,
+    include_chunks: bool = True,
+    created_after: datetime | None = None,
+    created_before: datetime | None = None,
 ) -> dict[str, Any]:
     """
     Search memories using TEMPR retrieval.
@@ -224,25 +236,28 @@ async def tool_recall(
         tags: Filter by tags (includes untagged memories)
         tags_match: How to match tags - "any" (OR), "all" (AND), or "exact"
         connection_budget: Max DB connections for this recall (default 1 for internal ops)
-        max_chunk_tokens: Maximum tokens for raw source chunk text (default 1000, always included)
+        max_chunk_tokens: Maximum tokens for raw source chunk text (default 1000)
         fact_types: Optional filter for fact types to retrieve. Defaults to ["experience", "world"].
+        include_chunks: Whether to fetch raw chunk text alongside facts (default True).
 
     Returns:
-        Dict with list of matching memories including raw chunk text
+        Dict with list of matching memories including raw chunk text (when include_chunks)
     """
     # Only world/experience are valid for raw recall (observation is handled by search_observations)
     recall_fact_type = [ft for ft in (fact_types or ["experience", "world"]) if ft in ("world", "experience")]
-    include_chunks = True
+    internal_ctx = replace(request_context, internal=True)
     result = await memory_engine.recall_async(
         bank_id=bank_id,
         query=query,
         fact_type=recall_fact_type,
         max_tokens=max_tokens,
         enable_trace=False,
-        request_context=request_context,
+        request_context=internal_ctx,
         tags=tags,
         tags_match=tags_match,
         tag_groups=tag_groups,
+        created_after=created_after,
+        created_before=created_before,
         _connection_budget=connection_budget,
         _quiet=True,  # Suppress logging for internal operations
         include_chunks=include_chunks,

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

@@ -10,7 +10,6 @@ from typing import Any
 
 from pydantic import BaseModel, ConfigDict, Field, field_validator
 
-# Valid fact types for recall operations (excludes 'opinion' which is deprecated)
 VALID_RECALL_FACT_TYPES = frozenset(["world", "experience", "observation"])
 
 

hindsight-api-slim/hindsight_api/engine/retain/bank_utils.py

@@ -113,6 +113,22 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
     Returns:
         BankProfile with name, typed DispositionTraits, and mission
     """
+    profile, _ = await get_or_create_bank_profile(pool, bank_id)
+    return profile
+
+
+async def get_or_create_bank_profile(pool, bank_id: str) -> tuple[BankProfile, bool]:
+    """
+    Get bank profile, auto-creating with defaults if it doesn't exist.
+
+    Same as get_bank_profile, but also returns a flag indicating whether the
+    bank was freshly created on this call. Used by the memory engine to apply
+    the HINDSIGHT_API_DEFAULT_BANK_TEMPLATE hook on first bank creation.
+
+    Returns:
+        Tuple of (BankProfile, created) where created is True if the bank
+        did not exist before this call.
+    """
     async with acquire_with_retry(pool) as conn:
         # Try to get existing bank
         row = await conn.fetchrow(
@@ -129,15 +145,18 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
             if isinstance(disposition_data, str):
                 disposition_data = json.loads(disposition_data)
 
-            return BankProfile(
-                name=row["name"],
-                disposition=DispositionTraits(**disposition_data),
-                mission=row["mission"] or "",
+            return (
+                BankProfile(
+                    name=row["name"],
+                    disposition=DispositionTraits(**disposition_data),
+                    mission=row["mission"] or "",
+                ),
+                False,
             )
 
         # Bank doesn't exist, create with defaults.
         # Generate internal_id here so we control the value and can use it
-        # immediately for HNSW index creation without a RETURNING round-trip.
+        # immediately for vector index creation without a RETURNING round-trip.
         internal_id = uuid.uuid4()
         inserted = await conn.fetchval(
             f"""
@@ -153,11 +172,15 @@ async def get_bank_profile(pool, bank_id: str) -> BankProfile:
             internal_id,
         )
 
-        if inserted:
+        created = inserted is not None
+        if created:
             # Fresh insert — create per-bank vector indexes (instant on empty bank)
             await create_bank_vector_indexes(conn, bank_id, str(internal_id))
 
-        return BankProfile(name=bank_id, disposition=DispositionTraits(**DEFAULT_DISPOSITION), mission="")
+        return (
+            BankProfile(name=bank_id, disposition=DispositionTraits(**DEFAULT_DISPOSITION), mission=""),
+            created,
+        )
 
 
 async def update_bank_disposition(pool, bank_id: str, disposition: dict[str, int]) -> None:

hindsight-api-slim/hindsight_api/engine/retain/chunk_storage.py

@@ -100,11 +100,20 @@ async def store_chunks_batch(conn, bank_id: str, document_id: str, chunks: list[
         content_hashes.append(compute_chunk_hash(chunk.chunk_text))
         chunk_id_map[chunk.chunk_index] = chunk_id
 
-    # Batch insert all chunks
+    # Batch upsert all chunks. ON CONFLICT makes this idempotent: re-submitting
+    # a retain under the same document_id (the pattern in vectorize-io/hindsight#977)
+    # may produce chunk_ids that already exist when upstream cascade-delete or
+    # delta-retain paths don't run (or race with a concurrent task). Overwriting
+    # is the correct behavior per the document_id grouping semantics — the caller
+    # intends this chunk to hold the latest content at that (document_id, index).
     await conn.execute(
         f"""
         INSERT INTO {fq_table("chunks")} (chunk_id, document_id, bank_id, chunk_text, chunk_index, content_hash)
         SELECT * FROM unnest($1::text[], $2::text[], $3::text[], $4::text[], $5::integer[], $6::text[])
+        ON CONFLICT (chunk_id) DO UPDATE SET
+            chunk_text = EXCLUDED.chunk_text,
+            chunk_index = EXCLUDED.chunk_index,
+            content_hash = EXCLUDED.content_hash
         """,
         chunk_ids,
         [document_id] * len(chunk_texts),

hindsight-api-slim/hindsight_api/engine/retain/embedding_utils.py

@@ -47,6 +47,16 @@ async def generate_embeddings_batch(embeddings_backend, texts: list[str]) -> lis
             embeddings_backend.encode,
             texts,
         )
-        return embeddings
     except Exception as e:
         raise Exception(f"Failed to generate batch embeddings: {str(e)}")
+
+    # Guarantee 1:1 alignment with input texts. A silent length mismatch here
+    # propagates downstream as zip() drops items, eventually surfacing as an
+    # IndexError in retain mapping (see issue #1037).
+    if len(embeddings) != len(texts):
+        raise RuntimeError(
+            f"Embeddings backend returned {len(embeddings)} vectors for {len(texts)} input texts; "
+            "expected exact 1:1 alignment"
+        )
+
+    return embeddings

hindsight-api-slim/hindsight_api/engine/retain/entity_processing.py

@@ -12,61 +12,27 @@ from .types import EntityLink, ProcessedFact
 logger = logging.getLogger(__name__)
 
 
-async def process_entities_batch(
-    entity_resolver,
-    conn,
-    bank_id: str,
-    unit_ids: list[str],
+def _prepare_facts_for_entity_processing(
     facts: list[ProcessedFact],
-    log_buffer: list[str] = None,
-    user_entities_per_content: dict[int, list[dict]] = None,
-    entity_labels: list | None = None,
-) -> list[EntityLink]:
+    user_entities_per_content: dict[int, list[dict]] | None = None,
+) -> tuple[list[str], list, list[list[dict]]]:
     """
-    Process entities for all facts and create entity links.
-
-    This function:
-    1. Extracts entity mentions from fact texts
-    2. Merges user-provided entities with LLM-extracted entities
-    3. Resolves entity names to canonical entities
-    4. Creates entity records in the database
-    5. Returns entity links ready for insertion
-
-    Args:
-        entity_resolver: EntityResolver instance for entity resolution
-        conn: Database connection
-        bank_id: Bank identifier
-        unit_ids: List of unit IDs (same length as facts)
-        facts: List of ProcessedFact objects
-        log_buffer: Optional buffer for detailed logging
-        user_entities_per_content: Dict mapping content_index to list of user-provided entities
+    Extract fact texts, dates, and merged entity lists from ProcessedFact objects.
 
     Returns:
-        List of EntityLink objects for batch insertion
+        Tuple of (fact_texts, fact_dates, entities_per_fact)
     """
-    if not unit_ids or not facts:
-        return []
-
-    if len(unit_ids) != len(facts):
-        raise ValueError(f"Mismatch between unit_ids ({len(unit_ids)}) and facts ({len(facts)})")
-
     user_entities_per_content = user_entities_per_content or {}
 
-    # Extract data for link_utils function
     fact_texts = [fact.fact_text for fact in facts]
-    # Use occurred_start if available, otherwise use mentioned_at for entity timestamps
     fact_dates = [fact.occurred_start if fact.occurred_start is not None else fact.mentioned_at for fact in facts]
 
-    # Convert EntityRef objects to dict format and merge with user-provided entities
     entities_per_fact = []
     for fact in facts:
-        # Start with LLM-extracted entities
         llm_entities = [{"text": entity.name, "type": "CONCEPT"} for entity in (fact.entities or [])]
 
-        # Get user entities for this content (use content_index from fact)
         user_entities = user_entities_per_content.get(fact.content_index, [])
 
-        # Merge with case-insensitive deduplication
         seen_texts = {e["text"].lower() for e in llm_entities}
         for user_entity in user_entities:
             if user_entity["text"].lower() not in seen_texts:
@@ -80,8 +46,48 @@ async def process_entities_batch(
 
         entities_per_fact.append(llm_entities)
 
-    # Use existing link_utils function for entity processing
-    entity_links = await link_utils.extract_entities_batch_optimized(
+    return fact_texts, fact_dates, entities_per_fact
+
+
+async def resolve_entities(
+    entity_resolver,
+    conn,
+    bank_id: str,
+    unit_ids: list[str],
+    facts: list[ProcessedFact],
+    log_buffer: list[str] = None,
+    user_entities_per_content: dict[int, list[dict]] = None,
+    entity_labels: list | None = None,
+) -> tuple[list[str], list[tuple], dict[str, list[str]]]:
+    """
+    Phase 1: Resolve entity names to canonical IDs (read-heavy).
+
+    Should be called on a SEPARATE connection OUTSIDE the main write transaction
+    to avoid holding the transaction open during expensive trigram scans.
+
+    Args:
+        entity_resolver: EntityResolver instance
+        conn: Database connection (separate from the main write transaction)
+        bank_id: Bank identifier
+        unit_ids: Placeholder unit IDs (used only for grouping)
+        facts: List of ProcessedFact objects
+        log_buffer: Optional buffer for detailed logging
+        user_entities_per_content: Dict mapping content_index to user-provided entities
+        entity_labels: Optional entity label taxonomy
+
+    Returns:
+        Tuple of (resolved_entity_ids, entity_to_unit, unit_to_entity_ids)
+        to pass to build_entity_links().
+    """
+    if not unit_ids or not facts:
+        return [], [], {}
+
+    if len(unit_ids) != len(facts):
+        raise ValueError(f"Mismatch between unit_ids ({len(unit_ids)}) and facts ({len(facts)})")
+
+    fact_texts, fact_dates, entities_per_fact = _prepare_facts_for_entity_processing(facts, user_entities_per_content)
+
+    return await link_utils.resolve_entities_only(
         entity_resolver,
         conn,
         bank_id,
@@ -90,11 +96,55 @@ async def process_entities_batch(
         "",  # context (not used in current implementation)
         fact_dates,
         entities_per_fact,
-        log_buffer,  # Pass log_buffer for detailed logging
+        log_buffer,
         entity_labels=entity_labels,
     )
 
-    return entity_links
+
+async def build_entity_links(
+    entity_resolver,
+    conn,
+    bank_id: str,
+    unit_ids: list[str],
+    resolved_entity_ids: list[str],
+    entity_to_unit: list[tuple],
+    unit_to_entity_ids: dict[str, list[str]],
+    log_buffer: list[str] = None,
+    skip_unit_entities_insert: bool = False,
+) -> list[EntityLink]:
+    """
+    Build entity links for UI graph visualization.
+
+    Queries unit_entities to find shared entities between new and existing units,
+    then generates EntityLink objects. When called from Phase 3 (post-transaction),
+    set skip_unit_entities_insert=True since unit_entities were already inserted
+    in Phase 2.
+
+    Args:
+        entity_resolver: EntityResolver instance
+        conn: Database connection
+        bank_id: Bank identifier
+        unit_ids: Actual unit IDs (must already be inserted in the DB)
+        resolved_entity_ids: From resolve_entities()
+        entity_to_unit: From resolve_entities()
+        unit_to_entity_ids: From resolve_entities()
+        log_buffer: Optional buffer for detailed logging
+        skip_unit_entities_insert: Skip unit_entities INSERT (already done in Phase 2)
+
+    Returns:
+        List of EntityLink objects for batch insertion
+    """
+    return await link_utils.build_entity_links_from_resolved(
+        entity_resolver,
+        conn,
+        bank_id,
+        unit_ids,
+        resolved_entity_ids,
+        entity_to_unit,
+        unit_to_entity_ids,
+        log_buffer,
+        skip_unit_entities_insert=skip_unit_entities_insert,
+    )
 
 
 async def insert_entity_links_batch(conn, entity_links: list[EntityLink], bank_id: str) -> None:

hindsight-api-slim/hindsight_api/engine/retain/fact_extraction.py

@@ -87,7 +87,7 @@ class Fact(BaseModel):
 
     # Required fields
     fact: str = Field(description="Combined fact text: what | when | where | who | why")
-    fact_type: Literal["world", "experience", "opinion"] = Field(description="Perspective: world/experience/opinion")
+    fact_type: Literal["world", "experience"] = Field(description="Perspective: world/experience")
 
     # Optional temporal fields
     occurred_start: str | None = None
@@ -909,6 +909,7 @@ def _build_user_message(
     event_date: datetime | None,
     context: str,
     metadata: dict[str, str] | None = None,
+    agent_name: str | None = None,
 ) -> str:
     """Build user message for fact extraction."""
     from .orchestrator import parse_datetime_flexible
@@ -927,11 +928,15 @@ def _build_user_message(
         metadata_lines = "\n".join(f"  {k}: {v}" for k, v in metadata.items())
         metadata_section = f"\nMetadata:\n{metadata_lines}"
 
+    narrator_section = ""
+    if agent_name:
+        narrator_section = f'\nNarrator: {agent_name} (AI agent — first-person statements like "I did X" are the agent\'s own actions; classify as "assistant")'
+
     return f"""Extract facts from the following text chunk.
 
 Chunk: {chunk_index + 1}/{total_chunks}
 Event Date: {event_date_str}
-Context: {sanitized_context}{metadata_section}
+Context: {sanitized_context}{metadata_section}{narrator_section}
 
 Text:
 {sanitized_chunk}"""
@@ -995,7 +1000,7 @@ async def _extract_facts_from_chunk(
     extract_causal_links = config.retain_extract_causal_links
 
     # Build user message using helper function
-    user_message = _build_user_message(chunk, chunk_index, total_chunks, event_date, context, metadata)
+    user_message = _build_user_message(chunk, chunk_index, total_chunks, event_date, context, metadata, agent_name)
 
     # Retry logic for JSON validation errors
     # Use retain-specific overrides if set, otherwise fall back to global LLM config
@@ -1055,7 +1060,7 @@ async def _extract_facts_from_chunk(
                     f"LLM response missing 'facts' field or returned empty list. "
                     f"Response: {extraction_response_json}. "
                     f"Input: "
-                    f"date: {event_date.isoformat()}, "
+                    f"date: {event_date.isoformat() if event_date else 'unset'}, "
                     f"context: {context if context else 'none'}, "
                     f"text: {chunk}"
                 )
@@ -1464,28 +1469,76 @@ async def extract_facts_from_text(
             f"chunk_size={config.retain_chunk_size:,}) - starting parallel LLM extraction"
         )
 
-    tasks = [
-        _extract_facts_with_auto_split(
-            chunk=chunk,
-            chunk_index=i,
-            total_chunks=len(chunks),
-            event_date=event_date,
-            context=context,
-            llm_config=llm_config,
-            config=config,
-            agent_name=agent_name,
-            metadata=metadata,
-        )
-        for i, chunk in enumerate(chunks)
-    ]
-    chunk_results = await asyncio.gather(*tasks)
+    # Per-chunk retry wrapper: each chunk gets up to MAX_CHUNK_RETRIES attempts.
+    # This handles transient LLM failures (timeouts, rate limits, malformed responses)
+    # without discarding the entire batch. If a chunk still fails after all retries,
+    # the ENTIRE retain fails — we do not accept partial extraction.
+    MAX_CHUNK_RETRIES = 3
+    CHUNK_RETRY_BASE_DELAY = 2.0  # seconds, doubles each retry
+
+    async def _extract_chunk_with_retry(chunk: str, chunk_index: int) -> tuple:
+        """Extract facts from a single chunk with retries on failure."""
+        last_exception = None
+        for attempt in range(MAX_CHUNK_RETRIES):
+            try:
+                return await _extract_facts_with_auto_split(
+                    chunk=chunk,
+                    chunk_index=chunk_index,
+                    total_chunks=len(chunks),
+                    event_date=event_date,
+                    context=context,
+                    llm_config=llm_config,
+                    config=config,
+                    agent_name=agent_name,
+                    metadata=metadata,
+                )
+            except Exception as e:
+                last_exception = e
+                if attempt < MAX_CHUNK_RETRIES - 1:
+                    delay = CHUNK_RETRY_BASE_DELAY * (2**attempt)
+                    logger.warning(
+                        f"Chunk {chunk_index}/{len(chunks)} extraction failed "
+                        f"(attempt {attempt + 1}/{MAX_CHUNK_RETRIES}): "
+                        f"{type(e).__name__}. Retrying in {delay:.0f}s..."
+                    )
+                    await asyncio.sleep(delay)
+                else:
+                    logger.error(
+                        f"Chunk {chunk_index}/{len(chunks)} extraction failed after "
+                        f"{MAX_CHUNK_RETRIES} attempts: {type(e).__name__}: {e}"
+                    )
+        raise last_exception
+
+    tasks = [_extract_chunk_with_retry(chunk, i) for i, chunk in enumerate(chunks)]
+
+    # return_exceptions=True so we can collect all results even if some chunks
+    # exhausted their retries. We check for failures below and fail the retain
+    # if ANY chunk could not be extracted — partial extraction is not acceptable.
+    chunk_results = await asyncio.gather(*tasks, return_exceptions=True)
+
     all_facts = []
     chunk_metadata = []  # [(chunk_text, fact_count), ...]
     total_usage = TokenUsage()
-    for chunk, (chunk_facts, chunk_usage) in zip(chunks, chunk_results):
+    failed_chunks = []
+    for i, (chunk, result) in enumerate(zip(chunks, chunk_results)):
+        if isinstance(result, Exception):
+            failed_chunks.append((i, result))
+            continue
+        chunk_facts, chunk_usage = result
         all_facts.extend(chunk_facts)
         chunk_metadata.append((chunk, len(chunk_facts)))
         total_usage = total_usage + chunk_usage
+
+    if failed_chunks:
+        # Fail the entire retain — partial extraction is not acceptable.
+        # All successfully extracted facts are discarded because the transaction
+        # hasn't committed yet. The worker poller will retry the entire task.
+        failed_summary = ", ".join(f"chunk {idx}: {type(err).__name__}" for idx, err in failed_chunks[:5])
+        raise RuntimeError(
+            f"Fact extraction failed: {len(failed_chunks)}/{len(chunks)} chunks failed "
+            f"after {MAX_CHUNK_RETRIES} retries each. First failures: {failed_summary}"
+        )
+
     return all_facts, chunk_metadata, total_usage
 
 
@@ -1584,7 +1637,13 @@ async def extract_facts_from_contents_batch_api(
 
             # Build user message using helper function
             user_message = _build_user_message(
-                chunk, chunk_index_in_content, len(chunks), item.event_date, item.context, item.metadata or None
+                chunk,
+                chunk_index_in_content,
+                len(chunks),
+                item.event_date,
+                item.context,
+                item.metadata or None,
+                agent_name,
             )
 
             # Build request body using helper function
@@ -1919,7 +1978,7 @@ async def extract_facts_from_contents_batch_api(
         for fact_from_llm in chunk_facts:
             extracted_fact = ExtractedFactType(
                 fact_text=fact_from_llm.fact,
-                fact_type="experience" if fact_from_llm.fact_type == "assistant" else "world",
+                fact_type=fact_from_llm.fact_type,
                 entities=[e.text for e in (fact_from_llm.entities or [])],
                 occurred_start=_parse_datetime(fact_from_llm.occurred_start) if fact_from_llm.occurred_start else None,
                 occurred_end=_parse_datetime(fact_from_llm.occurred_end) if fact_from_llm.occurred_end else None,
@@ -2055,8 +2114,9 @@ async def extract_facts_from_contents(
         )
         fact_extraction_tasks.append(task)
 
-    # Step 2: Wait for all fact extractions to complete
-    all_fact_results = await asyncio.gather(*fact_extraction_tasks)
+    # Step 2: Wait for all fact extractions to complete.
+    # Use return_exceptions=True so one content item failure doesn't discard the rest.
+    all_fact_results = await asyncio.gather(*fact_extraction_tasks, return_exceptions=True)
 
     # Step 3: Flatten and convert to typed objects
     extracted_facts: list[ExtractedFactType] = []
@@ -2066,9 +2126,16 @@ async def extract_facts_from_contents(
     global_chunk_idx = 0
     global_fact_idx = 0
 
-    for content_index, (content, (facts_from_llm, chunks_from_llm, content_usage)) in enumerate(
-        zip(contents, all_fact_results)
-    ):
+    # Filter out failed content items
+    valid_results = []
+    for content, result in zip(contents, all_fact_results):
+        if isinstance(result, Exception):
+            logger.warning(f"Content extraction failed (skipping): {type(result).__name__}: {result}")
+            valid_results.append((content, ([], [], TokenUsage())))
+        else:
+            valid_results.append((content, result))
+
+    for content_index, (content, (facts_from_llm, chunks_from_llm, content_usage)) in enumerate(valid_results):
         total_usage = total_usage + content_usage
         chunk_start_idx = global_chunk_idx
 
@@ -2096,7 +2163,7 @@ async def extract_facts_from_contents(
                     # mentioned_at is always the event_date (when the conversation/document occurred)
                     extracted_fact = ExtractedFactType(
                         fact_text=fact_from_llm.fact,
-                        fact_type="experience" if fact_from_llm.fact_type == "assistant" else "world",
+                        fact_type=fact_from_llm.fact_type,
                         entities=[e.text for e in (fact_from_llm.entities or [])],
                         # occurred_start/end: from LLM only, leave None if not provided
                         occurred_start=_parse_datetime(fact_from_llm.occurred_start)

hindsight-api-slim/hindsight_api/engine/retain/fact_storage.py

@@ -7,6 +7,7 @@ Handles insertion of facts into the database.
 import json
 import logging
 import uuid
+from datetime import datetime
 
 from ...config import get_config
 from ..memory_engine import fq_table
@@ -17,6 +18,23 @@ from .types import ProcessedFact
 logger = logging.getLogger(__name__)
 
 
+async def get_document_content(
+    conn,
+    bank_id: str,
+    document_id: str,
+) -> str | None:
+    """Fetch the original_text of an existing document.
+
+    Returns None if the document does not exist.
+    """
+    row = await conn.fetchval(
+        f"SELECT original_text FROM {fq_table('documents')} WHERE id = $1 AND bank_id = $2",
+        document_id,
+        bank_id,
+    )
+    return row
+
+
 async def insert_facts_batch(
     conn, bank_id: str, facts: list[ProcessedFact], document_id: str | None = None
 ) -> list[str]:
@@ -44,7 +62,6 @@ async def insert_facts_batch(
     mentioned_ats = []
     contexts = []
     fact_types = []
-    confidence_scores = []
     metadata_jsons = []
     chunk_ids = []
     document_ids = []
@@ -64,8 +81,6 @@ async def insert_facts_batch(
         mentioned_ats.append(fact.mentioned_at)
         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)
         metadata_jsons.append(json.dumps(fact.metadata))
         chunk_ids.append(fact.chunk_id)
         # Use per-fact document_id if available, otherwise fallback to batch-level document_id
@@ -103,18 +118,18 @@ async def insert_facts_batch(
             WITH input_data AS (
                 SELECT * FROM unnest(
                     $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
-                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[], $15::jsonb[], $16::text[]
+                    $8::text[], $9::text[], $10::jsonb[], $11::text[], $12::text[], $13::jsonb[], $14::jsonb[], $15::text[]
                 ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json,
+                       context, fact_type, metadata, chunk_id, document_id, tags_json,
                        observation_scopes_json, text_signals)
             )
             INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags,
+                                     context, fact_type, metadata, chunk_id, document_id, tags,
                                      observation_scopes, text_signals, search_vector)
             SELECT
                 $1,
                 text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                context, fact_type, confidence_score, metadata, chunk_id, document_id,
+                context, fact_type, metadata, chunk_id, document_id,
                 COALESCE(
                     (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
                     '{{}}'::varchar[]
@@ -135,18 +150,18 @@ async def insert_facts_batch(
             WITH input_data AS (
                 SELECT * FROM unnest(
                     $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
-                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[], $15::jsonb[], $16::text[]
+                    $8::text[], $9::text[], $10::jsonb[], $11::text[], $12::text[], $13::jsonb[], $14::jsonb[], $15::text[]
                 ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json,
+                       context, fact_type, metadata, chunk_id, document_id, tags_json,
                        observation_scopes_json, text_signals)
             )
             INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags,
+                                     context, fact_type, metadata, chunk_id, document_id, tags,
                                      observation_scopes, text_signals)
             SELECT
                 $1,
                 text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                context, fact_type, confidence_score, metadata, chunk_id, document_id,
+                context, fact_type, metadata, chunk_id, document_id,
                 COALESCE(
                     (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
                     '{{}}'::varchar[]
@@ -168,7 +183,6 @@ async def insert_facts_batch(
         mentioned_ats,
         contexts,
         fact_types,
-        confidence_scores,
         metadata_jsons,
         chunk_ids,
         document_ids,
@@ -211,6 +225,85 @@ async def ensure_bank_exists(conn, bank_id: str) -> None:
         await create_bank_vector_indexes(conn, bank_id, str(internal_id))
 
 
+async def delete_stale_observations_for_memories(
+    conn,
+    bank_id: str,
+    fact_ids: "list[str | uuid.UUID]",
+) -> int:
+    """Delete observations whose source memories are about to be removed.
+
+    Mirrors the cleanup performed by ``MemoryEngine.delete_document`` so that
+    every code path that removes ``memory_units`` also removes the
+    observations derived from them. Without this, ingesting a fresh version
+    of a document via the retain pipeline (which does a full-replace
+    ``DELETE FROM documents`` cascade) used to leave orphan observations
+    pointing at memory IDs that no longer existed.
+
+    For each observation referencing any of ``fact_ids``:
+    1. Delete the observation row (its text is stale once even one source
+       memory disappears).
+    2. Reset ``consolidated_at = NULL`` on the surviving source memories so
+       they get re-consolidated under fresh observations on the next run.
+
+    Must be called within an active transaction, before the source memories
+    are deleted.
+
+    Returns the number of observations deleted.
+    """
+    if not fact_ids:
+        return 0
+
+    fact_uuids = [uuid.UUID(str(fid)) if not isinstance(fid, uuid.UUID) else fid for fid in fact_ids]
+
+    affected_obs = await conn.fetch(
+        f"""
+        SELECT id, source_memory_ids
+        FROM {fq_table("memory_units")}
+        WHERE bank_id = $1
+          AND fact_type = 'observation'
+          AND source_memory_ids && $2::uuid[]
+        """,
+        bank_id,
+        fact_uuids,
+    )
+
+    if not affected_obs:
+        return 0
+
+    deleted_set = {str(uid) for uid in fact_uuids}
+    obs_ids = [obs["id"] for obs in affected_obs]
+    seen_remaining: set[str] = set()
+    remaining_source_ids: list[uuid.UUID] = []
+    for obs in affected_obs:
+        for src_id in obs["source_memory_ids"] or []:
+            src_str = str(src_id)
+            if src_str not in deleted_set and src_str not in seen_remaining:
+                remaining_source_ids.append(src_id)
+                seen_remaining.add(src_str)
+
+    await conn.execute(
+        f"DELETE FROM {fq_table('memory_units')} WHERE id = ANY($1::uuid[])",
+        obs_ids,
+    )
+
+    if remaining_source_ids:
+        await conn.execute(
+            f"""
+            UPDATE {fq_table("memory_units")}
+            SET consolidated_at = NULL
+            WHERE id = ANY($1::uuid[])
+              AND fact_type IN ('experience', 'world')
+            """,
+            remaining_source_ids,
+        )
+
+    logger.info(
+        f"[OBSERVATIONS] Deleted {len(obs_ids)} observations, reset {len(remaining_source_ids)} "
+        f"source memories for re-consolidation in bank {bank_id}"
+    )
+    return len(obs_ids)
+
+
 async def handle_document_tracking(
     conn,
     bank_id: str,
@@ -241,17 +334,58 @@ async def handle_document_tracking(
     combined_content = _sanitize_text(combined_content) or ""
     content_hash = hashlib.sha256(combined_content.encode()).hexdigest()
 
-    # Delete old document first (cascades to units and links)
-    # Only delete on the first batch to avoid deleting data we just inserted
+    # Delete old document first (cascades to units and links).
+    # Only delete on the first batch to avoid deleting data we just inserted.
+    # Before the cascade, fan out to delete observations derived from the
+    # outgoing memory_units — otherwise the FK ON DELETE CASCADE removes the
+    # source memory_units but leaves observation rows pointing at IDs that
+    # no longer exist (consolidated_at on co-source memories also stays
+    # frozen). Same cleanup the explicit ``delete_document`` API performs.
+    preserved_created_at = None
     if is_first_batch:
-        await conn.fetchval(
-            f"DELETE FROM {fq_table('documents')} WHERE id = $1 AND bank_id = $2 RETURNING id",
+        existing_unit_rows = await conn.fetch(
+            f"""
+            SELECT id FROM {fq_table("memory_units")}
+            WHERE document_id = $1 AND fact_type IN ('experience', 'world')
+            """,
+            document_id,
+        )
+        existing_unit_ids = [row["id"] for row in existing_unit_rows]
+        if existing_unit_ids:
+            invalidated = await delete_stale_observations_for_memories(conn, bank_id, existing_unit_ids)
+            if invalidated:
+                logger.info(
+                    f"[RETAIN] Document {document_id} re-ingested: invalidated "
+                    f"{invalidated} observation(s) derived from {len(existing_unit_ids)} outgoing memory_units"
+                )
+        # Explicitly delete memory_units by document_id BEFORE deleting the
+        # document row. The CASCADE from documents→chunks→memory_units only
+        # catches units that have a non-NULL chunk_id FK. Units with chunk_id=NULL
+        # (e.g. from partial writes or edge cases) would survive the cascade.
+        # This explicit delete ensures complete cleanup.
+        await conn.execute(
+            f"DELETE FROM {fq_table('memory_units')} WHERE document_id = $1 AND bank_id = $2",
+            document_id,
+            bank_id,
+        )
+        # Capture created_at before deletion so re-ingestion preserves it.
+        preserved_created_at = await conn.fetchval(
+            f"DELETE FROM {fq_table('documents')} WHERE id = $1 AND bank_id = $2 RETURNING created_at",
             document_id,
             bank_id,
         )
 
     # Insert document (or update if exists from concurrent operations)
-    await _upsert_document_row(conn, bank_id, document_id, combined_content, content_hash, retain_params, document_tags)
+    await _upsert_document_row(
+        conn,
+        bank_id,
+        document_id,
+        combined_content,
+        content_hash,
+        retain_params,
+        document_tags,
+        preserved_created_at=preserved_created_at,
+    )
 
 
 async def upsert_document_metadata(
@@ -284,16 +418,22 @@ async def _upsert_document_row(
     content_hash: str,
     retain_params: dict | None = None,
     document_tags: list[str] | None = None,
+    preserved_created_at: datetime | None = None,
 ) -> None:
-    """Insert or update a document row."""
+    """Insert or update a document row.
+
+    When ``preserved_created_at`` is provided, it is used for ``created_at`` on
+    INSERT so that re-ingesting a document (which deletes + inserts the row)
+    keeps the original creation timestamp. ``updated_at`` is always set to
+    ``NOW()`` on both INSERT and the ON CONFLICT UPDATE branch.
+    """
     await conn.execute(
         f"""
-        INSERT INTO {fq_table("documents")} (id, bank_id, original_text, content_hash, metadata, retain_params, tags)
-        VALUES ($1, $2, $3, $4, $5, $6, $7)
+        INSERT INTO {fq_table("documents")} (id, bank_id, original_text, content_hash, retain_params, tags, created_at, updated_at)
+        VALUES ($1, $2, $3, $4, $5, $6, COALESCE($7, NOW()), NOW())
         ON CONFLICT (id, bank_id) DO UPDATE
         SET original_text = EXCLUDED.original_text,
             content_hash = EXCLUDED.content_hash,
-            metadata = EXCLUDED.metadata,
             retain_params = EXCLUDED.retain_params,
             tags = EXCLUDED.tags,
             updated_at = NOW()
@@ -302,9 +442,9 @@ async def _upsert_document_row(
         bank_id,
         combined_content,
         content_hash,
-        json.dumps({}),  # Empty metadata dict
         json.dumps(retain_params) if retain_params else None,
         document_tags or [],
+        preserved_created_at,
     )
 
 

hindsight-api-slim/hindsight_api/engine/retain/link_creation.py

@@ -32,17 +32,26 @@ async def create_temporal_links_batch(conn, bank_id: str, unit_ids: list[str]) -
     return await link_utils.create_temporal_links_batch_per_fact(conn, bank_id, unit_ids, log_buffer=[])
 
 
-async def create_semantic_links_batch(conn, bank_id: str, unit_ids: list[str], embeddings: list[list[float]]) -> int:
+async def create_semantic_links_batch(
+    conn,
+    bank_id: str,
+    unit_ids: list[str],
+    embeddings: list[list[float]],
+    pre_computed_ann_links: list[tuple] | None = None,
+) -> int:
     """
     Create semantic links between facts.
 
     Links facts that are semantically similar based on embeddings.
+    When pre_computed_ann_links are provided (from Phase 1), they are used
+    instead of running ANN queries inside the transaction.
 
     Args:
         conn: Database connection
         bank_id: Bank identifier
         unit_ids: List of unit IDs to create links for
         embeddings: List of embedding vectors (same length as unit_ids)
+        pre_computed_ann_links: Pre-computed ANN results from Phase 1
 
     Returns:
         Number of semantic links created
@@ -53,7 +62,9 @@ async def create_semantic_links_batch(conn, bank_id: str, unit_ids: list[str], e
     if len(unit_ids) != len(embeddings):
         raise ValueError(f"Mismatch between unit_ids ({len(unit_ids)}) and embeddings ({len(embeddings)})")
 
-    return await link_utils.create_semantic_links_batch(conn, bank_id, unit_ids, embeddings, log_buffer=[])
+    return await link_utils.create_semantic_links_batch(
+        conn, bank_id, unit_ids, embeddings, log_buffer=[], pre_computed_ann_links=pre_computed_ann_links
+    )
 
 
 async def create_causal_links_batch(conn, bank_id: str, unit_ids: list[str], facts: list[ProcessedFact]) -> int:

hindsight-api-slim/hindsight_api/engine/retain/types.py

@@ -25,6 +25,9 @@ class RetainContentDict(TypedDict, total=False):
         observation_scopes: How to scope observations for consolidation (optional).
             "per_tag" runs one pass per individual tag; "combined" (default) runs a
             single pass with all tags; a list[list[str]] specifies exact passes.
+        update_mode: How to handle existing documents with the same document_id (optional).
+            "replace" (default) deletes old data and reprocesses. "append" concatenates
+            new content to the existing document and reprocesses.
     """
 
     content: str  # Required
@@ -37,6 +40,7 @@ class RetainContentDict(TypedDict, total=False):
     observation_scopes: (
         Literal["per_tag", "combined", "all_combinations"] | list[list[str]]
     )  # Observation scopes for consolidation
+    update_mode: Literal["replace", "append"]
 
 
 @dataclass
@@ -107,7 +111,7 @@ class ExtractedFact:
     """
 
     fact_text: str
-    fact_type: str  # "world", "experience", "opinion", "observation"
+    fact_type: str  # "world", "experience", "observation"
     entities: list[str] = field(default_factory=list)
     occurred_start: datetime | None = None
     occurred_end: datetime | None = None
@@ -221,6 +225,45 @@ class ProcessedFact:
         )
 
 
+@dataclass
+class Phase3Context:
+    """
+    Data passed from Phase 2 to Phase 3 for entity link building.
+
+    Contains the unit IDs and entity resolution data needed to build
+    entity links for UI graph visualization after the write transaction commits.
+    """
+
+    unit_ids: list[str] = field(default_factory=list)
+    resolved_entity_ids: list[str] = field(default_factory=list)
+    entity_to_unit: list[tuple] = field(default_factory=list)
+    unit_to_entity_ids: dict[str, list[str]] = field(default_factory=dict)
+
+
+@dataclass
+class EntityResolutionResult:
+    """
+    Result of Phase 1 entity resolution.
+
+    Contains resolved entity IDs and the mapping data needed to remap
+    placeholder unit IDs to real IDs after fact insertion in Phase 2.
+    """
+
+    resolved_entity_ids: list[str]
+    entity_to_unit: list[tuple]
+    unit_to_entity_ids: dict[str, list[str]]
+
+
+@dataclass
+class Phase1Result:
+    """
+    Full result of Phase 1 (entity resolution + optional semantic ANN).
+    """
+
+    entities: EntityResolutionResult
+    semantic_ann_links: list[tuple]
+
+
 @dataclass
 class EntityLink:
     """
@@ -248,7 +291,6 @@ class RetainBatch:
     contents: list[RetainContent]
     document_id: str | None = None
     fact_type_override: str | None = None
-    confidence_score: float | None = None
     document_tags: list[str] = field(default_factory=list)  # Tags applied to all items
 
     # Extracted data (populated during processing)

hindsight-api-slim/hindsight_api/engine/search/__init__.py

@@ -3,12 +3,11 @@ Search module for memory retrieval.
 
 Provides modular search architecture:
 - Retrieval: 4-way parallel (semantic + BM25 + graph + temporal)
-- Graph retrieval: Pluggable strategies (BFS, PPR)
+- Graph retrieval: Link expansion strategy
 - Reranking: Pluggable strategies (heuristic, cross-encoder)
 """
 
-from .graph_retrieval import BFSGraphRetriever, GraphRetriever
-from .mpfp_retrieval import MPFPGraphRetriever
+from .graph_retrieval import GraphRetriever
 from .reranking import CrossEncoderReranker
 from .retrieval import (
     ParallelRetrievalResult,
@@ -21,7 +20,5 @@ __all__ = [
     "set_default_graph_retriever",
     "ParallelRetrievalResult",
     "GraphRetriever",
-    "BFSGraphRetriever",
-    "MPFPGraphRetriever",
     "CrossEncoderReranker",
 ]

hindsight-api-slim/hindsight_api/engine/search/graph_retrieval.py

@@ -2,17 +2,16 @@
 Graph retrieval strategies for memory recall.
 
 This module provides an abstraction for graph-based memory retrieval,
-allowing different algorithms (BFS spreading activation, PPR, etc.) to be
-swapped without changing the rest of the recall pipeline.
+allowing different algorithms to be swapped without changing the rest
+of the recall pipeline.
 """
 
 import logging
 from abc import ABC, abstractmethod
+from datetime import datetime
 
-from ..db_utils import acquire_with_retry
-from ..memory_engine import fq_table
-from .tags import TagGroup, TagsMatch, filter_results_by_tag_groups, filter_results_by_tags
-from .types import MPFPTimings, RetrievalResult
+from .tags import TagGroup, TagsMatch
+from .types import GraphRetrievalTimings, RetrievalResult
 
 logger = logging.getLogger(__name__)
 
@@ -29,7 +28,7 @@ class GraphRetriever(ABC):
     @property
     @abstractmethod
     def name(self) -> str:
-        """Return identifier for this retrieval strategy (e.g., 'bfs', 'mpfp')."""
+        """Return identifier for this retrieval strategy (e.g., 'link_expansion')."""
         pass
 
     @abstractmethod
@@ -47,7 +46,9 @@ class GraphRetriever(ABC):
         tags: list[str] | None = None,  # Visibility scope tags for filtering
         tags_match: TagsMatch = "any",  # How to match tags: 'any' (OR) or 'all' (AND)
         tag_groups: list[TagGroup] | None = None,  # Compound boolean tag filter groups
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
+        created_after: datetime | None = None,  # Only include memory_units created after this time
+        created_before: datetime | None = None,  # Only include memory_units created before this time
+    ) -> tuple[list[RetrievalResult], GraphRetrievalTimings | None]:
         """
         Retrieve relevant facts via graph traversal.
 
@@ -55,228 +56,15 @@ class GraphRetriever(ABC):
             pool: Database connection pool
             query_embedding_str: Query embedding as string (for finding entry points)
             bank_id: Memory bank identifier
-            fact_type: Fact type to filter ('world', 'experience', 'opinion', 'observation')
+            fact_type: Fact type to filter ('world', 'experience', 'observation')
             budget: Maximum number of nodes to explore/return
             query_text: Original query text (optional, for some strategies)
             semantic_seeds: Pre-computed semantic entry points (from semantic retrieval)
             temporal_seeds: Pre-computed temporal entry points (from temporal retrieval)
-            adjacency: Pre-loaded typed adjacency graph (optional, for MPFP)
+            adjacency: Pre-loaded typed adjacency graph (optional)
             tags: Optional list of tags for visibility filtering (OR matching)
 
         Returns:
             Tuple of (List of RetrievalResult with activation scores, optional timing info)
         """
         pass
-
-
-class BFSGraphRetriever(GraphRetriever):
-    """
-    Graph retrieval using BFS-style spreading activation.
-
-    Starting from semantic entry points, spreads activation through
-    the memory graph (entity, temporal, causal links) using breadth-first
-    traversal with decaying activation.
-
-    This is the original Hindsight graph retrieval algorithm.
-    """
-
-    def __init__(
-        self,
-        entry_point_limit: int = 5,
-        entry_point_threshold: float = 0.5,
-        activation_decay: float = 0.8,
-        min_activation: float = 0.1,
-        batch_size: int = 20,
-    ):
-        """
-        Initialize BFS graph retriever.
-
-        Args:
-            entry_point_limit: Maximum number of entry points to start from
-            entry_point_threshold: Minimum semantic similarity for entry points
-            activation_decay: Decay factor per hop (activation *= decay)
-            min_activation: Minimum activation to continue spreading
-            batch_size: Number of nodes to process per batch (for neighbor fetching)
-        """
-        self.entry_point_limit = entry_point_limit
-        self.entry_point_threshold = entry_point_threshold
-        self.activation_decay = activation_decay
-        self.min_activation = min_activation
-        self.batch_size = batch_size
-
-    @property
-    def name(self) -> str:
-        return "bfs"
-
-    async def retrieve(
-        self,
-        pool,
-        query_embedding_str: str,
-        bank_id: str,
-        fact_type: str,
-        budget: int,
-        query_text: str | None = None,
-        semantic_seeds: list[RetrievalResult] | None = None,
-        temporal_seeds: list[RetrievalResult] | None = None,
-        adjacency=None,  # Not used by BFS
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
-        tag_groups: list[TagGroup] | None = None,
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
-        """
-        Retrieve facts using BFS spreading activation.
-
-        Algorithm:
-        1. Find entry points (top semantic matches above threshold)
-        2. BFS traversal: visit neighbors, propagate decaying activation
-        3. Boost causal links (causes, enables, prevents)
-        4. Return visited nodes up to budget
-
-        Note: BFS finds its own entry points via embedding search.
-        The semantic_seeds, temporal_seeds, and adjacency parameters are accepted
-        for interface compatibility but not used.
-        """
-        async with acquire_with_retry(pool) as conn:
-            results = await self._retrieve_with_conn(
-                conn,
-                query_embedding_str,
-                bank_id,
-                fact_type,
-                budget,
-                tags=tags,
-                tags_match=tags_match,
-                tag_groups=tag_groups,
-            )
-            return results, None
-
-    async def _retrieve_with_conn(
-        self,
-        conn,
-        query_embedding_str: str,
-        bank_id: str,
-        fact_type: str,
-        budget: int,
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
-        tag_groups: list[TagGroup] | None = None,
-    ) -> list[RetrievalResult]:
-        """Internal implementation with connection."""
-        from .tags import build_tag_groups_where_clause, build_tags_where_clause_simple
-
-        tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
-        tag_groups_param_start = 6 + (1 if tags else 0)
-        groups_clause, groups_params, _ = build_tag_groups_where_clause(tag_groups, tag_groups_param_start)
-        params = [query_embedding_str, bank_id, fact_type, self.entry_point_threshold, self.entry_point_limit]
-        if tags:
-            params.append(tags)
-        params.extend(groups_params)
-
-        # Step 1: Find entry points
-        entry_points = await conn.fetch(
-            f"""
-            SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, fact_type, document_id, chunk_id, tags,
-                   1 - (embedding <=> $1::vector) AS similarity
-            FROM {fq_table("memory_units")}
-            WHERE bank_id = $2
-              AND embedding IS NOT NULL
-              AND fact_type = $3
-              AND (1 - (embedding <=> $1::vector)) >= $4
-              {tags_clause}
-              {groups_clause}
-            ORDER BY embedding <=> $1::vector
-            LIMIT $5
-            """,
-            *params,
-        )
-
-        if not entry_points:
-            logger.debug(
-                f"[BFS] No entry points found for fact_type={fact_type} (tags={tags}, tags_match={tags_match})"
-            )
-            return []
-
-        logger.debug(
-            f"[BFS] Found {len(entry_points)} entry points for fact_type={fact_type} "
-            f"(tags={tags}, tags_match={tags_match})"
-        )
-
-        # Step 2: BFS spreading activation
-        visited = set()
-        results = []
-        queue = [(RetrievalResult.from_db_row(dict(r)), r["similarity"]) for r in entry_points]
-        budget_remaining = budget
-
-        while queue and budget_remaining > 0:
-            # Collect a batch of nodes to process
-            batch_nodes = []
-            batch_activations = {}
-
-            while queue and len(batch_nodes) < self.batch_size and budget_remaining > 0:
-                current, activation = queue.pop(0)
-                unit_id = current.id
-
-                if unit_id not in visited:
-                    visited.add(unit_id)
-                    budget_remaining -= 1
-                    current.activation = activation
-                    results.append(current)
-                    batch_nodes.append(current.id)
-                    batch_activations[unit_id] = activation
-
-            # Batch fetch neighbors
-            if batch_nodes and budget_remaining > 0:
-                max_neighbors = len(batch_nodes) * 20
-                neighbors = await conn.fetch(
-                    f"""
-                    SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.occurred_end,
-                           mu.mentioned_at, mu.fact_type,
-                           mu.document_id, mu.chunk_id, mu.tags, mu.metadata,
-                           ml.weight, ml.link_type, ml.from_unit_id
-                    FROM {fq_table("memory_links")} ml
-                    JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
-                    WHERE ml.from_unit_id = ANY($1::uuid[])
-                      AND ml.weight >= $2
-                      AND mu.fact_type = $3
-                    ORDER BY ml.weight DESC
-                    LIMIT $4
-                    """,
-                    batch_nodes,
-                    self.min_activation,
-                    fact_type,
-                    max_neighbors,
-                )
-
-                for n in neighbors:
-                    neighbor_id = str(n["id"])
-                    if neighbor_id not in visited:
-                        parent_id = str(n["from_unit_id"])
-                        parent_activation = batch_activations.get(parent_id, 0.5)
-
-                        # Boost causal links
-                        link_type = n["link_type"]
-                        base_weight = n["weight"]
-
-                        if link_type in ("causes", "caused_by"):
-                            causal_boost = 2.0
-                        elif link_type in ("enables", "prevents"):
-                            causal_boost = 1.5
-                        else:
-                            causal_boost = 1.0
-
-                        effective_weight = base_weight * causal_boost
-                        new_activation = parent_activation * effective_weight * self.activation_decay
-
-                        if new_activation > self.min_activation:
-                            neighbor_result = RetrievalResult.from_db_row(dict(n))
-                            queue.append((neighbor_result, new_activation))
-
-        # Apply tags filtering (BFS may traverse into memories that don't match tags criteria)
-        if tags:
-            results = filter_results_by_tags(results, tags, match=tags_match)
-
-        # Apply compound tag group filtering (post-traversal)
-        if tag_groups:
-            results = filter_results_by_tag_groups(results, tag_groups)
-
-        return results

hindsight-api-slim/hindsight_api/engine/search/link_expansion_retrieval.py

@@ -4,32 +4,39 @@ Link Expansion graph retrieval.
 Expands from semantic/temporal seeds through three parallel, first-class signals
 stored in memory_links:
 
-1. Entity links  — precomputed co-occurrence graph (created at retain time, bounded to
-                   MAX_LINKS_PER_ENTITY per entity). Score = number of distinct shared
-                   entities between the seed set and each candidate.
+1. Entity links  — query-time self-join through unit_entities. Score = number of distinct
+                   shared entities between the seed set and each candidate, computed via
+                   COUNT(DISTINCT entity_id). Uses a LATERAL per-entity cap
+                   (graph_per_entity_limit, default 200) to prevent high-fanout entities
+                   from exploding the self-join intermediate rows.
 2. Semantic links — precomputed kNN graph (each new fact linked to its top-5 most
                     similar existing facts at insert time, similarity >= 0.7). Checked
                     in both directions since the graph is not symmetric. Score = weight.
 3. Causal links  — explicit causal chains (causes/caused_by/enables/prevents).
                    Score = weight + 1.0 (boosted as highest-quality signal).
 
-All three signals are bounded at retain time, so no LATERAL fan-out caps are needed
-at query time. Each expansion is a simple aggregation over a small result set.
+Entity expansion is bounded by graph_per_entity_limit (LATERAL cap per entity).
+A timeout fallback (graph_expansion_timeout) drops entity expansion entirely if the
+query still exceeds the budget.
 
 For non-observation fact types the three expansions are issued as a single CTE query
 (one roundtrip, one connection) with a `source` discriminator column so the Python
 merge step can apply per-signal score transformations.
 """
 
+import asyncio
 import logging
 import math
 import time
+from datetime import datetime
+from typing import Any
 
+from ...config import get_config
 from ..db_utils import acquire_with_retry
 from ..memory_engine import fq_table
 from .graph_retrieval import GraphRetriever
 from .tags import TagGroup, TagsMatch, filter_results_by_tag_groups, filter_results_by_tags
-from .types import MPFPTimings, RetrievalResult
+from .types import GraphRetrievalTimings, RetrievalResult
 
 logger = logging.getLogger(__name__)
 
@@ -44,6 +51,8 @@ async def _find_semantic_seeds(
     tags: list[str] | None = None,
     tags_match: TagsMatch = "any",
     tag_groups: list[TagGroup] | None = None,
+    created_after: datetime | None = None,
+    created_before: datetime | None = None,
 ) -> list[RetrievalResult]:
     """Find semantic seeds via embedding search."""
     from .tags import build_tag_groups_where_clause, build_tags_where_clause_simple
@@ -51,15 +60,29 @@ async def _find_semantic_seeds(
     tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
     tag_groups_param_start = 6 + (1 if tags else 0)
     groups_clause, groups_params, _ = build_tag_groups_where_clause(tag_groups, tag_groups_param_start)
+
+    _next_idx = tag_groups_param_start + len(groups_params)
+    created_range_clause = ""
+    created_range_params: list[Any] = []
+    if created_after is not None:
+        created_range_params.append(created_after)
+        created_range_clause += f" AND updated_at > ${_next_idx}"
+        _next_idx += 1
+    if created_before is not None:
+        created_range_params.append(created_before)
+        created_range_clause += f" AND updated_at < ${_next_idx}"
+        _next_idx += 1
+
     params = [query_embedding_str, bank_id, fact_type, threshold, limit]
     if tags:
         params.append(tags)
     params.extend(groups_params)
+    params.extend(created_range_params)
 
     rows = await conn.fetch(
         f"""
         SELECT id, text, context, event_date, occurred_start, occurred_end,
-               mentioned_at, fact_type, document_id, chunk_id, tags,
+               mentioned_at, fact_type, document_id, chunk_id, tags, proof_count,
                1 - (embedding <=> $1::vector) AS similarity
         FROM {fq_table("memory_units")}
         WHERE bank_id = $2
@@ -68,6 +91,7 @@ async def _find_semantic_seeds(
           AND (1 - (embedding <=> $1::vector)) >= $4
           {tags_clause}
           {groups_clause}
+          {created_range_clause}
         ORDER BY embedding <=> $1::vector
         LIMIT $5
         """,
@@ -116,7 +140,9 @@ class LinkExpansionRetriever(GraphRetriever):
         tags: list[str] | None = None,
         tags_match: TagsMatch = "any",
         tag_groups: list[TagGroup] | None = None,
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
+        created_after: "datetime | None" = None,
+        created_before: "datetime | None" = None,
+    ) -> tuple[list[RetrievalResult], GraphRetrievalTimings | None]:
         """
         Retrieve facts by expanding links from seeds.
 
@@ -136,7 +162,7 @@ class LinkExpansionRetriever(GraphRetriever):
             Tuple of (results, timings)
         """
         start_time = time.time()
-        timings = MPFPTimings(fact_type=fact_type)
+        timings = GraphRetrievalTimings(fact_type=fact_type)
 
         async with acquire_with_retry(pool) as conn:
             # Find seeds if not provided
@@ -154,6 +180,8 @@ class LinkExpansionRetriever(GraphRetriever):
                     tags=tags,
                     tags_match=tags_match,
                     tag_groups=tag_groups,
+                    created_after=created_after,
+                    created_before=created_before,
                 )
                 timings.seeds_time = time.time() - seeds_start
                 logger.debug(
@@ -262,31 +290,48 @@ class LinkExpansionRetriever(GraphRetriever):
                     idx_memory_links_to_type_weight (to_unit_id, link_type, weight DESC)
                     → replaces costly BitmapAnd of two separate scans
         """
+        config = get_config()
         ml = fq_table("memory_links")
         mu = fq_table("memory_units")
-        all_rows = await conn.fetch(
-            f"""
-            WITH entity_expanded AS (
-                -- Entity co-occurrence: seeds → their precomputed entity-link neighbors.
-                -- Score = distinct shared entities (bounded at retain time to
-                -- MAX_LINKS_PER_ENTITY=50).  GROUP BY mu.id is sufficient because mu.id
-                -- is the primary key and functionally determines all other mu columns.
-                SELECT
-                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                    mu.occurred_end, mu.mentioned_at,
-                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                    COUNT(DISTINCT ml.entity_id)::float AS score,
-                    'entity'::text AS source
-                FROM {ml} ml
-                JOIN {mu} mu ON mu.id = ml.to_unit_id
-                WHERE ml.from_unit_id = ANY($1::uuid[])
-                  AND ml.link_type = 'entity'
-                  AND mu.fact_type = $2
-                  AND mu.id != ALL($1::uuid[])
+        ue = fq_table("unit_entities")
+
+        per_entity_limit = config.link_expansion_per_entity_limit
+
+        # Entity CTE with LATERAL fanout cap.
+        # Every seed entity (including high-frequency ones) is kept, but each
+        # entity's expansion is capped to per_entity_limit target units.  The
+        # LATERAL subquery orders by unit_id DESC so the most recently inserted
+        # units are preferred (a recency proxy that is free — it rides the PK
+        # index with no extra sort).
+        entity_cte = f"""
+            seed_entities AS (
+                SELECT DISTINCT ue.entity_id
+                FROM {ue} ue
+                WHERE ue.unit_id = ANY($1::uuid[])
+            ),
+            entity_expanded AS (
+                SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                       mu.occurred_end, mu.mentioned_at,
+                       mu.fact_type, mu.document_id, mu.chunk_id, mu.tags, mu.proof_count,
+                       COUNT(DISTINCT se.entity_id)::float AS score,
+                       'entity'::text AS source
+                FROM seed_entities se
+                CROSS JOIN LATERAL (
+                    SELECT ue_target.unit_id
+                    FROM {ue} ue_target
+                    WHERE ue_target.entity_id = se.entity_id
+                      AND ue_target.unit_id != ALL($1::uuid[])
+                    ORDER BY ue_target.unit_id DESC
+                    LIMIT {per_entity_limit}
+                ) t
+                JOIN {mu} mu ON mu.id = t.unit_id
+                WHERE mu.fact_type = $2
                 GROUP BY mu.id
                 ORDER BY score DESC
                 LIMIT $3
-            ),
+            )"""
+
+        semantic_causal_cte = f"""
             semantic_expanded AS (
                 -- Semantic kNN: both outgoing (seeds → their kNN at insert time) and
                 -- incoming (facts inserted after seeds that found seeds as kNN).
@@ -294,14 +339,14 @@ class LinkExpansionRetriever(GraphRetriever):
                 SELECT
                     id, text, context, event_date, occurred_start,
                     occurred_end, mentioned_at,
-                    fact_type, document_id, chunk_id, tags,
+                    fact_type, document_id, chunk_id, tags, proof_count,
                     MAX(weight) AS score,
                     'semantic'::text AS source
                 FROM (
                     SELECT
                         mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                         mu.occurred_end, mu.mentioned_at,
-                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags, mu.proof_count,
                         ml.weight
                     FROM {ml} ml
                     JOIN {mu} mu ON mu.id = ml.to_unit_id
@@ -313,7 +358,7 @@ class LinkExpansionRetriever(GraphRetriever):
                     SELECT
                         mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                         mu.occurred_end, mu.mentioned_at,
-                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags, mu.proof_count,
                         ml.weight
                     FROM {ml} ml
                     JOIN {mu} mu ON mu.id = ml.from_unit_id
@@ -324,7 +369,7 @@ class LinkExpansionRetriever(GraphRetriever):
                 ) sem_raw
                 GROUP BY id, text, context, event_date, occurred_start,
                          occurred_end, mentioned_at,
-                         fact_type, document_id, chunk_id, tags
+                         fact_type, document_id, chunk_id, tags, proof_count
                 ORDER BY score DESC
                 LIMIT $3
             ),
@@ -335,7 +380,7 @@ class LinkExpansionRetriever(GraphRetriever):
                 SELECT DISTINCT ON (mu.id)
                     mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                     mu.occurred_end, mu.mentioned_at,
-                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags, mu.proof_count,
                     ml.weight AS score,
                     'causal'::text AS source
                 FROM {ml} ml
@@ -346,18 +391,37 @@ class LinkExpansionRetriever(GraphRetriever):
                   AND mu.fact_type = $2
                 ORDER BY mu.id, ml.weight DESC
                 LIMIT $3
-            )
+            )"""
+
+        full_query = f"""
+            WITH {entity_cte},
+            {semantic_causal_cte}
             SELECT * FROM entity_expanded
             UNION ALL
             SELECT * FROM semantic_expanded
             UNION ALL
             SELECT * FROM causal_expanded
-            """,
-            seed_ids,
-            fact_type,
-            budget,
-            self.causal_weight_threshold,
-        )
+            """
+
+        params = [seed_ids, fact_type, budget, self.causal_weight_threshold]
+
+        try:
+            all_rows = await asyncio.wait_for(
+                conn.fetch(full_query, *params),
+                timeout=config.link_expansion_timeout,
+            )
+        except asyncio.TimeoutError:
+            logger.warning(
+                f"[LinkExpansion] Entity expansion timed out after {config.link_expansion_timeout}s "
+                f"for fact_type={fact_type}, falling back to semantic+causal only"
+            )
+            fallback_query = f"""
+                WITH {semantic_causal_cte}
+                SELECT * FROM semantic_expanded
+                UNION ALL
+                SELECT * FROM causal_expanded
+                """
+            all_rows = await conn.fetch(fallback_query, *params)
 
         entity_rows = [r for r in all_rows if r["source"] == "entity"]
         semantic_rows = [r for r in all_rows if r["source"] == "semantic"]
@@ -397,6 +461,33 @@ class LinkExpansionRetriever(GraphRetriever):
                 f"{len(source_ids_found)} source_memory_ids found"
             )
 
+        config = get_config()
+        ue = fq_table("unit_entities")
+        per_entity_limit = config.link_expansion_per_entity_limit
+
+        connected_sources_cte = f"""
+            source_entities AS (
+                SELECT DISTINCT ue_seed.entity_id
+                FROM seed_sources ss
+                JOIN {ue} ue_seed ON ue_seed.unit_id = ss.source_id
+            ),
+            connected_sources AS (
+                -- Find sources sharing entities with seed observation sources
+                -- via LATERAL-capped self-join (prevents hub entity fanout).
+                SELECT DISTINCT t.unit_id AS source_id
+                FROM source_entities se
+                CROSS JOIN LATERAL (
+                    SELECT ue_target.unit_id
+                    FROM {ue} ue_target
+                    WHERE ue_target.entity_id = se.entity_id
+                    ORDER BY ue_target.unit_id DESC
+                    LIMIT {per_entity_limit}
+                ) t
+                WHERE NOT EXISTS (
+                    SELECT 1 FROM seed_sources ss WHERE ss.source_id = t.unit_id
+                )
+            )"""
+
         entity_rows = await conn.fetch(
             f"""
             WITH seed_sources AS (
@@ -405,22 +496,14 @@ class LinkExpansionRetriever(GraphRetriever):
                 WHERE id = ANY($1::uuid[])
                   AND source_memory_ids IS NOT NULL
             ),
-            connected_sources AS (
-                -- Mirror the non-observation entity expansion: follow pre-bounded entity
-                -- links in memory_links (capped to MAX_LINKS_PER_ENTITY=50 at retain time).
-                -- Score = number of distinct shared entities, same as the non-obs path.
-                SELECT DISTINCT ml.to_unit_id AS source_id
-                FROM seed_sources ss
-                JOIN {fq_table("memory_links")} ml ON ml.from_unit_id = ss.source_id
-                WHERE ml.link_type = 'entity'
-            ),
+            {connected_sources_cte},
             connected_array AS (
                 SELECT array_agg(source_id) AS source_ids FROM connected_sources
             )
             SELECT
                 mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                 mu.occurred_end, mu.mentioned_at,
-                mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                mu.fact_type, mu.document_id, mu.chunk_id, mu.tags, mu.proof_count,
                 (SELECT COUNT(DISTINCT s) FROM unnest(mu.source_memory_ids) s WHERE s = ANY(ca.source_ids))::float AS score
             FROM {fq_table("memory_units")} mu, connected_array ca
             WHERE mu.fact_type = 'observation'
@@ -444,13 +527,13 @@ class LinkExpansionRetriever(GraphRetriever):
                 SELECT
                     id, text, context, event_date, occurred_start,
                     occurred_end, mentioned_at,
-                    fact_type, document_id, chunk_id, tags,
+                    fact_type, document_id, chunk_id, tags, proof_count,
                     MAX(weight) AS score,
                     'semantic'::text AS source
                 FROM (
                     SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                            mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id,
-                           mu.chunk_id, mu.tags, ml.weight
+                           mu.chunk_id, mu.tags, mu.proof_count, ml.weight
                     FROM {ml} ml JOIN {mu} mu ON mu.id = ml.to_unit_id
                     WHERE ml.from_unit_id = ANY($1::uuid[])
                       AND ml.link_type = 'semantic' AND mu.fact_type = 'observation'
@@ -458,21 +541,21 @@ class LinkExpansionRetriever(GraphRetriever):
                     UNION ALL
                     SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                            mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id,
-                           mu.chunk_id, mu.tags, ml.weight
+                           mu.chunk_id, mu.tags, mu.proof_count, ml.weight
                     FROM {ml} ml JOIN {mu} mu ON mu.id = ml.from_unit_id
                     WHERE ml.to_unit_id = ANY($1::uuid[])
                       AND ml.link_type = 'semantic' AND mu.fact_type = 'observation'
                       AND mu.id != ALL($1::uuid[])
                 ) sem_raw
                 GROUP BY id, text, context, event_date, occurred_start, occurred_end,
-                         mentioned_at, fact_type, document_id, chunk_id, tags
+                         mentioned_at, fact_type, document_id, chunk_id, tags, proof_count
                 ORDER BY score DESC LIMIT $2
             ),
             causal_expanded AS (
                 SELECT DISTINCT ON (mu.id)
                     mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                     mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id,
-                    mu.chunk_id, mu.tags, ml.weight AS score, 'causal'::text AS source
+                    mu.chunk_id, mu.tags, mu.proof_count, ml.weight AS score, 'causal'::text AS source
                 FROM {ml} ml JOIN {mu} mu ON ml.to_unit_id = mu.id
                 WHERE ml.from_unit_id = ANY($1::uuid[])
                   AND ml.link_type IN ('causes', 'caused_by', 'enables', 'prevents')

hindsight-api-slim/hindsight_api/engine/search/mpfp_retrieval.py

@@ -1,702 +0,0 @@
-"""
-Meta-Path Forward Push (MPFP) graph retrieval.
-
-A sublinear graph traversal algorithm for memory retrieval over heterogeneous
-graphs with multiple edge types (semantic, temporal, causal, entity).
-
-Combines meta-path patterns from HIN literature with Forward Push local
-propagation from Approximate PPR.
-
-Key properties:
-- Sublinear in graph size (threshold pruning bounds active nodes)
-- Lazy edge loading: only loads edges for frontier nodes, not entire graph
-- Predefined patterns capture different retrieval intents
-- All patterns run in parallel, results fused via RRF
-- No LLM in the loop during traversal
-"""
-
-import asyncio
-import logging
-from collections import defaultdict
-from dataclasses import dataclass, field
-
-from ..db_utils import acquire_with_retry
-from ..memory_engine import fq_table
-from .graph_retrieval import GraphRetriever
-from .tags import TagGroup, TagsMatch
-from .types import MPFPTimings, RetrievalResult
-
-logger = logging.getLogger(__name__)
-
-
-# -----------------------------------------------------------------------------
-# Data Classes
-# -----------------------------------------------------------------------------
-
-
-@dataclass
-class EdgeTarget:
-    """A neighbor node with its edge weight."""
-
-    node_id: str
-    weight: float
-
-
-@dataclass
-class EdgeCache:
-    """
-    Cache for lazily-loaded edges.
-
-    Grows per-hop as edges are loaded for frontier nodes.
-    Shared across patterns to avoid redundant loads.
-    Loads ALL edge types at once to minimize DB queries.
-    Thread-safe via asyncio lock to prevent redundant concurrent loads.
-    """
-
-    # edge_type -> from_node_id -> list of EdgeTarget
-    graphs: dict[str, dict[str, list[EdgeTarget]]] = field(default_factory=dict)
-    # Track which nodes have been fully loaded (all edge types)
-    _fully_loaded: set[str] = field(default_factory=set)
-    # Timing stats
-    db_queries: int = 0
-    edge_load_time: float = 0.0
-    # Detailed hop timing for debugging
-    hop_details: list[dict] = field(default_factory=list)
-    # Lock to prevent redundant concurrent loads
-    _lock: asyncio.Lock = field(default_factory=asyncio.Lock)
-
-    def get_neighbors(self, edge_type: str, node_id: str) -> list[EdgeTarget]:
-        """Get neighbors for a node via a specific edge type."""
-        return self.graphs.get(edge_type, {}).get(node_id, [])
-
-    def get_normalized_neighbors(self, edge_type: str, node_id: str, top_k: int) -> list[EdgeTarget]:
-        """Get top-k neighbors with weights normalized to sum to 1."""
-        neighbors = self.get_neighbors(edge_type, node_id)[:top_k]
-        if not neighbors:
-            return []
-
-        total = sum(n.weight for n in neighbors)
-        if total == 0:
-            return []
-
-        return [EdgeTarget(node_id=n.node_id, weight=n.weight / total) for n in neighbors]
-
-    def is_fully_loaded(self, node_id: str) -> bool:
-        """Check if all edges for this node have been loaded."""
-        return node_id in self._fully_loaded
-
-    def get_uncached(self, node_ids: list[str]) -> list[str]:
-        """Get node IDs that haven't been fully loaded yet."""
-        return [n for n in node_ids if not self.is_fully_loaded(n)]
-
-    def add_all_edges(self, edges_by_type: dict[str, dict[str, list[EdgeTarget]]], all_queried: list[str]):
-        """
-        Add loaded edges to the cache (all edge types at once).
-
-        Args:
-            edges_by_type: Dict mapping edge_type -> from_node_id -> list of EdgeTarget
-            all_queried: All node IDs that were queried (marks them as fully loaded)
-        """
-        for edge_type, edges in edges_by_type.items():
-            if edge_type not in self.graphs:
-                self.graphs[edge_type] = {}
-            for node_id, neighbors in edges.items():
-                self.graphs[edge_type][node_id] = neighbors
-
-        # Mark all queried nodes as fully loaded (even if they have no edges)
-        self._fully_loaded.update(all_queried)
-
-
-@dataclass
-class PatternResult:
-    """Result from a single pattern traversal."""
-
-    pattern: list[str]
-    scores: dict[str, float]  # node_id -> accumulated mass
-
-
-@dataclass
-class MPFPConfig:
-    """Configuration for MPFP algorithm."""
-
-    alpha: float = 0.15  # teleport/keep probability
-    threshold: float = 1e-6  # mass pruning threshold (lower = explore more)
-    top_k_neighbors: int = 20  # fan-out limit per node
-
-    # Patterns from semantic seeds
-    patterns_semantic: list[list[str]] = field(
-        default_factory=lambda: [
-            ["semantic", "semantic"],  # topic expansion
-            ["entity", "temporal"],  # entity timeline
-            ["semantic", "causes"],  # reasoning chains (forward)
-            ["semantic", "caused_by"],  # reasoning chains (backward)
-            ["entity", "semantic"],  # entity context
-        ]
-    )
-
-    # Patterns from temporal seeds
-    patterns_temporal: list[list[str]] = field(
-        default_factory=lambda: [
-            ["temporal", "semantic"],  # what was happening then
-            ["temporal", "entity"],  # who was involved then
-        ]
-    )
-
-
-@dataclass
-class SeedNode:
-    """An entry point node with its initial score."""
-
-    node_id: str
-    score: float  # initial mass (e.g., similarity score)
-
-
-# -----------------------------------------------------------------------------
-# Lazy Edge Loading
-# -----------------------------------------------------------------------------
-
-
-async def load_all_edges_for_frontier(
-    pool,
-    node_ids: list[str],
-    top_k_per_type: int = 20,
-) -> dict[str, dict[str, list[EdgeTarget]]]:
-    """
-    Load top-k edges per (node, edge_type) for frontier nodes.
-
-    Uses a LATERAL join to efficiently fetch only the top-k edges per type,
-    avoiding loading hundreds of entity edges when only 20 are needed.
-
-    Requires composite index: (from_unit_id, link_type, weight DESC)
-
-    Args:
-        pool: Database connection pool
-        node_ids: Frontier node IDs to load edges for
-        top_k_per_type: Max edges to load per (node, link_type) pair
-
-    Returns:
-        Dict mapping edge_type -> from_node_id -> list of EdgeTarget
-    """
-    if not node_ids:
-        return {}
-
-    async with acquire_with_retry(pool) as conn:
-        # Use LATERAL join to get top-k per (from_node, link_type)
-        # This leverages the composite index for efficient early termination
-        rows = await conn.fetch(
-            f"""
-            WITH frontier(node_id) AS (SELECT unnest($1::uuid[]))
-            SELECT f.node_id as from_unit_id, lt.link_type, edges.to_unit_id, edges.weight
-            FROM frontier f
-            CROSS JOIN (VALUES ('semantic'), ('temporal'), ('entity'), ('causes'), ('caused_by')) AS lt(link_type)
-            CROSS JOIN LATERAL (
-                SELECT ml.to_unit_id, ml.weight
-                FROM {fq_table("memory_links")} ml
-                WHERE ml.from_unit_id = f.node_id
-                  AND ml.link_type = lt.link_type
-                  AND ml.weight >= 0.1
-                ORDER BY ml.weight DESC
-                LIMIT $2
-            ) edges
-            """,
-            node_ids,
-            top_k_per_type,
-        )
-
-    # Group by edge_type -> from_node -> neighbors
-    result: dict[str, dict[str, list[EdgeTarget]]] = defaultdict(lambda: defaultdict(list))
-    for row in rows:
-        edge_type = row["link_type"]
-        from_id = str(row["from_unit_id"])
-        to_id = str(row["to_unit_id"])
-        weight = row["weight"]
-        result[edge_type][from_id].append(EdgeTarget(node_id=to_id, weight=weight))
-
-    # Convert nested defaultdicts to regular dicts
-    return {edge_type: dict(edges) for edge_type, edges in result.items()}
-
-
-# -----------------------------------------------------------------------------
-# Core Algorithm (Async with Lazy Loading)
-# -----------------------------------------------------------------------------
-
-
-@dataclass
-class PatternState:
-    """State for a pattern traversal between hops."""
-
-    pattern: list[str]
-    hop_index: int
-    scores: dict[str, float]
-    frontier: dict[str, float]
-
-
-def _init_pattern_state(seeds: list[SeedNode], pattern: list[str]) -> PatternState:
-    """Initialize pattern state from seeds."""
-    if not seeds:
-        return PatternState(pattern=pattern, hop_index=0, scores={}, frontier={})
-
-    total_seed_score = sum(s.score for s in seeds)
-    if total_seed_score == 0:
-        total_seed_score = len(seeds)
-
-    frontier = {s.node_id: s.score / total_seed_score for s in seeds}
-    return PatternState(pattern=pattern, hop_index=0, scores={}, frontier=frontier)
-
-
-def _execute_hop(state: PatternState, cache: EdgeCache, config: MPFPConfig) -> set[str]:
-    """
-    Execute ONE hop of traversal, return frontier nodes for next hop.
-
-    This is a pure function that uses cached edges (no DB access).
-    Returns set of uncached nodes needed for next hop.
-    """
-    if state.hop_index >= len(state.pattern):
-        return set()
-
-    edge_type = state.pattern[state.hop_index]
-
-    # Collect active nodes above threshold
-    active_nodes = [node_id for node_id, mass in state.frontier.items() if mass >= config.threshold]
-    if not active_nodes:
-        state.frontier = {}
-        return set()
-
-    # Propagate mass using cached edges
-    next_frontier: dict[str, float] = {}
-    uncached_for_next: set[str] = set()
-
-    for node_id, mass in state.frontier.items():
-        if mass < config.threshold:
-            continue
-
-        # Keep α portion for this node
-        state.scores[node_id] = state.scores.get(node_id, 0) + config.alpha * mass
-
-        # Push (1-α) to neighbors
-        push_mass = (1 - config.alpha) * mass
-        neighbors = cache.get_normalized_neighbors(edge_type, node_id, config.top_k_neighbors)
-
-        for neighbor in neighbors:
-            next_frontier[neighbor.node_id] = next_frontier.get(neighbor.node_id, 0) + push_mass * neighbor.weight
-            # Track if we'll need edges for this node in the next hop
-            if not cache.is_fully_loaded(neighbor.node_id):
-                uncached_for_next.add(neighbor.node_id)
-
-    state.frontier = next_frontier
-    state.hop_index += 1
-
-    return uncached_for_next
-
-
-def _finalize_pattern(state: PatternState, config: MPFPConfig) -> PatternResult:
-    """Finalize pattern by adding remaining frontier mass to scores."""
-    for node_id, mass in state.frontier.items():
-        if mass >= config.threshold:
-            state.scores[node_id] = state.scores.get(node_id, 0) + mass
-
-    return PatternResult(pattern=state.pattern, scores=state.scores)
-
-
-async def mpfp_traverse_hop_synchronized(
-    pool,
-    pattern_jobs: list[tuple[list[SeedNode], list[str]]],
-    config: MPFPConfig,
-    cache: EdgeCache,
-) -> list[PatternResult]:
-    """
-    Execute ALL patterns with hop-synchronized edge loading.
-
-    Instead of running each pattern independently (causing multiple DB queries),
-    this function:
-    1. Runs hop 1 for ALL patterns (using pre-warmed seed edges)
-    2. Collects ALL unique hop-2 frontier nodes across patterns
-    3. Pre-warms hop-2 edges in ONE query
-    4. Runs hop 2 for ALL patterns
-
-    This reduces DB queries from O(patterns * hops) to O(hops).
-
-    Args:
-        pool: Database connection pool
-        pattern_jobs: List of (seeds, pattern) tuples
-        config: Algorithm parameters
-        cache: Shared edge cache (should be pre-warmed with seed edges)
-
-    Returns:
-        List of PatternResult for each pattern
-    """
-    import time
-
-    # Initialize all pattern states
-    states = [_init_pattern_state(seeds, pattern) for seeds, pattern in pattern_jobs]
-
-    # Determine max hops (all patterns should be same length, but be safe)
-    max_hops = max((len(p) for _, p in pattern_jobs), default=0)
-
-    # Detailed timing for debugging
-    hop_times: list[dict] = []
-
-    # Execute hop-by-hop across ALL patterns
-    for hop in range(max_hops):
-        hop_start = time.time()
-        hop_timing = {"hop": hop, "patterns_executed": 0, "uncached_count": 0, "load_time": 0.0}
-
-        # Execute this hop for all patterns, collect uncached nodes for next hop
-        all_uncached: set[str] = set()
-        exec_start = time.time()
-        for state in states:
-            if state.hop_index < len(state.pattern):
-                uncached = _execute_hop(state, cache, config)
-                all_uncached.update(uncached)
-                hop_timing["patterns_executed"] += 1
-        hop_timing["exec_time"] = time.time() - exec_start
-
-        # Pre-warm edges for ALL uncached nodes before next hop
-        hop_timing["uncached_count"] = len(all_uncached)
-        if all_uncached:
-            uncached_list = list(all_uncached - cache._fully_loaded)
-            hop_timing["uncached_after_filter"] = len(uncached_list)
-            if uncached_list:
-                load_start = time.time()
-                edges_by_type = await load_all_edges_for_frontier(pool, uncached_list, config.top_k_neighbors)
-                hop_timing["load_time"] = time.time() - load_start
-                cache.edge_load_time += hop_timing["load_time"]
-                cache.db_queries += 1
-                cache.add_all_edges(edges_by_type, uncached_list)
-                hop_timing["edges_loaded"] = sum(
-                    len(neighbors) for edges in edges_by_type.values() for neighbors in edges.values()
-                )
-
-        hop_timing["total_time"] = time.time() - hop_start
-        hop_times.append(hop_timing)
-
-    # Store hop timing details in cache for logging
-    cache.hop_details = hop_times
-
-    # Finalize all patterns
-    return [_finalize_pattern(state, config) for state in states]
-
-
-async def mpfp_traverse_async(
-    pool,
-    seeds: list[SeedNode],
-    pattern: list[str],
-    config: MPFPConfig,
-    cache: EdgeCache,
-) -> PatternResult:
-    """
-    Async Forward Push traversal with lazy edge loading.
-
-    NOTE: For better performance with multiple patterns, use mpfp_traverse_hop_synchronized().
-    This function is kept for single-pattern use cases.
-    """
-    if not seeds:
-        return PatternResult(pattern=pattern, scores={})
-
-    results = await mpfp_traverse_hop_synchronized(pool, [(seeds, pattern)], config, cache)
-    return results[0] if results else PatternResult(pattern=pattern, scores={})
-
-
-def rrf_fusion(
-    results: list[PatternResult],
-    k: int = 60,
-    top_k: int = 50,
-) -> list[tuple[str, float]]:
-    """
-    Reciprocal Rank Fusion to combine pattern results.
-
-    Args:
-        results: List of pattern results
-        k: RRF constant (higher = more uniform weighting)
-        top_k: Number of results to return
-
-    Returns:
-        List of (node_id, fused_score) tuples, sorted by score descending
-    """
-    fused: dict[str, float] = {}
-
-    for result in results:
-        if not result.scores:
-            continue
-
-        # Rank nodes by their score in this pattern
-        ranked = sorted(result.scores.keys(), key=lambda n: result.scores[n], reverse=True)
-
-        for rank, node_id in enumerate(ranked):
-            fused[node_id] = fused.get(node_id, 0) + 1.0 / (k + rank + 1)
-
-    # Sort by fused score and return top-k
-    sorted_results = sorted(fused.items(), key=lambda x: x[1], reverse=True)
-
-    return sorted_results[:top_k]
-
-
-# -----------------------------------------------------------------------------
-# Database Loading
-# -----------------------------------------------------------------------------
-
-
-async def fetch_memory_units_by_ids(
-    pool,
-    node_ids: list[str],
-    fact_type: str,
-) -> list[RetrievalResult]:
-    """Fetch full memory unit details for a list of node IDs."""
-    if not node_ids:
-        return []
-
-    async with acquire_with_retry(pool) as conn:
-        rows = await conn.fetch(
-            f"""
-            SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, fact_type, document_id, chunk_id, tags, metadata
-            FROM {fq_table("memory_units")}
-            WHERE id = ANY($1::uuid[])
-              AND fact_type = $2
-            """,
-            node_ids,
-            fact_type,
-        )
-
-    return [RetrievalResult.from_db_row(dict(r)) for r in rows]
-
-
-# -----------------------------------------------------------------------------
-# Graph Retriever Implementation
-# -----------------------------------------------------------------------------
-
-
-class MPFPGraphRetriever(GraphRetriever):
-    """
-    Graph retrieval using Meta-Path Forward Push with lazy edge loading.
-
-    Runs predefined patterns in parallel from semantic and temporal seeds,
-    loading edges on-demand per hop instead of loading entire graph upfront.
-    """
-
-    def __init__(self, config: MPFPConfig | None = None):
-        """
-        Initialize MPFP retriever.
-
-        Args:
-            config: Algorithm configuration (uses defaults if None)
-        """
-        if config is None:
-            # Read top_k_neighbors from global config
-            from ...config import get_config
-
-            global_config = get_config()
-            config = MPFPConfig(top_k_neighbors=global_config.mpfp_top_k_neighbors)
-        self.config = config
-
-    @property
-    def name(self) -> str:
-        return "mpfp"
-
-    async def retrieve(
-        self,
-        pool,
-        query_embedding_str: str,
-        bank_id: str,
-        fact_type: str,
-        budget: int,
-        query_text: str | None = None,
-        semantic_seeds: list[RetrievalResult] | None = None,
-        temporal_seeds: list[RetrievalResult] | None = None,
-        adjacency=None,  # Ignored - kept for interface compatibility
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
-        tag_groups: list[TagGroup] | None = None,
-    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
-        """
-        Retrieve facts using MPFP algorithm with lazy edge loading.
-
-        Args:
-            pool: Database connection pool
-            query_embedding_str: Query embedding (used for fallback seed finding)
-            bank_id: Memory bank ID
-            fact_type: Fact type to filter
-            budget: Maximum results to return
-            query_text: Original query text (optional)
-            semantic_seeds: Pre-computed semantic entry points
-            temporal_seeds: Pre-computed temporal entry points
-            adjacency: Ignored (kept for interface compatibility)
-            tags: Optional list of tags for visibility filtering (OR matching)
-
-        Returns:
-            Tuple of (List of RetrievalResult with activation scores, MPFPTimings)
-        """
-        import time
-
-        timings = MPFPTimings(fact_type=fact_type)
-
-        # Convert seeds to SeedNode format
-        semantic_seed_nodes = self._convert_seeds(semantic_seeds, "similarity")
-        temporal_seed_nodes = self._convert_seeds(temporal_seeds, "temporal_score")
-
-        # If no semantic seeds provided, fall back to finding our own
-        if not semantic_seed_nodes:
-            seeds_start = time.time()
-            semantic_seed_nodes = await self._find_semantic_seeds(
-                pool,
-                query_embedding_str,
-                bank_id,
-                fact_type,
-                tags=tags,
-                tags_match=tags_match,
-                tag_groups=tag_groups,
-            )
-            timings.seeds_time = time.time() - seeds_start
-            logger.debug(
-                f"[MPFP] Found {len(semantic_seed_nodes)} semantic seeds for fact_type={fact_type} (tags={tags}, tags_match={tags_match})"
-            )
-
-        # Collect all pattern jobs
-        pattern_jobs = []
-
-        # Patterns from semantic seeds
-        for pattern in self.config.patterns_semantic:
-            if semantic_seed_nodes:
-                pattern_jobs.append((semantic_seed_nodes, pattern))
-
-        # Patterns from temporal seeds
-        for pattern in self.config.patterns_temporal:
-            if temporal_seed_nodes:
-                pattern_jobs.append((temporal_seed_nodes, pattern))
-
-        if not pattern_jobs:
-            logger.debug(
-                f"[MPFP] No pattern jobs (semantic_seeds={len(semantic_seed_nodes)}, temporal_seeds={len(temporal_seed_nodes)})"
-            )
-            return [], timings
-
-        timings.pattern_count = len(pattern_jobs)
-
-        # Shared edge cache across all patterns
-        cache = EdgeCache()
-
-        # Pre-warm cache with ALL seed node edges BEFORE running patterns
-        # This prevents redundant DB queries at hop 1
-        all_seed_ids = list({s.node_id for seeds, _ in pattern_jobs for s in seeds})
-        if all_seed_ids:
-            import time as time_module
-
-            prewarm_start = time_module.time()
-            edges_by_type = await load_all_edges_for_frontier(pool, all_seed_ids, self.config.top_k_neighbors)
-            cache.edge_load_time += time_module.time() - prewarm_start
-            cache.db_queries += 1
-            cache.add_all_edges(edges_by_type, all_seed_ids)
-
-        # Run all patterns with HOP-SYNCHRONIZED edge loading
-        # This batches hop-2 edge loads across ALL patterns into ONE query
-        # Reduces DB queries from O(patterns * hops) to O(hops)
-        step_start = time.time()
-        pattern_results = await mpfp_traverse_hop_synchronized(pool, pattern_jobs, self.config, cache)
-        timings.traverse = time.time() - step_start
-
-        # Record edge loading stats from cache
-        timings.edge_count = sum(len(neighbors) for g in cache.graphs.values() for neighbors in g.values())
-        timings.db_queries = cache.db_queries
-        timings.edge_load_time = cache.edge_load_time
-        timings.hop_details = cache.hop_details
-
-        # Fuse results
-        step_start = time.time()
-        fused = rrf_fusion(pattern_results, top_k=budget)
-        timings.fusion = time.time() - step_start
-
-        if not fused:
-            logger.debug(f"[MPFP] No fused results after RRF fusion (pattern_count={len(pattern_results)})")
-            return [], timings
-
-        # Get top result IDs
-        result_ids = [node_id for node_id, score in fused][:budget]
-
-        # Fetch full details
-        step_start = time.time()
-        results = await fetch_memory_units_by_ids(pool, result_ids, fact_type)
-        timings.fetch = time.time() - step_start
-
-        # Filter results by tags (graph traversal may have picked up unfiltered memories)
-        if tags:
-            from .tags import filter_results_by_tags
-
-            results = filter_results_by_tags(results, tags, match=tags_match)
-
-        # Apply compound tag group filtering (post-traversal)
-        if tag_groups:
-            from .tags import filter_results_by_tag_groups
-
-            results = filter_results_by_tag_groups(results, tag_groups)
-
-        timings.result_count = len(results)
-
-        # Add activation scores from fusion
-        score_map = {node_id: score for node_id, score in fused}
-        for result in results:
-            result.activation = score_map.get(result.id, 0.0)
-
-        # Sort by activation
-        results.sort(key=lambda r: r.activation or 0, reverse=True)
-
-        return results, timings
-
-    def _convert_seeds(
-        self,
-        seeds: list[RetrievalResult] | None,
-        score_attr: str,
-    ) -> list[SeedNode]:
-        """Convert RetrievalResult seeds to SeedNode format."""
-        if not seeds:
-            return []
-
-        result = []
-        for seed in seeds:
-            score = getattr(seed, score_attr, None)
-            if score is None:
-                score = seed.activation or seed.similarity or 1.0
-            result.append(SeedNode(node_id=seed.id, score=score))
-
-        return result
-
-    async def _find_semantic_seeds(
-        self,
-        pool,
-        query_embedding_str: str,
-        bank_id: str,
-        fact_type: str,
-        limit: int = 20,
-        threshold: float = 0.3,
-        tags: list[str] | None = None,
-        tags_match: TagsMatch = "any",
-        tag_groups: list[TagGroup] | None = None,
-    ) -> list[SeedNode]:
-        """Fallback: find semantic seeds via embedding search."""
-        from .tags import build_tag_groups_where_clause, build_tags_where_clause_simple
-
-        tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
-        tag_groups_param_start = 6 + (1 if tags else 0)
-        groups_clause, groups_params, _ = build_tag_groups_where_clause(tag_groups, tag_groups_param_start)
-        params = [query_embedding_str, bank_id, fact_type, threshold, limit]
-        if tags:
-            params.append(tags)
-        params.extend(groups_params)
-
-        async with acquire_with_retry(pool) as conn:
-            rows = await conn.fetch(
-                f"""
-                SELECT id, 1 - (embedding <=> $1::vector) AS similarity
-                FROM {fq_table("memory_units")}
-                WHERE bank_id = $2
-                  AND embedding IS NOT NULL
-                  AND fact_type = $3
-                  AND (1 - (embedding <=> $1::vector)) >= $4
-                  {tags_clause}
-                  {groups_clause}
-                ORDER BY embedding <=> $1::vector
-                LIMIT $5
-                """,
-                *params,
-            )
-
-        return [SeedNode(node_id=str(r["id"]), score=r["similarity"]) for r in rows]

hindsight-api-slim/hindsight_api/engine/search/reranking.py

@@ -2,6 +2,7 @@
 Cross-encoder neural reranking for search results.
 """
 
+import math
 from datetime import datetime, timezone
 
 from .types import MergedCandidate, ScoredResult
@@ -13,6 +14,7 @@ UTC = timezone.utc
 # so the max combined boost is (1 + alpha/2)^2 ≈ +21% and min is (1 - alpha/2)^2 ≈ -19%.
 _RECENCY_ALPHA: float = 0.2
 _TEMPORAL_ALPHA: float = 0.2
+_PROOF_COUNT_ALPHA: float = 0.1  # Conservative: max ±5% for evidence strength
 
 
 def apply_combined_scoring(
@@ -20,32 +22,81 @@ def apply_combined_scoring(
     now: datetime,
     recency_alpha: float = _RECENCY_ALPHA,
     temporal_alpha: float = _TEMPORAL_ALPHA,
+    proof_count_alpha: float = _PROOF_COUNT_ALPHA,
+    is_passthrough_reranker: bool = False,
 ) -> None:
     """Apply combined scoring to a list of ScoredResults in-place.
 
-    Uses the cross-encoder score as the primary relevance signal, with recency
-    and temporal proximity applied as multiplicative boosts. This ensures the
-    influence of these secondary signals is always proportional to the base
-    relevance score, regardless of the cross-encoder model's score calibration.
+    Uses the cross-encoder score as the primary relevance signal, with recency,
+    temporal proximity, and proof count applied as multiplicative boosts. This
+    ensures the influence of these secondary signals is always proportional to
+    the base relevance score, regardless of the cross-encoder model's score
+    calibration.
 
     Formula::
 
-        recency_boost  = 1 + recency_alpha  * (recency  - 0.5)   # in [1-α/2, 1+α/2]
-        temporal_boost = 1 + temporal_alpha * (temporal - 0.5)   # in [1-α/2, 1+α/2]
-        combined_score = cross_encoder_score_normalized * recency_boost * temporal_boost
+        recency_boost     = 1 + recency_alpha     * (recency     - 0.5)   # in [1-α/2, 1+α/2]
+        temporal_boost    = 1 + temporal_alpha    * (temporal    - 0.5)   # in [1-α/2, 1+α/2]
+        proof_count_boost = 1 + proof_count_alpha * (proof_norm  - 0.5)   # in [1-α/2, 1+α/2]
+        combined_score    = CE_normalized * recency_boost * temporal_boost * proof_count_boost
+
+    proof_norm maps proof_count using a smooth logarithmic curve centered at 0.5,
+    clamped to [0, 1]:
+      proof_count=1 → 0.5 + 0 = 0.5 (neutral multiplier)
+      proof_count=150 → clamped to 1.0 (max +5% boost)
 
     Temporal proximity is treated as neutral (0.5) when not set by temporal retrieval,
     so temporal_boost collapses to 1.0 for non-temporal queries.
 
+    Proof count is treated as neutral (0.5) when not available (non-observation facts),
+    so proof_count_boost collapses to 1.0 for world/experience/opinion facts.
+
     Args:
         scored_results: Results from the cross-encoder reranker. Mutated in place.
         now: Current UTC datetime for recency calculation.
         recency_alpha: Max relative recency adjustment (default 0.2 → ±10%).
         temporal_alpha: Max relative temporal adjustment (default 0.2 → ±10%).
+        proof_count_alpha: Max relative proof count adjustment (default 0.1 → ±5%).
     """
     if now.tzinfo is None:
         now = now.replace(tzinfo=UTC)
 
+    # When the configured cross-encoder is a passthrough (e.g.
+    # RRFPassthroughCrossEncoder used by slim deployments), every
+    # cross_encoder_score_normalized is identical and provides no relevance
+    # signal. In that case the multiplicative recency / temporal / proof_count
+    # boosts below become the *only* ranking signal — making the final order a
+    # pure recency sort regardless of how relevant a candidate actually is.
+    #
+    # Detect that case and seed cross_encoder_score_normalized from the RRF
+    # rank instead, so the boosts modulate a meaningful base score rather than
+    # replacing it. This is a no-op for real cross-encoders, which produce
+    # diverse scores.
+    # When the reranker is a passthrough (e.g. RRFPassthroughCrossEncoder used
+    # by slim deployments), every cross_encoder_score_normalized is identical
+    # and provides no relevance signal. The multiplicative recency / temporal /
+    # proof_count boosts below would then become the *only* ranking signal,
+    # making the final order a pure recency sort regardless of how relevant a
+    # candidate actually is.
+    #
+    # Seed cross_encoder_score_normalized from the RRF rank instead, so the
+    # boosts modulate a meaningful base score. Caller passes is_passthrough
+    # explicitly because "all scores identical" is too fragile a heuristic —
+    # a real reranker can also tie scores (especially in tests with synthetic
+    # data) and we'd corrupt legitimate single-result reranks.
+    if is_passthrough_reranker and scored_results:
+        n = len(scored_results)
+        sorted_by_rrf = sorted(
+            scored_results,
+            key=lambda s: getattr(getattr(s, "candidate", None), "rrf_score", 0.0),
+            reverse=True,
+        )
+        denom = max(1, n - 1)
+        for new_rank, sr in enumerate(sorted_by_rrf):
+            # Map rank → [0.1, 1.0] so the recency boost can still nudge
+            # ordering between adjacent candidates without overpowering RRF.
+            sr.cross_encoder_score_normalized = 1.0 - (0.9 * new_rank / denom)
+
     for sr in scored_results:
         # Recency: linear decay over 365 days → [0.1, 1.0]; neutral 0.5 if no date.
         sr.recency = 0.5
@@ -59,13 +110,23 @@ def apply_combined_scoring(
         # Temporal proximity: meaningful only for temporal queries; neutral otherwise.
         sr.temporal = sr.retrieval.temporal_proximity if sr.retrieval.temporal_proximity is not None else 0.5
 
+        # Proof count: log-normalized evidence strength; neutral for non-observations.
+        proof_count = sr.retrieval.proof_count
+        if proof_count is not None and proof_count >= 1:
+            # Clamp to [0, 1] so extreme counts stay within documented ±5% range
+            proof_norm = min(1.0, max(0.0, 0.5 + (math.log(proof_count) / 10.0)))
+        else:
+            # Neutral baseline is precisely 0.5, ensuring neutral multiplier (1.0)
+            proof_norm = 0.5
+
         # RRF: kept at 0.0 for trace continuity but excluded from scoring.
         # RRF is batch-relative (min-max normalised) and redundant after reranking.
         sr.rrf_normalized = 0.0
 
         recency_boost = 1.0 + recency_alpha * (sr.recency - 0.5)
         temporal_boost = 1.0 + temporal_alpha * (sr.temporal - 0.5)
-        sr.combined_score = sr.cross_encoder_score_normalized * recency_boost * temporal_boost
+        proof_count_boost = 1.0 + proof_count_alpha * (proof_norm - 0.5)
+        sr.combined_score = sr.cross_encoder_score_normalized * recency_boost * temporal_boost * proof_count_boost
         sr.weight = sr.combined_score
 
 

hindsight-api-slim/hindsight_api/engine/search/retrieval.py

@@ -13,16 +13,15 @@ import logging
 import re
 from dataclasses import dataclass, field
 from datetime import UTC, datetime
-from typing import Optional
+from typing import Any, Optional
 
 from ...config import get_config
 from ..db_utils import acquire_with_retry
 from ..memory_engine import fq_table
-from .graph_retrieval import BFSGraphRetriever, GraphRetriever
+from .graph_retrieval import GraphRetriever
 from .link_expansion_retrieval import LinkExpansionRetriever
-from .mpfp_retrieval import MPFPGraphRetriever
 from .tags import TagGroup, TagsMatch, build_tag_groups_where_clause, build_tags_where_clause_simple
-from .types import MPFPTimings, RetrievalResult
+from .types import GraphRetrievalTimings, RetrievalResult
 
 logger = logging.getLogger(__name__)
 
@@ -46,7 +45,9 @@ class ParallelRetrievalResult:
     temporal: list[RetrievalResult] | None
     timings: dict[str, float] = field(default_factory=dict)
     temporal_constraint: tuple | None = None  # (start_date, end_date)
-    mpfp_timings: list[MPFPTimings] = field(default_factory=list)  # MPFP sub-step timings per fact type
+    graph_timings: list[GraphRetrievalTimings] = field(
+        default_factory=list
+    )  # Graph retrieval sub-step timings per fact type
     max_conn_wait: float = 0.0  # Maximum connection acquisition wait time across all methods
 
 
@@ -72,15 +73,7 @@ def get_default_graph_retriever() -> GraphRetriever:
     if _default_graph_retriever is None:
         config = get_config()
         retriever_type = config.graph_retriever.lower()
-        if retriever_type == "mpfp":
-            _default_graph_retriever = MPFPGraphRetriever()
-            logger.info(
-                f"Using MPFP graph retriever (top_k_neighbors={_default_graph_retriever.config.top_k_neighbors})"
-            )
-        elif retriever_type == "bfs":
-            _default_graph_retriever = BFSGraphRetriever()
-            logger.info("Using BFS graph retriever")
-        elif retriever_type == "link_expansion":
+        if retriever_type == "link_expansion":
             _default_graph_retriever = LinkExpansionRetriever()
             logger.info("Using LinkExpansion graph retriever")
         else:
@@ -105,6 +98,8 @@ async def retrieve_semantic_bm25_combined(
     tags: list[str] | None = None,
     tags_match: TagsMatch = "any",
     tag_groups: list[TagGroup] | None = None,
+    created_after: datetime | None = None,
+    created_before: datetime | None = None,
 ) -> dict[str, tuple[list[RetrievalResult], list[RetrievalResult]]]:
     """
     Combined semantic + BM25 retrieval for multiple fact types in a single query.
@@ -148,7 +143,7 @@ async def retrieve_semantic_bm25_combined(
 
     cols = (
         "id, text, context, event_date, occurred_start, occurred_end, mentioned_at, "
-        "fact_type, document_id, chunk_id, tags, metadata"
+        "fact_type, document_id, chunk_id, tags, metadata, proof_count"
     )
     table = fq_table("memory_units")
 
@@ -170,6 +165,21 @@ async def retrieve_semantic_bm25_combined(
     tag_groups_param_start = tags_param_idx + (1 if tags else 0)
     groups_clause, groups_params, _ = build_tag_groups_where_clause(tag_groups, tag_groups_param_start)
 
+    # --- created_at time range filter (appended after tags/groups) ---
+    # Param indices are computed relative to the final params list built below,
+    # so we pre-compute the next available index after all preceding params.
+    _next_idx = tag_groups_param_start + len(groups_params)
+    created_range_clause = ""
+    created_range_params: list[Any] = []
+    if created_after is not None:
+        created_range_params.append(created_after)
+        created_range_clause += f" AND updated_at > ${_next_idx}"
+        _next_idx += 1
+    if created_before is not None:
+        created_range_params.append(created_before)
+        created_range_clause += f" AND updated_at < ${_next_idx}"
+        _next_idx += 1
+
     # --- Semantic UNION ALL arms (one per fact_type) ---
     # Each arm has its own ORDER BY embedding <=> $1 LIMIT {hnsw_fetch}, which
     # lets the planner use the partial HNSW index for that fact_type.
@@ -187,6 +197,7 @@ async def retrieve_semantic_bm25_combined(
             f"   AND (1 - (embedding <=> $1::vector)) >= 0.3"
             f"   {tags_clause}"
             f"   {groups_clause}"
+            f"   {created_range_clause}"
             f" ORDER BY embedding <=> $1::vector"
             f" LIMIT {hnsw_fetch})"
         )
@@ -227,6 +238,7 @@ async def retrieve_semantic_bm25_combined(
                 f"   {bm25_where_filter}"
                 f"   {tags_clause}"
                 f"   {groups_clause}"
+                f"   {created_range_clause}"
                 f" ORDER BY {bm25_order_by}"
                 f" LIMIT $3)"
             )
@@ -240,6 +252,7 @@ async def retrieve_semantic_bm25_combined(
     if tags:
         params.append(tags)
     params.extend(groups_params)
+    params.extend(created_range_params)
 
     rows = await conn.fetch(query, *params)
 
@@ -273,6 +286,8 @@ async def retrieve_temporal_combined(
     tags: list[str] | None = None,
     tags_match: TagsMatch = "any",
     tag_groups: list[TagGroup] | None = None,
+    created_after: datetime | None = None,
+    created_before: datetime | None = None,
 ) -> dict[str, list[RetrievalResult]]:
     """
     Temporal retrieval for multiple fact types in a single query.
@@ -306,10 +321,25 @@ async def retrieve_temporal_combined(
     tags_clause = build_tags_where_clause_simple(tags, 7, match=tags_match)
     tag_groups_param_start = 7 + (1 if tags else 0)
     groups_clause, groups_params, _ = build_tag_groups_where_clause(tag_groups, tag_groups_param_start)
+
+    # created_at time range filter (after tags/groups)
+    _next_idx = tag_groups_param_start + len(groups_params)
+    created_range_clause = ""
+    created_range_params: list[Any] = []
+    if created_after is not None:
+        created_range_params.append(created_after)
+        created_range_clause += f" AND updated_at > ${_next_idx}"
+        _next_idx += 1
+    if created_before is not None:
+        created_range_params.append(created_before)
+        created_range_clause += f" AND updated_at < ${_next_idx}"
+        _next_idx += 1
+
     params: list = [query_emb_str, bank_id, fact_types, start_date, end_date, semantic_threshold]
     if tags:
         params.append(tags)
     params.extend(groups_params)
+    params.extend(created_range_params)
 
     # Two-phase entry point query:
     # Phase 1 (date_ranked): rank by date only — no embedding computation — for all units in
@@ -341,9 +371,10 @@ async def retrieve_temporal_combined(
               )
               {tags_clause}
               {groups_clause}
+              {created_range_clause}
         ),
         sim_ranked AS (
-            SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id, mu.chunk_id, mu.tags, mu.metadata,
+            SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.proof_count, mu.document_id, mu.chunk_id, mu.tags, mu.metadata,
                    1 - (mu.embedding <=> $1::vector) AS similarity,
                    ROW_NUMBER() OVER (PARTITION BY mu.fact_type ORDER BY mu.embedding <=> $1::vector) AS sim_rn
             FROM date_ranked dr
@@ -351,7 +382,7 @@ async def retrieve_temporal_combined(
             WHERE dr.rn <= 50
               AND (1 - (mu.embedding <=> $1::vector)) >= $6
         )
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, document_id, chunk_id, tags, metadata, similarity
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, proof_count, document_id, chunk_id, tags, metadata, similarity
         FROM sim_ranked
         WHERE sim_rn <= 10
         """,
@@ -543,6 +574,8 @@ async def retrieve_all_fact_types_parallel(
     tags: list[str] | None = None,
     tags_match: TagsMatch = "any",
     tag_groups: list[TagGroup] | None = None,
+    created_after: datetime | None = None,
+    created_before: datetime | None = None,
 ) -> MultiFactTypeRetrievalResult:
     """
     Optimized retrieval for multiple fact types using batched queries.
@@ -601,6 +634,8 @@ async def retrieve_all_fact_types_parallel(
             tags=tags,
             tags_match=tags_match,
             tag_groups=tag_groups,
+            created_after=created_after,
+            created_before=created_before,
         )
         semantic_bm25_time = time.time() - semantic_bm25_start
 
@@ -620,6 +655,8 @@ async def retrieve_all_fact_types_parallel(
                 tags=tags,
                 tags_match=tags_match,
                 tag_groups=tag_groups,
+                created_after=created_after,
+                created_before=created_before,
             )
             temporal_time = time.time() - temporal_start
 
@@ -627,9 +664,11 @@ async def retrieve_all_fact_types_parallel(
     timings["temporal_combined"] = temporal_time
 
     # Step 3: Run graph retrieval for each fact type in parallel
-    async def run_graph_for_fact_type(ft: str) -> tuple[str, list[RetrievalResult], float, MPFPTimings | None]:
+    async def run_graph_for_fact_type(
+        ft: str,
+    ) -> tuple[str, list[RetrievalResult], float, GraphRetrievalTimings | None]:
         graph_start = time.time()
-        results, mpfp_timing = await retriever.retrieve(
+        results, graph_timing = await retriever.retrieve(
             pool=pool,
             query_embedding_str=query_embedding_str,
             bank_id=bank_id,
@@ -641,8 +680,10 @@ async def retrieve_all_fact_types_parallel(
             tags=tags,
             tags_match=tags_match,
             tag_groups=tag_groups,
+            created_after=created_after,
+            created_before=created_before,
         )
-        return ft, results, time.time() - graph_start, mpfp_timing
+        return ft, results, time.time() - graph_start, graph_timing
 
     # Run graph for all fact types in parallel
     graph_tasks = [run_graph_for_fact_type(ft) for ft in fact_types]
@@ -651,7 +692,7 @@ async def retrieve_all_fact_types_parallel(
     # Organize results by fact type
     results_by_fact_type: dict[str, ParallelRetrievalResult] = {}
     max_conn_wait = conn_wait  # Single connection for semantic+bm25+temporal
-    all_mpfp_timings: list[MPFPTimings] = []
+    all_graph_timings: list[GraphRetrievalTimings] = []
 
     for ft in fact_types:
         # Get semantic + bm25 results for this fact type
@@ -660,14 +701,14 @@ async def retrieve_all_fact_types_parallel(
         # Find graph results for this fact type
         graph_results = []
         graph_time = 0.0
-        mpfp_timing = None
+        graph_timing = None
         for gr in graph_results_list:
             if gr[0] == ft:
                 graph_results = gr[1]
                 graph_time = gr[2]
-                mpfp_timing = gr[3]
-                if mpfp_timing:
-                    all_mpfp_timings.append(mpfp_timing)
+                graph_timing = gr[3]
+                if graph_timing:
+                    all_graph_timings.append(graph_timing)
                 break
 
         # Get temporal results for this fact type from combined result
@@ -688,7 +729,7 @@ async def retrieve_all_fact_types_parallel(
                 "temporal_extraction": temporal_extraction_time,
             },
             temporal_constraint=temporal_constraint,
-            mpfp_timings=[mpfp_timing] if mpfp_timing else [],
+            graph_timings=[graph_timing] if graph_timing else [],
             max_conn_wait=max_conn_wait,
         )
 

hindsight-api-slim/hindsight_api/engine/search/think_utils.py

@@ -62,17 +62,18 @@ def format_facts_for_prompt(facts: list[MemoryFact]) -> str:
         if fact.context:
             fact_obj["context"] = fact.context
 
-        # Add occurred_start if available (when the fact occurred)
-        if fact.occurred_start:
-            occurred_start = fact.occurred_start
-            if isinstance(occurred_start, str):
-                fact_obj["occurred_start"] = occurred_start
-            elif isinstance(occurred_start, datetime):
-                fact_obj["occurred_start"] = occurred_start.strftime("%Y-%m-%d %H:%M:%S")
+        # Add temporal fields if available
+        for field_name in ("occurred_start", "occurred_end", "mentioned_at"):
+            value = getattr(fact, field_name, None)
+            if value:
+                if isinstance(value, str):
+                    fact_obj[field_name] = value
+                elif isinstance(value, datetime):
+                    fact_obj[field_name] = value.strftime("%Y-%m-%d %H:%M:%S")
 
         formatted.append(fact_obj)
 
-    return json.dumps(formatted, indent=2)
+    return json.dumps(formatted, indent=2, ensure_ascii=False)
 
 
 def format_entity_summaries_for_prompt(entities: dict) -> str:
@@ -110,11 +111,7 @@ def build_think_prompt(
     context: str | None = None,
     entity_summaries_text: str | None = None,
 ) -> str:
-    """Build the think prompt for the LLM.
-
-    Note: opinion_facts_text parameter removed - opinions are now stored as mental models
-    and included via entity_summaries_text.
-    """
+    """Build the think prompt for the LLM."""
     disposition_desc = build_disposition_description(disposition)
 
     name_section = f"""

hindsight-api-slim/hindsight_api/engine/search/trace.py

@@ -131,7 +131,7 @@ class RetrievalResult(BaseModel):
     text: str = Field(description="Memory unit text content")
     context: str = Field(default="", description="Memory unit context")
     event_date: datetime | None = Field(default=None, description="When the memory occurred")
-    fact_type: str | None = Field(default=None, description="Fact type (world, experience, opinion)")
+    fact_type: str | None = Field(default=None, description="Fact type (world, experience)")
     score: float = Field(description="Score from this retrieval method")
     score_name: str = Field(description="Name of the score (e.g., 'similarity', 'bm25_score', 'activation')")
 
@@ -140,9 +140,7 @@ class RetrievalMethodResults(BaseModel):
     """Results from a single retrieval method."""
 
     method_name: Literal["semantic", "bm25", "graph", "temporal"] = Field(description="Name of retrieval method")
-    fact_type: str | None = Field(
-        default=None, description="Fact type this retrieval was for (world, experience, opinion)"
-    )
+    fact_type: str | None = Field(default=None, description="Fact type this retrieval was for (world, experience)")
     results: list[RetrievalResult] = Field(description="Retrieved results with ranks")
     duration_seconds: float = Field(description="Time taken for this retrieval")
     metadata: dict[str, Any] = Field(default_factory=dict, description="Method-specific metadata")

hindsight-api-slim/hindsight_api/engine/search/tracer.py

@@ -319,7 +319,7 @@ class SearchTracer:
             duration_seconds: Time taken for this retrieval
             score_field: Field name containing the score in data dict
             metadata: Optional metadata about this retrieval method
-            fact_type: Fact type this retrieval was for (world, experience, opinion)
+            fact_type: Fact type this retrieval was for (world, experience)
         """
         retrieval_results = []
         for rank, (doc_id, data) in enumerate(results, start=1):

hindsight-api-slim/hindsight_api/engine/search/types.py

@@ -11,8 +11,8 @@ from typing import Any
 
 
 @dataclass
-class MPFPTimings:
-    """Timing breakdown for a single MPFP retrieval call."""
+class GraphRetrievalTimings:
+    """Timing breakdown for a single graph retrieval call."""
 
     fact_type: str
     edge_count: int = 0  # Total edges loaded
@@ -48,6 +48,7 @@ class RetrievalResult:
     chunk_id: str | None = None
     tags: list[str] | None = None  # Visibility scope tags
     metadata: dict[str, str] | None = None  # User-provided metadata
+    proof_count: int | None = None  # Number of supporting memories (observations only)
 
     # Retrieval-specific scores (only one will be set depending on retrieval method)
     similarity: float | None = None  # Semantic retrieval
@@ -72,6 +73,7 @@ class RetrievalResult:
             chunk_id=row.get("chunk_id"),
             tags=row.get("tags"),
             metadata=row.get("metadata"),
+            proof_count=row.get("proof_count"),
             similarity=row.get("similarity"),
             bm25_score=row.get("bm25_score"),
             activation=row.get("activation"),

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

@@ -82,20 +82,16 @@ class TaskBackend(ABC):
 
         Args:
             task_dict: Task dictionary to execute
+
+        Raises:
+            Exception: Re-raised from executor on failure.
         """
         if self._executor is None:
             task_type = task_dict.get("type", "unknown")
             logger.warning(f"No executor registered, skipping task {task_type}")
             return
 
-        try:
-            await self._executor(task_dict)
-        except Exception as e:
-            task_type = task_dict.get("type", "unknown")
-            logger.error(f"Error executing task {task_type}: {e}")
-            import traceback
-
-            traceback.print_exc()
+        await self._executor(task_dict)
 
 
 class SyncTaskBackend(TaskBackend):
@@ -197,17 +193,21 @@ class BrokerTaskBackend(TaskBackend):
         table = fq_table("async_operations", schema)
 
         if operation_id:
-            # Update existing operation with task payload
+            # Callers now include task_payload in the same INSERT that creates the
+            # async_operations row (see MemoryEngine._submit_async_operation). The
+            # WHERE clause guards against overwriting that payload — the UPDATE is a
+            # no-op when the row is already claimable, and only fills in a NULL payload
+            # for any legacy caller that still creates the row first.
             await pool.execute(
                 f"""
                 UPDATE {table}
                 SET task_payload = $1::jsonb, updated_at = now()
-                WHERE operation_id = $2
+                WHERE operation_id = $2 AND task_payload IS NULL
                 """,
                 payload_json,
                 operation_id,
             )
-            logger.debug(f"Updated task payload for operation {operation_id}")
+            logger.debug(f"submit_task UPDATE for operation {operation_id} (no-op if payload already set)")
         else:
             # Insert new operation (for tasks without pre-created records)
             # e.g., access_count_update tasks

hindsight-api-slim/hindsight_api/extensions/__init__.py

@@ -55,6 +55,7 @@ from hindsight_api.extensions.tenant import (
     TenantExtension,
 )
 from hindsight_api.models import RequestContext
+from hindsight_api.worker.exceptions import DeferOperation
 
 __all__ = [
     # Base
@@ -68,6 +69,7 @@ __all__ = [
     # MCP Extension
     "MCPExtension",
     # Operation Validator - Core
+    "DeferOperation",
     "OperationValidationError",
     "OperationValidatorExtension",
     "RecallContext",

hindsight-api-slim/hindsight_api/extensions/operation_validator.py

@@ -96,7 +96,6 @@ class RetainContext:
     request_context: "RequestContext"
     document_id: str | None = None
     fact_type_override: str | None = None
-    confidence_score: float | None = None
 
 
 @dataclass
@@ -169,7 +168,6 @@ class RetainResult:
     request_context: "RequestContext"
     document_id: str | None
     fact_type_override: str | None
-    confidence_score: float | None
     # Result
     unit_ids: list[list[str]]  # List of unit IDs per content item
     success: bool = True
@@ -378,6 +376,16 @@ class OperationValidatorExtension(Extension, ABC):
         2. [operation executes]
         3. on_*_complete (post-operation)
 
+    Outcomes for `validate_*` hooks:
+        - accept: return `ValidationResult.accept()` (or `accept_with(...)`)
+        - reject: return `ValidationResult.reject(reason, status_code)`
+          (raises `OperationValidationError` upstream)
+        - defer: raise `DeferOperation(exec_date, reason)` from
+          `hindsight_api.worker.exceptions` to requeue the task for a
+          future time without bumping `retry_count`. Worker-only — do
+          not raise from `validate_recall` / `validate_reflect` in
+          synchronous HTTP request paths, where it surfaces as a 500.
+
     Supported operations:
         - retain, recall, reflect (core memory operations)
         - consolidate (mental models consolidation)
@@ -402,7 +410,6 @@ class OperationValidatorExtension(Extension, ABC):
                 - request_context: Request context with auth info
                 - document_id: Optional document ID
                 - fact_type_override: Optional fact type override
-                - confidence_score: Optional confidence score
 
         Returns:
             ValidationResult indicating whether the operation is allowed.

hindsight-api-slim/hindsight_api/mcp_tools.py

@@ -29,6 +29,7 @@ from hindsight_api.models import RequestContext
 _ALL_TOOLS: frozenset[str] = frozenset(
     {
         "retain",
+        "sync_retain",
         "recall",
         "reflect",
         "list_banks",
@@ -139,6 +140,7 @@ def build_content_dict(
     metadata: dict[str, str] | None = None,
     document_id: str | None = None,
     strategy: str | None = None,
+    update_mode: str | None = None,
 ) -> tuple[dict[str, Any], str | None]:
     """Build a content dict for retain operations.
 
@@ -150,6 +152,7 @@ def build_content_dict(
         metadata: Optional key-value metadata to attach to the memory
         document_id: Optional document ID to associate the memory with
         strategy: Optional named retain strategy override (e.g., 'exact', 'verbose')
+        update_mode: How to handle existing documents ('replace' or 'append')
 
     Returns:
         Tuple of (content_dict, error_message). error_message is None if successful.
@@ -184,6 +187,8 @@ def build_content_dict(
         content_dict["document_id"] = document_id
     if strategy is not None:
         content_dict["strategy"] = strategy
+    if update_mode is not None:
+        content_dict["update_mode"] = update_mode
 
     return content_dict, None
 
@@ -202,6 +207,7 @@ def register_mcp_tools(
     """
     tools_to_register = config.tools or {
         "retain",
+        "sync_retain",
         "recall",
         "reflect",
         "list_banks",
@@ -235,6 +241,9 @@ def register_mcp_tools(
     if "retain" in tools_to_register:
         _register_retain(mcp, memory, config)
 
+    if "sync_retain" in tools_to_register:
+        _register_sync_retain(mcp, memory, config)
+
     if "recall" in tools_to_register:
         _register_recall(mcp, memory, config)
 
@@ -482,7 +491,7 @@ def _apply_audit_logging(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsCon
     if hasattr(mcp, "_tool_manager"):
         # FastMCP 2.x
         try:
-            for name, tool in mcp._tool_manager._tools.items():
+            for name, tool in mcp._tool_manager._tools.items():  # type: ignore[unresolved-attribute]  # FastMCP 2.x internal; guarded by hasattr
                 if name in _AUDITABLE_MCP_TOOLS:
                     object.__setattr__(tool, "run", _wrap_tool_run(name, tool.run))
         except (AttributeError, KeyError) as e:
@@ -539,6 +548,7 @@ def _register_retain(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig)
             document_id: str | None = None,
             bank_id: str | None = None,
             strategy: str | None = None,
+            update_mode: str | None = None,
         ) -> dict:
             """
             Args:
@@ -550,12 +560,15 @@ def _register_retain(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig)
                 document_id: Optional document ID to associate this memory with
                 bank_id: Optional bank to store in (defaults to session bank). Use for cross-bank operations.
                 strategy: Optional named retain strategy (e.g., 'exact' for verbatim storage). Strategies are defined in the bank config.
+                update_mode: How to handle existing documents with the same document_id. 'replace' (default) or 'append' (concatenates new content to existing).
             """
             target_bank = bank_id or config.bank_id_resolver()
             if target_bank is None:
                 return {"status": "error", "message": "No bank_id configured"}
 
-            content_dict, error = build_content_dict(content, context, timestamp, tags, metadata, document_id, strategy)
+            content_dict, error = build_content_dict(
+                content, context, timestamp, tags, metadata, document_id, strategy, update_mode
+            )
             if error:
                 return {"status": "error", "message": error}
 
@@ -590,6 +603,7 @@ def _register_retain(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig)
             metadata: dict[str, str] | None = None,
             document_id: str | None = None,
             strategy: str | None = None,
+            update_mode: str | None = None,
         ) -> dict:
             """
             Args:
@@ -600,12 +614,15 @@ def _register_retain(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig)
                 metadata: Optional key-value metadata to attach (e.g., {'source': 'slack', 'channel': 'general'})
                 document_id: Optional document ID to associate this memory with
                 strategy: Optional named retain strategy (e.g., 'exact' for verbatim storage). Strategies are defined in the bank config.
+                update_mode: How to handle existing documents with the same document_id. 'replace' (default) or 'append' (concatenates new content to existing).
             """
             target_bank = config.bank_id_resolver()
             if target_bank is None:
                 return {"status": "error", "message": "No bank_id configured"}
 
-            content_dict, error = build_content_dict(content, context, timestamp, tags, metadata, document_id, strategy)
+            content_dict, error = build_content_dict(
+                content, context, timestamp, tags, metadata, document_id, strategy, update_mode
+            )
             if error:
                 return {"status": "error", "message": error}
 
@@ -630,6 +647,124 @@ def _register_retain(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig)
                 return {"status": "error", "message": str(e)}
 
 
+def _register_sync_retain(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
+    """Register the sync_retain tool (synchronous retain that waits for completion)."""
+
+    if config.include_bank_id_param:
+
+        @mcp.tool()
+        async def sync_retain(
+            content: str,
+            context: str = "general",
+            timestamp: str | None = None,
+            tags: list[str] | None = None,
+            metadata: dict[str, str] | None = None,
+            document_id: str | None = None,
+            bank_id: str | None = None,
+            strategy: str | None = None,
+        ) -> dict:
+            """Store information to long-term memory and wait for completion.
+
+            Unlike retain (which is asynchronous), this tool blocks until the memory
+            is fully stored and immediately available for recall.
+
+            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'
+                timestamp: When this event/fact occurred (ISO format, e.g., '2024-01-15T10:30:00Z'). Useful for timeline tracking.
+                tags: Optional tags for scoped visibility filtering (e.g., ['project:alpha', 'user:123'])
+                metadata: Optional key-value metadata to attach (e.g., {'source': 'slack', 'channel': 'general'})
+                document_id: Optional document ID to associate this memory with
+                bank_id: Optional bank to store in (defaults to session bank). Use for cross-bank operations.
+                strategy: Optional named retain strategy (e.g., 'exact' for verbatim storage). Strategies are defined in the bank config.
+            """
+            target_bank = bank_id or config.bank_id_resolver()
+            if target_bank is None:
+                return {"status": "error", "message": "No bank_id configured"}
+
+            content_dict, error = build_content_dict(content, context, timestamp, tags, metadata, document_id, strategy)
+            if error:
+                return {"status": "error", "message": error}
+
+            request_context = _get_request_context(config)
+
+            try:
+                result = await memory.retain_batch_async(
+                    bank_id=target_bank,
+                    contents=[content_dict],
+                    request_context=request_context,
+                    strategy=content_dict.pop("strategy", None),
+                )
+                memory_ids = [uid for batch in result for uid in batch]
+                return {
+                    "status": "completed",
+                    "message": "Memory stored successfully",
+                    "memory_ids": memory_ids,
+                }
+            except OperationValidationError as e:
+                logger.warning(f"Sync retain rejected: {e}")
+                return {"status": "error", "message": str(e)}
+            except Exception as e:
+                logger.error(f"Error in sync retain: {e}", exc_info=True)
+                return {"status": "error", "message": str(e)}
+
+    else:
+
+        @mcp.tool()
+        async def sync_retain(
+            content: str,
+            context: str = "general",
+            timestamp: str | None = None,
+            tags: list[str] | None = None,
+            metadata: dict[str, str] | None = None,
+            document_id: str | None = None,
+            strategy: str | None = None,
+        ) -> dict:
+            """Store information to long-term memory and wait for completion.
+
+            Unlike retain (which is asynchronous), this tool blocks until the memory
+            is fully stored and immediately available for recall.
+
+            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'
+                timestamp: When this event/fact occurred (ISO format, e.g., '2024-01-15T10:30:00Z'). Useful for timeline tracking.
+                tags: Optional tags for scoped visibility filtering (e.g., ['project:alpha', 'user:123'])
+                metadata: Optional key-value metadata to attach (e.g., {'source': 'slack', 'channel': 'general'})
+                document_id: Optional document ID to associate this memory with
+                strategy: Optional named retain strategy (e.g., 'exact' for verbatim storage). Strategies are defined in the bank config.
+            """
+            target_bank = config.bank_id_resolver()
+            if target_bank is None:
+                return {"status": "error", "message": "No bank_id configured"}
+
+            content_dict, error = build_content_dict(content, context, timestamp, tags, metadata, document_id, strategy)
+            if error:
+                return {"status": "error", "message": error}
+
+            request_context = _get_request_context(config)
+
+            try:
+                result = await memory.retain_batch_async(
+                    bank_id=target_bank,
+                    contents=[content_dict],
+                    request_context=request_context,
+                    strategy=content_dict.pop("strategy", None),
+                )
+                memory_ids = [uid for batch in result for uid in batch]
+                return {
+                    "status": "completed",
+                    "message": "Memory stored successfully",
+                    "memory_ids": memory_ids,
+                }
+            except OperationValidationError as e:
+                logger.warning(f"Sync retain rejected: {e}")
+                return {"status": "error", "message": str(e)}
+            except Exception as e:
+                logger.error(f"Error in sync retain: {e}", exc_info=True)
+                return {"status": "error", "message": str(e)}
+
+
 def _register_recall(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
     """Register the recall tool."""
     description = config.recall_description or DEFAULT_MCP_RECALL_DESCRIPTION
@@ -1002,6 +1137,7 @@ def _register_list_mental_models(mcp: FastMCP, memory: MemoryEngine, config: MCP
         @mcp.tool()
         async def list_mental_models(
             tags: list[str] | None = None,
+            detail: str = "full",
             bank_id: str | None = None,
         ) -> str:
             """
@@ -1013,6 +1149,7 @@ def _register_list_mental_models(mcp: FastMCP, memory: MemoryEngine, config: MCP
 
             Args:
                 tags: Optional tags to filter by (returns models matching any tag)
+                detail: Detail level - 'metadata' (names/tags only), 'content' (adds content/config), 'full' (includes reflect_response). Default: 'full'
                 bank_id: Optional bank to list from (defaults to session bank). Use for cross-bank operations.
             """
             try:
@@ -1023,6 +1160,7 @@ def _register_list_mental_models(mcp: FastMCP, memory: MemoryEngine, config: MCP
                 models = await memory.list_mental_models(
                     bank_id=target_bank,
                     tags=tags,
+                    detail=detail,
                     request_context=_get_request_context(config),
                 )
                 return json.dumps({"items": models}, indent=2, default=str)
@@ -1038,6 +1176,7 @@ def _register_list_mental_models(mcp: FastMCP, memory: MemoryEngine, config: MCP
         @mcp.tool()
         async def list_mental_models(
             tags: list[str] | None = None,
+            detail: str = "full",
         ) -> dict:
             """
             List mental models (pinned reflections) for this memory bank.
@@ -1048,6 +1187,7 @@ def _register_list_mental_models(mcp: FastMCP, memory: MemoryEngine, config: MCP
 
             Args:
                 tags: Optional tags to filter by (returns models matching any tag)
+                detail: Detail level - 'metadata' (names/tags only), 'content' (adds content/config), 'full' (includes reflect_response). Default: 'full'
             """
             try:
                 target_bank = config.bank_id_resolver()
@@ -1057,6 +1197,7 @@ def _register_list_mental_models(mcp: FastMCP, memory: MemoryEngine, config: MCP
                 models = await memory.list_mental_models(
                     bank_id=target_bank,
                     tags=tags,
+                    detail=detail,
                     request_context=_get_request_context(config),
                 )
                 return {"items": models}
@@ -1076,16 +1217,18 @@ def _register_get_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPTo
         @mcp.tool()
         async def get_mental_model(
             mental_model_id: str,
+            detail: str = "full",
             bank_id: str | None = None,
         ) -> str:
             """
             Get a specific mental model by ID.
 
-            Returns the full mental model including its generated content, source query,
-            and metadata. Use list_mental_models first to discover available model IDs.
+            Returns the mental model with the requested detail level. Use list_mental_models
+            first to discover available model IDs.
 
             Args:
                 mental_model_id: The ID of the mental model to retrieve
+                detail: Detail level - 'metadata' (names/tags only), 'content' (adds content/config), 'full' (includes reflect_response). Default: 'full'
                 bank_id: Optional bank (defaults to session bank). Use for cross-bank operations.
             """
             try:
@@ -1096,6 +1239,7 @@ def _register_get_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPTo
                 model = await memory.get_mental_model(
                     bank_id=target_bank,
                     mental_model_id=mental_model_id,
+                    detail=detail,
                     request_context=_get_request_context(config),
                 )
                 if model is None:
@@ -1113,15 +1257,17 @@ def _register_get_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPTo
         @mcp.tool()
         async def get_mental_model(
             mental_model_id: str,
+            detail: str = "full",
         ) -> dict:
             """
             Get a specific mental model by ID.
 
-            Returns the full mental model including its generated content, source query,
-            and metadata. Use list_mental_models first to discover available model IDs.
+            Returns the mental model with the requested detail level. Use list_mental_models
+            first to discover available model IDs.
 
             Args:
                 mental_model_id: The ID of the mental model to retrieve
+                detail: Detail level - 'metadata' (names/tags only), 'content' (adds content/config), 'full' (includes reflect_response). Default: 'full'
             """
             try:
                 target_bank = config.bank_id_resolver()
@@ -1131,6 +1277,7 @@ def _register_get_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPTo
                 model = await memory.get_mental_model(
                     bank_id=target_bank,
                     mental_model_id=mental_model_id,
+                    detail=detail,
                     request_context=_get_request_context(config),
                 )
                 if model is None:
@@ -2707,6 +2854,44 @@ def _register_get_bank_stats(mcp: FastMCP, memory: MemoryEngine, config: MCPTool
             return f'{{"error": "{e}"}}'
 
 
+async def _do_update_bank(
+    memory: MemoryEngine,
+    target_bank: str,
+    request_context: RequestContext,
+    *,
+    name: str | None = None,
+    mission: str | None = None,
+    config_updates: dict[str, Any] | None = None,
+) -> dict[str, Any]:
+    """Shared implementation for update_bank MCP tool variants.
+
+    Args:
+        name: Display name (stored in banks table).
+        mission: Deprecated alias for reflect_mission — mapped into config_updates.
+        config_updates: Arbitrary config overrides passed to config_resolver.update_bank_config().
+            Supports all configurable fields (retain_mission, disposition_*, etc.).
+            The config resolver validates keys and rejects non-configurable/credential fields.
+    """
+    # Update display name via engine (stored in DB banks table)
+    if name is not None:
+        await memory.update_bank(
+            target_bank,
+            name=name,
+            request_context=request_context,
+        )
+
+    # Merge deprecated mission alias into config_updates as reflect_mission
+    effective_config: dict[str, Any] = dict(config_updates) if config_updates else {}
+    if mission is not None and "reflect_mission" not in effective_config:
+        effective_config["reflect_mission"] = mission
+
+    if effective_config:
+        await memory._config_resolver.update_bank_config(target_bank, effective_config, request_context)
+
+    # Return updated profile
+    return await memory.get_bank_profile(target_bank, request_context=request_context)
+
+
 def _register_update_bank(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
     """Register the update_bank tool."""
 
@@ -2716,16 +2901,37 @@ def _register_update_bank(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsCo
         async def update_bank(
             name: str | None = None,
             mission: str | None = None,
+            config_updates: dict[str, Any] | None = None,
             bank_id: str | None = None,
         ) -> str:
             """
-            Update a memory bank's metadata.
+            Update a memory bank's configuration.
 
-            Changes the name or mission of an existing bank.
+            Updates the bank's name and/or any bank-level configuration fields.
+            Only provided fields will be updated; omitted fields remain unchanged.
 
             Args:
-                name: New human-friendly name for the bank
-                mission: New mission describing who the agent is and what they're trying to accomplish
+                name: Human-friendly display name for the bank.
+                mission: Deprecated alias for config_updates.reflect_mission.
+                config_updates: Dictionary of configuration fields to update. Supports all
+                    bank-configurable fields including:
+                    - reflect_mission: Mission/context for Reflect operations.
+                    - retain_mission: Steers what gets extracted during retain().
+                    - retain_extraction_mode: 'concise' (default), 'verbose', or 'custom'.
+                    - retain_custom_instructions: Custom extraction prompt (active when mode is 'custom').
+                    - retain_chunk_size: Maximum token size for each content chunk.
+                    - retain_chunk_batch_size: Number of chunks to process in parallel.
+                    - enable_observations: Toggle observation consolidation after retain().
+                    - observations_mission: Controls observation synthesis rules.
+                    - disposition_skepticism: Critical evaluation level (1-5).
+                    - disposition_literalism: Literal vs. abstract interpretation (1-5).
+                    - disposition_empathy: Emotional context consideration (1-5).
+                    - entity_labels: Controlled vocabulary for entity classification.
+                    - entities_allow_free_form: Allow labels outside entity_labels.
+                    - recall_include_chunks: Include raw chunks in recall results.
+                    - recall_max_tokens: Max tokens for recall results.
+                    - mcp_enabled_tools: Tool allowlist for this bank.
+                    Any configurable field name is accepted (use Python field names).
                 bank_id: Optional bank (defaults to session bank). Use for cross-bank operations.
             """
             try:
@@ -2733,14 +2939,16 @@ def _register_update_bank(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsCo
                 if target_bank is None:
                     return '{"error": "No bank_id configured"}'
 
-                result = await memory.update_bank(
+                result = await _do_update_bank(
+                    memory,
                     target_bank,
+                    _get_request_context(config),
                     name=name,
                     mission=mission,
-                    request_context=_get_request_context(config),
+                    config_updates=config_updates,
                 )
                 return json.dumps(result, indent=2, default=str)
-            except OperationValidationError as e:
+            except (OperationValidationError, ValueError) as e:
                 logger.warning(f"Operation rejected: {e}")
                 return json.dumps({"error": str(e)})
             except Exception as e:
@@ -2753,29 +2961,52 @@ def _register_update_bank(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsCo
         async def update_bank(
             name: str | None = None,
             mission: str | None = None,
+            config_updates: dict[str, Any] | None = None,
         ) -> dict:
             """
-            Update this memory bank's metadata.
+            Update this memory bank's configuration.
 
-            Changes the name or mission of the bank.
+            Updates the bank's name and/or any bank-level configuration fields.
+            Only provided fields will be updated; omitted fields remain unchanged.
 
             Args:
-                name: New human-friendly name for the bank
-                mission: New mission describing who the agent is and what they're trying to accomplish
+                name: Human-friendly display name for the bank.
+                mission: Deprecated alias for config_updates.reflect_mission.
+                config_updates: Dictionary of configuration fields to update. Supports all
+                    bank-configurable fields including:
+                    - reflect_mission: Mission/context for Reflect operations.
+                    - retain_mission: Steers what gets extracted during retain().
+                    - retain_extraction_mode: 'concise' (default), 'verbose', or 'custom'.
+                    - retain_custom_instructions: Custom extraction prompt (active when mode is 'custom').
+                    - retain_chunk_size: Maximum token size for each content chunk.
+                    - retain_chunk_batch_size: Number of chunks to process in parallel.
+                    - enable_observations: Toggle observation consolidation after retain().
+                    - observations_mission: Controls observation synthesis rules.
+                    - disposition_skepticism: Critical evaluation level (1-5).
+                    - disposition_literalism: Literal vs. abstract interpretation (1-5).
+                    - disposition_empathy: Emotional context consideration (1-5).
+                    - entity_labels: Controlled vocabulary for entity classification.
+                    - entities_allow_free_form: Allow labels outside entity_labels.
+                    - recall_include_chunks: Include raw chunks in recall results.
+                    - recall_max_tokens: Max tokens for recall results.
+                    - mcp_enabled_tools: Tool allowlist for this bank.
+                    Any configurable field name is accepted (use Python field names).
             """
             try:
                 target_bank = config.bank_id_resolver()
                 if target_bank is None:
                     return {"error": "No bank_id configured"}
 
-                result = await memory.update_bank(
+                result = await _do_update_bank(
+                    memory,
                     target_bank,
+                    _get_request_context(config),
                     name=name,
                     mission=mission,
-                    request_context=_get_request_context(config),
+                    config_updates=config_updates,
                 )
                 return result
-            except OperationValidationError as e:
+            except (OperationValidationError, ValueError) as e:
                 logger.warning(f"Operation rejected: {e}")
                 return {"error": str(e)}
             except Exception as e:
@@ -2873,6 +3104,7 @@ def _register_clear_memories(mcp: FastMCP, memory: MemoryEngine, config: MCPTool
                 result = await memory.delete_bank(
                     target_bank,
                     fact_type=type,
+                    delete_bank_profile=False,
                     request_context=_get_request_context(config),
                 )
                 return json.dumps({"status": "cleared", "bank_id": target_bank, **result}, default=str)
@@ -2905,6 +3137,7 @@ def _register_clear_memories(mcp: FastMCP, memory: MemoryEngine, config: MCPTool
                 result = await memory.delete_bank(
                     target_bank,
                     fact_type=type,
+                    delete_bank_profile=False,
                     request_context=_get_request_context(config),
                 )
                 return {"status": "cleared", "bank_id": target_bank, **result}

hindsight-api-slim/hindsight_api/metrics.py

@@ -252,6 +252,9 @@ class MetricsCollector(MetricsCollectorBase):
 
     def __init__(self):
         self.meter = get_meter()
+        from .config import get_config
+
+        self._include_bank_id = get_config().metrics_include_bank_id
 
         # Operation latency histogram (in seconds)
         # Records duration of retain, recall, reflect operations
@@ -332,10 +335,11 @@ class MetricsCollector(MetricsCollectorBase):
         start_time = time.time()
         attributes = {
             "operation": operation,
-            "bank_id": bank_id,
             "source": source,
             "tenant": _get_tenant(),
         }
+        if self._include_bank_id:
+            attributes["bank_id"] = bank_id
         if budget:
             attributes["budget"] = budget
         if max_tokens:

hindsight-api-slim/hindsight_api/models.py

@@ -62,7 +62,6 @@ class Document(Base):
     bank_id: Mapped[str] = mapped_column(Text, primary_key=True)
     original_text: Mapped[str | None] = mapped_column(Text)
     content_hash: Mapped[str | None] = mapped_column(Text)
-    doc_metadata: Mapped[dict] = mapped_column("metadata", JSONB, server_default=sql_text("'{}'::jsonb"))
     created_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
     updated_at: Mapped[datetime] = mapped_column(TIMESTAMP(timezone=True), server_default=func.now())
 
@@ -97,7 +96,6 @@ class MemoryUnit(Base):
     occurred_end: Mapped[datetime | None] = mapped_column(TIMESTAMP(timezone=True))  # When fact occurred (range end)
     mentioned_at: Mapped[datetime | None] = mapped_column(TIMESTAMP(timezone=True))  # When fact was mentioned
     fact_type: Mapped[str] = mapped_column(Text, nullable=False, server_default="world")
-    confidence_score: Mapped[float | None] = mapped_column(Float)
     unit_metadata: Mapped[dict] = mapped_column(
         "metadata", JSONB, server_default=sql_text("'{}'::jsonb")
     )  # User-defined metadata (str->str)
@@ -121,14 +119,7 @@ class MemoryUnit(Base):
             name="memory_units_document_fkey",
             ondelete="CASCADE",
         ),
-        CheckConstraint("fact_type IN ('world', 'experience', 'opinion', 'observation')"),
-        CheckConstraint("confidence_score IS NULL OR (confidence_score >= 0.0 AND confidence_score <= 1.0)"),
-        CheckConstraint(
-            "(fact_type = 'opinion' AND confidence_score IS NOT NULL) OR "
-            "(fact_type = 'observation') OR "
-            "(fact_type NOT IN ('opinion', 'observation') AND confidence_score IS NULL)",
-            name="confidence_score_fact_type_check",
-        ),
+        CheckConstraint("fact_type IN ('world', 'experience', 'observation')"),
         Index("idx_memory_units_bank_id", "bank_id"),
         Index("idx_memory_units_document_id", "document_id"),
         Index("idx_memory_units_event_date", "event_date", postgresql_ops={"event_date": "DESC"}),
@@ -142,20 +133,6 @@ class MemoryUnit(Base):
             "event_date",
             postgresql_ops={"event_date": "DESC"},
         ),
-        Index(
-            "idx_memory_units_opinion_confidence",
-            "bank_id",
-            "confidence_score",
-            postgresql_where=sql_text("fact_type = 'opinion'"),
-            postgresql_ops={"confidence_score": "DESC"},
-        ),
-        Index(
-            "idx_memory_units_opinion_date",
-            "bank_id",
-            "event_date",
-            postgresql_where=sql_text("fact_type = 'opinion'"),
-            postgresql_ops={"event_date": "DESC"},
-        ),
         Index(
             "idx_memory_units_observation_date",
             "bank_id",

hindsight-api-slim/hindsight_api/worker/exceptions.py

@@ -7,3 +7,24 @@ class RetryTaskAt(Exception):
     def __init__(self, retry_at: datetime, message: str = ""):
         self.retry_at = retry_at
         super().__init__(message)
+
+
+class DeferOperation(Exception):
+    """Raise from an extension hook (or task handler) to requeue the
+    operation for execution at a later time, without counting as a retry.
+
+    Unlike `RetryTaskAt`, this is not a failure: `retry_count` is not
+    incremented and `error_message` is not written. Use this for
+    backpressure / "not yet, try later" decisions made before or during
+    task execution (e.g. quota windows, warming dependencies, upstream
+    rate limits).
+
+    Worker-only: raising this from a hook called in HTTP request context
+    (e.g. `validate_recall` for a synchronous recall) will surface as an
+    unhandled 500 — there is no queue to defer to.
+    """
+
+    def __init__(self, exec_date: datetime, reason: str = ""):
+        self.exec_date = exec_date
+        self.reason = reason
+        super().__init__(reason)