tests

docs
doc
2026-01-19 18:34:32 +01:00 · 2026-01-19 17:12:55 +01:00 · 2026-01-19 16:44:45 +01:00 · 2026-01-19 16:32:18 +01:00 · 2026-01-19 14:27:31 +01:00 · 2026-01-19 14:18:33 +01:00
209 changed files with 35392 additions and 4449 deletions
@@ -153,8 +153,15 @@ jobs:
    - name: Build docs
      run: npm run build --workspace=hindsight-docs

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

    steps:
    - uses: actions/checkout@v4
@@ -171,6 +178,10 @@ jobs:
          hindsight-cli/target
        key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}

+    - name: Run unit tests
+      working-directory: hindsight-cli
+      run: cargo test
+
    - name: Build CLI
      working-directory: hindsight-cli
      run: cargo build --release
@@ -182,29 +193,6 @@ jobs:
        path: hindsight-cli/target/release/hindsight
        retention-days: 1

-  test-rust-cli:
-    runs-on: ubuntu-latest
-    needs: build-rust-cli
-    env:
-      HINDSIGHT_API_LLM_PROVIDER: groq
-      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
-      HINDSIGHT_API_URL: http://localhost:8888
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Download CLI artifact
-      uses: actions/download-artifact@v4
-      with:
-        name: hindsight-cli
-        path: /tmp/cli
-
-    - name: Make CLI executable
-      run: chmod +x /tmp/cli/hindsight
-
    - name: Install uv
      uses: astral-sh/setup-uv@v5
      with:
@@ -251,7 +239,7 @@ jobs:

    - name: Run CLI smoke test
      run: |
-        HINDSIGHT_CLI=/tmp/cli/hindsight ./hindsight-cli/smoke-test.sh
+        HINDSIGHT_CLI=hindsight-cli/target/release/hindsight ./hindsight-cli/smoke-test.sh

    - name: Show API server logs
      if: always()
@@ -777,7 +765,7 @@ jobs:

  test-doc-examples:
    runs-on: ubuntu-latest
-    needs: build-rust-cli
+    needs: test-rust-cli
    env:
      HINDSIGHT_API_LLM_PROVIDER: groq
      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
@@ -29,7 +29,7 @@ nltk_data/

 # Monitoring stack (Prometheus/Grafana binaries and data)
 .monitoring/
-.pgbouncer
+.pgbouncer/

 # Large benchmark datasets (will be downloaded automatically)
 **/longmemeval_s_cleaned.json
@@ -1,153 +1,3 @@
 # AGENTS.md

-This document captures architectural decisions and coding conventions for the Hindsight project.
-
-## Documentation
-
- **Main documentation**: [hindsight-docs/docs/developer/](./hindsight-docs/docs/developer/)
- **Use case patterns**: [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/)
- **API reference**: Auto-generated from OpenAPI spec
-
-## Project Structure
-
-```
-hindsight/              # Python package for embedded usage
-hindsight-api/          # FastAPI server (core memory engine)
-hindsight-cli/          # Rust CLI client
-hindsight-embed/        # Embedded CLI (no server needed)
-hindsight-control-plane/ # Next.js admin UI
-hindsight-docs/         # Docusaurus documentation site
-hindsight-dev/          # Development tools and benchmarks
-hindsight-integrations/ # Framework integrations (LangChain, etc.)
-hindsight-clients/      # Generated API clients (Python, TypeScript, Rust)
-```
-
-## Core Concepts
-
-### Memory Banks
- Each bank is an isolated memory store (like a "brain" for one user/agent)
- Banks contain: memory units (facts), entities, documents, entity links
- Banks have a **disposition** (personality traits) and **background** (context)
- Bank isolation is strict - no cross-bank data leakage
-
-### Memory Types
- **World facts**: General knowledge ("The sky is blue")
- **Experience facts**: Personal experiences ("I visited Paris in 2023")
- **Opinion facts**: Beliefs with confidence scores ("Paris is beautiful" - 0.9 confidence)
-
-### Operations
- **Retain**: Store new memories (extracts facts, entities, relationships)
- **Recall**: Retrieve memories (semantic, BM25, graph, temporal search)
- **Reflect**: Deep analysis to form new insights/opinions
-
-## API Design Decisions
-
-### Single Bank Per Request
- All API endpoints (`recall`, `reflect`, `retain`) operate on a single bank
- Multi-bank queries are the **client/agent's responsibility** to orchestrate
- This keeps the API simple and the isolation model clear
-
-### Disposition Traits (3-trait system)
- **Skepticism** (1-5): How skeptical vs trusting when forming opinions
- **Literalism** (1-5): How literally to interpret information
- **Empathy** (1-5): How much to consider emotional context
- These influence the `reflect` operation, not `recall`
- Background info also only affects `reflect` (opinion formation)
-
-## Multi-Bank Architecture Patterns
-
-See [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/) for detailed guides:
-
- **Per-User Memory**: One bank per user, simplest pattern
- **Support Agent + Shared Knowledge**: User bank + shared docs bank, client orchestrates
-
-## Developer Guide
-
-### Running the API Server
-
-```bash
-# From project root
-./scripts/dev/start-api.sh
-
-# With options
-./scripts/dev/start-api.sh --reload --port 8888 --log-level debug
-```
-
-### Running Tests
-
-```bash
-# API tests
-cd hindsight-api
-uv run pytest tests/
-
-# Specific test
-uv run pytest tests/test_http_api_integration.py -v
-```
-
-### Generating OpenAPI Spec
-
-After changing API endpoints, regenerate the OpenAPI spec and docs:
-
-```bash
-./scripts/generate-openapi.sh
-```
-
-This will:
-1. Generate `openapi.json` at project root
-2. Copy to `hindsight-docs/openapi.json`
-3. Regenerate API reference documentation
-
-### Generating API Clients
-
-After updating the OpenAPI spec, regenerate all clients:
-
-```bash
-./scripts/generate-clients.sh
-```
-
-This generates:
- **Rust client**: `hindsight-clients/rust/` (via progenitor in build.rs)
- **Python client**: `hindsight-clients/python/` (via openapi-generator Docker)
- **TypeScript client**: `hindsight-clients/typescript/` (via @hey-api/openapi-ts)
-
-Note: The maintained wrapper `hindsight_client.py` and `README.md` are preserved during regeneration.
-
-### Running the Documentation Site
-
-```bash
-./scripts/dev/start-docs.sh
-```
-
-### Running the Control Plane
-
-```bash
-./scripts/dev/start-control-plane.sh
-```
-
-## Code Style
-
-### Python (hindsight-api)
- Use `uv` for package management
- Async throughout (asyncpg, async FastAPI endpoints)
- Pydantic models for request/response validation
- No py files at project root - maintain clean directory structure
-
-### TypeScript (control-plane, clients)
- Next.js with App Router for control plane
- Tailwind CSS with shadcn/ui components
-
-### Rust (CLI)
- Async with tokio
- reqwest for HTTP client
- progenitor for API client generation
-
-## Database
-
- PostgreSQL with pgvector extension
- Schema managed via Alembic migrations in `hindsight-api/alembic/`, db migrations happen during api startup, no manual commands
- Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
-
-# Branding
-## Colors
- Primary: gradient from #0074d9 to #009296
-
+See [CLAUDE.md](./CLAUDE.md) for project documentation and coding conventions.
@@ -174,12 +174,63 @@ This runs the same checks as the pre-commit hook (Ruff for Python, ESLint/Pretti
 - Multi-bank queries are client responsibility to orchestrate
 - Disposition traits only affect reflect, not recall

+### Control Plane API Routes
+
+When adding or modifying parameters in the dataplane API (hindsight-api), you must also update the control plane routes that proxy to it:
+
+1. **API Routes** (`hindsight-control-plane/src/app/api/`):
+   - `recall/route.ts` - proxies to `/v1/default/banks/{bank_id}/memories/recall`
+   - `reflect/route.ts` - proxies to `/v1/default/banks/{bank_id}/reflect`
+   - `memories/retain/route.ts` - proxies to `/v1/default/banks/{bank_id}/memories/retain`
+   - Other routes follow the same pattern
+
+2. **Client types** (`hindsight-control-plane/src/lib/api.ts`):
+   - Update the TypeScript type definitions for `recall()`, `reflect()`, `retain()` etc.
+
+3. **Checklist when adding new API parameters**:
+   - Add parameter extraction in the route handler (destructure from `body`)
+   - Pass the parameter to the SDK call
+   - Update the client type definition in `lib/api.ts`
+   - Update any UI components that need to use the new parameter
+
 ### Python Style
 - Python 3.11+, type hints required
 - Async throughout (asyncpg, async FastAPI)
 - Pydantic models for request/response
 - Ruff for linting (line-length 120)
 - No Python files at project root - maintain clean directory structure
+- **Never use multi-item tuple return values** - prefer dataclass or Pydantic model for structured returns
+
+### Type Safety with Pydantic Models
+**NEVER use raw `dict` types for structured data.** Always use Pydantic models:
+- Use Pydantic `BaseModel` for all data structures passed between functions
+- Add `@field_validator` for type coercion (e.g., ensuring datetimes are timezone-aware)
+- Avoid `dict.get()` patterns - use typed model attributes instead
+- Parse external data (JSON, API responses) into Pydantic models at the boundary
+- This catches type errors at parse time, not deep in business logic
+
+```python
+# BAD - error-prone dict access
+def process(data: dict) -> str:
+    return data.get("name", "")  # No validation, silent failures
+
+# GOOD - typed and validated
+class UserData(BaseModel):
+    name: str
+    created_at: datetime
+
+    @field_validator("created_at", mode="before")
+    @classmethod
+    def ensure_tz_aware(cls, v):
+        if isinstance(v, str):
+            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
+        if v.tzinfo is None:
+            return v.replace(tzinfo=timezone.utc)
+        return v
+
+def process(data: UserData) -> str:
+    return data.name  # Type-safe, validated at construction
+```

 ### TypeScript Style
 - Next.js App Router for control plane
@@ -5,7 +5,7 @@
 [Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://vectorize.io/hindsight/cloud)

 [![CI](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml)
-[![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
+[![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 ![PyPI - Downloads](https://img.shields.io/pypi/dm/hindsight-api?label=PyPI)
 ![NPM Downloads](https://img.shields.io/npm/dm/%40vectorize-io%2Fhindsight-client?logoColor=orange&label=NPM&color=blue&link=https%3A%2F%2Fwww.npmjs.com%2Fpackage%2F%40vectorize-io%2Fhindsight-client)
@@ -242,7 +242,7 @@ client.reflect(bank_id="my-bank", query="What should I know about Alice?")
 - [CLI](https://hindsight.vectorize.io/sdks/cli)

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

 ---
@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.2.1
-appVersion: "0.2.1"
+version: 0.3.0
+appVersion: "0.3.0"
 keywords:
  - ai
  - memory
@@ -80,6 +80,22 @@ Control plane selector labels
 app.kubernetes.io/component: control-plane
 {{- end }}

+{{/*
+Worker labels
+*/}}
+{{- define "hindsight.worker.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: worker
+{{- end }}
+
+{{/*
+Worker selector labels
+*/}}
+{{- define "hindsight.worker.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: worker
+{{- end }}
+
 {{/*
 Create the name of the service account to use
 */}}
@@ -55,6 +55,11 @@ spec:
        {{- end }}
        - name: HINDSIGHT_API_DATABASE_URL
          value: {{ include "hindsight.databaseUrl" . | quote }}
+        {{- /* Disable internal worker when dedicated workers are enabled */}}
+        {{- if .Values.worker.enabled }}
+        - name: HINDSIGHT_API_WORKER_ENABLED
+          value: "false"
+        {{- end }}
        {{- range $key, $value := .Values.api.env }}
        - name: {{ $key }}
          value: {{ $value | quote }}
@@ -0,0 +1,25 @@
+{{- if .Values.worker.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-worker
+  labels:
+    {{- include "hindsight.worker.labels" . | nindent 4 }}
+  {{- if .Values.podAnnotations }}
+  annotations:
+    {{- /* Common Prometheus annotations for metrics scraping */}}
+    prometheus.io/scrape: "true"
+    prometheus.io/port: {{ .Values.worker.service.port | quote }}
+    prometheus.io/path: "/metrics"
+  {{- end }}
+spec:
+  # Headless service for StatefulSet (enables stable DNS names like worker-0.worker.namespace)
+  clusterIP: None
+  ports:
+  - port: {{ .Values.worker.service.port }}
+    targetPort: {{ .Values.worker.service.targetPort }}
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.worker.selectorLabels" . | nindent 4 }}
+{{- end }}
@@ -0,0 +1,110 @@
+{{- if .Values.worker.enabled }}
+apiVersion: apps/v1
+kind: StatefulSet
+metadata:
+  name: {{ include "hindsight.fullname" . }}-worker
+  labels:
+    {{- include "hindsight.worker.labels" . | nindent 4 }}
+spec:
+  serviceName: {{ include "hindsight.fullname" . }}-worker
+  replicas: {{ .Values.worker.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.worker.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      annotations:
+        {{- if not .Values.existingSecret }}
+        checksum/secret: {{ include (print $.Template.BasePath "/secret.yaml") . | sha256sum }}
+        {{- end }}
+        {{- with .Values.podAnnotations }}
+        {{- toYaml . | nindent 8 }}
+        {{- end }}
+      labels:
+        {{- include "hindsight.worker.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: worker
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version }}"
+        imagePullPolicy: {{ .Values.worker.image.pullPolicy }}
+        command: ["hindsight-worker"]
+        ports:
+        - name: http
+          containerPort: {{ .Values.worker.service.targetPort }}
+          protocol: TCP
+        {{- if .Values.existingSecret }}
+        envFrom:
+        - secretRef:
+            name: {{ .Values.existingSecret }}
+        {{- end }}
+        env:
+        {{- /* POSTGRES_PASSWORD must be defined before DATABASE_URL for $(VAR) interpolation */}}
+        {{- if not .Values.postgresql.enabled }}
+        - name: POSTGRES_PASSWORD
+          valueFrom:
+            secretKeyRef:
+              name: {{ include "hindsight.secretName" . }}
+              key: postgres-password
+        {{- end }}
+        - name: HINDSIGHT_API_DATABASE_URL
+          value: {{ include "hindsight.databaseUrl" . | quote }}
+        {{- /* Worker ID uses pod name (StatefulSet provides stable names like worker-0, worker-1) */}}
+        - name: HINDSIGHT_API_WORKER_ID
+          valueFrom:
+            fieldRef:
+              fieldPath: metadata.name
+        {{- /* Inherit LLM config from api.env */}}
+        {{- range $key, $value := .Values.api.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        {{- /* Worker-specific env vars */}}
+        {{- range $key, $value := .Values.worker.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        {{- /* Only use secrets when not using existingSecret */}}
+        {{- if not .Values.existingSecret }}
+        {{- /* Inherit secrets from api.secrets */}}
+        {{- range $key, $value := .Values.api.secrets }}
+        - name: {{ $key }}
+          valueFrom:
+            secretKeyRef:
+              name: {{ include "hindsight.secretName" $ }}
+              key: {{ $key }}
+        {{- end }}
+        {{- /* Worker-specific secrets (can override api.secrets) */}}
+        {{- range $key, $value := .Values.worker.secrets }}
+        - name: {{ $key }}
+          valueFrom:
+            secretKeyRef:
+              name: {{ include "hindsight.secretName" $ }}
+              key: {{ $key }}
+        {{- end }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.worker.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.worker.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.worker.resources | nindent 10 }}
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+{{- end }}
@@ -67,6 +67,63 @@ api:
    # HINDSIGHT_API_LLM_API_KEY: "your-api-key"
    # HINDSIGHT_API_LLM_BASE_URL: "https://api.groq.com/openai/v1"

+# Worker settings (distributed task processing)
+# When enabled, dedicated worker pods process tasks and the API's internal worker is disabled
+worker:
+  enabled: false
+  replicaCount: 2
+  image:
+    repository: ghcr.io/vectorize-io/hindsight-api
+    pullPolicy: IfNotPresent
+    # tag defaults to .Values.version if not specified
+
+  service:
+    # Service for metrics scraping (headless for StatefulSet)
+    port: 8889
+    targetPort: 8889
+
+  # Resource limits and requests
+  resources:
+    limits:
+      cpu: 2000m
+      memory: 4Gi
+    requests:
+      cpu: 500m
+      memory: 1Gi
+
+  # Liveness and readiness probes
+  livenessProbe:
+    httpGet:
+      path: /health
+      port: 8889
+    initialDelaySeconds: 30
+    periodSeconds: 10
+    timeoutSeconds: 5
+    failureThreshold: 3
+
+  readinessProbe:
+    httpGet:
+      path: /health
+      port: 8889
+    initialDelaySeconds: 10
+    periodSeconds: 5
+    timeoutSeconds: 3
+    failureThreshold: 3
+
+  # Worker-specific environment variables
+  env:
+    # Poll interval in milliseconds (how often to check for new tasks)
+    HINDSIGHT_API_WORKER_POLL_INTERVAL_MS: "500"
+    # Number of tasks to claim per poll cycle
+    HINDSIGHT_API_WORKER_BATCH_SIZE: "10"
+    # Max retries before marking a task as failed
+    HINDSIGHT_API_WORKER_MAX_RETRIES: "3"
+    # HTTP port for metrics/health (matches service.targetPort)
+    HINDSIGHT_API_WORKER_HTTP_PORT: "8889"
+
+  # Secret environment variables (inherited from api.secrets if not specified)
+  secrets: {}
+
 # Image settings for control plane
 controlPlane:
  enabled: true
@@ -244,6 +244,65 @@ def run_db_migration(
    typer.echo("Database migrations completed successfully")


+async def _decommission_worker(db_url: str, worker_id: str, schema: str = "public") -> int:
+    """Release all tasks owned by a worker, setting them back to pending status."""
+    is_pg0, instance_name, _ = parse_pg0_url(db_url)
+    if is_pg0:
+        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
+    resolved_url = await resolve_database_url(db_url)
+
+    conn = await asyncpg.connect(resolved_url)
+    try:
+        table = _fq_table("async_operations", schema)
+        result = await conn.fetch(
+            f"""
+            UPDATE {table}
+            SET status = 'pending', worker_id = NULL, claimed_at = NULL, updated_at = now()
+            WHERE worker_id = $1 AND status = 'processing'
+            RETURNING operation_id
+            """,
+            worker_id,
+        )
+        return len(result)
+    finally:
+        await conn.close()
+
+
+@app.command(name="decommission-worker")
+def decommission_worker(
+    worker_id: str = typer.Argument(..., help="Worker ID to decommission"),
+    schema: str = typer.Option("public", "--schema", "-s", help="Database schema"),
+    yes: bool = typer.Option(False, "--yes", "-y", help="Skip confirmation prompt"),
+):
+    """Release all tasks owned by a worker (sets status back to pending).
+
+    Use this command when a worker has crashed or been removed without graceful shutdown.
+    All tasks that were being processed by the worker will be released back to the queue
+    so other workers can pick them up.
+    """
+    config = HindsightConfig.from_env()
+
+    if not config.database_url:
+        typer.echo("Error: Database URL not configured.", err=True)
+        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
+        raise typer.Exit(1)
+
+    if not yes:
+        typer.confirm(
+            f"This will release all tasks owned by worker '{worker_id}' back to pending. Continue?",
+            abort=True,
+        )
+
+    typer.echo(f"Decommissioning worker '{worker_id}' (schema: {schema})...")
+
+    count = asyncio.run(_decommission_worker(config.database_url, worker_id, schema))
+
+    if count > 0:
+        typer.echo(f"Released {count} task(s) from worker '{worker_id}'")
+    else:
+        typer.echo(f"No tasks found for worker '{worker_id}'")
+
+
 def main():
    app()

@@ -0,0 +1,48 @@
+"""add_tags_column
+
+Revision ID: g2a3b4c5d6e7
+Revises: f1a2b3c4d5e6
+Create Date: 2025-01-13
+
+Add tags column to memory_units and documents tables for visibility scoping.
+Tags enable filtering memories by scope (e.g., user IDs, session IDs) during recall/reflect.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "g2a3b4c5d6e7"
+down_revision: str | Sequence[str] | None = "f1a2b3c4d5e6"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (e.g., 'tenant_x.' or '' for public)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add tags column to memory_units and documents tables."""
+    schema = _get_schema_prefix()
+
+    # Add tags column to memory_units table
+    op.execute(f"ALTER TABLE {schema}memory_units ADD COLUMN IF NOT EXISTS tags VARCHAR[] NOT NULL DEFAULT '{{}}'")
+
+    # Create GIN index for efficient array containment queries (tags && ARRAY['x'])
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_memory_units_tags ON {schema}memory_units USING GIN (tags)")
+
+    # Add tags column to documents table for document-level tags
+    op.execute(f"ALTER TABLE {schema}documents ADD COLUMN IF NOT EXISTS tags VARCHAR[] NOT NULL DEFAULT '{{}}'")
+
+
+def downgrade() -> None:
+    """Remove tags columns and index."""
+    schema = _get_schema_prefix()
+
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_tags")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS tags")
+    op.execute(f"ALTER TABLE {schema}documents DROP COLUMN IF EXISTS tags")
@@ -0,0 +1,112 @@
+"""mental_models_v4
+
+Revision ID: h3c4d5e6f7g8
+Revises: g2a3b4c5d6e7
+Create Date: 2026-01-08 00:00:00.000000
+
+This migration implements the v4 mental models system:
+1. Deletes existing observation memory_units (observations now in mental models)
+2. Adds mission column to banks (replacing background)
+3. Creates mental_models table with final schema
+
+Mental models can reference entities when an entity is "promoted" to a mental model.
+Summary content is stored as JSONB observations with per-observation fact attribution.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "h3c4d5e6f7g8"
+down_revision: str | Sequence[str] | None = "g2a3b4c5d6e7"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Apply mental models v4 changes."""
+    schema = _get_schema_prefix()
+
+    # Step 1: Delete observation memory_units (cascades to unit_entities links)
+    # Observations are now handled through mental models, not memory_units
+    op.execute(f"DELETE FROM {schema}memory_units WHERE fact_type = 'observation'")
+
+    # Step 2: Drop observation-specific index (if it exists)
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_observation_date")
+
+    # Step 3: Add mission column to banks (replacing background)
+    op.execute(f"ALTER TABLE {schema}banks ADD COLUMN IF NOT EXISTS mission TEXT")
+
+    # Migrate: copy background to mission if background column exists
+    # Use DO block to check column existence first (idempotent for re-runs)
+    schema_name = context.config.get_main_option("target_schema") or "public"
+    op.execute(f"""
+        DO $$
+        BEGIN
+            IF EXISTS (
+                SELECT 1 FROM information_schema.columns
+                WHERE table_schema = '{schema_name}' AND table_name = 'banks' AND column_name = 'background'
+            ) THEN
+                UPDATE {schema}banks
+                SET mission = background
+                WHERE mission IS NULL;
+            END IF;
+        END $$;
+    """)
+
+    # Remove background column (replaced by mission)
+    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS background")
+
+    # Step 4: Create mental_models table with final v4 schema (if not exists)
+    op.execute(f"""
+        CREATE TABLE IF NOT EXISTS {schema}mental_models (
+            id VARCHAR(64) NOT NULL,
+            bank_id VARCHAR(64) NOT NULL,
+            subtype VARCHAR(32) NOT NULL,
+            name VARCHAR(256) NOT NULL,
+            description TEXT NOT NULL,
+            entity_id UUID,
+            observations JSONB DEFAULT '{{"observations": []}}'::jsonb,
+            links VARCHAR[],
+            tags VARCHAR[] DEFAULT '{{}}',
+            last_updated TIMESTAMP WITH TIME ZONE,
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            PRIMARY KEY (id, bank_id),
+            FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE,
+            FOREIGN KEY (entity_id) REFERENCES {schema}entities(id) ON DELETE SET NULL,
+            CONSTRAINT ck_mental_models_subtype CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned'))
+        )
+    """)
+
+    # Step 5: Create indexes for efficient queries (if not exist)
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_bank_id ON {schema}mental_models(bank_id)")
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_subtype ON {schema}mental_models(bank_id, subtype)")
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_entity_id ON {schema}mental_models(entity_id)")
+    # GIN index for efficient tags array filtering
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_tags ON {schema}mental_models USING GIN(tags)")
+
+
+def downgrade() -> None:
+    """Revert mental models v4 changes."""
+    schema = _get_schema_prefix()
+
+    # Drop mental_models table (cascades to indexes)
+    op.execute(f"DROP TABLE IF EXISTS {schema}mental_models CASCADE")
+
+    # Add back background column to banks
+    op.execute(f"ALTER TABLE {schema}banks ADD COLUMN IF NOT EXISTS background TEXT")
+
+    # Migrate mission back to background
+    op.execute(f"UPDATE {schema}banks SET background = mission WHERE background IS NULL")
+
+    # Remove mission column
+    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS mission")
+
+    # Note: Cannot restore deleted observations - they are lost on downgrade
@@ -0,0 +1,41 @@
+"""delete_opinions
+
+Revision ID: i4d5e6f7g8h9
+Revises: h3c4d5e6f7g8
+Create Date: 2026-01-15 00:00:00.000000
+
+This migration removes opinion facts from memory_units.
+Opinions are no longer a separate fact type - they are now represented
+through mental model observations with confidence scores.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "i4d5e6f7g8h9"
+down_revision: str | Sequence[str] | None = "h3c4d5e6f7g8"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Delete opinion memory_units."""
+    schema = _get_schema_prefix()
+
+    # Delete opinion memory_units (cascades to unit_entities links)
+    # Opinions are now handled through mental model observations
+    op.execute(f"DELETE FROM {schema}memory_units WHERE fact_type = 'opinion'")
+
+
+def downgrade() -> None:
+    """Cannot restore deleted opinions."""
+    # Note: Cannot restore deleted opinions - they are lost on downgrade
+    pass
@@ -0,0 +1,95 @@
+"""mental_model_versions
+
+Revision ID: j5e6f7g8h9i0
+Revises: i4d5e6f7g8h9
+Create Date: 2026-01-16 00:00:00.000000
+
+This migration adds versioning support for mental models:
+1. Creates mental_model_versions table to store observation snapshots
+2. Adds version column to mental_models for tracking current version
+
+This enables changelog/diff functionality for mental model observations.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "j5e6f7g8h9i0"
+down_revision: str | Sequence[str] | None = "i4d5e6f7g8h9"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Create mental_model_versions table and add version tracking."""
+    schema = _get_schema_prefix()
+
+    # Create mental_model_versions table for storing observation snapshots
+    op.execute(f"""
+        CREATE TABLE {schema}mental_model_versions (
+            id SERIAL PRIMARY KEY,
+            mental_model_id VARCHAR(64) NOT NULL,
+            bank_id VARCHAR(64) NOT NULL,
+            version INT NOT NULL,
+            observations JSONB NOT NULL DEFAULT '{{"observations": []}}'::jsonb,
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            FOREIGN KEY (mental_model_id, bank_id)
+                REFERENCES {schema}mental_models(id, bank_id) ON DELETE CASCADE,
+            UNIQUE (mental_model_id, bank_id, version)
+        )
+    """)
+
+    # Index for efficient version queries (get latest, list versions)
+    op.execute(f"""
+        CREATE INDEX idx_mental_model_versions_lookup
+        ON {schema}mental_model_versions(mental_model_id, bank_id, version DESC)
+    """)
+
+    # Add version column to mental_models to track current version
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD COLUMN IF NOT EXISTS version INT NOT NULL DEFAULT 0
+    """)
+
+    # Migrate existing mental models: create version 1 for any that have observations
+    op.execute(f"""
+        INSERT INTO {schema}mental_model_versions (mental_model_id, bank_id, version, observations, created_at)
+        SELECT id, bank_id, 1, observations, COALESCE(last_updated, created_at)
+        FROM {schema}mental_models
+        WHERE observations IS NOT NULL
+          AND observations != '{{"observations": []}}'::jsonb
+          AND (observations->'observations') IS NOT NULL
+          AND jsonb_array_length(observations->'observations') > 0
+    """)
+
+    # Update version to 1 for migrated mental models
+    op.execute(f"""
+        UPDATE {schema}mental_models
+        SET version = 1
+        WHERE observations IS NOT NULL
+          AND observations != '{{"observations": []}}'::jsonb
+          AND (observations->'observations') IS NOT NULL
+          AND jsonb_array_length(observations->'observations') > 0
+    """)
+
+
+def downgrade() -> None:
+    """Remove mental_model_versions table and version column."""
+    schema = _get_schema_prefix()
+
+    # Drop index
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_mental_model_versions_lookup")
+
+    # Drop versions table
+    op.execute(f"DROP TABLE IF EXISTS {schema}mental_model_versions")
+
+    # Remove version column from mental_models
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS version")
@@ -0,0 +1,58 @@
+"""add_directive_subtype
+
+Revision ID: k6f7g8h9i0j1
+Revises: j5e6f7g8h9i0
+Create Date: 2026-01-16 00:00:00.000000
+
+This migration adds 'directive' to the mental_models subtype constraint.
+Directives are hard rules with user-provided observations that the reflect agent must follow.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "k6f7g8h9i0j1"
+down_revision: str | Sequence[str] | None = "j5e6f7g8h9i0"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add 'directive' to mental_models subtype constraint."""
+    schema = _get_schema_prefix()
+
+    # Drop existing constraint
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
+
+    # Create new constraint with 'directive' added
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT ck_mental_models_subtype
+        CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned', 'directive'))
+    """)
+
+
+def downgrade() -> None:
+    """Remove 'directive' from mental_models subtype constraint."""
+    schema = _get_schema_prefix()
+
+    # First delete any directives (cannot downgrade if they exist)
+    op.execute(f"DELETE FROM {schema}mental_models WHERE subtype = 'directive'")
+
+    # Drop constraint with directive
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
+
+    # Recreate original constraint without directive
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT ck_mental_models_subtype
+        CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned'))
+    """)
@@ -0,0 +1,109 @@
+"""add_worker_columns
+
+Revision ID: l7g8h9i0j1k2
+Revises: k6f7g8h9i0j1
+Create Date: 2026-01-19 00:00:00.000000
+
+This migration adds columns to async_operations for distributed worker support:
+- worker_id: ID of the worker that claimed the task
+- claimed_at: When the task was claimed
+- retry_count: Number of retry attempts
+- task_payload: The serialized task dictionary
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import context, op
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = "l7g8h9i0j1k2"
+down_revision: str | Sequence[str] | None = "k6f7g8h9i0j1"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add worker columns to async_operations."""
+    schema = _get_schema_prefix()
+
+    # Add worker_id column (ID of worker that claimed the task)
+    op.add_column(
+        "async_operations",
+        sa.Column("worker_id", sa.Text(), nullable=True),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add claimed_at column (when task was claimed by worker)
+    op.add_column(
+        "async_operations",
+        sa.Column("claimed_at", postgresql.TIMESTAMP(timezone=True), nullable=True),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add retry_count column (number of retry attempts)
+    op.add_column(
+        "async_operations",
+        sa.Column("retry_count", sa.Integer(), server_default="0", nullable=False),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add task_payload column (serialized task dictionary)
+    op.add_column(
+        "async_operations",
+        sa.Column(
+            "task_payload",
+            postgresql.JSONB(astext_type=sa.Text()),
+            nullable=True,
+        ),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add index for efficient worker polling (pending tasks ordered by creation time)
+    op.execute(
+        f"CREATE INDEX idx_async_operations_pending_claim ON {schema}async_operations (status, created_at) "
+        f"WHERE status = 'pending' AND task_payload IS NOT NULL"
+    )
+
+    # Add index for finding tasks by worker_id (for decommissioning)
+    op.execute(
+        f"CREATE INDEX idx_async_operations_worker_id ON {schema}async_operations (worker_id) WHERE worker_id IS NOT NULL"
+    )
+
+
+def downgrade() -> None:
+    """Remove worker columns from async_operations."""
+    schema = _get_schema_prefix()
+
+    # Drop indexes
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_pending_claim")
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_worker_id")
+
+    # Drop columns
+    op.drop_column(
+        "async_operations",
+        "task_payload",
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+    op.drop_column(
+        "async_operations",
+        "retry_count",
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+    op.drop_column(
+        "async_operations",
+        "claimed_at",
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+    op.drop_column(
+        "async_operations",
+        "worker_id",
+        schema=context.config.get_main_option("target_schema") or None,
+    )
@@ -196,7 +196,7 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
        Each bank is an isolated memory store (like a separate "brain").

        Returns:
-            JSON list of banks with their IDs, names, dispositions, and backgrounds.
+            JSON list of banks with their IDs, names, dispositions, and missions.
        """
        try:
            banks = await memory.list_banks(request_context=RequestContext())
@@ -206,7 +206,7 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
            return f'{{"error": "{e}", "banks": []}}'

    @mcp.tool()
-    async def create_bank(bank_id: str, name: str | None = None, background: str | None = None) -> str:
+    async def create_bank(bank_id: str, name: str | None = None, mission: str | None = None) -> str:
        """
        Create a new memory bank or get an existing one.

@@ -216,18 +216,18 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
        Args:
            bank_id: Unique identifier for the bank (e.g., 'user-123', 'agent-alpha')
            name: Optional human-friendly name for the bank
-            background: Optional background context about the bank's owner/purpose
+            mission: Optional mission describing who the agent is and what they're trying to accomplish
        """
        try:
            # get_bank_profile auto-creates bank if it doesn't exist
            profile = await memory.get_bank_profile(bank_id, request_context=RequestContext())

-            # Update name/background if provided
-            if name is not None or background is not None:
+            # Update name/mission if provided
+            if name is not None or mission is not None:
                await memory.update_bank(
                    bank_id,
                    name=name,
-                    background=background,
+                    mission=mission,
                    request_context=RequestContext(),
                )
                # Fetch updated profile
@@ -4,9 +4,12 @@ Centralized configuration for Hindsight API.
 All environment variables and their defaults are defined here.
 """

+import json
 import logging
 import os
+import sys
 from dataclasses import dataclass
+from datetime import datetime, timezone

 from dotenv import find_dotenv, load_dotenv

@@ -41,10 +44,19 @@ ENV_EMBEDDINGS_LOCAL_MODEL = "HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL"
 ENV_EMBEDDINGS_TEI_URL = "HINDSIGHT_API_EMBEDDINGS_TEI_URL"
 ENV_EMBEDDINGS_OPENAI_API_KEY = "HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY"
 ENV_EMBEDDINGS_OPENAI_MODEL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_MODEL"
+ENV_EMBEDDINGS_OPENAI_BASE_URL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_BASE_URL"

 ENV_COHERE_API_KEY = "HINDSIGHT_API_COHERE_API_KEY"
 ENV_EMBEDDINGS_COHERE_MODEL = "HINDSIGHT_API_EMBEDDINGS_COHERE_MODEL"
+ENV_EMBEDDINGS_COHERE_BASE_URL = "HINDSIGHT_API_EMBEDDINGS_COHERE_BASE_URL"
 ENV_RERANKER_COHERE_MODEL = "HINDSIGHT_API_RERANKER_COHERE_MODEL"
+ENV_RERANKER_COHERE_BASE_URL = "HINDSIGHT_API_RERANKER_COHERE_BASE_URL"
+
+# LiteLLM gateway configuration (for embeddings and reranker via LiteLLM proxy)
+ENV_LITELLM_API_BASE = "HINDSIGHT_API_LITELLM_API_BASE"
+ENV_LITELLM_API_KEY = "HINDSIGHT_API_LITELLM_API_KEY"
+ENV_EMBEDDINGS_LITELLM_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_MODEL"
+ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"

 ENV_RERANKER_PROVIDER = "HINDSIGHT_API_RERANKER_PROVIDER"
 ENV_RERANKER_LOCAL_MODEL = "HINDSIGHT_API_RERANKER_LOCAL_MODEL"
@@ -59,6 +71,7 @@ ENV_RERANKER_FLASHRANK_CACHE_DIR = "HINDSIGHT_API_RERANKER_FLASHRANK_CACHE_DIR"
 ENV_HOST = "HINDSIGHT_API_HOST"
 ENV_PORT = "HINDSIGHT_API_PORT"
 ENV_LOG_LEVEL = "HINDSIGHT_API_LOG_LEVEL"
+ENV_LOG_FORMAT = "HINDSIGHT_API_LOG_FORMAT"
 ENV_WORKERS = "HINDSIGHT_API_WORKERS"
 ENV_MCP_ENABLED = "HINDSIGHT_API_MCP_ENABLED"
 ENV_GRAPH_RETRIEVER = "HINDSIGHT_API_GRAPH_RETRIEVER"
@@ -67,6 +80,7 @@ ENV_RECALL_MAX_CONCURRENT = "HINDSIGHT_API_RECALL_MAX_CONCURRENT"
 ENV_RECALL_CONNECTION_BUDGET = "HINDSIGHT_API_RECALL_CONNECTION_BUDGET"
 ENV_MCP_LOCAL_BANK_ID = "HINDSIGHT_API_MCP_LOCAL_BANK_ID"
 ENV_MCP_INSTRUCTIONS = "HINDSIGHT_API_MCP_INSTRUCTIONS"
+ENV_MENTAL_MODEL_REFRESH_CONCURRENCY = "HINDSIGHT_API_MENTAL_MODEL_REFRESH_CONCURRENCY"

 # Observation thresholds
 ENV_OBSERVATION_MIN_FACTS = "HINDSIGHT_API_OBSERVATION_MIN_FACTS"
@@ -92,10 +106,16 @@ ENV_DB_POOL_MAX_SIZE = "HINDSIGHT_API_DB_POOL_MAX_SIZE"
 ENV_DB_COMMAND_TIMEOUT = "HINDSIGHT_API_DB_COMMAND_TIMEOUT"
 ENV_DB_ACQUIRE_TIMEOUT = "HINDSIGHT_API_DB_ACQUIRE_TIMEOUT"

-# Background task processing
-ENV_TASK_BACKEND = "HINDSIGHT_API_TASK_BACKEND"
-ENV_TASK_BACKEND_MEMORY_BATCH_SIZE = "HINDSIGHT_API_TASK_BACKEND_MEMORY_BATCH_SIZE"
-ENV_TASK_BACKEND_MEMORY_BATCH_INTERVAL = "HINDSIGHT_API_TASK_BACKEND_MEMORY_BATCH_INTERVAL"
+# Worker configuration (distributed task processing)
+ENV_WORKER_ENABLED = "HINDSIGHT_API_WORKER_ENABLED"
+ENV_WORKER_ID = "HINDSIGHT_API_WORKER_ID"
+ENV_WORKER_POLL_INTERVAL_MS = "HINDSIGHT_API_WORKER_POLL_INTERVAL_MS"
+ENV_WORKER_MAX_RETRIES = "HINDSIGHT_API_WORKER_MAX_RETRIES"
+ENV_WORKER_BATCH_SIZE = "HINDSIGHT_API_WORKER_BATCH_SIZE"
+ENV_WORKER_HTTP_PORT = "HINDSIGHT_API_WORKER_HTTP_PORT"
+
+# Reflect agent settings
+ENV_REFLECT_MAX_ITERATIONS = "HINDSIGHT_API_REFLECT_MAX_ITERATIONS"

 # Default values
 DEFAULT_DATABASE_URL = "pg0"
@@ -121,9 +141,15 @@ DEFAULT_RERANKER_FLASHRANK_CACHE_DIR = None  # Use default cache directory
 DEFAULT_EMBEDDINGS_COHERE_MODEL = "embed-english-v3.0"
 DEFAULT_RERANKER_COHERE_MODEL = "rerank-english-v3.0"

+# LiteLLM defaults
+DEFAULT_LITELLM_API_BASE = "http://localhost:4000"
+DEFAULT_EMBEDDINGS_LITELLM_MODEL = "text-embedding-3-small"
+DEFAULT_RERANKER_LITELLM_MODEL = "cohere/rerank-english-v3.0"
+
 DEFAULT_HOST = "0.0.0.0"
 DEFAULT_PORT = 8888
 DEFAULT_LOG_LEVEL = "info"
+DEFAULT_LOG_FORMAT = "text"  # Options: "text", "json"
 DEFAULT_WORKERS = 1
 DEFAULT_MCP_ENABLED = True
 DEFAULT_GRAPH_RETRIEVER = "link_expansion"  # Options: "link_expansion", "mpfp", "bfs"
@@ -131,6 +157,7 @@ DEFAULT_MPFP_TOP_K_NEIGHBORS = 20  # Fan-out limit per node in MPFP graph traver
 DEFAULT_RECALL_MAX_CONCURRENT = 32  # Max concurrent recall operations per worker
 DEFAULT_RECALL_CONNECTION_BUDGET = 4  # Max concurrent DB connections per recall operation
 DEFAULT_MCP_LOCAL_BANK_ID = "mcp"
+DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY = 8  # Max concurrent mental model refreshes

 # Observation thresholds
 DEFAULT_OBSERVATION_MIN_FACTS = 5  # Min facts required to generate entity observations
@@ -153,10 +180,16 @@ DEFAULT_DB_POOL_MAX_SIZE = 100
 DEFAULT_DB_COMMAND_TIMEOUT = 60  # seconds
 DEFAULT_DB_ACQUIRE_TIMEOUT = 30  # seconds

-# Background task processing
-DEFAULT_TASK_BACKEND = "memory"  # Options: "memory", "noop"
-DEFAULT_TASK_BACKEND_MEMORY_BATCH_SIZE = 10
-DEFAULT_TASK_BACKEND_MEMORY_BATCH_INTERVAL = 1.0  # seconds
+# Worker configuration (distributed task processing)
+DEFAULT_WORKER_ENABLED = True  # API runs worker by default (standalone mode)
+DEFAULT_WORKER_ID = None  # Will use hostname if not specified
+DEFAULT_WORKER_POLL_INTERVAL_MS = 500  # Poll database every 500ms
+DEFAULT_WORKER_MAX_RETRIES = 3  # Max retries before marking task failed
+DEFAULT_WORKER_BATCH_SIZE = 10  # Tasks to claim per poll cycle
+DEFAULT_WORKER_HTTP_PORT = 8889  # HTTP port for worker metrics/health
+
+# Reflect agent settings
+DEFAULT_REFLECT_MAX_ITERATIONS = 10  # Max tool call iterations before forcing response

 # Default MCP tool descriptions (can be customized via env vars)
 DEFAULT_MCP_RETAIN_DESCRIPTION = """Store important information to long-term memory.
@@ -182,6 +215,36 @@ Use this tool PROACTIVELY to:
 EMBEDDING_DIMENSION = DEFAULT_EMBEDDING_DIMENSION


+class JsonFormatter(logging.Formatter):
+    """JSON formatter for structured logging.
+
+    Outputs logs in JSON format with a 'severity' field that cloud logging
+    systems (GCP, AWS CloudWatch, etc.) can parse to correctly categorize log levels.
+    """
+
+    SEVERITY_MAP = {
+        logging.DEBUG: "DEBUG",
+        logging.INFO: "INFO",
+        logging.WARNING: "WARNING",
+        logging.ERROR: "ERROR",
+        logging.CRITICAL: "CRITICAL",
+    }
+
+    def format(self, record: logging.LogRecord) -> str:
+        log_entry = {
+            "severity": self.SEVERITY_MAP.get(record.levelno, "DEFAULT"),
+            "message": record.getMessage(),
+            "timestamp": datetime.now(timezone.utc).isoformat(),
+            "logger": record.name,
+        }
+
+        # Add exception info if present
+        if record.exc_info:
+            log_entry["exception"] = self.formatException(record.exc_info)
+
+        return json.dumps(log_entry)
+
+
 def _validate_extraction_mode(mode: str) -> str:
    """Validate and normalize extraction mode."""
    mode_lower = mode.lower()
@@ -224,6 +287,8 @@ class HindsightConfig:
    embeddings_provider: str
    embeddings_local_model: str
    embeddings_tei_url: str | None
+    embeddings_openai_base_url: str | None
+    embeddings_cohere_base_url: str | None

    # Reranker
    reranker_provider: str
@@ -232,11 +297,13 @@ class HindsightConfig:
    reranker_tei_batch_size: int
    reranker_tei_max_concurrent: int
    reranker_max_candidates: int
+    reranker_cohere_base_url: str | None

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

    # Recall
@@ -244,6 +311,7 @@ class HindsightConfig:
    mpfp_top_k_neighbors: int
    recall_max_concurrent: int
    recall_connection_budget: int
+    mental_model_refresh_concurrency: int

    # Observation thresholds
    observation_min_facts: int
@@ -269,10 +337,16 @@ class HindsightConfig:
    db_command_timeout: int
    db_acquire_timeout: int

-    # Background task processing
-    task_backend: str
-    task_backend_memory_batch_size: int
-    task_backend_memory_batch_interval: float
+    # Worker configuration (distributed task processing)
+    worker_enabled: bool
+    worker_id: str | None
+    worker_poll_interval_ms: int
+    worker_max_retries: int
+    worker_batch_size: int
+    worker_http_port: int
+
+    # Reflect agent settings
+    reflect_max_iterations: int

    @classmethod
    def from_env(cls) -> "HindsightConfig":
@@ -300,6 +374,8 @@ class HindsightConfig:
            embeddings_provider=os.getenv(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER),
            embeddings_local_model=os.getenv(ENV_EMBEDDINGS_LOCAL_MODEL, DEFAULT_EMBEDDINGS_LOCAL_MODEL),
            embeddings_tei_url=os.getenv(ENV_EMBEDDINGS_TEI_URL),
+            embeddings_openai_base_url=os.getenv(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None,
+            embeddings_cohere_base_url=os.getenv(ENV_EMBEDDINGS_COHERE_BASE_URL) or None,
            # Reranker
            reranker_provider=os.getenv(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER),
            reranker_local_model=os.getenv(ENV_RERANKER_LOCAL_MODEL, DEFAULT_RERANKER_LOCAL_MODEL),
@@ -309,10 +385,12 @@ class HindsightConfig:
                os.getenv(ENV_RERANKER_TEI_MAX_CONCURRENT, str(DEFAULT_RERANKER_TEI_MAX_CONCURRENT))
            ),
            reranker_max_candidates=int(os.getenv(ENV_RERANKER_MAX_CANDIDATES, str(DEFAULT_RERANKER_MAX_CANDIDATES))),
+            reranker_cohere_base_url=os.getenv(ENV_RERANKER_COHERE_BASE_URL) or None,
            # Server
            host=os.getenv(ENV_HOST, DEFAULT_HOST),
            port=int(os.getenv(ENV_PORT, DEFAULT_PORT)),
            log_level=os.getenv(ENV_LOG_LEVEL, DEFAULT_LOG_LEVEL),
+            log_format=os.getenv(ENV_LOG_FORMAT, DEFAULT_LOG_FORMAT).lower(),
            mcp_enabled=os.getenv(ENV_MCP_ENABLED, str(DEFAULT_MCP_ENABLED)).lower() == "true",
            # Recall
            graph_retriever=os.getenv(ENV_GRAPH_RETRIEVER, DEFAULT_GRAPH_RETRIEVER),
@@ -321,6 +399,9 @@ class HindsightConfig:
            recall_connection_budget=int(
                os.getenv(ENV_RECALL_CONNECTION_BUDGET, str(DEFAULT_RECALL_CONNECTION_BUDGET))
            ),
+            mental_model_refresh_concurrency=int(
+                os.getenv(ENV_MENTAL_MODEL_REFRESH_CONCURRENCY, str(DEFAULT_MENTAL_MODEL_REFRESH_CONCURRENCY))
+            ),
            # Optimization flags
            skip_llm_verification=os.getenv(ENV_SKIP_LLM_VERIFICATION, "false").lower() == "true",
            lazy_reranker=os.getenv(ENV_LAZY_RERANKER, "false").lower() == "true",
@@ -352,14 +433,15 @@ class HindsightConfig:
            db_pool_max_size=int(os.getenv(ENV_DB_POOL_MAX_SIZE, str(DEFAULT_DB_POOL_MAX_SIZE))),
            db_command_timeout=int(os.getenv(ENV_DB_COMMAND_TIMEOUT, str(DEFAULT_DB_COMMAND_TIMEOUT))),
            db_acquire_timeout=int(os.getenv(ENV_DB_ACQUIRE_TIMEOUT, str(DEFAULT_DB_ACQUIRE_TIMEOUT))),
-            # Background task processing
-            task_backend=os.getenv(ENV_TASK_BACKEND, DEFAULT_TASK_BACKEND),
-            task_backend_memory_batch_size=int(
-                os.getenv(ENV_TASK_BACKEND_MEMORY_BATCH_SIZE, str(DEFAULT_TASK_BACKEND_MEMORY_BATCH_SIZE))
-            ),
-            task_backend_memory_batch_interval=float(
-                os.getenv(ENV_TASK_BACKEND_MEMORY_BATCH_INTERVAL, str(DEFAULT_TASK_BACKEND_MEMORY_BATCH_INTERVAL))
-            ),
+            # Worker configuration
+            worker_enabled=os.getenv(ENV_WORKER_ENABLED, str(DEFAULT_WORKER_ENABLED)).lower() == "true",
+            worker_id=os.getenv(ENV_WORKER_ID) or DEFAULT_WORKER_ID,
+            worker_poll_interval_ms=int(os.getenv(ENV_WORKER_POLL_INTERVAL_MS, str(DEFAULT_WORKER_POLL_INTERVAL_MS))),
+            worker_max_retries=int(os.getenv(ENV_WORKER_MAX_RETRIES, str(DEFAULT_WORKER_MAX_RETRIES))),
+            worker_batch_size=int(os.getenv(ENV_WORKER_BATCH_SIZE, str(DEFAULT_WORKER_BATCH_SIZE))),
+            worker_http_port=int(os.getenv(ENV_WORKER_HTTP_PORT, str(DEFAULT_WORKER_HTTP_PORT))),
+            # Reflect agent settings
+            reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
        )

    def get_llm_base_url(self) -> str:
@@ -390,12 +472,28 @@ class HindsightConfig:
        return log_level_map.get(self.log_level.lower(), logging.INFO)

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

    def log_config(self) -> None:
        """Log the current configuration (without sensitive values)."""
@@ -15,18 +15,24 @@ from concurrent.futures import ThreadPoolExecutor
 import httpx

 from ..config import (
+    DEFAULT_LITELLM_API_BASE,
    DEFAULT_RERANKER_COHERE_MODEL,
    DEFAULT_RERANKER_FLASHRANK_CACHE_DIR,
    DEFAULT_RERANKER_FLASHRANK_MODEL,
+    DEFAULT_RERANKER_LITELLM_MODEL,
    DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT,
    DEFAULT_RERANKER_LOCAL_MODEL,
    DEFAULT_RERANKER_PROVIDER,
    DEFAULT_RERANKER_TEI_BATCH_SIZE,
    DEFAULT_RERANKER_TEI_MAX_CONCURRENT,
    ENV_COHERE_API_KEY,
+    ENV_LITELLM_API_BASE,
+    ENV_LITELLM_API_KEY,
+    ENV_RERANKER_COHERE_BASE_URL,
    ENV_RERANKER_COHERE_MODEL,
    ENV_RERANKER_FLASHRANK_CACHE_DIR,
    ENV_RERANKER_FLASHRANK_MODEL,
+    ENV_RERANKER_LITELLM_MODEL,
    ENV_RERANKER_LOCAL_MAX_CONCURRENT,
    ENV_RERANKER_LOCAL_MODEL,
    ENV_RERANKER_PROVIDER,
@@ -392,6 +398,7 @@ class CohereCrossEncoder(CrossEncoderModel):
        self,
        api_key: str,
        model: str = DEFAULT_RERANKER_COHERE_MODEL,
+        base_url: str | None = None,
        timeout: float = 60.0,
    ):
        """
@@ -400,10 +407,12 @@ class CohereCrossEncoder(CrossEncoderModel):
        Args:
            api_key: Cohere API key
            model: Cohere rerank model name (default: rerank-english-v3.0)
+            base_url: Custom base URL for Cohere-compatible API (e.g., Azure-hosted endpoint)
            timeout: Request timeout in seconds (default: 60.0)
        """
        self.api_key = api_key
        self.model = model
+        self.base_url = base_url
        self.timeout = timeout
        self._client = None

@@ -421,8 +430,14 @@ class CohereCrossEncoder(CrossEncoderModel):
        except ImportError:
            raise ImportError("cohere is required for CohereCrossEncoder. Install it with: pip install cohere")

-        logger.info(f"Reranker: initializing Cohere provider with model {self.model}")
-        self._client = cohere.Client(api_key=self.api_key, timeout=self.timeout)
+        base_url_msg = f" at {self.base_url}" if self.base_url else ""
+        logger.info(f"Reranker: initializing Cohere provider with model {self.model}{base_url_msg}")
+
+        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
+        client_kwargs = {"api_key": self.api_key, "timeout": self.timeout}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        self._client = cohere.Client(**client_kwargs)
        logger.info("Reranker: Cohere provider initialized")

    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
@@ -641,6 +656,116 @@ class FlashRankCrossEncoder(CrossEncoderModel):
        return await loop.run_in_executor(FlashRankCrossEncoder._executor, self._predict_sync, pairs)


+class LiteLLMCrossEncoder(CrossEncoderModel):
+    """
+    LiteLLM cross-encoder implementation using LiteLLM proxy's /rerank endpoint.
+
+    LiteLLM provides a unified interface for multiple reranking providers via
+    the Cohere-compatible /rerank endpoint.
+    See: https://docs.litellm.ai/docs/rerank
+
+    Supported providers via LiteLLM:
+    - Cohere (rerank-english-v3.0, etc.) - prefix with cohere/
+    - Together AI - prefix with together_ai/
+    - Azure AI - prefix with azure_ai/
+    - Jina AI - prefix with jina_ai/
+    - AWS Bedrock - prefix with bedrock/
+    - Voyage AI - prefix with voyage/
+    """
+
+    def __init__(
+        self,
+        api_base: str = DEFAULT_LITELLM_API_BASE,
+        api_key: str | None = None,
+        model: str = DEFAULT_RERANKER_LITELLM_MODEL,
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize LiteLLM cross-encoder client.
+
+        Args:
+            api_base: Base URL of the LiteLLM proxy (default: http://localhost:4000)
+            api_key: API key for the LiteLLM proxy (optional, depends on proxy config)
+            model: Reranking model name (default: cohere/rerank-english-v3.0)
+                   Use provider prefix (e.g., cohere/, together_ai/, voyage/)
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        self.api_base = api_base.rstrip("/")
+        self.api_key = api_key
+        self.model = model
+        self.timeout = timeout
+        self._async_client: httpx.AsyncClient | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "litellm"
+
+    async def initialize(self) -> None:
+        """Initialize the async HTTP client."""
+        if self._async_client is not None:
+            return
+
+        logger.info(f"Reranker: initializing LiteLLM provider at {self.api_base} with model {self.model}")
+
+        headers = {"Content-Type": "application/json"}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+
+        self._async_client = httpx.AsyncClient(timeout=self.timeout, headers=headers)
+        logger.info("Reranker: LiteLLM provider initialized")
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Score query-document pairs using the LiteLLM proxy's /rerank endpoint.
+
+        Args:
+            pairs: List of (query, document) tuples to score
+
+        Returns:
+            List of relevance scores
+        """
+        if self._async_client is None:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        if not pairs:
+            return []
+
+        # Group pairs by query (LiteLLM rerank expects one query with multiple documents)
+        query_groups: dict[str, list[tuple[int, str]]] = {}
+        for idx, (query, text) in enumerate(pairs):
+            if query not in query_groups:
+                query_groups[query] = []
+            query_groups[query].append((idx, text))
+
+        all_scores = [0.0] * len(pairs)
+
+        for query, indexed_texts in query_groups.items():
+            texts = [text for _, text in indexed_texts]
+            indices = [idx for idx, _ in indexed_texts]
+
+            # LiteLLM /rerank follows Cohere API format
+            response = await self._async_client.post(
+                f"{self.api_base}/rerank",
+                json={
+                    "model": self.model,
+                    "query": query,
+                    "documents": texts,
+                    "top_n": len(texts),  # Return all scores
+                },
+            )
+            response.raise_for_status()
+            result = response.json()
+
+            # Map scores back to original positions
+            # Response format: {"results": [{"index": 0, "relevance_score": 0.9}, ...]}
+            for item in result.get("results", []):
+                original_idx = item["index"]
+                score = item.get("relevance_score", item.get("score", 0.0))
+                all_scores[indices[original_idx]] = score
+
+        return all_scores
+
+
 def create_cross_encoder_from_env() -> CrossEncoderModel:
    """
    Create a CrossEncoderModel instance based on environment variables.
@@ -671,14 +796,20 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
        if not api_key:
            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'cohere'")
        model = os.environ.get(ENV_RERANKER_COHERE_MODEL, DEFAULT_RERANKER_COHERE_MODEL)
-        return CohereCrossEncoder(api_key=api_key, model=model)
+        base_url = os.environ.get(ENV_RERANKER_COHERE_BASE_URL) or None
+        return CohereCrossEncoder(api_key=api_key, model=model, base_url=base_url)
    elif provider == "flashrank":
        model = os.environ.get(ENV_RERANKER_FLASHRANK_MODEL, DEFAULT_RERANKER_FLASHRANK_MODEL)
        cache_dir = os.environ.get(ENV_RERANKER_FLASHRANK_CACHE_DIR, DEFAULT_RERANKER_FLASHRANK_CACHE_DIR)
        return FlashRankCrossEncoder(model_name=model, cache_dir=cache_dir)
+    elif provider == "litellm":
+        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
+        api_key = os.environ.get(ENV_LITELLM_API_KEY)
+        model = os.environ.get(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL)
+        return LiteLLMCrossEncoder(api_base=api_base, api_key=api_key, model=model)
    elif provider == "rrf":
        return RRFPassthroughCrossEncoder()
    else:
        raise ValueError(
-            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'flashrank', 'rrf'"
+            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'flashrank', 'litellm', 'rrf'"
        )
@@ -17,16 +17,23 @@ import httpx

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

@@ -322,6 +329,7 @@ class OpenAIEmbeddings(Embeddings):
        self,
        api_key: str,
        model: str = DEFAULT_EMBEDDINGS_OPENAI_MODEL,
+        base_url: str | None = None,
        batch_size: int = 100,
        max_retries: int = 3,
    ):
@@ -331,11 +339,13 @@ class OpenAIEmbeddings(Embeddings):
        Args:
            api_key: OpenAI API key
            model: OpenAI embedding model name (default: text-embedding-3-small)
+            base_url: Custom base URL for OpenAI-compatible API (e.g., Azure OpenAI endpoint)
            batch_size: Maximum batch size for embedding requests (default: 100)
            max_retries: Maximum number of retries for failed requests (default: 3)
        """
        self.api_key = api_key
        self.model = model
+        self.base_url = base_url
        self.batch_size = batch_size
        self.max_retries = max_retries
        self._client = None
@@ -361,8 +371,14 @@ class OpenAIEmbeddings(Embeddings):
        except ImportError:
            raise ImportError("openai is required for OpenAIEmbeddings. Install it with: pip install openai")

-        logger.info(f"Embeddings: initializing OpenAI provider with model {self.model}")
-        self._client = OpenAI(api_key=self.api_key, max_retries=self.max_retries)
+        base_url_msg = f" at {self.base_url}" if self.base_url else ""
+        logger.info(f"Embeddings: initializing OpenAI provider with model {self.model}{base_url_msg}")
+
+        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
+        client_kwargs = {"api_key": self.api_key, "max_retries": self.max_retries}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        self._client = OpenAI(**client_kwargs)

        # Try to get dimension from known models, otherwise do a test embedding
        if self.model in self.MODEL_DIMENSIONS:
@@ -435,6 +451,7 @@ class CohereEmbeddings(Embeddings):
        self,
        api_key: str,
        model: str = DEFAULT_EMBEDDINGS_COHERE_MODEL,
+        base_url: str | None = None,
        batch_size: int = 96,
        timeout: float = 60.0,
        input_type: str = "search_document",
@@ -445,6 +462,7 @@ class CohereEmbeddings(Embeddings):
        Args:
            api_key: Cohere API key
            model: Cohere embedding model name (default: embed-english-v3.0)
+            base_url: Custom base URL for Cohere-compatible API (e.g., Azure-hosted endpoint)
            batch_size: Maximum batch size for embedding requests (default: 96, Cohere's limit)
            timeout: Request timeout in seconds (default: 60.0)
            input_type: Input type for embeddings (default: search_document).
@@ -452,6 +470,7 @@ class CohereEmbeddings(Embeddings):
        """
        self.api_key = api_key
        self.model = model
+        self.base_url = base_url
        self.batch_size = batch_size
        self.timeout = timeout
        self.input_type = input_type
@@ -478,8 +497,14 @@ class CohereEmbeddings(Embeddings):
        except ImportError:
            raise ImportError("cohere is required for CohereEmbeddings. Install it with: pip install cohere")

-        logger.info(f"Embeddings: initializing Cohere provider with model {self.model}")
-        self._client = cohere.Client(api_key=self.api_key, timeout=self.timeout)
+        base_url_msg = f" at {self.base_url}" if self.base_url else ""
+        logger.info(f"Embeddings: initializing Cohere provider with model {self.model}{base_url_msg}")
+
+        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
+        client_kwargs = {"api_key": self.api_key, "timeout": self.timeout}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        self._client = cohere.Client(**client_kwargs)

        # Try to get dimension from known models, otherwise do a test embedding
        if self.model in self.MODEL_DIMENSIONS:
@@ -529,6 +554,123 @@ class CohereEmbeddings(Embeddings):
        return all_embeddings


+class LiteLLMEmbeddings(Embeddings):
+    """
+    LiteLLM embeddings implementation using LiteLLM proxy's /embeddings endpoint.
+
+    LiteLLM provides a unified interface for multiple embedding providers.
+    The proxy exposes an OpenAI-compatible /embeddings endpoint.
+    See: https://docs.litellm.ai/docs/embedding/supported_embedding
+
+    Supported providers via LiteLLM:
+    - OpenAI (text-embedding-3-small, text-embedding-ada-002, etc.)
+    - Cohere (embed-english-v3.0, etc.) - prefix with cohere/
+    - Vertex AI (textembedding-gecko, etc.) - prefix with vertex_ai/
+    - HuggingFace, Mistral, Voyage AI, etc.
+
+    The embedding dimension is auto-detected from the model at initialization.
+    """
+
+    def __init__(
+        self,
+        api_base: str = DEFAULT_LITELLM_API_BASE,
+        api_key: str | None = None,
+        model: str = DEFAULT_EMBEDDINGS_LITELLM_MODEL,
+        batch_size: int = 100,
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize LiteLLM embeddings client.
+
+        Args:
+            api_base: Base URL of the LiteLLM proxy (default: http://localhost:4000)
+            api_key: API key for the LiteLLM proxy (optional, depends on proxy config)
+            model: Embedding model name (default: text-embedding-3-small)
+                   Use provider prefix for non-OpenAI models (e.g., cohere/embed-english-v3.0)
+            batch_size: Maximum batch size for embedding requests (default: 100)
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        self.api_base = api_base.rstrip("/")
+        self.api_key = api_key
+        self.model = model
+        self.batch_size = batch_size
+        self.timeout = timeout
+        self._client: httpx.Client | None = None
+        self._dimension: int | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "litellm"
+
+    @property
+    def dimension(self) -> int:
+        if self._dimension is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+        return self._dimension
+
+    async def initialize(self) -> None:
+        """Initialize the HTTP client and detect embedding dimension."""
+        if self._client is not None:
+            return
+
+        logger.info(f"Embeddings: initializing LiteLLM provider at {self.api_base} with model {self.model}")
+
+        headers = {"Content-Type": "application/json"}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+
+        self._client = httpx.Client(timeout=self.timeout, headers=headers)
+
+        # Do a test embedding to detect dimension
+        try:
+            response = self._client.post(
+                f"{self.api_base}/embeddings",
+                json={"model": self.model, "input": ["test"]},
+            )
+            response.raise_for_status()
+            result = response.json()
+            if result.get("data") and len(result["data"]) > 0:
+                self._dimension = len(result["data"][0]["embedding"])
+            logger.info(f"Embeddings: LiteLLM provider initialized (model: {self.model}, dim: {self._dimension})")
+        except httpx.HTTPError as e:
+            raise RuntimeError(f"Failed to connect to LiteLLM proxy at {self.api_base}: {e}")
+
+    def encode(self, texts: list[str]) -> list[list[float]]:
+        """
+        Generate embeddings using the LiteLLM proxy.
+
+        Args:
+            texts: List of text strings to encode
+
+        Returns:
+            List of embedding vectors
+        """
+        if self._client is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+
+        if not texts:
+            return []
+
+        all_embeddings = []
+
+        # Process in batches
+        for i in range(0, len(texts), self.batch_size):
+            batch = texts[i : i + self.batch_size]
+
+            response = self._client.post(
+                f"{self.api_base}/embeddings",
+                json={"model": self.model, "input": batch},
+            )
+            response.raise_for_status()
+            result = response.json()
+
+            # Sort by index to ensure correct order
+            batch_embeddings = sorted(result["data"], key=lambda x: x["index"])
+            all_embeddings.extend([e["embedding"] for e in batch_embeddings])
+
+        return all_embeddings
+
+
 def create_embeddings_from_env() -> Embeddings:
    """
    Create an Embeddings instance based on environment variables.
@@ -558,12 +700,21 @@ def create_embeddings_from_env() -> Embeddings:
                f"when {ENV_EMBEDDINGS_PROVIDER} is 'openai'"
            )
        model = os.environ.get(ENV_EMBEDDINGS_OPENAI_MODEL, DEFAULT_EMBEDDINGS_OPENAI_MODEL)
-        return OpenAIEmbeddings(api_key=api_key, model=model)
+        base_url = os.environ.get(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None
+        return OpenAIEmbeddings(api_key=api_key, model=model, base_url=base_url)
    elif provider == "cohere":
        api_key = os.environ.get(ENV_COHERE_API_KEY)
        if not api_key:
            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'cohere'")
        model = os.environ.get(ENV_EMBEDDINGS_COHERE_MODEL, DEFAULT_EMBEDDINGS_COHERE_MODEL)
-        return CohereEmbeddings(api_key=api_key, model=model)
+        base_url = os.environ.get(ENV_EMBEDDINGS_COHERE_BASE_URL) or None
+        return CohereEmbeddings(api_key=api_key, model=model, base_url=base_url)
+    elif provider == "litellm":
+        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
+        api_key = os.environ.get(ENV_LITELLM_API_KEY)
+        model = os.environ.get(ENV_EMBEDDINGS_LITELLM_MODEL, DEFAULT_EMBEDDINGS_LITELLM_MODEL)
+        return LiteLLMEmbeddings(api_base=api_base, api_key=api_key, model=model)
    else:
-        raise ValueError(f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai', 'cohere'")
+        raise ValueError(
+            f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai', 'cohere', 'litellm'"
+        )
@@ -160,14 +160,14 @@ class MemoryEngineInterface(ABC):
        request_context: "RequestContext",
    ) -> dict[str, Any]:
        """
-        Get bank profile including disposition and background.
+        Get bank profile including disposition and mission.

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

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

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

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

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

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

@@ -518,7 +537,7 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        request_context: "RequestContext",
-    ) -> list[dict[str, Any]]:
+    ) -> dict[str, Any]:
        """
        List async operations for a bank.

@@ -527,7 +546,7 @@ class MemoryEngineInterface(ABC):
            request_context: Request context for authentication.

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

@@ -561,16 +580,16 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        name: str | None = None,
-        background: str | None = None,
+        mission: str | None = None,
        request_context: "RequestContext",
    ) -> dict[str, Any]:
        """
-        Update bank name and/or background.
+        Update bank name and/or mission.

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

        Returns:
@@ -209,10 +209,10 @@ class LLMProvider:
            OutputTooLongError: If output exceeds token limits.
            Exception: Re-raises API errors after retries exhausted.
        """
-        queue_start_time = time.time()
+        semaphore_start = time.time()
        async with _global_llm_semaphore:
+            semaphore_wait_time = time.time() - semaphore_start
            start_time = time.time()
-            semaphore_wait_time = start_time - queue_start_time

            # Handle Mock provider (for testing)
            if self.provider == "mock":
@@ -318,43 +318,44 @@ class LLMProvider:

            last_exception = None

+            # Prepare response format ONCE before the retry loop
+            # (to avoid appending schema to messages on every retry)
+            if response_format is not None:
+                schema = None
+                if hasattr(response_format, "model_json_schema"):
+                    schema = response_format.model_json_schema()
+
+                if strict_schema and schema is not None:
+                    # Use OpenAI's strict JSON schema enforcement
+                    # This guarantees all required fields are returned
+                    call_params["response_format"] = {
+                        "type": "json_schema",
+                        "json_schema": {
+                            "name": "response",
+                            "strict": True,
+                            "schema": schema,
+                        },
+                    }
+                else:
+                    # Soft enforcement: add schema to prompt and use json_object mode
+                    if schema is not None:
+                        schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+
+                        if call_params["messages"] and call_params["messages"][0].get("role") == "system":
+                            call_params["messages"][0]["content"] += schema_msg
+                        elif call_params["messages"]:
+                            call_params["messages"][0]["content"] = (
+                                schema_msg + "\n\n" + call_params["messages"][0]["content"]
+                            )
+                    if self.provider not in ("lmstudio", "ollama"):
+                        # LM Studio and Ollama don't support json_object response format reliably
+                        # We rely on the schema in the system message instead
+                        call_params["response_format"] = {"type": "json_object"}
+
            for attempt in range(max_retries + 1):
                try:
                    if response_format is not None:
-                        schema = None
-                        if hasattr(response_format, "model_json_schema"):
-                            schema = response_format.model_json_schema()
-
-                        if strict_schema and schema is not None:
-                            # Use OpenAI's strict JSON schema enforcement
-                            # This guarantees all required fields are returned
-                            call_params["response_format"] = {
-                                "type": "json_schema",
-                                "json_schema": {
-                                    "name": "response",
-                                    "strict": True,
-                                    "schema": schema,
-                                },
-                            }
-                        else:
-                            # Soft enforcement: add schema to prompt and use json_object mode
-                            if schema is not None:
-                                schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
-
-                                if call_params["messages"] and call_params["messages"][0].get("role") == "system":
-                                    call_params["messages"][0]["content"] += schema_msg
-                                elif call_params["messages"]:
-                                    call_params["messages"][0]["content"] = (
-                                        schema_msg + "\n\n" + call_params["messages"][0]["content"]
-                                    )
-                            if self.provider not in ("lmstudio", "ollama"):
-                                # LM Studio and Ollama don't support json_object response format reliably
-                                # We rely on the schema in the system message instead
-                                call_params["response_format"] = {"type": "json_object"}
-
-                        logger.debug(f"Sending request to {self.provider}/{self.model} (timeout={self.timeout})")
                        response = await self._client.chat.completions.create(**call_params)
-                        logger.debug(f"Received response from {self.provider}/{self.model}")

                        content = response.choices[0].message.content

@@ -467,13 +468,11 @@ class LLMProvider:

                except APIConnectionError as e:
                    last_exception = e
+                    status_code = getattr(e, "status_code", None) or getattr(
+                        getattr(e, "response", None), "status_code", None
+                    )
+                    logger.warning(f"APIConnectionError (HTTP {status_code}), attempt {attempt + 1}: {str(e)[:200]}")
                    if attempt < max_retries:
-                        status_code = getattr(e, "status_code", None) or getattr(
-                            getattr(e, "response", None), "status_code", None
-                        )
-                        logger.warning(
-                            f"Connection error, retrying... (attempt {attempt + 1}/{max_retries + 1}) - status_code={status_code}, message={e}"
-                        )
                        backoff = min(initial_backoff * (2**attempt), max_backoff)
                        await asyncio.sleep(backoff)
                        continue
@@ -487,6 +486,45 @@ class LLMProvider:
                        logger.error(f"Auth error (HTTP {e.status_code}), not retrying: {str(e)}")
                        raise

+                    # Handle tool_use_failed error - model outputted in tool call format
+                    # Convert to expected JSON format and continue
+                    if e.status_code == 400 and response_format is not None:
+                        try:
+                            error_body = e.body if hasattr(e, "body") else {}
+                            if isinstance(error_body, dict):
+                                error_info: dict[str, Any] = error_body.get("error") or {}
+                                if error_info.get("code") == "tool_use_failed":
+                                    failed_gen = error_info.get("failed_generation", "")
+                                    if failed_gen:
+                                        # Parse the tool call format and convert to actions format
+                                        tool_call = json.loads(failed_gen)
+                                        tool_name = tool_call.get("name", "")
+                                        tool_args = tool_call.get("arguments", {})
+                                        # Convert to actions format: {"actions": [{"tool": "name", ...args}]}
+                                        converted = {"actions": [{"tool": tool_name, **tool_args}]}
+                                        if skip_validation:
+                                            result = converted
+                                        else:
+                                            result = response_format.model_validate(converted)
+
+                                        # Record metrics for this successful recovery
+                                        duration = time.time() - start_time
+                                        metrics = get_metrics_collector()
+                                        metrics.record_llm_call(
+                                            provider=self.provider,
+                                            model=self.model,
+                                            scope=scope,
+                                            duration=duration,
+                                            input_tokens=0,
+                                            output_tokens=0,
+                                            success=True,
+                                        )
+                                        if return_usage:
+                                            return result, TokenUsage(input_tokens=0, output_tokens=0, total_tokens=0)
+                                        return result
+                        except (json.JSONDecodeError, KeyError, TypeError):
+                            pass  # Failed to parse tool_use_failed, continue with normal retry
+
                    last_exception = e
                    if attempt < max_retries:
                        backoff = min(initial_backoff * (2**attempt), max_backoff)
@@ -497,14 +535,416 @@ class LLMProvider:
                        logger.error(f"API error after {max_retries + 1} attempts: {str(e)}")
                        raise

-                except Exception as e:
-                    logger.error(f"Unexpected error during LLM call: {type(e).__name__}: {str(e)}")
+                except Exception:
                    raise

            if last_exception:
                raise last_exception
            raise RuntimeError("LLM call failed after all retries with no exception captured")

+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> "LLMToolCallResult":
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or {"type": "function", "function": {"name": "..."}}
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        from .response_models import LLMToolCall, LLMToolCallResult
+
+        async with _global_llm_semaphore:
+            start_time = time.time()
+
+            # Handle Mock provider
+            if self.provider == "mock":
+                return await self._call_with_tools_mock(messages, tools, scope)
+
+            # Handle Anthropic separately (uses different tool format)
+            if self.provider == "anthropic":
+                return await self._call_with_tools_anthropic(
+                    messages, tools, max_completion_tokens, max_retries, initial_backoff, max_backoff, start_time, scope
+                )
+
+            # Handle Gemini (convert to Gemini tool format)
+            if self.provider == "gemini":
+                return await self._call_with_tools_gemini(
+                    messages, tools, max_retries, initial_backoff, max_backoff, start_time, scope
+                )
+
+            # OpenAI-compatible providers (OpenAI, Groq, Ollama, LMStudio)
+            call_params: dict[str, Any] = {
+                "model": self.model,
+                "messages": messages,
+                "tools": tools,
+                "tool_choice": tool_choice,
+            }
+
+            if max_completion_tokens is not None:
+                call_params["max_completion_tokens"] = max_completion_tokens
+            if temperature is not None:
+                call_params["temperature"] = temperature
+
+            # Provider-specific parameters
+            if self.provider == "groq":
+                call_params["seed"] = DEFAULT_LLM_SEED
+
+            last_exception = None
+
+            for attempt in range(max_retries + 1):
+                try:
+                    response = await self._client.chat.completions.create(**call_params)
+
+                    message = response.choices[0].message
+                    finish_reason = response.choices[0].finish_reason
+
+                    # Extract tool calls if present
+                    tool_calls: list[LLMToolCall] = []
+                    if message.tool_calls:
+                        for tc in message.tool_calls:
+                            try:
+                                args = json.loads(tc.function.arguments) if tc.function.arguments else {}
+                            except json.JSONDecodeError:
+                                args = {"_raw": tc.function.arguments}
+                            tool_calls.append(LLMToolCall(id=tc.id, name=tc.function.name, arguments=args))
+
+                    content = message.content
+
+                    # Record metrics
+                    duration = time.time() - start_time
+                    usage = response.usage
+                    input_tokens = usage.prompt_tokens or 0 if usage else 0
+                    output_tokens = usage.completion_tokens or 0 if usage else 0
+
+                    metrics = get_metrics_collector()
+                    metrics.record_llm_call(
+                        provider=self.provider,
+                        model=self.model,
+                        scope=scope,
+                        duration=duration,
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        success=True,
+                    )
+
+                    return LLMToolCallResult(content=content, tool_calls=tool_calls, finish_reason=finish_reason)
+
+                except APIConnectionError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                        continue
+                    raise
+
+                except APIStatusError as e:
+                    if e.status_code in (401, 403):
+                        raise
+                    last_exception = e
+                    if attempt < max_retries:
+                        await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                        continue
+                    raise
+
+                except Exception:
+                    raise
+
+            if last_exception:
+                raise last_exception
+            raise RuntimeError("Tool call failed after all retries")
+
+    async def _call_with_tools_mock(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        scope: str,
+    ) -> "LLMToolCallResult":
+        """Handle mock tool calls for testing."""
+        from .response_models import LLMToolCallResult
+
+        call_record = {
+            "provider": self.provider,
+            "model": self.model,
+            "messages": messages,
+            "tools": [t.get("function", {}).get("name") for t in tools],
+            "scope": scope,
+        }
+        self._mock_calls.append(call_record)
+
+        if self._mock_response is not None:
+            if isinstance(self._mock_response, LLMToolCallResult):
+                return self._mock_response
+            # Allow setting just tool calls as a list
+            if isinstance(self._mock_response, list):
+                from .response_models import LLMToolCall
+
+                return LLMToolCallResult(
+                    tool_calls=[
+                        LLMToolCall(id=f"mock_{i}", name=tc["name"], arguments=tc.get("arguments", {}))
+                        for i, tc in enumerate(self._mock_response)
+                    ],
+                    finish_reason="tool_calls",
+                )
+
+        return LLMToolCallResult(content="mock response", finish_reason="stop")
+
+    async def _call_with_tools_anthropic(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None,
+        max_retries: int,
+        initial_backoff: float,
+        max_backoff: float,
+        start_time: float,
+        scope: str,
+    ) -> "LLMToolCallResult":
+        """Handle Anthropic tool calling."""
+        from anthropic import APIConnectionError, APIStatusError
+
+        from .response_models import LLMToolCall, LLMToolCallResult
+
+        # Convert OpenAI tool format to Anthropic format
+        anthropic_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            anthropic_tools.append(
+                {
+                    "name": func.get("name", ""),
+                    "description": func.get("description", ""),
+                    "input_schema": func.get("parameters", {"type": "object", "properties": {}}),
+                }
+            )
+
+        # Convert messages - handle tool results
+        system_prompt = None
+        anthropic_messages = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt = (system_prompt + "\n\n" + content) if system_prompt else content
+            elif role == "tool":
+                # Anthropic uses tool_result blocks
+                anthropic_messages.append(
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "tool_result", "tool_use_id": msg.get("tool_call_id", ""), "content": content}
+                        ],
+                    }
+                )
+            elif role == "assistant" and msg.get("tool_calls"):
+                # Convert assistant tool calls
+                tool_use_blocks = []
+                for tc in msg["tool_calls"]:
+                    tool_use_blocks.append(
+                        {
+                            "type": "tool_use",
+                            "id": tc.get("id", ""),
+                            "name": tc.get("function", {}).get("name", ""),
+                            "input": json.loads(tc.get("function", {}).get("arguments", "{}")),
+                        }
+                    )
+                anthropic_messages.append({"role": "assistant", "content": tool_use_blocks})
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "tools": anthropic_tools,
+            "max_tokens": max_completion_tokens or 4096,
+        }
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._anthropic_client.messages.create(**call_params)
+
+                # Extract content and tool calls
+                content_parts = []
+                tool_calls: list[LLMToolCall] = []
+
+                for block in response.content:
+                    if block.type == "text":
+                        content_parts.append(block.text)
+                    elif block.type == "tool_use":
+                        tool_calls.append(LLMToolCall(id=block.id, name=block.name, arguments=block.input or {}))
+
+                content = "".join(content_parts) if content_parts else None
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Record metrics
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=time.time() - start_time,
+                    input_tokens=response.usage.input_tokens or 0,
+                    output_tokens=response.usage.output_tokens or 0,
+                    success=True,
+                )
+
+                return LLMToolCallResult(content=content, tool_calls=tool_calls, finish_reason=finish_reason)
+
+            except (APIConnectionError, APIStatusError) as e:
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic tool call failed")
+
+    async def _call_with_tools_gemini(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_retries: int,
+        initial_backoff: float,
+        max_backoff: float,
+        start_time: float,
+        scope: str,
+    ) -> "LLMToolCallResult":
+        """Handle Gemini tool calling."""
+        from .response_models import LLMToolCall, LLMToolCallResult
+
+        # Convert tools to Gemini format
+        gemini_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            gemini_tools.append(
+                genai_types.Tool(
+                    function_declarations=[
+                        genai_types.FunctionDeclaration(
+                            name=func.get("name", ""),
+                            description=func.get("description", ""),
+                            parameters=func.get("parameters"),
+                        )
+                    ]
+                )
+            )
+
+        # Convert messages
+        system_instruction = None
+        gemini_contents = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction = (system_instruction + "\n\n" + content) if system_instruction else content
+            elif role == "tool":
+                # Gemini uses function_response
+                gemini_contents.append(
+                    genai_types.Content(
+                        role="user",
+                        parts=[
+                            genai_types.Part(
+                                function_response=genai_types.FunctionResponse(
+                                    name=msg.get("name", ""),
+                                    response={"result": content},
+                                )
+                            )
+                        ],
+                    )
+                )
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+
+        config = genai_types.GenerateContentConfig(
+            system_instruction=system_instruction,
+            tools=gemini_tools,
+        )
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._gemini_client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=config,
+                )
+
+                # Extract content and tool calls
+                content = None
+                tool_calls: list[LLMToolCall] = []
+
+                if response.candidates and response.candidates[0].content:
+                    for part in response.candidates[0].content.parts:
+                        if hasattr(part, "text") and part.text:
+                            content = part.text
+                        if hasattr(part, "function_call") and part.function_call:
+                            fc = part.function_call
+                            tool_calls.append(
+                                LLMToolCall(
+                                    id=f"gemini_{len(tool_calls)}",
+                                    name=fc.name,
+                                    arguments=dict(fc.args) if fc.args else {},
+                                )
+                            )
+
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Record metrics
+                metrics = get_metrics_collector()
+                input_tokens = response.usage_metadata.prompt_token_count if response.usage_metadata else 0
+                output_tokens = response.usage_metadata.candidates_token_count if response.usage_metadata else 0
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=time.time() - start_time,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                return LLMToolCallResult(content=content, tool_calls=tool_calls, finish_reason=finish_reason)
+
+            except genai_errors.APIError as e:
+                if e.code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Gemini tool call failed")
+
    async def _call_anthropic(
        self,
        messages: list[dict[str, str]],
@@ -0,0 +1,18 @@
+"""
+Mental models module for Hindsight.
+
+Mental models are synthesized summaries that represent understanding. They come
+in different subtypes based on how they were created:
+
+- Structural: Derived from the bank's mission (e.g., "Be a PM for engineering team")
+  These are created upfront based on what any agent with this role would need.
+
+- Emergent: Discovered from data patterns (named entities, temporal clusters, etc.)
+  These surface organically as facts are retained.
+
+- Pinned: User-defined models that persist across refreshes.
+"""
+
+from .models import MentalModel, MentalModelSubtype
+
+__all__ = ["MentalModel", "MentalModelSubtype"]
@@ -0,0 +1,311 @@
+"""
+Emergent mental model detection and promotion.
+
+Emergent models are discovered from data patterns:
+- Named entity extraction (people, projects, systems)
+- Temporal clustering (events with multiple references)
+- Causal patterns ("Because X, we do Y")
+- Behavioral anchors ("After X, we started Y")
+- Reference frequency (anything mentioned repeatedly)
+
+When a pattern is detected, it goes through a mission filter to check relevance,
+and if relevant, is promoted to a mental model.
+"""
+
+import logging
+from typing import TYPE_CHECKING
+
+from pydantic import BaseModel, Field
+
+from .models import EmergentCandidate
+
+if TYPE_CHECKING:
+    from ..llm_wrapper import LLMConfig
+
+logger = logging.getLogger(__name__)
+
+
+class MissionFilterCandidate(BaseModel):
+    """Result of mission filtering for a single candidate."""
+
+    name: str
+    promote: bool = Field(description="True if this is a specific named entity worth tracking")
+    reason: str = Field(description="Brief explanation for the decision")
+
+
+class MissionFilterResponse(BaseModel):
+    """Response from LLM for mission filtering."""
+
+    candidates: list[MissionFilterCandidate] = Field(description="Filtering decision for each candidate")
+
+
+def build_mission_filter_prompt(mission: str, candidates: list[EmergentCandidate]) -> str:
+    """Build the prompt for filtering candidates by mission relevance."""
+    candidate_list = "\n".join(
+        [f"- {c.name} (mentions: {c.mention_count}, method: {c.detection_method})" for c in candidates]
+    )
+
+    return f"""Filter these detected entities. For each one, decide: promote=true or promote=false.
+
+MISSION: {mission}
+
+DETECTED ENTITIES:
+{candidate_list}
+
+=== DECISION RULES ===
+
+Set promote=true ONLY for specific, named entities:
+- Person names: "John", "Maria", "Alice Chen", "Dr. Smith"
+- Named organizations: "Google", "Acme Corp", "Frontend Team"
+- Named places: "Central Park Zoo", "NYC Office", "Building A"
+- Named projects: "Project Phoenix", "Auth Service v2"
+
+Set promote=false for EVERYTHING ELSE, including:
+- Common English words: user, support, help, family, kids, parents, friends, people, team, photo, nature, park, office, home, work, school, joy, love, hope, fear, anger, gratitude, kindness, passion, motivation, inspiration, encouragement, positivity, energy, community, connection, commitment, collaboration, growth, impact, difference, success, progress, change, education, volunteering, veterans, homeless, shelter, meeting, project, system, process, event
+- Generic categories (even capitalized): Users, Customers, Team, Family, Kids, Veterans, Community
+- Abstract concepts: motivation, inspiration, gratitude, commitment, resilience
+
+THE TEST: Is this a specific name you'd find in a contact list or org chart?
+- "John" → YES (promote=true)
+- "kids" → NO (promote=false)
+- "community" → NO (promote=false)
+- "Maria" → YES (promote=true)
+- "park" → NO (promote=false)
+
+When in doubt, set promote=false."""
+
+
+def get_mission_filter_system_message() -> str:
+    """System message for mission filtering."""
+    return """You filter entities for promotion. Output JSON with 'candidates' array.
+
+Rules:
+- promote=true ONLY for specific names (people, organizations, named places/projects)
+- promote=false for common words, generic categories, abstract concepts
+
+Examples:
+- "John" → promote=true (person name)
+- "kids" → promote=false (generic category)
+- "community" → promote=false (abstract concept)
+- "Google" → promote=true (organization name)
+- "motivation" → promote=false (abstract concept)
+
+When in doubt, promote=false. Most entities should be rejected."""
+
+
+async def filter_candidates_by_mission(
+    llm_config: "LLMConfig",
+    mission: str,
+    candidates: list[EmergentCandidate],
+) -> list[EmergentCandidate]:
+    """
+    Filter emergent candidates to keep only specific, named entities.
+
+    Args:
+        llm_config: LLM configuration
+        mission: The bank's mission (used for context)
+        candidates: List of detected candidates
+
+    Returns:
+        Filtered list of candidates that are specific named entities
+    """
+    if not candidates:
+        return []
+
+    if not mission:
+        # No mission = no filtering, keep all candidates
+        logger.debug("[EMERGENT] No mission set, skipping filter")
+        return candidates
+
+    prompt = build_mission_filter_prompt(mission, candidates)
+
+    try:
+        result = await llm_config.call(
+            messages=[
+                {"role": "system", "content": get_mission_filter_system_message()},
+                {"role": "user", "content": prompt},
+            ],
+            response_format=MissionFilterResponse,
+            scope="mental_model_mission_filter",
+        )
+
+        # Build name -> promote map
+        promote_map = {c.name: c.promote for c in result.candidates}
+
+        # Filter candidates
+        filtered = []
+        for candidate in candidates:
+            if candidate.name in promote_map:
+                if promote_map[candidate.name]:
+                    filtered.append(candidate)
+                    logger.debug(f"[EMERGENT] Promoting '{candidate.name}'")
+                else:
+                    logger.debug(f"[EMERGENT] Rejecting '{candidate.name}'")
+            else:
+                # Candidate not in response - reject by default
+                logger.debug(f"[EMERGENT] '{candidate.name}' not in response, rejecting")
+
+        logger.info(f"[EMERGENT] Mission filter: {len(filtered)}/{len(candidates)} candidates promoted")
+        return filtered
+
+    except Exception as e:
+        logger.warning(f"[EMERGENT] Mission filter failed, rejecting all candidates: {e}")
+        return []
+
+
+async def evaluate_emergent_models(
+    llm_config: "LLMConfig",
+    models: list[dict],
+) -> list[str]:
+    """
+    Evaluate existing emergent models to check if they should be kept.
+
+    This re-evaluates emergent models using the same filtering criteria
+    as new candidates. Models that are generic/abstract will be removed.
+
+    Args:
+        llm_config: LLM configuration
+        models: List of existing emergent model dicts with 'name', 'id'
+
+    Returns:
+        List of model IDs that should be REMOVED (no longer valid)
+    """
+    if not models:
+        return []
+
+    # Convert existing models to candidates for evaluation
+    candidates = [
+        EmergentCandidate(
+            name=m["name"],
+            detection_method="existing_emergent_model",
+            mention_count=0,
+        )
+        for m in models
+    ]
+
+    # Build a simple prompt for re-evaluation
+    names_list = "\n".join([f"- {m['name']}" for m in models])
+    prompt = f"""Re-evaluate these existing mental models. For each one, decide: promote=true (keep) or promote=false (remove).
+
+EXISTING MODELS:
+{names_list}
+
+=== DECISION RULES ===
+
+Set promote=true ONLY for specific, named entities:
+- Person names: "John", "Maria", "Alice Chen", "Dr. Smith"
+- Named organizations: "Google", "Acme Corp", "Frontend Team"
+- Named places: "Central Park Zoo", "NYC Office", "Building A"
+- Named projects: "Project Phoenix", "Auth Service v2"
+
+Set promote=false for EVERYTHING ELSE, including:
+- Common English words: user, support, help, family, kids, parents, friends, people, team, photo, nature, park, office, home, work, school, joy, love, hope, fear, anger, gratitude, kindness, passion, motivation, inspiration, encouragement, positivity, energy, community, connection, commitment, collaboration, growth, impact, difference, success, progress, change, education, volunteering, veterans, homeless, shelter, meeting, project, system, process, event
+- Generic categories (even capitalized): Users, Customers, Team, Family, Kids, Veterans, Community
+- Abstract concepts: motivation, inspiration, gratitude, commitment, resilience
+
+THE TEST: Is this a specific name you'd find in a contact list or org chart?
+- "John" → YES (promote=true)
+- "kids" → NO (promote=false)
+- "community" → NO (promote=false)
+
+When in doubt, set promote=false."""
+
+    try:
+        result = await llm_config.call(
+            messages=[
+                {"role": "system", "content": get_mission_filter_system_message()},
+                {"role": "user", "content": prompt},
+            ],
+            response_format=MissionFilterResponse,
+            scope="mental_model_emergent_evaluation",
+        )
+
+        # Build name -> promote map
+        promote_map = {c.name: c.promote for c in result.candidates}
+
+        # Find models to remove
+        models_to_remove = []
+        for model in models:
+            name = model["name"]
+            if name in promote_map:
+                if not promote_map[name]:
+                    models_to_remove.append(model["id"])
+                else:
+                    logger.debug(f"[EMERGENT] Keeping '{name}'")
+            else:
+                # Model not in response - remove to be safe
+                logger.info(f"[EMERGENT] '{name}' not in evaluation response, marking for removal")
+                models_to_remove.append(model["id"])
+
+        logger.info(f"[EMERGENT] Evaluation: {len(models_to_remove)}/{len(models)} emergent models marked for removal")
+        return models_to_remove
+
+    except Exception as e:
+        logger.warning(f"[EMERGENT] Evaluation failed, keeping all models: {e}")
+        return []
+
+
+async def detect_entity_candidates(
+    pool,
+    bank_id: str,
+    min_mentions: int = 5,
+    top_percent: int = 20,
+) -> list[EmergentCandidate]:
+    """
+    Detect entities that are candidates for promotion to mental models.
+
+    Args:
+        pool: Database connection pool
+        bank_id: Bank identifier
+        min_mentions: Minimum mention count to consider
+        top_percent: Only consider top X% by mention count
+
+    Returns:
+        List of entity candidates
+    """
+    from ..db_utils import acquire_with_retry
+    from ..memory_engine import fq_table
+
+    candidates = []
+
+    async with acquire_with_retry(pool) as conn:
+        # Get entities that meet criteria and don't already have mental models
+        rows = await conn.fetch(
+            f"""
+            WITH ranked AS (
+                SELECT
+                    e.id,
+                    e.canonical_name,
+                    e.mention_count,
+                    PERCENT_RANK() OVER (ORDER BY e.mention_count DESC) as rank_pct
+                FROM {fq_table("entities")} e
+                LEFT JOIN {fq_table("mental_models")} mm
+                    ON mm.entity_id = e.id AND mm.bank_id = e.bank_id
+                WHERE e.bank_id = $1
+                  AND e.mention_count >= $2
+                  AND mm.id IS NULL  -- Not already a mental model
+            )
+            SELECT id, canonical_name, mention_count
+            FROM ranked
+            WHERE rank_pct <= $3
+            ORDER BY mention_count DESC
+            LIMIT 50
+            """,
+            bank_id,
+            min_mentions,
+            top_percent / 100.0,
+        )
+
+        for row in rows:
+            candidates.append(
+                EmergentCandidate(
+                    name=row["canonical_name"],
+                    detection_method="named_entity_extraction",
+                    mention_count=row["mention_count"],
+                    entity_id=str(row["id"]),
+                    relevance_score=0.0,
+                )
+            )
+
+    logger.debug(f"[EMERGENT] Detected {len(candidates)} entity candidates")
+    return candidates
@@ -0,0 +1,98 @@
+"""
+Pydantic models for mental models.
+"""
+
+from datetime import datetime, timezone
+from enum import Enum
+
+from pydantic import BaseModel, Field
+
+
+class MentalModelSubtype(str, Enum):
+    """Subtype of mental model - how it was created."""
+
+    STRUCTURAL = "structural"  # Derived from mission, created upfront
+    EMERGENT = "emergent"  # Discovered from data patterns
+    LEARNED = "learned"  # Formed through reflection
+    PINNED = "pinned"  # User-defined topic, observations LLM-generated
+    DIRECTIVE = "directive"  # User-defined hard rules, observations user-provided
+
+
+class MentalModel(BaseModel):
+    """
+    A mental model representing synthesized understanding.
+
+    Mental models are the agent's consolidated knowledge. Unlike raw facts,
+    mental models provide:
+    - A one-liner description for quick scanning/retrieval
+    - A full summary for deep understanding
+    - Links to related mental models
+    """
+
+    id: str = Field(description="Unique identifier within the bank")
+    bank_id: str = Field(description="Bank this mental model belongs to")
+    subtype: MentalModelSubtype = Field(description="How this model was created")
+    name: str = Field(description="Human-readable name")
+    description: str = Field(description="One-liner for quick scanning and retrieval matching")
+    summary: str | None = Field(default=None, description="Full synthesized understanding")
+
+    # References
+    entity_id: str | None = Field(default=None, description="Reference to entities table when type=entity")
+    source_facts: list[str] = Field(default_factory=list, description="Fact IDs used to generate summary")
+    links: list[str] = Field(default_factory=list, description="Related mental model IDs")
+
+    # Tags for scoped visibility (similar to document tags)
+    tags: list[str] = Field(default_factory=list, description="Tags for scoped visibility filtering")
+
+    # Timestamps
+    last_updated: datetime | None = Field(default=None, description="When summary was last regenerated")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this model was created"
+    )
+
+
+class StructuralModelTemplate(BaseModel):
+    """
+    A template for a structural mental model.
+
+    Generated by LLM based on the bank's mission. Represents what any agent
+    with this role would need to track.
+    """
+
+    id: str = Field(default="", description="Existing model ID to keep, or empty for new models")
+    name: str = Field(description="Human-readable name")
+    description: str = Field(description="What this model should track")
+    initial_probes: list[str] = Field(default_factory=list, description="Initial search queries to populate this model")
+
+
+class StructuralModelDerivationResponse(BaseModel):
+    """Response from LLM for structural model derivation."""
+
+    templates: list[StructuralModelTemplate] = Field(description="Structural model templates derived from the mission")
+
+
+class EmergentCandidate(BaseModel):
+    """
+    A candidate for promotion to emergent mental model.
+
+    Detected through pattern analysis of facts.
+    """
+
+    name: str = Field(description="Name of the detected pattern/entity")
+    detection_method: str = Field(description="How this candidate was detected")
+    mention_count: int = Field(default=0, description="How many times referenced")
+    entity_id: str | None = Field(default=None, description="Entity ID if detected as entity")
+    relevance_score: float = Field(default=0.0, description="Score from mission filter (0-1)")
+
+
+class ResearchResult(BaseModel):
+    """
+    Result from the research endpoint.
+
+    Contains the answer along with the mental models and facts used.
+    """
+
+    answer: str = Field(description="The synthesized answer")
+    mental_models_used: list[str] = Field(default_factory=list, description="IDs of mental models that contributed")
+    facts_used: list[str] = Field(default_factory=list, description="Fact IDs that contributed")
+    question_type: str | None = Field(default=None, description="Detected question type (WHO, WHAT, HOW, etc.)")
@@ -0,0 +1,228 @@
+"""
+Structural mental model derivation from bank mission.
+
+Structural models are derived from the bank's mission - they represent what
+any agent with this role would need to track. For example:
+
+Mission: "Be a PM for engineering team"
+Structural models:
+  - Team Structure (who's on the team, roles)
+  - Project Overview (current projects, status)
+  - Processes (how releases work, how decisions are made)
+  - Key Systems (what we own, dependencies)
+"""
+
+import logging
+from typing import TYPE_CHECKING
+
+from pydantic import BaseModel, Field
+
+from .models import StructuralModelTemplate
+
+if TYPE_CHECKING:
+    from ..llm_wrapper import LLMConfig
+
+logger = logging.getLogger(__name__)
+
+
+class StructuralDerivationResponse(BaseModel):
+    """Response from LLM for structural model derivation."""
+
+    templates: list[StructuralModelTemplate] = Field(description="Structural model templates derived from the mission")
+
+
+class StructuralRelevanceResult(BaseModel):
+    """Result of evaluating a structural model's relevance to the mission."""
+
+    name: str
+    relevant: bool
+    reason: str
+
+
+class StructuralRelevanceResponse(BaseModel):
+    """Response from LLM for structural model relevance evaluation."""
+
+    models: list[StructuralRelevanceResult] = Field(description="Relevance evaluation for each model")
+
+
+def build_structural_derivation_prompt(mission: str, existing_models: list[dict] | None = None) -> str:
+    """Build the prompt for deriving structural models from a mission."""
+    existing_section = ""
+    if existing_models:
+        model_list = "\n".join([f"- id='{m['id']}' name='{m['name']}': {m['description']}" for m in existing_models])
+        existing_section = f"""
+EXISTING STRUCTURAL MODELS:
+{model_list}
+
+IMPORTANT: If keeping an existing model, you MUST return its EXACT 'id' value.
+Models not included in your output will be REMOVED.
+"""
+
+    return f"""Given this agent mission, identify the KEY THINGS to track to achieve it.
+
+MISSION: {mission}
+{existing_section}
+IMPORTANT CONSTRAINTS:
+- Return 0-3 structural models MAXIMUM (less is better!)
+- Only include models for SPECIFIC, CONCRETE things the agent needs to track
+- Each model must be DIRECTLY tied to achieving the mission
+- If the mission is simple, return 0 models (empty array is fine)
+- If existing models are provided and you want to keep one, use its EXACT id
+- Do NOT create near-duplicates (e.g., don't create "topic-map" if "topic-connections" exists)
+
+GOOD examples (specific, actionable):
+- Mission: "Be a PM for engineering team" → "Team Members" (track who's on the team)
+- Mission: "Track customer feedback" → "Customer Issues" (track specific complaints/requests)
+- Mission: "Manage project X" → "Project X Milestones" (track progress)
+
+BAD examples (too generic, don't create these):
+- "Processes", "Workflows", "Key Systems", "Important Events"
+- "Communication", "Collaboration", "Progress", "Status"
+- Generic role-based models not tied to the specific mission
+
+For each model:
+1. id: Use EXACT existing id if keeping a model, or leave empty for new models
+2. name: Short, specific name (e.g., "Team Members", "Sprint Goals")
+3. description: One line describing what to track
+4. initial_probes: 2-3 search queries to find relevant information
+
+Return ONLY the models that should exist. Existing models not in your output will be deleted."""
+
+
+def get_structural_derivation_system_message() -> str:
+    """System message for structural model derivation."""
+    return """You identify the key things to track for a mission. Be VERY selective.
+
+Rules:
+- Maximum 3 models (prefer fewer)
+- Only SPECIFIC, CONCRETE things - not generic categories
+- Each must DIRECTLY help achieve the mission
+- Empty array is valid if no models are truly needed
+- If existing models are shown and you want to keep one, return its EXACT id
+- Never create duplicates - if a similar model exists, keep the existing one
+
+Output JSON with 'templates' array (can be empty)."""
+
+
+def _normalize_id(text: str) -> str:
+    """Normalize a string to a canonical form for comparison.
+
+    Removes common suffixes, pluralization, and normalizes separators.
+    """
+    # Lowercase and normalize separators
+    normalized = text.lower().replace(" ", "-").replace("_", "-")
+
+    # Remove common suffixes that indicate the same concept
+    suffixes_to_remove = ["-map", "-list", "-overview", "-tracker", "-s"]
+    for suffix in suffixes_to_remove:
+        if normalized.endswith(suffix) and len(normalized) > len(suffix):
+            normalized = normalized[: -len(suffix)]
+
+    return normalized
+
+
+def _find_similar_existing_id(new_id: str, existing_models: list[dict]) -> str | None:
+    """Find an existing model ID that is similar to the new ID.
+
+    Returns the existing ID if a similar one is found, None otherwise.
+    """
+    if not existing_models:
+        return None
+
+    new_normalized = _normalize_id(new_id)
+
+    for model in existing_models:
+        existing_id = model.get("id", "")
+        existing_normalized = _normalize_id(existing_id)
+
+        # Check if one is a prefix of the other (normalized)
+        if new_normalized.startswith(existing_normalized) or existing_normalized.startswith(new_normalized):
+            return existing_id
+
+        # Check if they're the same when normalized
+        if new_normalized == existing_normalized:
+            return existing_id
+
+    return None
+
+
+async def derive_structural_models(
+    llm_config: "LLMConfig",
+    mission: str,
+    existing_models: list[dict] | None = None,
+) -> tuple[list[StructuralModelTemplate], list[str]]:
+    """
+    Derive structural model templates from a bank's mission.
+
+    This combines derivation and evaluation in one call. The LLM sees existing
+    models and decides which to keep. Any existing model not in the output
+    will be marked for removal.
+
+    Args:
+        llm_config: LLM configuration for calling the model
+        mission: The bank's mission (e.g., "Be a PM for engineering team")
+        existing_models: Optional list of existing model dicts with 'name', 'description', 'id'
+
+    Returns:
+        Tuple of (templates to create/keep, IDs of existing models to remove)
+
+    Raises:
+        Exception: If LLM call fails
+    """
+    prompt = build_structural_derivation_prompt(mission, existing_models)
+
+    result = await llm_config.call(
+        messages=[
+            {"role": "system", "content": get_structural_derivation_system_message()},
+            {"role": "user", "content": prompt},
+        ],
+        response_format=StructuralDerivationResponse,
+        scope="mental_model_structural_derivation",
+    )
+
+    templates = result.templates
+    logger.info(f"[STRUCTURAL] LLM returned {len(templates)} structural models")
+
+    # Build set of existing IDs for quick lookup
+    existing_ids = {m["id"] for m in existing_models} if existing_models else set()
+
+    # Process templates: validate IDs, deduplicate, assign stable IDs
+    processed_templates: list[StructuralModelTemplate] = []
+    kept_existing_ids: set[str] = set()
+
+    for template in templates:
+        # If LLM returned an ID, check if it's a valid existing ID
+        if template.id and template.id in existing_ids:
+            # LLM is keeping an existing model
+            kept_existing_ids.add(template.id)
+            processed_templates.append(template)
+            logger.info(f"[STRUCTURAL] Keeping existing model: {template.id}")
+        else:
+            # New model or LLM didn't return a valid ID
+            # Generate ID from name
+            generated_id = template.name.lower().replace(" ", "-").replace("_", "-")
+
+            # Check for similar existing models to prevent near-duplicates
+            similar_id = _find_similar_existing_id(generated_id, existing_models)
+            if similar_id and similar_id not in kept_existing_ids:
+                # Use the existing similar model instead of creating a new one
+                logger.info(f"[STRUCTURAL] Detected near-duplicate: '{generated_id}' matches existing '{similar_id}'")
+                template.id = similar_id
+                kept_existing_ids.add(similar_id)
+            else:
+                template.id = generated_id
+
+            processed_templates.append(template)
+
+    # Find existing models to remove (not kept in LLM output)
+    models_to_remove = []
+    if existing_models:
+        for model in existing_models:
+            if model["id"] not in kept_existing_ids:
+                logger.info(f"[STRUCTURAL] Marking '{model['name']}' (id={model['id']}) for removal")
+                models_to_remove.append(model["id"])
+
+    if models_to_remove:
+        logger.info(f"[STRUCTURAL] {len(models_to_remove)} existing models will be removed")
+
+    return processed_templates, models_to_remove
@@ -0,0 +1,20 @@
+"""
+Reflect agent module for agentic reflection with tools.
+
+The reflect agent uses an iterative loop with tools to:
+1. Lookup mental models (existing knowledge)
+2. Recall facts (semantic + temporal search)
+3. Learn new insights (create/update mental models)
+4. Expand memories (get chunk/document context)
+"""
+
+from .agent import ReflectAgentResult, run_reflect_agent
+from .models import MentalModelInput, ReflectAction, ReflectActionBatch
+
+__all__ = [
+    "run_reflect_agent",
+    "ReflectAgentResult",
+    "ReflectAction",
+    "ReflectActionBatch",
+    "MentalModelInput",
+]
@@ -0,0 +1,723 @@
+"""
+Reflect agent - agentic loop for reflection with native tool calling.
+"""
+
+import asyncio
+import json
+import logging
+import time
+from typing import TYPE_CHECKING, Any, Awaitable, Callable
+
+from .models import DirectiveInfo, LLMCall, MentalModelInput, ReflectAgentResult, ToolCall
+from .prompts import FINAL_SYSTEM_PROMPT, _extract_directive_rules, build_final_prompt, build_system_prompt_for_tools
+from .tools_schema import get_reflect_tools
+
+
+def _build_directives_applied(directives: list[dict[str, Any]] | None) -> list[DirectiveInfo]:
+    """Build list of DirectiveInfo from directive mental models."""
+    if not directives:
+        return []
+
+    result = []
+    for directive in directives:
+        directive_id = directive.get("id", "")
+        directive_name = directive.get("name", "")
+        observations = directive.get("observations", [])
+
+        rules = []
+        for obs in observations:
+            # Support both Pydantic Observation objects and dicts
+            if hasattr(obs, "content"):
+                rules.append(obs.content)
+            elif isinstance(obs, dict) and obs.get("content"):
+                rules.append(obs["content"])
+
+        result.append(DirectiveInfo(id=directive_id, name=directive_name, rules=rules))
+
+    return result
+
+
+if TYPE_CHECKING:
+    from ..llm_wrapper import LLMProvider
+    from ..response_models import LLMToolCall
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_MAX_ITERATIONS = 10
+
+
+async def _generate_structured_output(
+    answer: str,
+    response_schema: dict,
+    llm_config: "LLMProvider",
+    reflect_id: str,
+) -> dict[str, Any] | None:
+    """Generate structured output from an answer using the provided JSON schema.
+
+    Args:
+        answer: The text answer to extract structured data from
+        response_schema: JSON Schema for the expected output structure
+        llm_config: LLM provider for making the extraction call
+        reflect_id: Reflect ID for logging
+
+    Returns:
+        Structured output dict if successful, None otherwise
+    """
+    try:
+        from typing import Any as TypingAny
+
+        from pydantic import create_model
+
+        def _json_schema_type_to_python(field_schema: dict) -> type:
+            """Map JSON schema type to Python type for better LLM guidance."""
+            json_type = field_schema.get("type", "string")
+            if json_type == "array":
+                return list
+            elif json_type == "object":
+                return dict
+            elif json_type == "integer":
+                return int
+            elif json_type == "number":
+                return float
+            elif json_type == "boolean":
+                return bool
+            else:
+                return str
+
+        # Build fields from JSON schema properties
+        schema_props = response_schema.get("properties", {})
+        required_fields = set(response_schema.get("required", []))
+        fields: dict[str, TypingAny] = {}
+        for field_name, field_schema in schema_props.items():
+            field_type = _json_schema_type_to_python(field_schema)
+            default = ... if field_name in required_fields else None
+            fields[field_name] = (field_type, default)
+
+        if not fields:
+            return None
+
+        DynamicModel = create_model("StructuredResponse", **fields)
+
+        # Include the full schema in the prompt for better LLM guidance
+        schema_str = json.dumps(response_schema, indent=2)
+
+        # Call LLM with the answer to extract structured data
+        structured_prompt = f"""Based on this answer, extract the information into the requested structured format.
+
+Answer: {answer}
+
+JSON Schema to follow:
+```json
+{schema_str}
+```
+
+Return ONLY a valid JSON object that matches this exact schema. Pay special attention to field types:
+- "type": "array" means the value must be a JSON array/list, NOT a string
+- "type": "string" means the value must be a string
+- "type": "object" means the value must be a JSON object
+
+Do not include any explanation, only the JSON object."""
+
+        structured_result = await llm_config.call(
+            messages=[
+                {
+                    "role": "system",
+                    "content": "Extract structured data from the given answer. Return only valid JSON matching the provided schema exactly.",
+                },
+                {"role": "user", "content": structured_prompt},
+            ],
+            response_format=DynamicModel,
+            scope="reflect_structured",
+            skip_validation=True,  # We'll handle the dict ourselves
+        )
+
+        # Convert to dict
+        if hasattr(structured_result, "model_dump"):
+            structured_output = structured_result.model_dump()
+        elif isinstance(structured_result, dict):
+            structured_output = structured_result
+        else:
+            # Try to parse as JSON
+            structured_output = json.loads(str(structured_result))
+
+        logger.info(f"[REFLECT {reflect_id}] Generated structured output with {len(structured_output)} fields")
+        return structured_output
+
+    except Exception as e:
+        logger.warning(f"[REFLECT {reflect_id}] Failed to generate structured output: {e}")
+        return None
+
+
+async def run_reflect_agent(
+    llm_config: "LLMProvider",
+    bank_id: str,
+    query: str,
+    bank_profile: dict[str, Any],
+    lookup_fn: Callable[[str | None], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
+    learn_fn: Callable[[MentalModelInput], Awaitable[dict[str, Any]]] | None = None,
+    context: str | None = None,
+    max_iterations: int = DEFAULT_MAX_ITERATIONS,
+    max_tokens: int | None = None,
+    response_schema: dict | None = None,
+    directives: list[dict[str, Any]] | None = None,
+) -> ReflectAgentResult:
+    """
+    Execute the reflect agent loop using native tool calling.
+
+    The agent iteratively calls tools to gather information and learn,
+    then provides a final answer via the done() tool.
+
+    Args:
+        llm_config: LLM provider for agent calls
+        bank_id: Bank identifier
+        query: Question to answer
+        bank_profile: Bank profile with name and mission
+        lookup_fn: Tool callback for lookup (model_id) -> result
+        recall_fn: Tool callback for recall (query, max_tokens) -> result
+        expand_fn: Tool callback for expand (memory_id, depth) -> result
+        learn_fn: Optional tool callback for learn (MentalModelInput) -> result.
+                  If None, learn tool is disabled.
+        context: Optional additional context
+        max_iterations: Maximum number of iterations before forcing response
+        max_tokens: Maximum tokens for the final response
+        response_schema: Optional JSON Schema for structured output in final response
+        directives: Optional list of directive mental models to inject as hard rules
+
+    Returns:
+        ReflectAgentResult with final answer and metadata
+    """
+    enable_learn = learn_fn is not None
+    reflect_id = f"{bank_id[:8]}-{int(time.time() * 1000) % 100000}"
+    start_time = time.time()
+
+    # Build directives_applied for the trace
+    directives_applied = _build_directives_applied(directives)
+
+    # Extract directive rules for tool schema (if any)
+    directive_rules = _extract_directive_rules(directives) if directives else None
+
+    # Get tools for this agent (with directive compliance field if directives exist)
+    tools = get_reflect_tools(enable_learn=enable_learn, directive_rules=directive_rules)
+
+    # Build initial messages (directives are injected into system prompt at START and END)
+    system_prompt = build_system_prompt_for_tools(bank_profile, context, directives=directives)
+    messages: list[dict[str, Any]] = [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": query},
+    ]
+
+    # Tracking
+    mental_models_created: list[str] = []
+    total_tools_called = 0
+    tool_trace: list[ToolCall] = []
+    tool_trace_summary: list[dict[str, Any]] = []
+    llm_trace: list[dict[str, Any]] = []
+    context_history: list[dict[str, Any]] = []  # For final prompt fallback
+
+    # Track available IDs for validation (prevents hallucinated citations)
+    available_memory_ids: set[str] = set()
+    available_model_ids: set[str] = set()
+
+    # Pre-fetch mental models so the agent always starts with this knowledge
+    prefetch_start = time.time()
+    models_result = await lookup_fn(None)  # List all mental models
+    prefetch_duration = int((time.time() - prefetch_start) * 1000)
+
+    # Track available model IDs
+    if isinstance(models_result, dict) and "models" in models_result:
+        for model in models_result["models"]:
+            if "id" in model:
+                available_model_ids.add(model["id"])
+
+    # Add to context history for the agent
+    context_history.append({"tool": "list_mental_models", "output": models_result})
+
+    # Add to tool trace
+    tool_trace.append(
+        ToolCall(
+            tool="list_mental_models",
+            input={"tool": "list_mental_models"},
+            output=models_result,
+            duration_ms=prefetch_duration,
+            iteration=0,
+        )
+    )
+    tool_trace_summary.append(
+        {
+            "tool": "list_mental_models",
+            "input_summary": "(prefetch)",
+            "duration_ms": prefetch_duration,
+            "output_chars": len(json.dumps(models_result, default=str)),
+        }
+    )
+    total_tools_called += 1
+
+    # Include in the user message so the agent sees it
+    models_info = json.dumps(models_result, indent=2, default=str)
+    messages[1]["content"] = f"{query}\n\n## Available Mental Models (pre-fetched)\n```json\n{models_info}\n```"
+
+    def _get_llm_trace() -> list[LLMCall]:
+        return [LLMCall(scope=c["scope"], duration_ms=c["duration_ms"]) for c in llm_trace]
+
+    def _log_completion(answer: str, iterations: int, forced: bool = False):
+        elapsed_ms = int((time.time() - start_time) * 1000)
+        tools_summary = (
+            ", ".join(
+                f"{t['tool']}({t['input_summary']})={t['duration_ms']}ms/{t.get('output_chars', 0)}c"
+                for t in tool_trace_summary
+            )
+            or "none"
+        )
+        llm_summary = ", ".join(f"{c['scope']}={c['duration_ms']}ms" for c in llm_trace) or "none"
+        total_llm_ms = sum(c["duration_ms"] for c in llm_trace)
+        total_tools_ms = sum(t["duration_ms"] for t in tool_trace_summary)
+
+        answer_preview = answer[:100] + "..." if len(answer) > 100 else answer
+        mode = "forced" if forced else "done"
+        logger.info(
+            f"[REFLECT {reflect_id}] {mode} | "
+            f"query='{query[:50]}...' | "
+            f"iterations={iterations} | "
+            f"llm=[{llm_summary}] ({total_llm_ms}ms) | "
+            f"tools=[{tools_summary}] ({total_tools_ms}ms) | "
+            f"answer='{answer_preview}' | "
+            f"total={elapsed_ms}ms"
+        )
+
+    for iteration in range(max_iterations):
+        is_last = iteration == max_iterations - 1
+
+        if is_last:
+            # Force text response on last iteration - no tools
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            llm_start = time.time()
+            response = await llm_config.call(
+                messages=[
+                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "user", "content": prompt},
+                ],
+                scope="reflect_agent_final",
+                max_completion_tokens=max_tokens,
+            )
+            llm_trace.append({"scope": "final", "duration_ms": int((time.time() - llm_start) * 1000)})
+            answer = response.strip()
+
+            # Generate structured output if schema provided
+            structured_output = None
+            if response_schema and answer:
+                structured_output = await _generate_structured_output(answer, response_schema, llm_config, reflect_id)
+
+            _log_completion(answer, iteration + 1, forced=True)
+            return ReflectAgentResult(
+                text=answer,
+                structured_output=structured_output,
+                iterations=iteration + 1,
+                tools_called=total_tools_called,
+                mental_models_created=mental_models_created,
+                tool_trace=tool_trace,
+                llm_trace=_get_llm_trace(),
+                directives_applied=directives_applied,
+            )
+
+        # Call LLM with tools
+        llm_start = time.time()
+
+        try:
+            result = await llm_config.call_with_tools(
+                messages=messages,
+                tools=tools,
+                scope="reflect_agent",
+                tool_choice="required" if iteration == 0 else "auto",  # Force tool use on first iteration
+            )
+            llm_duration = int((time.time() - llm_start) * 1000)
+            llm_trace.append({"scope": f"agent_{iteration + 1}", "duration_ms": llm_duration})
+
+        except Exception:
+            llm_trace.append(
+                {"scope": f"agent_{iteration + 1}_err", "duration_ms": int((time.time() - llm_start) * 1000)}
+            )
+            # Guardrail: If no evidence gathered yet, retry
+            has_gathered_evidence = bool(available_memory_ids) or bool(available_model_ids)
+            if not has_gathered_evidence and iteration < max_iterations - 1:
+                continue
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            llm_start = time.time()
+            response = await llm_config.call(
+                messages=[
+                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "user", "content": prompt},
+                ],
+                scope="reflect_agent_final",
+                max_completion_tokens=max_tokens,
+            )
+            llm_trace.append({"scope": "final", "duration_ms": int((time.time() - llm_start) * 1000)})
+            answer = response.strip()
+
+            # Generate structured output if schema provided
+            structured_output = None
+            if response_schema and answer:
+                structured_output = await _generate_structured_output(answer, response_schema, llm_config, reflect_id)
+
+            _log_completion(answer, iteration + 1, forced=True)
+            return ReflectAgentResult(
+                text=answer,
+                structured_output=structured_output,
+                iterations=iteration + 1,
+                tools_called=total_tools_called,
+                mental_models_created=mental_models_created,
+                tool_trace=tool_trace,
+                llm_trace=_get_llm_trace(),
+                directives_applied=directives_applied,
+            )
+
+        # No tool calls - LLM wants to respond with text
+        if not result.tool_calls:
+            if result.content:
+                answer = result.content.strip()
+
+                # Generate structured output if schema provided
+                structured_output = None
+                if response_schema and answer:
+                    structured_output = await _generate_structured_output(
+                        answer, response_schema, llm_config, reflect_id
+                    )
+
+                _log_completion(answer, iteration + 1)
+                return ReflectAgentResult(
+                    text=answer,
+                    structured_output=structured_output,
+                    iterations=iteration + 1,
+                    tools_called=total_tools_called,
+                    mental_models_created=mental_models_created,
+                    tool_trace=tool_trace,
+                    llm_trace=_get_llm_trace(),
+                    directives_applied=directives_applied,
+                )
+            # Empty response, force final
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            llm_start = time.time()
+            response = await llm_config.call(
+                messages=[
+                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "user", "content": prompt},
+                ],
+                scope="reflect_agent_final",
+                max_completion_tokens=max_tokens,
+            )
+            llm_trace.append({"scope": "final", "duration_ms": int((time.time() - llm_start) * 1000)})
+            answer = response.strip()
+
+            # Generate structured output if schema provided
+            structured_output = None
+            if response_schema and answer:
+                structured_output = await _generate_structured_output(answer, response_schema, llm_config, reflect_id)
+
+            _log_completion(answer, iteration + 1, forced=True)
+            return ReflectAgentResult(
+                text=answer,
+                structured_output=structured_output,
+                iterations=iteration + 1,
+                tools_called=total_tools_called,
+                mental_models_created=mental_models_created,
+                tool_trace=tool_trace,
+                llm_trace=_get_llm_trace(),
+                directives_applied=directives_applied,
+            )
+
+        # Check for done tool call (handle both 'done' and 'functions.done')
+        done_call = next((tc for tc in result.tool_calls if tc.name == "done" or tc.name == "functions.done"), None)
+        if done_call:
+            # Guardrail: Require evidence before done
+            has_gathered_evidence = bool(available_memory_ids) or bool(available_model_ids)
+            if not has_gathered_evidence and iteration < max_iterations - 1:
+                # Add assistant message and fake tool result asking for evidence
+                messages.append(
+                    {
+                        "role": "assistant",
+                        "tool_calls": [_tool_call_to_dict(done_call)],
+                    }
+                )
+                messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": done_call.id,
+                        "content": json.dumps(
+                            {
+                                "error": "You must call recall() or list_mental_models() to gather evidence before providing your final answer."
+                            }
+                        ),
+                    }
+                )
+                continue
+
+            # Process done tool
+            return await _process_done_tool(
+                done_call,
+                available_memory_ids,
+                available_model_ids,
+                iteration + 1,
+                total_tools_called,
+                mental_models_created,
+                tool_trace,
+                _get_llm_trace(),
+                _log_completion,
+                reflect_id,
+                directives_applied=directives_applied,
+                llm_config=llm_config,
+                response_schema=response_schema,
+            )
+
+        # Execute other tools in parallel (exclude done and functions.done)
+        other_tools = [tc for tc in result.tool_calls if tc.name not in ("done", "functions.done")]
+        if other_tools:
+            # Add assistant message with tool calls
+            messages.append(
+                {
+                    "role": "assistant",
+                    "tool_calls": [_tool_call_to_dict(tc) for tc in other_tools],
+                }
+            )
+
+            # Execute tools in parallel
+            tool_tasks = [
+                _execute_tool_with_timing(tc, lookup_fn, recall_fn, expand_fn, learn_fn) for tc in other_tools
+            ]
+            tool_results = await asyncio.gather(*tool_tasks, return_exceptions=True)
+            total_tools_called += len(other_tools)
+
+            # Process results and add to messages
+            for tc, result_data in zip(other_tools, tool_results):
+                if isinstance(result_data, Exception):
+                    # Tool execution failed - log and raise to fail the request
+                    logger.error(f"[REFLECT {reflect_id}] Tool {tc.name} failed with exception: {result_data}")
+                    raise RuntimeError(f"Reflect tool '{tc.name}' failed: {result_data}")
+
+                output, duration_ms = result_data
+
+                # Check if tool returned an error response
+                if isinstance(output, dict) and "error" in output:
+                    logger.error(f"[REFLECT {reflect_id}] Tool {tc.name} returned error: {output['error']}")
+                    raise RuntimeError(f"Reflect tool '{tc.name}' error: {output['error']}")
+
+                # Track created mental models
+                if tc.name == "learn" and isinstance(output, dict) and "model_id" in output:
+                    mental_models_created.append(output["model_id"])
+
+                # Track available memory IDs from recall
+                if tc.name == "recall" and isinstance(output, dict) and "memories" in output:
+                    for memory in output["memories"]:
+                        if "id" in memory:
+                            available_memory_ids.add(memory["id"])
+
+                # Track available model IDs
+                if tc.name in ("list_mental_models", "get_mental_model") and isinstance(output, dict):
+                    if output.get("found") and "model" in output:
+                        model_id = output["model"].get("id")
+                        if model_id:
+                            available_model_ids.add(model_id)
+                    elif "models" in output:
+                        for model in output["models"]:
+                            if "id" in model:
+                                available_model_ids.add(model["id"])
+
+                # Add tool result message
+                messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": tc.id,
+                        "content": json.dumps(output, default=str),
+                    }
+                )
+
+                # Track for logging and context history
+                input_dict = {"tool": tc.name, **tc.arguments}
+                input_summary = _summarize_input(tc.name, tc.arguments)
+
+                tool_trace.append(
+                    ToolCall(
+                        tool=tc.name, input=input_dict, output=output, duration_ms=duration_ms, iteration=iteration + 1
+                    )
+                )
+
+                try:
+                    output_chars = len(json.dumps(output))
+                except (TypeError, ValueError):
+                    output_chars = len(str(output))
+
+                tool_trace_summary.append(
+                    {
+                        "tool": tc.name,
+                        "input_summary": input_summary,
+                        "duration_ms": duration_ms,
+                        "output_chars": output_chars,
+                    }
+                )
+
+                # Keep context history for fallback final prompt
+                context_history.append({"tool": tc.name, "input": input_dict, "output": output})
+
+    # Should not reach here
+    answer = "I was unable to formulate a complete answer within the iteration limit."
+    _log_completion(answer, max_iterations, forced=True)
+    return ReflectAgentResult(
+        text=answer,
+        iterations=max_iterations,
+        tools_called=total_tools_called,
+        mental_models_created=mental_models_created,
+        tool_trace=tool_trace,
+        llm_trace=_get_llm_trace(),
+        directives_applied=directives_applied,
+    )
+
+
+def _tool_call_to_dict(tc: "LLMToolCall") -> dict[str, Any]:
+    """Convert LLMToolCall to OpenAI message format."""
+    return {
+        "id": tc.id,
+        "type": "function",
+        "function": {
+            "name": tc.name,
+            "arguments": json.dumps(tc.arguments),
+        },
+    }
+
+
+async def _process_done_tool(
+    done_call: "LLMToolCall",
+    available_memory_ids: set[str],
+    available_model_ids: set[str],
+    iterations: int,
+    total_tools_called: int,
+    mental_models_created: list[str],
+    tool_trace: list[ToolCall],
+    llm_trace: list[LLMCall],
+    log_completion: Callable,
+    reflect_id: str,
+    directives_applied: list[DirectiveInfo],
+    llm_config: "LLMProvider | None" = None,
+    response_schema: dict | None = None,
+) -> ReflectAgentResult:
+    """Process the done tool call and return the result."""
+    args = done_call.arguments
+
+    answer = args.get("answer", "").strip()
+    if not answer:
+        answer = "No answer provided."
+
+    # Validate IDs
+    used_memory_ids = [mid for mid in args.get("memory_ids", []) if mid in available_memory_ids]
+    used_model_ids = [mid for mid in args.get("model_ids", []) if mid in available_model_ids]
+
+    # Generate structured output if schema provided
+    structured_output = None
+    if response_schema and llm_config and answer:
+        structured_output = await _generate_structured_output(answer, response_schema, llm_config, reflect_id)
+
+    log_completion(answer, iterations)
+    return ReflectAgentResult(
+        text=answer,
+        structured_output=structured_output,
+        iterations=iterations,
+        tools_called=total_tools_called,
+        mental_models_created=mental_models_created,
+        tool_trace=tool_trace,
+        llm_trace=llm_trace,
+        used_memory_ids=used_memory_ids,
+        used_model_ids=used_model_ids,
+        directives_applied=directives_applied,
+    )
+
+
+async def _execute_tool_with_timing(
+    tc: "LLMToolCall",
+    lookup_fn: Callable[[str | None], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
+    learn_fn: Callable[[MentalModelInput], Awaitable[dict[str, Any]]] | None = None,
+) -> tuple[dict[str, Any], int]:
+    """Execute a tool call and return result with timing."""
+    start = time.time()
+    result = await _execute_tool(tc.name, tc.arguments, lookup_fn, recall_fn, expand_fn, learn_fn)
+    duration_ms = int((time.time() - start) * 1000)
+    return result, duration_ms
+
+
+async def _execute_tool(
+    tool_name: str,
+    args: dict[str, Any],
+    lookup_fn: Callable[[str | None], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
+    learn_fn: Callable[[MentalModelInput], Awaitable[dict[str, Any]]] | None = None,
+) -> dict[str, Any]:
+    """Execute a single tool by name."""
+    # Normalize tool name - some LLMs return 'functions.done' instead of 'done'
+    if tool_name.startswith("functions."):
+        tool_name = tool_name[len("functions.") :]
+
+    if tool_name == "list_mental_models":
+        return await lookup_fn(None)
+
+    elif tool_name == "get_mental_model":
+        model_id = args.get("model_id")
+        if not model_id:
+            return {"error": "get_mental_model requires model_id"}
+        return await lookup_fn(model_id)
+
+    elif tool_name == "recall":
+        query = args.get("query")
+        if not query:
+            return {"error": "recall requires a query parameter"}
+        max_tokens = max(args.get("max_tokens") or 2048, 1000)  # Default 2048, min 1000
+        return await recall_fn(query, max_tokens)
+
+    elif tool_name == "learn":
+        if learn_fn is None:
+            return {"error": "learn tool is not available"}
+        name = args.get("name")
+        description = args.get("description")
+        if not name or not description:
+            return {"error": "learn requires name and description"}
+        return await learn_fn(MentalModelInput(name=name, description=description))
+
+    elif tool_name == "expand":
+        memory_ids = args.get("memory_ids", [])
+        if not memory_ids:
+            return {"error": "expand requires memory_ids"}
+        depth = args.get("depth", "chunk")
+        return await expand_fn(memory_ids, depth)
+
+    else:
+        return {"error": f"Unknown tool: {tool_name}"}
+
+
+def _summarize_input(tool_name: str, args: dict[str, Any]) -> str:
+    """Create a summary of tool input for logging, showing all params."""
+    if tool_name == "list_mental_models":
+        return "()"
+    elif tool_name == "get_mental_model":
+        return f"(model_id={args.get('model_id', '?')})"
+    elif tool_name == "recall":
+        query = args.get("query", "")
+        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
+        # Show actual value used (default 2048, min 1000)
+        max_tokens = max(args.get("max_tokens") or 2048, 1000)
+        return f"(query={query_preview}, max_tokens={max_tokens})"
+    elif tool_name == "learn":
+        name = args.get("name", "?")
+        desc = args.get("description", "")
+        desc_preview = f"'{desc[:20]}...'" if len(desc) > 20 else f"'{desc}'"
+        return f"(name='{name}', description={desc_preview})"
+    elif tool_name == "expand":
+        memory_ids = args.get("memory_ids", [])
+        depth = args.get("depth", "chunk")
+        return f"(memory_ids=[{len(memory_ids)} ids], depth={depth})"
+    elif tool_name == "done":
+        answer = args.get("answer", "")
+        answer_preview = f"'{answer[:30]}...'" if len(answer) > 30 else f"'{answer}'"
+        memory_ids = args.get("memory_ids", [])
+        model_ids = args.get("model_ids", [])
+        return f"(answer={answer_preview}, memory_ids={len(memory_ids)}, model_ids={len(model_ids)})"
+    return str(args)
@@ -0,0 +1,114 @@
+"""
+Pydantic models for the reflect agent.
+"""
+
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+
+class MentalModelObservation(BaseModel):
+    """An observation within a mental model with its supporting memories."""
+
+    title: str = Field(description="Observation header (can be empty for intro)")
+    text: str = Field(description="Observation content - no headers, use lists/tables/bold")
+    memory_ids: list[str] = Field(default_factory=list, description="Memory IDs supporting this observation")
+
+
+class MentalModelInput(BaseModel):
+    """Input for the learn tool to create a mental model placeholder.
+
+    The agent only specifies name and description - the actual content/observations
+    are generated during refresh, similar to pinned models.
+    """
+
+    name: str = Field(description="Human-readable name for the mental model")
+    description: str = Field(description="What to track - used as prompt for content generation during refresh")
+    entity_id: str | None = Field(default=None, description="Optional link to existing entity ID")
+
+
+class AnswerSection(BaseModel):
+    """A section of the answer with its supporting evidence (DEPRECATED)."""
+
+    title: str = Field(description="Section header/title")
+    text: str = Field(description="Section content")
+    memory_ids: list[str] = Field(default_factory=list, description="Memory IDs supporting this section")
+    model_ids: list[str] = Field(default_factory=list, description="Mental model IDs supporting this section")
+
+
+class ReflectAction(BaseModel):
+    """Single action the reflect agent can take."""
+
+    tool: Literal["list_mental_models", "get_mental_model", "recall", "learn", "expand", "done"] = Field(
+        description="Tool to invoke: list_mental_models, get_mental_model, recall, learn, expand, or done"
+    )
+    # Tool-specific parameters
+    model_id: str | None = Field(default=None, description="Mental model ID for get_mental_model")
+    query: str | None = Field(default=None, description="Search query for recall")
+    max_tokens: int | None = Field(default=None, description="Max tokens for recall results (default 2048)")
+    mental_model: MentalModelInput | None = Field(default=None, description="Mental model to create/update for learn")
+    memory_ids: list[str] | None = Field(default=None, description="Memory unit IDs for expand (batched)")
+    depth: Literal["chunk", "document"] | None = Field(default=None, description="Expansion depth for expand")
+    sections: list[AnswerSection] | None = Field(default=None, description="DEPRECATED: Use answer field instead")
+    observations: list[MentalModelObservation] | None = Field(
+        default=None, description="Observations for done action (when output_mode=observations)"
+    )
+    # Plain text answer fields (for output_mode=answer)
+    answer: str | None = Field(default=None, description="Plain text answer for done action (no markdown)")
+    answer_memory_ids: list[str] | None = Field(
+        default=None, description="Memory IDs supporting the answer", alias="memory_ids"
+    )
+    answer_model_ids: list[str] | None = Field(
+        default=None, description="Mental model IDs supporting the answer", alias="model_ids"
+    )
+    reasoning: str | None = Field(default=None, description="Brief reasoning for this action")
+
+
+class ReflectActionBatch(BaseModel):
+    """Batch of actions for parallel execution."""
+
+    actions: list[ReflectAction] = Field(description="List of actions to execute in parallel")
+
+
+class ToolCall(BaseModel):
+    """A single tool call made during reflect."""
+
+    tool: str = Field(description="Tool name: lookup, recall, learn, expand")
+    input: dict = Field(description="Tool input parameters")
+    output: dict = Field(description="Tool output/result")
+    duration_ms: int = Field(description="Execution time in milliseconds")
+    iteration: int = Field(default=0, description="Iteration number (1-based) when this tool was called")
+
+
+class LLMCall(BaseModel):
+    """A single LLM call made during reflect."""
+
+    scope: str = Field(description="Call scope: agent_1, agent_2, final, etc.")
+    duration_ms: int = Field(description="Execution time in milliseconds")
+
+
+class DirectiveInfo(BaseModel):
+    """Information about a directive that was applied during reflect."""
+
+    id: str = Field(description="Directive mental model ID")
+    name: str = Field(description="Directive name")
+    rules: list[str] = Field(default_factory=list, description="Directive rules/observations that were applied")
+
+
+class ReflectAgentResult(BaseModel):
+    """Result from the reflect agent."""
+
+    text: str = Field(description="Final answer text")
+    structured_output: dict[str, Any] | None = Field(
+        default=None, description="Structured output parsed according to provided response_schema"
+    )
+    iterations: int = Field(default=0, description="Number of iterations taken")
+    tools_called: int = Field(default=0, description="Total number of tool calls made")
+    mental_models_created: list[str] = Field(default_factory=list, description="IDs of mental models created/updated")
+    tool_trace: list[ToolCall] = Field(default_factory=list, description="Trace of all tool calls made")
+    llm_trace: list[LLMCall] = Field(default_factory=list, description="Trace of all LLM calls made")
+    used_memory_ids: list[str] = Field(default_factory=list, description="Validated memory IDs actually used in answer")
+    used_model_ids: list[str] = Field(default_factory=list, description="Validated model IDs actually used in answer")
+    directives_applied: list[DirectiveInfo] = Field(
+        default_factory=list, description="Directive mental models that affected this reflection"
+    )
@@ -0,0 +1,248 @@
+"""
+Models and utilities for evidence-grounded observations with computed trends.
+
+Observations are part of mental models and represent patterns/beliefs derived
+from memories. Each observation must be grounded in specific evidence (quotes)
+from memories, and trends are computed algorithmically from evidence timestamps.
+"""
+
+from datetime import datetime, timedelta, timezone
+from enum import Enum
+
+from pydantic import BaseModel, Field, computed_field, field_validator
+
+
+class Trend(str, Enum):
+    """Computed trend for an observation based on evidence timestamps.
+
+    Trends indicate how an observation's evidence is distributed over time:
+    - STABLE: Evidence spread across time, continues to present
+    - STRENGTHENING: More/denser evidence recently than before
+    - WEAKENING: Evidence mostly old, sparse recently
+    - NEW: All evidence within recent window
+    - STALE: No evidence in recent window (may no longer apply)
+    """
+
+    STABLE = "stable"
+    STRENGTHENING = "strengthening"
+    WEAKENING = "weakening"
+    NEW = "new"
+    STALE = "stale"
+
+
+class ObservationEvidence(BaseModel):
+    """A single piece of evidence supporting an observation.
+
+    Each evidence item must include an exact quote from the source memory
+    to ensure observations are grounded and verifiable.
+    """
+
+    memory_id: str = Field(description="ID of the memory unit this evidence comes from")
+    quote: str = Field(description="Exact quote from the memory supporting the observation")
+    relevance: str = Field(default="", description="Brief explanation of how this quote supports the observation")
+    timestamp: datetime = Field(description="When the source memory was created")
+
+    @field_validator("timestamp", mode="before")
+    @classmethod
+    def ensure_timezone_aware(cls, v: datetime | str | None) -> datetime:
+        """Ensure timestamp is always timezone-aware UTC."""
+        if v is None:
+            return datetime.now(timezone.utc)
+        if isinstance(v, str):
+            # Parse ISO format string, handling 'Z' suffix
+            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
+        if isinstance(v, datetime):
+            if v.tzinfo is None:
+                return v.replace(tzinfo=timezone.utc)
+            return v
+        raise ValueError(f"Invalid timestamp type: {type(v)}")
+
+
+class Observation(BaseModel):
+    """A single observation within a mental model.
+
+    Observations represent patterns, preferences, beliefs, or other insights
+    derived from memories. Each observation must be grounded in evidence
+    with exact quotes from source memories.
+    """
+
+    title: str = Field(description="Short summary title for the observation (5-10 words)")
+    content: str = Field(description="The observation content - detailed explanation of what we believe to be true")
+    evidence: list[ObservationEvidence] = Field(default_factory=list, description="Supporting evidence with quotes")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this observation was first created"
+    )
+
+    @field_validator("created_at", mode="before")
+    @classmethod
+    def ensure_created_at_timezone_aware(cls, v: datetime | str | None) -> datetime:
+        """Ensure created_at is always timezone-aware UTC."""
+        if v is None:
+            return datetime.now(timezone.utc)
+        if isinstance(v, str):
+            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
+        if isinstance(v, datetime):
+            if v.tzinfo is None:
+                return v.replace(tzinfo=timezone.utc)
+            return v
+        raise ValueError(f"Invalid created_at type: {type(v)}")
+
+    @computed_field
+    @property
+    def trend(self) -> Trend:
+        """Compute trend from evidence timestamps."""
+        return compute_trend(self.evidence)
+
+    @computed_field
+    @property
+    def evidence_span(self) -> dict[str, str | None]:
+        """Get the time span covered by evidence."""
+        if not self.evidence:
+            return {"from": None, "to": None}
+        timestamps = [e.timestamp for e in self.evidence]
+        return {
+            "from": min(timestamps).isoformat(),
+            "to": max(timestamps).isoformat(),
+        }
+
+    @computed_field
+    @property
+    def evidence_count(self) -> int:
+        """Number of evidence items supporting this observation."""
+        return len(self.evidence)
+
+
+def compute_trend(
+    evidence: list[ObservationEvidence],
+    now: datetime | None = None,
+    recent_days: int = 30,
+    old_days: int = 90,
+) -> Trend:
+    """Compute the trend for an observation based on evidence timestamps.
+
+    The trend indicates how the evidence is distributed over time:
+    - STABLE: Evidence spread across time, continues to present
+    - STRENGTHENING: More evidence recently than historically
+    - WEAKENING: Evidence mostly old, sparse recently
+    - NEW: All evidence is recent (within recent_days)
+    - STALE: No evidence in recent window
+
+    Args:
+        evidence: List of evidence items with timestamps
+        now: Reference time for calculations (defaults to current UTC time)
+        recent_days: Number of days to consider "recent" (default 30)
+        old_days: Number of days to consider "old" (default 90)
+
+    Returns:
+        Computed Trend enum value
+    """
+    if now is None:
+        now = datetime.now(timezone.utc)
+
+    # Ensure now is timezone-aware
+    if now.tzinfo is None:
+        now = now.replace(tzinfo=timezone.utc)
+
+    if not evidence:
+        return Trend.STALE
+
+    recent_cutoff = now - timedelta(days=recent_days)
+    old_cutoff = now - timedelta(days=old_days)
+
+    # Normalize timestamps to UTC for comparison
+    def normalize_ts(ts: datetime) -> datetime:
+        if ts.tzinfo is None:
+            return ts.replace(tzinfo=timezone.utc)
+        return ts
+
+    recent = [e for e in evidence if normalize_ts(e.timestamp) > recent_cutoff]
+    old = [e for e in evidence if normalize_ts(e.timestamp) < old_cutoff]
+    middle = [e for e in evidence if old_cutoff <= normalize_ts(e.timestamp) <= recent_cutoff]
+
+    # No recent evidence = stale
+    if not recent:
+        return Trend.STALE
+
+    # All evidence is recent = new
+    if not old and not middle:
+        return Trend.NEW
+
+    # Compare density (evidence per day)
+    recent_density = len(recent) / recent_days if recent_days > 0 else 0
+    older_period = old_days - recent_days
+    older_density = (len(old) + len(middle)) / older_period if older_period > 0 else 0
+
+    # Avoid division by zero
+    if older_density == 0:
+        return Trend.NEW
+
+    ratio = recent_density / older_density
+
+    if ratio > 1.5:
+        return Trend.STRENGTHENING
+    elif ratio < 0.5:
+        return Trend.WEAKENING
+    else:
+        return Trend.STABLE
+
+
+class CandidateObservation(BaseModel):
+    """A candidate observation generated during the seed phase.
+
+    Candidates are preliminary observations that need evidence validation
+    before becoming full observations.
+    """
+
+    content: str = Field(description="The proposed observation content")
+    seed_memory_ids: list[str] = Field(default_factory=list, description="Memory IDs that inspired this candidate")
+
+
+class CandidateWithEvidence(BaseModel):
+    """A candidate observation with gathered supporting and contradicting evidence."""
+
+    candidate: CandidateObservation
+    supporting_memories: list[dict] = Field(default_factory=list, description="Memories that support this observation")
+    contradicting_memories: list[dict] = Field(
+        default_factory=list, description="Memories that contradict this observation"
+    )
+
+
+class MentalModelSnapshot(BaseModel):
+    """A versioned snapshot of a mental model's observations.
+
+    Used for tracking changes over time and enabling diff views.
+    """
+
+    version: int = Field(description="Version number (1-indexed)")
+    observations: list[Observation] = Field(default_factory=list, description="Observations at this version")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this version was created"
+    )
+    reflect_summary: str | None = Field(default=None, description="Summary of changes in this version")
+
+
+def verify_evidence_quotes(
+    observation: Observation,
+    memories: dict[str, str],
+) -> tuple[bool, list[str]]:
+    """Verify that all evidence quotes exist in the referenced memories.
+
+    Args:
+        observation: The observation to verify
+        memories: Dict mapping memory_id to memory content
+
+    Returns:
+        Tuple of (is_valid, list of error messages)
+    """
+    errors = []
+
+    for evidence in observation.evidence:
+        memory_content = memories.get(evidence.memory_id)
+        if memory_content is None:
+            errors.append(f"Memory {evidence.memory_id} not found")
+            continue
+
+        if evidence.quote not in memory_content:
+            errors.append(f"Quote not found in memory {evidence.memory_id}: '{evidence.quote[:50]}...'")
+
+    return len(errors) == 0, errors
@@ -0,0 +1,762 @@
+"""
+System prompts for the reflect agent.
+"""
+
+import json
+from typing import Any
+
+
+def _extract_directive_rules(directives: list[dict[str, Any]]) -> list[str]:
+    """
+    Extract directive rules as a list of strings.
+
+    Args:
+        directives: List of directive mental models with observations
+
+    Returns:
+        List of directive rule strings
+    """
+    rules = []
+    for directive in directives:
+        directive_name = directive.get("name", "")
+        observations = directive.get("observations", [])
+        if observations:
+            for obs in observations:
+                # Support both Pydantic Observation objects and dicts
+                if hasattr(obs, "title"):
+                    title = obs.title
+                    content = obs.content
+                else:
+                    title = obs.get("title", "")
+                    content = obs.get("content", "")
+                if title and content:
+                    rules.append(f"**{title}**: {content}")
+                elif content:
+                    rules.append(content)
+        elif directive_name:
+            # Fallback to description if no observations
+            desc = directive.get("description", "")
+            if desc:
+                rules.append(f"**{directive_name}**: {desc}")
+    return rules
+
+
+def build_directives_section(directives: list[dict[str, Any]]) -> str:
+    """
+    Build the directives section for the system prompt.
+
+    Directives are hard rules that MUST be followed in all responses.
+
+    Args:
+        directives: List of directive mental models with observations
+    """
+    if not directives:
+        return ""
+
+    rules = _extract_directive_rules(directives)
+    if not rules:
+        return ""
+
+    parts = [
+        "## DIRECTIVES (MANDATORY)",
+        "These are hard rules you MUST follow in ALL responses:",
+        "",
+    ]
+
+    for rule in rules:
+        parts.append(f"- {rule}")
+
+    parts.extend(
+        [
+            "",
+            "NEVER violate these directives, even if other context suggests otherwise.",
+            "IMPORTANT: Do NOT explain or justify how you handled directives in your answer. Just follow them silently.",
+            "",
+        ]
+    )
+    return "\n".join(parts)
+
+
+def build_directives_reminder(directives: list[dict[str, Any]]) -> str:
+    """
+    Build a reminder section for directives to place at the end of the prompt.
+
+    Args:
+        directives: List of directive mental models with observations
+    """
+    if not directives:
+        return ""
+
+    rules = _extract_directive_rules(directives)
+    if not rules:
+        return ""
+
+    parts = [
+        "",
+        "## REMINDER: MANDATORY DIRECTIVES",
+        "Before responding, ensure your answer complies with ALL of these directives:",
+        "",
+    ]
+
+    for i, rule in enumerate(rules, 1):
+        parts.append(f"{i}. {rule}")
+
+    parts.append("")
+    parts.append("Your response will be REJECTED if it violates any directive above.")
+    parts.append("Do NOT include any commentary about how you handled directives - just follow them.")
+    return "\n".join(parts)
+
+
+def build_system_prompt_for_tools(
+    bank_profile: dict[str, Any],
+    context: str | None = None,
+    directives: list[dict[str, Any]] | None = None,
+) -> str:
+    """
+    Build the system prompt for tool-calling reflect agent.
+
+    This is a simplified prompt since tools are defined separately via the tools parameter.
+
+    Args:
+        bank_profile: Bank profile with name and mission
+        context: Optional additional context
+        directives: Optional list of directive mental models to inject as hard rules
+    """
+    name = bank_profile.get("name", "Assistant")
+    mission = bank_profile.get("mission", "")
+
+    no_info_rule = (
+        "- Only say 'I don't have information' AFTER trying list_mental_models AND recall with no relevant results"
+    )
+
+    parts = []
+
+    # Inject directives at the VERY START for maximum prominence
+    if directives:
+        parts.append(build_directives_section(directives))
+
+    parts.extend(
+        [
+            "You are a reflection agent that answers questions by reasoning over retrieved memories.",
+            "",
+        ]
+    )
+
+    parts.extend(
+        [
+            "## CRITICAL RULES",
+            "- You must NEVER fabricate information that has no basis in retrieved data",
+            "- You SHOULD synthesize, infer, and reason from the retrieved memories",
+            "- You MUST call recall() before saying you don't have information",
+            no_info_rule,
+            "",
+            "## How to Reason",
+            "- If memories mention someone did an activity, you can infer they likely enjoyed it",
+            "- Synthesize a coherent narrative from related memories",
+            "- Be a thoughtful interpreter, not just a literal repeater",
+            "- When the exact answer isn't stated, use what IS stated to give the best answer",
+            "",
+            "## Query Strategy (IMPORTANT)",
+            "recall() uses semantic search. NEVER just echo the user's question - decompose it into targeted searches:",
+            "",
+            "BAD: User asks 'recurring lesson themes between students' → recall('recurring lesson themes between students')",
+            "GOOD: Break it down into component searches:",
+            "  1. recall('lessons') - find all lesson-related memories",
+            "  2. recall('teaching sessions') - alternative phrasing",
+            "  3. recall('student progress') - find student-related memories",
+            "  4. recall('topics taught') - find subject matter",
+            "",
+            "Think: What ENTITIES and CONCEPTS does this question involve? Search for each separately.",
+            "- Questions about patterns → search for the individual instances first",
+            "- Questions comparing things → search for each thing separately",
+            "- Questions about relationships → search for each party involved",
+            "",
+            "## Workflow",
+        ]
+    )
+
+    # Answer mode: include mental model lookup in workflow
+    parts.extend(
+        [
+            "1. Review the pre-fetched mental models for relevant synthesized knowledge",
+            "2. If relevant, call get_mental_model(model_id) for full observations",
+            "3. DECOMPOSE the question into component searches (see Query Strategy above)",
+            "   - Identify entities and concepts in the question",
+            "   - Search for each separately with targeted queries",
+            "4. Run multiple recall() calls - don't just echo the user's question",
+            "5. Use expand() if you need more context on specific memories",
+            "6. BEFORE answering: Check if any person/project/concept from the memories deserves a mental model - use learn() if so",
+            "7. When ready, call done() with your answer and supporting memory_ids",
+            "",
+            "## When to Use learn() - IMPORTANT",
+            "ACTIVELY look for opportunities to use learn() when you discover:",
+            "- A person mentioned in 2+ memories who has no mental model yet",
+            "- A project or concept the user asks about that has no mental model",
+            "- A pattern or topic worth tracking for future questions",
+            "",
+            "DO NOT wait to be asked - proactively create models when you see the need.",
+            "Example: learn(name='Project Alpha', description='Track goals, status, and key decisions for Project Alpha')",
+            "",
+            "## Output Format: Plain Text Answer",
+            "Call done() with a plain text 'answer' field.",
+            "- Do NOT use markdown formatting",
+            "- NEVER include memory IDs, UUIDs, or 'Memory references' in the answer text",
+            "- Put memory IDs ONLY in the memory_ids array parameter, not in the answer",
+        ]
+    )
+
+    parts.append("")
+    parts.append(f"## Memory Bank: {name}")
+
+    if mission:
+        parts.append(f"Mission: {mission}")
+
+    # Disposition traits
+    disposition = bank_profile.get("disposition", {})
+    if disposition:
+        traits = []
+        if "skepticism" in disposition:
+            traits.append(f"skepticism={disposition['skepticism']}")
+        if "literalism" in disposition:
+            traits.append(f"literalism={disposition['literalism']}")
+        if "empathy" in disposition:
+            traits.append(f"empathy={disposition['empathy']}")
+        if traits:
+            parts.append(f"Disposition: {', '.join(traits)}")
+
+    if context:
+        parts.append(f"\n## Additional Context\n{context}")
+
+    # Add directive reminder at the END for recency effect
+    if directives:
+        parts.append(build_directives_reminder(directives))
+
+    return "\n".join(parts)
+
+
+def build_agent_prompt(
+    query: str,
+    context_history: list[dict],
+    bank_profile: dict,
+    additional_context: str | None = None,
+) -> str:
+    """Build the user prompt for the reflect agent."""
+    parts = []
+
+    # Bank identity
+    name = bank_profile.get("name", "Assistant")
+    mission = bank_profile.get("mission", "")
+
+    parts.append(f"## Memory Bank Context\nName: {name}")
+    if mission:
+        parts.append(f"Mission: {mission}")
+
+    # Disposition traits if present
+    disposition = bank_profile.get("disposition", {})
+    if disposition:
+        traits = []
+        if "skepticism" in disposition:
+            traits.append(f"skepticism={disposition['skepticism']}")
+        if "literalism" in disposition:
+            traits.append(f"literalism={disposition['literalism']}")
+        if "empathy" in disposition:
+            traits.append(f"empathy={disposition['empathy']}")
+        if traits:
+            parts.append(f"Disposition: {', '.join(traits)}")
+
+    # Additional context from caller
+    if additional_context:
+        parts.append(f"\n## Additional Context\n{additional_context}")
+
+    # Tool call history
+    if context_history:
+        parts.append("\n## Tool Results (synthesize and reason from this data)")
+        for i, entry in enumerate(context_history, 1):
+            tool = entry["tool"]
+            output = entry["output"]
+            # Format as proper JSON for LLM readability
+            try:
+                output_str = json.dumps(output, indent=2, default=str)
+            except (TypeError, ValueError):
+                output_str = str(output)
+            parts.append(f"\n### Call {i}: {tool}\n```json\n{output_str}\n```")
+
+    # The question
+    parts.append(f"\n## Question\n{query}")
+
+    # Instructions
+    if context_history:
+        parts.append(
+            "\n## Instructions\n"
+            "Based on the tool results above, either call more tools or provide your final answer. "
+            "Synthesize and reason from the data - make reasonable inferences when helpful. "
+            "If you have related information, use it to give the best possible answer."
+        )
+    else:
+        parts.append(
+            "\n## Instructions\n"
+            "Start by calling list_mental_models() to see available mental models - they contain pre-synthesized knowledge. "
+            "If a relevant model exists, use get_mental_model(model_id) to get its observations. "
+            "Then use recall(query) for specific details not covered by mental models."
+        )
+
+    return "\n".join(parts)
+
+
+def build_final_prompt(
+    query: str,
+    context_history: list[dict],
+    bank_profile: dict,
+    additional_context: str | None = None,
+) -> str:
+    """Build the final prompt when forcing a text response (no tools)."""
+    parts = []
+
+    # Bank identity
+    name = bank_profile.get("name", "Assistant")
+    mission = bank_profile.get("mission", "")
+
+    parts.append(f"## Memory Bank Context\nName: {name}")
+    if mission:
+        parts.append(f"Mission: {mission}")
+
+    # Disposition traits if present
+    disposition = bank_profile.get("disposition", {})
+    if disposition:
+        traits = []
+        if "skepticism" in disposition:
+            traits.append(f"skepticism={disposition['skepticism']}")
+        if "literalism" in disposition:
+            traits.append(f"literalism={disposition['literalism']}")
+        if "empathy" in disposition:
+            traits.append(f"empathy={disposition['empathy']}")
+        if traits:
+            parts.append(f"Disposition: {', '.join(traits)}")
+
+    # Additional context from caller
+    if additional_context:
+        parts.append(f"\n## Additional Context\n{additional_context}")
+
+    # Tool call history
+    if context_history:
+        parts.append("\n## Retrieved Data (synthesize and reason from this data)")
+        for entry in context_history:
+            tool = entry["tool"]
+            output = entry["output"]
+            # Format as proper JSON for LLM readability
+            try:
+                output_str = json.dumps(output, indent=2, default=str)
+            except (TypeError, ValueError):
+                output_str = str(output)
+            parts.append(f"\n### From {tool}:\n```json\n{output_str}\n```")
+    else:
+        parts.append("\n## Retrieved Data\nNo data was retrieved.")
+
+    # The question
+    parts.append(f"\n## Question\n{query}")
+
+    # Final instructions
+    parts.append(
+        "\n## Instructions\n"
+        "Provide a thoughtful answer by synthesizing and reasoning from the retrieved data above. "
+        "You can make reasonable inferences from the memories, but don't completely fabricate information."
+        "If the exact answer isn't stated, use what IS stated to give the best possible answer. "
+        "Only say 'I don't have information' if the retrieved data is truly unrelated to the question."
+    )
+
+    return "\n".join(parts)
+
+
+FINAL_SYSTEM_PROMPT = """You are a thoughtful assistant that synthesizes answers from retrieved memories.
+
+Your approach:
+- Reason over the retrieved memories to answer the question
+- Make reasonable inferences when the exact answer isn't explicitly stated
+- Connect related memories to form a complete picture
+- Be helpful - if you have related information, use it to give the best possible answer
+
+Only say "I don't have information" if the retrieved data is truly unrelated to the question.
+Do NOT fabricate information that has no basis in the retrieved data."""
+
+
+# =============================================================================
+# 4-Phase Mental Model Reflect Prompts
+# =============================================================================
+
+SEED_PHASE_SYSTEM_PROMPT = """You are analyzing memories to discover NEW patterns and generate candidate observations.
+
+Your task is to identify potential observations (beliefs, preferences, patterns, behaviors) that could be part of a mental model about this person/topic.
+
+## Important: Avoid Redundancy
+If existing observations are provided, DO NOT generate candidates that are essentially the same.
+Focus on discovering NEW patterns not already covered by existing observations.
+
+## Rules
+- Generate 5-15 candidate observations for NEW patterns only
+- Each candidate should be specific and testable (can be supported or contradicted by evidence)
+- Note which memory IDs inspired each candidate (these are seeds, not final evidence)
+- Focus on patterns that appear MULTIPLE TIMES across many memories - the more the better
+- The best candidates are ones you can find 10, 20, or even 50+ supporting memories for
+- Skip patterns that are already covered by existing observations
+
+## Output Format
+Return a JSON array of candidate observations:
+```json
+{
+  "candidates": [
+    {
+      "content": "The specific observation/belief/pattern - be detailed and specific",
+      "seed_memory_ids": ["memory_id_1", "memory_id_2", "memory_id_3"]
+    }
+  ]
+}
+```
+
+Focus on patterns that appear multiple times or have strong signals. Don't generate obvious or trivial observations.
+Prefer candidates with MORE seed memories - they're more likely to be real patterns.
+Return an empty candidates array if no genuinely new patterns are found."""
+
+
+def build_seed_phase_prompt(
+    memories: list[dict],
+    topic: str | None = None,
+    existing_observations: list[dict] | None = None,
+) -> str:
+    """Build the user prompt for the seed phase.
+
+    Args:
+        memories: List of memories to analyze
+        topic: Optional topic focus for the mental model
+        existing_observations: Optional list of existing observations to avoid rediscovering
+    """
+    parts = []
+
+    if topic:
+        parts.append(f"## Topic Focus\n{topic}\n")
+
+    # Include existing observations so we don't rediscover them
+    if existing_observations:
+        parts.append("## Existing Observations (DO NOT regenerate these)")
+        parts.append("These patterns are already tracked. Focus on discovering NEW patterns:\n")
+        for i, obs in enumerate(existing_observations, 1):
+            title = obs.get("title", "")
+            content = obs.get("content", "")
+            parts.append(f"{i}. **{title}**: {content}\n")
+        parts.append("")
+
+    parts.append("## Memories to Analyze")
+    parts.append("Review these memories and identify patterns, preferences, beliefs, and behaviors:\n")
+
+    for mem in memories:
+        mem_id = mem.get("id", "unknown")
+        content = mem.get("content", mem.get("text", ""))
+        timestamp = mem.get("timestamp", mem.get("created_at", ""))
+        parts.append(f"[{mem_id}] ({timestamp}): {content}\n")
+
+    parts.append("\n## Instructions")
+    if existing_observations:
+        parts.append("Generate candidate observations for NEW patterns not already covered above.")
+        parts.append("If all patterns are already covered by existing observations, return an empty candidates array.")
+    else:
+        parts.append("Generate candidate observations based on patterns you see in these memories.")
+    parts.append("Look for: recurring themes, stated preferences, behavioral patterns, beliefs, values, goals.")
+
+    return "\n".join(parts)
+
+
+VALIDATE_PHASE_SYSTEM_PROMPT = """You are validating candidate observations against evidence.
+
+For each candidate, you have:
+- Supporting memories (evidence FOR the observation)
+- Contradicting memories (evidence AGAINST the observation)
+
+## Your Task
+1. Evaluate each candidate based on the evidence
+2. For valid candidates, extract EXACT QUOTES from supporting memories
+3. Discard candidates with insufficient or contradicting evidence
+4. Merge similar candidates into single, refined observations
+
+## Rules for Quotes
+- Quotes must be EXACT text from the memory, not paraphrased
+- Each quote should directly support the observation
+- The MORE evidence quotes, the BETTER - don't limit yourself, include ALL relevant quotes (10, 20, 50+)
+- Observations with only 1-2 quotes are weak and should be discarded unless the evidence is exceptionally strong
+- Stronger observations have more supporting evidence - aim for comprehensive coverage
+
+## Output Format
+Return validated observations with evidence:
+```json
+{
+  "observations": [
+    {
+      "title": "Short descriptive title (3-8 words) - like a headline",
+      "content": "The full observation content - detailed explanation of the pattern/belief",
+      "evidence": [
+        {
+          "memory_id": "exact_memory_id",
+          "quote": "Exact quote from the memory text",
+          "relevance": "Brief explanation of how this supports the observation",
+          "timestamp": "2024-01-15T10:00:00Z"
+        }
+      ]
+    }
+  ],
+  "discarded": [
+    {
+      "content": "The discarded candidate",
+      "reason": "Why it was discarded (insufficient evidence, contradicted, etc.)"
+    }
+  ],
+  "merged": [
+    {
+      "from": ["candidate 1 content", "candidate 2 content"],
+      "into": "The merged observation content"
+    }
+  ]
+}
+```
+
+## Title Guidelines
+- Title should be a SHORT label (like "Prefers morning meetings" or "Coffee enthusiast")
+- NOT a truncated version of the content
+- Think of it as a category/tag for the observation
+
+Be rigorous: only keep observations with clear, verifiable evidence from multiple memories."""
+
+
+def build_validate_phase_prompt(candidates_with_evidence: list[dict]) -> str:
+    """Build the user prompt for the validate phase."""
+    parts = ["## Candidates to Validate\n"]
+
+    for i, item in enumerate(candidates_with_evidence, 1):
+        candidate = item.get("candidate", {})
+        supporting = item.get("supporting_memories", [])
+        contradicting = item.get("contradicting_memories", [])
+
+        parts.append(f"### Candidate {i}: {candidate.get('content', '')}")
+
+        if supporting:
+            parts.append("\n**Supporting Evidence:**")
+            for mem in supporting:
+                mem_id = mem.get("id", "unknown")
+                content = mem.get("content", mem.get("text", ""))
+                timestamp = mem.get("timestamp", mem.get("created_at", ""))
+                parts.append(f"- [{mem_id}] ({timestamp}): {content}")
+
+        if contradicting:
+            parts.append("\n**Contradicting Evidence:**")
+            for mem in contradicting:
+                mem_id = mem.get("id", "unknown")
+                content = mem.get("content", mem.get("text", ""))
+                timestamp = mem.get("timestamp", mem.get("created_at", ""))
+                parts.append(f"- [{mem_id}] ({timestamp}): {content}")
+
+        if not supporting and not contradicting:
+            parts.append("\n*No additional evidence found*")
+
+        parts.append("")
+
+    parts.append("## Instructions")
+    parts.append("1. Evaluate each candidate based on its evidence")
+    parts.append("2. Keep candidates with strong supporting evidence")
+    parts.append("3. Discard candidates with no evidence or strong contradictions")
+    parts.append("4. Merge similar candidates")
+    parts.append("5. Extract EXACT quotes (copy-paste from memory text) for evidence")
+
+    return "\n".join(parts)
+
+
+COMPARE_PHASE_SYSTEM_PROMPT = """You are merging new observations with an existing mental model.
+
+You have:
+- EXISTING observations (from the current mental model)
+- NEW observations (from this reflect cycle)
+
+## Your Task
+Produce the final, complete mental model by:
+1. Keeping existing observations that are still valid
+2. Updating existing observations with new evidence (ADD new evidence to existing)
+3. Adding new observations that don't overlap with existing
+4. Removing existing observations that are contradicted by new evidence
+5. Merging overlapping observations
+
+## Rules
+- The final model should have no contradictions
+- Each observation must have evidence with exact quotes
+- COMBINE evidence from both existing and new observations
+- If an existing observation has new supporting evidence, ADD ALL the new evidence to it
+- Include ALL relevant evidence - the more quotes the better (10, 20, 50+ is great)
+- Observations with more evidence are more reliable - don't limit the number of quotes
+
+## Output Format
+Return the complete, final mental model:
+```json
+{
+  "observations": [
+    {
+      "title": "Short descriptive title (3-8 words)",
+      "content": "Full observation content - detailed explanation",
+      "evidence": [
+        {
+          "memory_id": "id",
+          "quote": "exact quote",
+          "relevance": "explanation",
+          "timestamp": "ISO timestamp"
+        }
+      ],
+      "created_at": "ISO timestamp of when observation was first created"
+    }
+  ],
+  "changes": {
+    "kept": ["Observation that was kept unchanged"],
+    "updated": [{"from": "old content", "to": "new content", "reason": "why"}],
+    "added": ["New observation that was added"],
+    "removed": [{"content": "removed observation", "reason": "why removed"}],
+    "merged": [{"from": ["obs1", "obs2"], "into": "merged observation"}]
+  }
+}
+```"""
+
+
+def build_compare_phase_prompt(
+    existing_observations: list[dict],
+    new_observations: list[dict],
+) -> str:
+    """Build the user prompt for the compare phase."""
+    parts = []
+
+    parts.append("## Existing Mental Model Observations")
+    if existing_observations:
+        for i, obs in enumerate(existing_observations, 1):
+            title = obs.get("title", "")
+            content = obs.get("content", obs.get("text", ""))
+            evidence = obs.get("evidence", [])
+            parts.append(f"\n### Existing {i}: {title}")
+            parts.append(f"Content: {content}")
+            if evidence:
+                parts.append(f"Evidence ({len(evidence)} items):")
+                for ev in evidence[:5]:  # Show max 5 evidence items
+                    parts.append(f'  - [{ev.get("memory_id", "?")}]: "{ev.get("quote", "")}"')
+                if len(evidence) > 5:
+                    parts.append(f"  ... and {len(evidence) - 5} more")
+    else:
+        parts.append("*No existing observations*")
+
+    parts.append("\n## New Observations from This Reflect")
+    if new_observations:
+        for i, obs in enumerate(new_observations, 1):
+            title = obs.get("title", "")
+            content = obs.get("content", "")
+            evidence = obs.get("evidence", [])
+            parts.append(f"\n### New {i}: {title}")
+            parts.append(f"Content: {content}")
+            if evidence:
+                parts.append(f"Evidence ({len(evidence)} items):")
+                for ev in evidence:
+                    parts.append(f'  - [{ev.get("memory_id", "?")}]: "{ev.get("quote", "")}"')
+    else:
+        parts.append("*No new observations*")
+
+    parts.append("\n## Instructions")
+    parts.append("Merge these into a coherent, non-contradictory mental model.")
+    parts.append("Preserve all valid evidence. Remove stale or contradicted observations.")
+
+    return "\n".join(parts)
+
+
+# =============================================================================
+# UPDATE EXISTING Phase Prompts (for diff-based refresh)
+# =============================================================================
+
+UPDATE_EXISTING_SYSTEM_PROMPT = """You are updating existing observations with newly found evidence.
+
+For each existing observation, you have been given:
+- The original observation (title, content, existing evidence)
+- Newly found supporting memories
+- Newly found contradicting memories
+
+## Your Task
+1. Extract EXACT QUOTES from new supporting memories to add to the observation
+2. Flag observations with strong contradicting evidence for potential removal
+3. Keep existing evidence intact - only ADD new evidence
+
+## Rules for Quotes
+- Quotes must be EXACT text from the memory, not paraphrased
+- Each quote should directly support the observation
+- Include ALL relevant quotes from the new memories
+
+## Output Format
+Return updated observations with new evidence:
+```json
+{
+  "updated_observations": [
+    {
+      "title": "Original title",
+      "content": "Original content",
+      "existing_evidence_count": 5,
+      "new_evidence": [
+        {
+          "memory_id": "exact_memory_id",
+          "quote": "Exact quote from the memory text",
+          "relevance": "Brief explanation of how this supports the observation",
+          "timestamp": "2024-01-15T10:00:00Z"
+        }
+      ],
+      "has_contradiction": false,
+      "contradiction_note": null
+    }
+  ]
+}
+```
+
+If an observation has strong contradicting evidence, set has_contradiction=true and explain in contradiction_note."""
+
+
+def build_update_existing_prompt(observations_with_evidence: list[dict]) -> str:
+    """Build the user prompt for the update existing phase.
+
+    Args:
+        observations_with_evidence: List of existing observations with new evidence found
+    """
+    parts = ["## Existing Observations to Update\n"]
+
+    for i, item in enumerate(observations_with_evidence, 1):
+        obs = item.get("observation", {})
+        supporting = item.get("supporting_memories", [])
+        contradicting = item.get("contradicting_memories", [])
+
+        title = obs.get("title", "")
+        content = obs.get("content", "")
+        existing_evidence = obs.get("evidence", [])
+
+        parts.append(f"### Observation {i}: {title}")
+        parts.append(f"Content: {content}")
+        parts.append(f"Existing evidence count: {len(existing_evidence)}")
+
+        if supporting:
+            parts.append("\n**New Supporting Memories:**")
+            for mem in supporting:
+                mem_id = mem.get("id", "unknown")
+                mem_content = mem.get("content", mem.get("text", ""))
+                timestamp = mem.get("timestamp", mem.get("created_at", ""))
+                parts.append(f"- [{mem_id}] ({timestamp}): {mem_content}")
+
+        if contradicting:
+            parts.append("\n**New Contradicting Memories:**")
+            for mem in contradicting:
+                mem_id = mem.get("id", "unknown")
+                mem_content = mem.get("content", mem.get("text", ""))
+                timestamp = mem.get("timestamp", mem.get("created_at", ""))
+                parts.append(f"- [{mem_id}] ({timestamp}): {mem_content}")
+
+        if not supporting and not contradicting:
+            parts.append("\n*No new evidence found*")
+
+        parts.append("")
+
+    parts.append("## Instructions")
+    parts.append("1. Extract EXACT quotes from new supporting memories")
+    parts.append("2. Flag observations with strong contradictions")
+    parts.append("3. Return the updated observations with new evidence added")
+
+    return "\n".join(parts)
@@ -0,0 +1,450 @@
+"""
+Tool implementations for the reflect agent.
+"""
+
+import logging
+import re
+import uuid
+from datetime import datetime, timezone
+from typing import TYPE_CHECKING, Any
+
+from .models import MentalModelInput
+from .observations import Observation, ObservationEvidence, Trend
+
+if TYPE_CHECKING:
+    from asyncpg import Connection
+
+    from ...api.http import RequestContext
+    from ..memory_engine import MemoryEngine
+
+logger = logging.getLogger(__name__)
+
+
+def generate_model_id(name: str) -> str:
+    """Generate a stable ID from mental model name."""
+    # Normalize: lowercase, replace spaces/special chars with hyphens
+    normalized = re.sub(r"[^a-z0-9]+", "-", name.lower()).strip("-")
+    # Truncate to reasonable length
+    return normalized[:50]
+
+
+def _parse_observations(observations_raw: list) -> list[Observation]:
+    """Parse raw observation dicts into typed Observation models."""
+    observations: list[Observation] = []
+    for obs in observations_raw:
+        if not isinstance(obs, dict):
+            continue
+
+        try:
+            parsed = Observation(
+                title=obs.get("title", ""),
+                content=obs.get("content", ""),
+                evidence=[
+                    ObservationEvidence(
+                        memory_id=ev.get("memory_id", ""),
+                        quote=ev.get("quote", ""),
+                        relevance=ev.get("relevance", ""),
+                        timestamp=ev.get("timestamp"),
+                    )
+                    for ev in obs.get("evidence", [])
+                    if isinstance(ev, dict)
+                ],
+                created_at=obs.get("created_at"),
+            )
+            observations.append(parsed)
+        except Exception as e:
+            logger.warning(f"Failed to parse observation: {e}")
+            continue
+
+    return observations
+
+
+async def tool_lookup(
+    conn: "Connection",
+    bank_id: str,
+    model_id: str | None = None,
+    tags: list[str] | None = None,
+    tags_match: str = "any",
+) -> dict[str, Any]:
+    """
+    List or get mental models.
+
+    Args:
+        conn: Database connection
+        bank_id: Bank identifier
+        model_id: Optional specific model ID to get (if None, lists all)
+        tags: Optional tags to filter models (when listing)
+        tags_match: How to match tags - "any" (OR), "all" (AND)
+
+    Returns:
+        Dict with either a list of models or a single model's details
+    """
+    if model_id:
+        # Get specific mental model with full details including observations
+        row = await conn.fetchrow(
+            """
+            SELECT id, subtype, name, description, observations, entity_id, last_updated
+            FROM mental_models
+            WHERE id = $1 AND bank_id = $2
+            """,
+            model_id,
+            bank_id,
+        )
+        if row:
+            # Parse observations JSON
+            obs_data = row["observations"] or {"observations": []}
+            if isinstance(obs_data, str):
+                import json
+
+                obs_data = json.loads(obs_data)
+            observations_raw = obs_data.get("observations", []) if isinstance(obs_data, dict) else obs_data
+
+            # Parse observations into typed models
+            observations = _parse_observations(observations_raw)
+
+            return {
+                "found": True,
+                "model": {
+                    "id": row["id"],
+                    "subtype": row["subtype"],
+                    "name": row["name"],
+                    "description": row["description"],
+                    "observations": observations,
+                    "entity_id": str(row["entity_id"]) if row["entity_id"] else None,
+                    "last_updated": row["last_updated"].isoformat() if row["last_updated"] else None,
+                },
+            }
+        return {"found": False, "model_id": model_id}
+    else:
+        # List mental models (compact: id, name, description only)
+        # Full observations are retrieved via get_mental_model(model_id)
+        # NOTE: Directives (subtype='directive') are excluded from listing -
+        # they are injected into the system prompt, not discoverable via tools
+        # Filter by tags if provided
+        if tags:
+            if tags_match == "all":
+                # All tags must match
+                rows = await conn.fetch(
+                    """
+                    SELECT id, subtype, name, description
+                    FROM mental_models
+                    WHERE bank_id = $1 AND tags @> $2::varchar[] AND subtype != 'directive'
+                    ORDER BY last_updated DESC NULLS LAST, created_at DESC
+                    """,
+                    bank_id,
+                    tags,
+                )
+            else:
+                # Any tag matches (OR) - default
+                rows = await conn.fetch(
+                    """
+                    SELECT id, subtype, name, description
+                    FROM mental_models
+                    WHERE bank_id = $1 AND tags && $2::varchar[] AND subtype != 'directive'
+                    ORDER BY last_updated DESC NULLS LAST, created_at DESC
+                    """,
+                    bank_id,
+                    tags,
+                )
+        else:
+            rows = await conn.fetch(
+                """
+                SELECT id, subtype, name, description
+                FROM mental_models
+                WHERE bank_id = $1 AND subtype != 'directive'
+                ORDER BY last_updated DESC NULLS LAST, created_at DESC
+                """,
+                bank_id,
+            )
+
+        return {
+            "count": len(rows),
+            "models": [
+                {
+                    "id": row["id"],
+                    "subtype": row["subtype"],
+                    "name": row["name"],
+                    "description": row["description"],
+                }
+                for row in rows
+            ],
+        }
+
+
+async def tool_recall(
+    memory_engine: "MemoryEngine",
+    bank_id: str,
+    query: str,
+    request_context: "RequestContext",
+    max_tokens: int = 2048,
+    max_results: int = 50,
+    tags: list[str] | None = None,
+    tags_match: str = "any",
+    connection_budget: int = 1,
+) -> dict[str, Any]:
+    """
+    Search memories using TEMPR retrieval.
+
+    Args:
+        memory_engine: Memory engine instance
+        bank_id: Bank identifier
+        query: Search query
+        request_context: Request context for authentication
+        max_tokens: Maximum tokens for results (default 2048)
+        max_results: Maximum number of results
+        tags: Filter by tags (includes untagged memories)
+        tags_match: How to match tags - "any" (OR), "all" (AND), or "exact"
+        connection_budget: Max DB connections for this recall (default 1 for internal ops)
+
+    Returns:
+        Dict with list of matching memories
+    """
+    result = await memory_engine.recall_async(
+        bank_id=bank_id,
+        query=query,
+        fact_type=["experience", "world"],  # Exclude opinions
+        max_tokens=max_tokens,
+        enable_trace=False,
+        request_context=request_context,
+        tags=tags,
+        tags_match=tags_match,
+        _connection_budget=connection_budget,
+    )
+
+    memories = []
+    for m in result.results[:max_results]:
+        memories.append(
+            {
+                "id": str(m.id),
+                "text": m.text,
+                "type": m.fact_type,
+                "entities": m.entities or [],
+                "occurred": m.occurred_start,  # Already ISO format string
+            }
+        )
+
+    return {
+        "query": query,
+        "count": len(memories),
+        "memories": memories,
+    }
+
+
+async def tool_learn(
+    conn: "Connection",
+    bank_id: str,
+    input: MentalModelInput,
+    tags: list[str] | None = None,
+) -> dict[str, Any]:
+    """
+    Create a mental model placeholder with subtype='learned'.
+
+    The agent only specifies name and description - actual observations are generated
+    in the background via refresh, similar to pinned models.
+
+    Args:
+        conn: Database connection
+        bank_id: Bank identifier
+        input: Mental model input data (name, description, optional entity_id)
+        tags: Tags to apply to new mental models (from reflect context)
+
+    Returns:
+        Dict with created model info including model_id for background generation
+    """
+    model_id = generate_model_id(input.name)
+
+    # Parse entity_id if provided
+    entity_uuid = None
+    if input.entity_id:
+        try:
+            entity_uuid = uuid.UUID(input.entity_id)
+        except ValueError:
+            logger.warning(f"Invalid entity_id format: {input.entity_id}")
+
+    # Check if model exists
+    existing = await conn.fetchrow(
+        "SELECT id FROM mental_models WHERE id = $1 AND bank_id = $2",
+        model_id,
+        bank_id,
+    )
+
+    if existing:
+        # Update description only - observations will be regenerated
+        await conn.execute(
+            """
+            UPDATE mental_models SET
+                description = $3,
+                entity_id = $4
+            WHERE id = $1 AND bank_id = $2
+            """,
+            model_id,
+            bank_id,
+            input.description,
+            entity_uuid,
+        )
+        status = "updated"
+    else:
+        # Insert new model placeholder - observations will be generated in background
+        await conn.execute(
+            """
+            INSERT INTO mental_models (id, bank_id, subtype, name, description, observations, entity_id, tags, created_at)
+            VALUES ($1, $2, 'learned', $3, $4, '{}'::jsonb, $5, $6, NOW())
+            """,
+            model_id,
+            bank_id,
+            input.name,
+            input.description,
+            entity_uuid,
+            tags or [],
+        )
+        status = "created"
+
+    logger.info(f"[REFLECT] Mental model '{model_id}' {status} in bank {bank_id} - pending background generation")
+
+    return {
+        "status": status,
+        "model_id": model_id,
+        "name": input.name,
+        "pending_generation": True,
+    }
+
+
+async def tool_expand(
+    conn: "Connection",
+    bank_id: str,
+    memory_ids: list[str],
+    depth: str,
+) -> dict[str, Any]:
+    """
+    Expand multiple memories to get chunk or document context.
+
+    Args:
+        conn: Database connection
+        bank_id: Bank identifier
+        memory_ids: List of memory unit IDs
+        depth: "chunk" or "document"
+
+    Returns:
+        Dict with results array, each containing memory, chunk, and optionally document data
+    """
+    if not memory_ids:
+        return {"error": "memory_ids is required and must not be empty"}
+
+    # Validate and convert UUIDs
+    valid_uuids: list[uuid.UUID] = []
+    errors: dict[str, str] = {}
+    for mid in memory_ids:
+        try:
+            valid_uuids.append(uuid.UUID(mid))
+        except ValueError:
+            errors[mid] = f"Invalid memory_id format: {mid}"
+
+    if not valid_uuids:
+        return {"error": "No valid memory IDs provided", "details": errors}
+
+    # Batch fetch all memory units
+    memories = await conn.fetch(
+        """
+        SELECT id, text, chunk_id, document_id, fact_type, context
+        FROM memory_units
+        WHERE id = ANY($1) AND bank_id = $2
+        """,
+        valid_uuids,
+        bank_id,
+    )
+    memory_map = {row["id"]: row for row in memories}
+
+    # Collect chunk_ids and document_ids for batch fetching
+    chunk_ids = [m["chunk_id"] for m in memories if m["chunk_id"]]
+    doc_ids_from_chunks: set[str] = set()
+    doc_ids_direct: set[str] = set()
+
+    # Batch fetch all chunks
+    chunk_map: dict[str, Any] = {}
+    if chunk_ids:
+        chunks = await conn.fetch(
+            """
+            SELECT chunk_id, chunk_text, chunk_index, document_id
+            FROM chunks
+            WHERE chunk_id = ANY($1)
+            """,
+            chunk_ids,
+        )
+        chunk_map = {row["chunk_id"]: row for row in chunks}
+        if depth == "document":
+            doc_ids_from_chunks = {c["document_id"] for c in chunks if c["document_id"]}
+
+    # Collect direct document IDs (memories without chunks)
+    if depth == "document":
+        for m in memories:
+            if not m["chunk_id"] and m["document_id"]:
+                doc_ids_direct.add(m["document_id"])
+
+    # Batch fetch all documents
+    doc_map: dict[str, Any] = {}
+    all_doc_ids = list(doc_ids_from_chunks | doc_ids_direct)
+    if all_doc_ids:
+        docs = await conn.fetch(
+            """
+            SELECT id, original_text, metadata, retain_params
+            FROM documents
+            WHERE id = ANY($1) AND bank_id = $2
+            """,
+            all_doc_ids,
+            bank_id,
+        )
+        doc_map = {row["id"]: row for row in docs}
+
+    # Build results
+    results: list[dict[str, Any]] = []
+    for mid, mem_uuid in zip(memory_ids, valid_uuids):
+        if mid in errors:
+            results.append({"memory_id": mid, "error": errors[mid]})
+            continue
+
+        memory = memory_map.get(mem_uuid)
+        if not memory:
+            results.append({"memory_id": mid, "error": f"Memory not found: {mid}"})
+            continue
+
+        item: dict[str, Any] = {
+            "memory_id": mid,
+            "memory": {
+                "id": str(memory["id"]),
+                "text": memory["text"],
+                "type": memory["fact_type"],
+                "context": memory["context"],
+            },
+        }
+
+        # Add chunk if available
+        if memory["chunk_id"] and memory["chunk_id"] in chunk_map:
+            chunk = chunk_map[memory["chunk_id"]]
+            item["chunk"] = {
+                "id": chunk["chunk_id"],
+                "text": chunk["chunk_text"],
+                "index": chunk["chunk_index"],
+                "document_id": chunk["document_id"],
+            }
+            # Add document if depth=document
+            if depth == "document" and chunk["document_id"] in doc_map:
+                doc = doc_map[chunk["document_id"]]
+                item["document"] = {
+                    "id": doc["id"],
+                    "full_text": doc["original_text"],
+                    "metadata": doc["metadata"],
+                    "retain_params": doc["retain_params"],
+                }
+        elif memory["document_id"] and depth == "document" and memory["document_id"] in doc_map:
+            # No chunk, but has document_id
+            doc = doc_map[memory["document_id"]]
+            item["document"] = {
+                "id": doc["id"],
+                "full_text": doc["original_text"],
+                "metadata": doc["metadata"],
+                "retain_params": doc["retain_params"],
+            }
+
+        results.append(item)
+
+    return {"results": results, "count": len(results)}
@@ -0,0 +1,218 @@
+"""
+Tool schema definitions for the reflect agent.
+
+These are OpenAI-format tool definitions used with native tool calling.
+"""
+
+# Tool definitions in OpenAI format
+TOOL_LIST_MENTAL_MODELS = {
+    "type": "function",
+    "function": {
+        "name": "list_mental_models",
+        "description": "List all available mental models - your synthesized knowledge about entities, concepts, and events. Returns an array of models with id, name, and description.",
+        "parameters": {
+            "type": "object",
+            "properties": {},
+            "required": [],
+        },
+    },
+}
+
+TOOL_GET_MENTAL_MODEL = {
+    "type": "function",
+    "function": {
+        "name": "get_mental_model",
+        "description": "Get full details of a specific mental model including all observations and memory references.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "model_id": {
+                    "type": "string",
+                    "description": "ID of the mental model (from list_mental_models results)",
+                },
+            },
+            "required": ["model_id"],
+        },
+    },
+}
+
+TOOL_RECALL = {
+    "type": "function",
+    "function": {
+        "name": "recall",
+        "description": "Search memories using semantic + temporal retrieval. Returns relevant memories from experience and world knowledge, each with an 'id' you can reference.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "query": {
+                    "type": "string",
+                    "description": "Search query string",
+                },
+                "max_tokens": {
+                    "type": "integer",
+                    "description": "Optional limit on result size (default 2048). Use higher values for broader searches.",
+                },
+            },
+            "required": ["query"],
+        },
+    },
+}
+
+TOOL_LEARN = {
+    "type": "function",
+    "function": {
+        "name": "learn",
+        "description": "Create a new mental model to track an important recurring topic. Use when you discover a person, project, concept, or pattern that appears frequently and would benefit from synthesized knowledge. The model content will be generated automatically.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "name": {
+                    "type": "string",
+                    "description": "Human-readable name (e.g., 'Project Alpha', 'John Smith', 'Product Strategy')",
+                },
+                "description": {
+                    "type": "string",
+                    "description": "What to track and synthesize (e.g., 'Track goals, milestones, blockers, and key decisions for Project Alpha')",
+                },
+            },
+            "required": ["name", "description"],
+        },
+    },
+}
+
+TOOL_EXPAND = {
+    "type": "function",
+    "function": {
+        "name": "expand",
+        "description": "Get more context for one or more memories. Memory hierarchy: memory -> chunk -> document.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "memory_ids": {
+                    "type": "array",
+                    "items": {"type": "string"},
+                    "description": "Array of memory IDs from recall results (batch multiple for efficiency)",
+                },
+                "depth": {
+                    "type": "string",
+                    "enum": ["chunk", "document"],
+                    "description": "chunk: surrounding text chunk, document: full source document",
+                },
+            },
+            "required": ["memory_ids", "depth"],
+        },
+    },
+}
+
+TOOL_DONE_ANSWER = {
+    "type": "function",
+    "function": {
+        "name": "done",
+        "description": "Signal completion with your final answer. Use this when you have gathered enough information to answer the question.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "answer": {
+                    "type": "string",
+                    "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
+                },
+                "memory_ids": {
+                    "type": "array",
+                    "items": {"type": "string"},
+                    "description": "Array of memory IDs that support your answer (put IDs here, NOT in answer text)",
+                },
+                "model_ids": {
+                    "type": "array",
+                    "items": {"type": "string"},
+                    "description": "Array of mental model IDs that support your answer",
+                },
+            },
+            "required": ["answer"],
+        },
+    },
+}
+
+
+def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
+    """
+    Build the done tool schema with directive compliance field.
+
+    When directives are present, adds a required field that forces the agent
+    to confirm compliance with each directive before submitting.
+
+    Args:
+        directive_rules: List of directive rule strings
+    """
+    from typing import Any, cast
+
+    # Build rules list for description
+    rules_list = "\n".join(f"  {i + 1}. {rule}" for i, rule in enumerate(directive_rules))
+
+    # Build the tool with directive compliance field
+    return {
+        "type": "function",
+        "function": {
+            "name": "done",
+            "description": (
+                "Signal completion with your final answer. IMPORTANT: You must confirm directive compliance before submitting. "
+                "Your answer will be REJECTED if it violates any directive."
+            ),
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "answer": {
+                        "type": "string",
+                        "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
+                    },
+                    "memory_ids": {
+                        "type": "array",
+                        "items": {"type": "string"},
+                        "description": "Array of memory IDs that support your answer (put IDs here, NOT in answer text)",
+                    },
+                    "model_ids": {
+                        "type": "array",
+                        "items": {"type": "string"},
+                        "description": "Array of mental model IDs that support your answer",
+                    },
+                    "directive_compliance": {
+                        "type": "string",
+                        "description": f"REQUIRED: Confirm your answer complies with ALL directives. List each directive and how your answer follows it:\n{rules_list}\n\nFormat: 'Directive 1: [how answer complies]. Directive 2: [how answer complies]...'",
+                    },
+                },
+                "required": ["answer", "directive_compliance"],
+            },
+        },
+    }
+
+
+def get_reflect_tools(enable_learn: bool = True, directive_rules: list[str] | None = None) -> list[dict]:
+    """
+    Get the list of tools for the reflect agent.
+
+    Args:
+        enable_learn: Whether to include the learn tool
+        directive_rules: Optional list of directive rule strings. If provided,
+                        the done() tool will require directive compliance confirmation.
+
+    Returns:
+        List of tool definitions in OpenAI format
+    """
+    tools = []
+
+    # Include mental model tools for lookup
+    tools.append(TOOL_LIST_MENTAL_MODELS)
+    tools.append(TOOL_GET_MENTAL_MODEL)
+    tools.append(TOOL_RECALL)
+
+    if enable_learn:
+        tools.append(TOOL_LEARN)
+
+    tools.append(TOOL_EXPAND)
+
+    # Use directive-aware done tool if directives are present
+    if directive_rules:
+        tools.append(_build_done_tool_with_directives(directive_rules))
+    else:
+        tools.append(TOOL_DONE_ANSWER)
+
+    return tools
@@ -10,8 +10,60 @@ from typing import Any

 from pydantic import BaseModel, ConfigDict, Field

-# Valid fact types for recall operations (excludes 'observation' which is internal)
-VALID_RECALL_FACT_TYPES = frozenset(["world", "experience", "opinion"])
+# Valid fact types for recall operations (excludes 'observation' which is internal, and 'opinion' which is deprecated)
+VALID_RECALL_FACT_TYPES = frozenset(["world", "experience"])
+
+
+class LLMToolCall(BaseModel):
+    """A tool call requested by the LLM."""
+
+    id: str = Field(description="Unique identifier for this tool call")
+    name: str = Field(description="Name of the tool to call")
+    arguments: dict[str, Any] = Field(description="Arguments to pass to the tool")
+
+
+class LLMToolCallResult(BaseModel):
+    """Result from an LLM call that may include tool calls."""
+
+    content: str | None = Field(default=None, description="Text content if any")
+    tool_calls: list[LLMToolCall] = Field(default_factory=list, description="Tool calls requested by the LLM")
+    finish_reason: str | None = Field(default=None, description="Reason the LLM stopped: 'stop', 'tool_calls', etc.")
+
+
+class ToolCallTrace(BaseModel):
+    """A single tool call made during reflect."""
+
+    tool: str = Field(description="Tool name: lookup, recall, learn, expand")
+    input: dict = Field(description="Tool input parameters")
+    output: dict = Field(description="Tool output/result")
+    duration_ms: int = Field(description="Execution time in milliseconds")
+    iteration: int = Field(default=0, description="Iteration number (1-based) when this tool was called")
+
+
+class LLMCallTrace(BaseModel):
+    """A single LLM call made during reflect."""
+
+    scope: str = Field(description="Call scope: agent_1, agent_2, final, etc.")
+    duration_ms: int = Field(description="Execution time in milliseconds")
+
+
+class MentalModelRef(BaseModel):
+    """Reference to a mental model accessed during reflect."""
+
+    id: str = Field(description="Mental model ID")
+    name: str = Field(description="Mental model name")
+    type: str = Field(description="Mental model type: entity, concept, event")
+    subtype: str = Field(description="Mental model subtype: structural, emergent, learned")
+    description: str = Field(description="Brief description")
+    summary: str | None = Field(default=None, description="Full summary (when looked up in detail)")
+
+
+class DirectiveRef(BaseModel):
+    """Reference to a directive that was applied during reflect."""
+
+    id: str = Field(description="Directive mental model ID")
+    name: str = Field(description="Directive name")
+    rules: list[str] = Field(default_factory=list, description="Directive rules/observations that were applied")


 class TokenUsage(BaseModel):
@@ -85,6 +137,7 @@ class MemoryFact(BaseModel):
                "metadata": {"source": "slack"},
                "chunk_id": "bank123_session_abc123_0",
                "activation": 0.95,
+                "tags": ["user_a", "session_123"],
            }
        }
    )
@@ -102,6 +155,7 @@ class MemoryFact(BaseModel):
    chunk_id: str | None = Field(
        None, description="ID of the chunk this fact was extracted from (format: bank_id_document_id_chunk_index)"
    )
+    tags: list[str] | None = Field(None, description="Visibility scope tags associated with this fact")


 class ChunkInfo(BaseModel):
@@ -196,6 +250,22 @@ class ReflectResult(BaseModel):
        default=None,
        description="Token usage metrics for the LLM calls made during this reflect operation.",
    )
+    tool_trace: list[ToolCallTrace] = Field(
+        default_factory=list,
+        description="Trace of tool calls made during reflection. Only present when include.tool_calls is enabled.",
+    )
+    llm_trace: list[LLMCallTrace] = Field(
+        default_factory=list,
+        description="Trace of LLM calls made during reflection. Only present when include.tool_calls is enabled.",
+    )
+    mental_models: list[MentalModelRef] = Field(
+        default_factory=list,
+        description="Mental models accessed during reflection, including directives (subtype='directive').",
+    )
+    directives_applied: list[DirectiveRef] = Field(
+        default_factory=list,
+        description="Directive mental models that were applied during this reflection.",
+    )


 class Opinion(BaseModel):
@@ -259,3 +329,32 @@ class EntityState(BaseModel):
    observations: list[EntityObservation] = Field(
        default_factory=list, description="List of observations about this entity"
    )
+
+
+class MentalModel(BaseModel):
+    """
+    A manually configured mental model for tracking specific topics/areas.
+
+    Mental models are user-defined focus areas that the agent should track
+    and maintain summaries for, unlike auto-extracted entities.
+    """
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "id": "team-dynamics",
+                "name": "Team Dynamics",
+                "description": "Track how the team collaborates, communication patterns, conflicts, and resolutions",
+                "summary": "The team has strong collaboration...",
+                "summary_updated_at": "2024-01-15T10:30:00Z",
+                "created_at": "2024-01-10T08:00:00Z",
+            }
+        }
+    )
+
+    id: str = Field(description="Unique identifier (alphanumeric lowercase)")
+    name: str = Field(description="Display name for the mental model")
+    description: str = Field(description="Prompt/directions for what to track and summarize")
+    summary: str | None = Field(None, description="Generated summary based on relevant facts")
+    summary_updated_at: str | None = Field(None, description="ISO format date when summary was last updated")
+    created_at: str = Field(description="ISO format date when the mental model was created")
@@ -1,5 +1,5 @@
 """
-bank profile utilities for disposition and background management.
+bank profile utilities for disposition and mission management.
 """

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

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


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

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


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

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

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

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

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

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


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


-async def merge_bank_background(pool, llm_config, bank_id: str, new_info: str, update_disposition: bool = True) -> dict:
+async def set_bank_mission(pool, bank_id: str, mission: str) -> None:
    """
-    Merge new background information with existing background using LLM.
-    Normalizes to first person ("I") and resolves conflicts.
-    Optionally infers disposition traits from the merged background.
+    Set bank mission (replacing any existing mission).

    Args:
        pool: Database connection pool
-        llm_config: LLM configuration for background merging
        bank_id: bank IDentifier
-        new_info: New background information to add/merge
-        update_disposition: If True, infer Big Five traits from background (default: True)
+        mission: The mission text
+    """
+    # Ensure bank exists first
+    await get_bank_profile(pool, bank_id)
+
+    async with acquire_with_retry(pool) as conn:
+        await conn.execute(
+            f"""
+            UPDATE {fq_table("banks")}
+            SET mission = $2,
+                updated_at = NOW()
+            WHERE bank_id = $1
+            """,
+            bank_id,
+            mission,
+        )
+
+
+async def merge_bank_mission(pool, llm_config, bank_id: str, new_info: str) -> dict:
+    """
+    Merge new mission information with existing mission using LLM.
+    Normalizes to first person ("I") and resolves conflicts.
+
+    Args:
+        pool: Database connection pool
+        llm_config: LLM configuration for mission merging
+        bank_id: bank IDentifier
+        new_info: New mission information to add/merge

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

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

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

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

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


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

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

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

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

 New information to add: {new_info}

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

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

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

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

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

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

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

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


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

    Returns:
-        List of dicts with bank_id, name, disposition, background, created_at, updated_at
+        List of dicts with bank_id, name, disposition, mission, created_at, updated_at
    """
    async with acquire_with_retry(pool) as conn:
        rows = await conn.fetch(
            f"""
-            SELECT bank_id, name, disposition, background, created_at, updated_at
+            SELECT bank_id, name, disposition, mission, created_at, updated_at
            FROM {fq_table("banks")}
            ORDER BY updated_at DESC
            """
@@ -381,7 +259,7 @@ async def list_banks(pool) -> list:
                    "bank_id": row["bank_id"],
                    "name": row["name"],
                    "disposition": disposition_data,
-                    "background": row["background"],
+                    "mission": row["mission"] or "",
                    "created_at": row["created_at"].isoformat() if row["created_at"] else None,
                    "updated_at": row["updated_at"].isoformat() if row["updated_at"] else None,
                }
@@ -1268,6 +1268,7 @@ async def extract_facts_from_contents(
                        # mentioned_at: always the event_date (when the conversation/document occurred)
                        mentioned_at=content.event_date,
                        metadata=content.metadata,
+                        tags=content.tags,
                    )

                    extracted_facts.append(extracted_fact)
@@ -45,6 +45,7 @@ async def insert_facts_batch(
    metadata_jsons = []
    chunk_ids = []
    document_ids = []
+    tags_list = []

    for fact in facts:
        fact_texts.append(fact.fact_text)
@@ -65,16 +66,31 @@ async def insert_facts_batch(
        chunk_ids.append(fact.chunk_id)
        # Use per-fact document_id if available, otherwise fallback to batch-level document_id
        document_ids.append(fact.document_id if fact.document_id else document_id)
+        # Convert tags to JSON string for proper batch insertion (PostgreSQL unnest doesn't handle 2D arrays well)
+        tags_list.append(json.dumps(fact.tags if fact.tags else []))

    # Batch insert all facts
+    # Note: tags are passed as JSON strings and converted back to varchar[] via jsonb_array_elements_text + array_agg
    results = await conn.fetch(
        f"""
-        INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                                 context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id)
-        SELECT $1, * FROM unnest(
-            $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
-            $8::text[], $9::text[], $10::float[], $11::int[], $12::jsonb[], $13::text[], $14::text[]
+        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::int[], $12::jsonb[], $13::text[], $14::text[], $15::jsonb[]
+            ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                   context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id, tags_json)
        )
+        INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                                 context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id, tags)
+        SELECT
+            $1,
+            text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+            context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id,
+            COALESCE(
+                (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
+                '{{}}'::varchar[]
+            )
+        FROM input_data
        RETURNING id
        """,
        bank_id,
@@ -91,6 +107,7 @@ async def insert_facts_batch(
        metadata_jsons,
        chunk_ids,
        document_ids,
+        tags_list,
    )

    unit_ids = [str(row["id"]) for row in results]
@@ -109,7 +126,7 @@ async def ensure_bank_exists(conn, bank_id: str) -> None:
    """
    await conn.execute(
        f"""
-        INSERT INTO {fq_table("banks")} (bank_id, disposition, background)
+        INSERT INTO {fq_table("banks")} (bank_id, disposition, mission)
        VALUES ($1, $2::jsonb, $3)
        ON CONFLICT (bank_id) DO UPDATE
        SET updated_at = NOW()
@@ -121,7 +138,13 @@ async def ensure_bank_exists(conn, bank_id: str) -> None:


 async def handle_document_tracking(
-    conn, bank_id: str, document_id: str, combined_content: str, is_first_batch: bool, retain_params: dict | None = None
+    conn,
+    bank_id: str,
+    document_id: str,
+    combined_content: str,
+    is_first_batch: bool,
+    retain_params: dict | None = None,
+    document_tags: list[str] | None = None,
 ) -> None:
    """
    Handle document tracking in the database.
@@ -133,6 +156,7 @@ async def handle_document_tracking(
        combined_content: Combined content text from all content items
        is_first_batch: Whether this is the first batch (for chunked operations)
        retain_params: Optional parameters passed during retain (context, event_date, etc.)
+        document_tags: Optional list of tags to associate with the document
    """
    import hashlib

@@ -149,13 +173,14 @@ async def handle_document_tracking(
    # Insert document (or update if exists from concurrent operations)
    await conn.execute(
        f"""
-        INSERT INTO {fq_table("documents")} (id, bank_id, original_text, content_hash, metadata, retain_params)
-        VALUES ($1, $2, $3, $4, $5, $6)
+        INSERT INTO {fq_table("documents")} (id, bank_id, original_text, content_hash, metadata, retain_params, tags)
+        VALUES ($1, $2, $3, $4, $5, $6, $7)
        ON CONFLICT (id, bank_id) DO UPDATE
        SET original_text = EXCLUDED.original_text,
            content_hash = EXCLUDED.content_hash,
            metadata = EXCLUDED.metadata,
            retain_params = EXCLUDED.retain_params,
+            tags = EXCLUDED.tags,
            updated_at = NOW()
        """,
        document_id,
@@ -164,4 +189,5 @@ async def handle_document_tracking(
        content_hash,
        json.dumps({}),  # Empty metadata dict
        json.dumps(retain_params) if retain_params else None,
+        document_tags or [],
    )
@@ -1,254 +0,0 @@
-"""
-Observation regeneration for retain pipeline.
-
-Regenerates entity observations as part of the retain transaction.
-"""
-
-import logging
-import time
-import uuid
-from datetime import UTC, datetime
-
-from ...config import get_config
-from ..memory_engine import fq_table
-from ..search import observation_utils
-from . import embedding_utils
-from .types import EntityLink
-
-logger = logging.getLogger(__name__)
-
-
-def utcnow():
-    """Get current UTC time."""
-    return datetime.now(UTC)
-
-
-# Simple dataclass-like container for facts (avoid importing from memory_engine)
-class MemoryFactForObservation:
-    def __init__(self, id: str, text: str, fact_type: str, context: str, occurred_start: str | None):
-        self.id = id
-        self.text = text
-        self.fact_type = fact_type
-        self.context = context
-        self.occurred_start = occurred_start
-
-
-async def regenerate_observations_batch(
-    conn, embeddings_model, llm_config, bank_id: str, entity_links: list[EntityLink], log_buffer: list[str] = None
-) -> None:
-    """
-    Regenerate observations for top entities in this batch.
-
-    Called INSIDE the retain transaction for atomicity - if observations
-    fail, the entire retain batch is rolled back.
-
-    Args:
-        conn: Database connection (from the retain transaction)
-        embeddings_model: Embeddings model for generating observation embeddings
-        llm_config: LLM configuration for observation extraction
-        bank_id: Bank identifier
-        entity_links: Entity links from this batch
-        log_buffer: Optional log buffer for timing
-    """
-    config = get_config()
-    TOP_N_ENTITIES = config.observation_top_entities
-    MIN_FACTS_THRESHOLD = config.observation_min_facts
-
-    if not entity_links:
-        return
-
-    # Count mentions per entity in this batch
-    entity_mention_counts: dict[str, int] = {}
-    for link in entity_links:
-        if link.entity_id:
-            entity_id = str(link.entity_id)
-            entity_mention_counts[entity_id] = entity_mention_counts.get(entity_id, 0) + 1
-
-    if not entity_mention_counts:
-        return
-
-    # Sort by mention count descending and take top N
-    sorted_entities = sorted(entity_mention_counts.items(), key=lambda x: x[1], reverse=True)
-    entities_to_process = [e[0] for e in sorted_entities[:TOP_N_ENTITIES]]
-
-    obs_start = time.time()
-
-    # Convert to UUIDs
-    entity_uuids = [uuid.UUID(eid) if isinstance(eid, str) else eid for eid in entities_to_process]
-
-    # Batch query for entity names
-    entity_rows = await conn.fetch(
-        f"""
-        SELECT id, canonical_name FROM {fq_table("entities")}
-        WHERE id = ANY($1) AND bank_id = $2
-        """,
-        entity_uuids,
-        bank_id,
-    )
-    entity_names = {row["id"]: row["canonical_name"] for row in entity_rows}
-
-    # Batch query for fact counts
-    fact_counts = await conn.fetch(
-        f"""
-        SELECT ue.entity_id, COUNT(*) as cnt
-        FROM {fq_table("unit_entities")} ue
-        JOIN {fq_table("memory_units")} mu ON ue.unit_id = mu.id
-        WHERE ue.entity_id = ANY($1) AND mu.bank_id = $2
-        GROUP BY ue.entity_id
-        """,
-        entity_uuids,
-        bank_id,
-    )
-    entity_fact_counts = {row["entity_id"]: row["cnt"] for row in fact_counts}
-
-    # Filter entities that meet the threshold
-    entities_with_names = []
-    for entity_id in entities_to_process:
-        entity_uuid = uuid.UUID(entity_id) if isinstance(entity_id, str) else entity_id
-        if entity_uuid not in entity_names:
-            continue
-        fact_count = entity_fact_counts.get(entity_uuid, 0)
-        if fact_count >= MIN_FACTS_THRESHOLD:
-            entities_with_names.append((entity_id, entity_names[entity_uuid]))
-
-    if not entities_with_names:
-        return
-
-    # Process entities SEQUENTIALLY (asyncpg doesn't allow concurrent queries on same connection)
-    # We must use the same connection to stay in the retain transaction
-    total_observations = 0
-
-    for entity_id, entity_name in entities_with_names:
-        try:
-            obs_ids = await _regenerate_entity_observations(
-                conn, embeddings_model, llm_config, bank_id, entity_id, entity_name
-            )
-            total_observations += len(obs_ids)
-        except Exception as e:
-            logger.error(f"[OBSERVATIONS] Error processing entity {entity_id}: {e}")
-
-    obs_time = time.time() - obs_start
-    if log_buffer is not None:
-        log_buffer.append(
-            f"[11] Observations: {total_observations} observations for {len(entities_with_names)} entities in {obs_time:.3f}s"
-        )
-
-
-async def _regenerate_entity_observations(
-    conn, embeddings_model, llm_config, bank_id: str, entity_id: str, entity_name: str
-) -> list[str]:
-    """
-    Regenerate observations for a single entity.
-
-    Uses the provided connection (part of retain transaction).
-
-    Args:
-        conn: Database connection (from the retain transaction)
-        embeddings_model: Embeddings model
-        llm_config: LLM configuration
-        bank_id: Bank identifier
-        entity_id: Entity UUID
-        entity_name: Canonical name of the entity
-
-    Returns:
-        List of created observation IDs
-    """
-    entity_uuid = uuid.UUID(entity_id) if isinstance(entity_id, str) else entity_id
-
-    # Get all facts mentioning this entity (exclude observations themselves)
-    rows = await conn.fetch(
-        f"""
-        SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.fact_type
-        FROM {fq_table("memory_units")} mu
-        JOIN {fq_table("unit_entities")} ue ON mu.id = ue.unit_id
-        WHERE mu.bank_id = $1
-          AND ue.entity_id = $2
-          AND mu.fact_type IN ('world', 'experience')
-        ORDER BY mu.occurred_start DESC
-        LIMIT 50
-        """,
-        bank_id,
-        entity_uuid,
-    )
-
-    if not rows:
-        return []
-
-    # Convert to fact objects for observation extraction
-    facts = []
-    for row in rows:
-        occurred_start = row["occurred_start"].isoformat() if row["occurred_start"] else None
-        facts.append(
-            MemoryFactForObservation(
-                id=str(row["id"]),
-                text=row["text"],
-                fact_type=row["fact_type"],
-                context=row["context"],
-                occurred_start=occurred_start,
-            )
-        )
-
-    # Extract observations using LLM
-    observations = await observation_utils.extract_observations_from_facts(llm_config, entity_name, facts)
-
-    if not observations:
-        return []
-
-    # Delete old observations for this entity
-    await conn.execute(
-        f"""
-        DELETE FROM {fq_table("memory_units")}
-        WHERE id IN (
-            SELECT mu.id
-            FROM {fq_table("memory_units")} mu
-            JOIN {fq_table("unit_entities")} ue ON mu.id = ue.unit_id
-            WHERE mu.bank_id = $1
-              AND mu.fact_type = 'observation'
-              AND ue.entity_id = $2
-        )
-        """,
-        bank_id,
-        entity_uuid,
-    )
-
-    # Generate embeddings for new observations
-    embeddings = await embedding_utils.generate_embeddings_batch(embeddings_model, observations)
-
-    # Insert new observations
-    current_time = utcnow()
-    created_ids = []
-
-    for obs_text, embedding in zip(observations, embeddings):
-        result = await conn.fetchrow(
-            f"""
-            INSERT INTO {fq_table("memory_units")} (
-                bank_id, text, embedding, context, event_date,
-                occurred_start, occurred_end, mentioned_at,
-                fact_type, access_count
-            )
-            VALUES ($1, $2, $3, $4, $5, $6, $7, $8, 'observation', 0)
-            RETURNING id
-            """,
-            bank_id,
-            obs_text,
-            str(embedding),
-            f"observation about {entity_name}",
-            current_time,
-            current_time,
-            current_time,
-            current_time,
-        )
-        obs_id = str(result["id"])
-        created_ids.append(obs_id)
-
-        # Link observation to entity
-        await conn.execute(
-            f"""
-            INSERT INTO {fq_table("unit_entities")} (unit_id, entity_id)
-            VALUES ($1, $2)
-            """,
-            uuid.UUID(obs_id),
-            entity_uuid,
-        )
-
-    return created_ids
@@ -9,7 +9,6 @@ import time
 import uuid
 from datetime import UTC, datetime

-from ...config import get_config
 from ..db_utils import acquire_with_retry
 from . import bank_utils

@@ -28,9 +27,8 @@ from . import (
    fact_extraction,
    fact_storage,
    link_creation,
-    observation_regeneration,
 )
-from .types import ExtractedFact, ProcessedFact, RetainContent, RetainContentDict
+from .types import EntityLink, ExtractedFact, ProcessedFact, RetainContent, RetainContentDict

 logger = logging.getLogger(__name__)

@@ -40,7 +38,6 @@ async def retain_batch(
    embeddings_model,
    llm_config,
    entity_resolver,
-    task_backend,
    format_date_fn,
    duplicate_checker_fn,
    bank_id: str,
@@ -49,6 +46,7 @@ async def retain_batch(
    is_first_batch: bool = True,
    fact_type_override: str | None = None,
    confidence_score: float | None = None,
+    document_tags: list[str] | None = None,
 ) -> tuple[list[list[str]], TokenUsage]:
    """
    Process a batch of content through the retain pipeline.
@@ -58,7 +56,6 @@ async def retain_batch(
        embeddings_model: Embeddings model for generating embeddings
        llm_config: LLM configuration for fact extraction
        entity_resolver: Entity resolver for entity processing
-        task_backend: Task backend for background jobs
        format_date_fn: Function to format datetime to readable string
        duplicate_checker_fn: Function to check for duplicate facts
        bank_id: Bank identifier
@@ -67,6 +64,7 @@ async def retain_batch(
        is_first_batch: Whether this is the first batch
        fact_type_override: Override fact type for all facts
        confidence_score: Confidence score for opinions
+        document_tags: Tags applied to all items in this batch

    Returns:
        Tuple of (unit ID lists, token usage for fact extraction)
@@ -88,12 +86,16 @@ async def retain_batch(
    # Convert dicts to RetainContent objects
    contents = []
    for item in contents_dicts:
+        # Merge item-level tags with document-level tags
+        item_tags = item.get("tags", []) or []
+        merged_tags = list(set(item_tags + (document_tags or [])))
        content = RetainContent(
            content=item["content"],
            context=item.get("context", ""),
            event_date=item.get("event_date") or utcnow(),
            metadata=item.get("metadata", {}),
            entities=item.get("entities", []),
+            tags=merged_tags,
        )
        contents.append(content)

@@ -131,7 +133,7 @@ async def retain_batch(
                        if first_item.get("metadata"):
                            retain_params["metadata"] = first_item["metadata"]
                    await fact_storage.handle_document_tracking(
-                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params
+                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params, document_tags
                    )
                else:
                    # Check for per-item document_ids
@@ -159,7 +161,7 @@ async def retain_batch(
                            if first_item.get("metadata"):
                                retain_params["metadata"] = first_item["metadata"]
                        await fact_storage.handle_document_tracking(
-                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params
+                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params, document_tags
                        )

        total_time = time.time() - start_time
@@ -225,7 +227,7 @@ async def retain_batch(
                        retain_params["metadata"] = first_item["metadata"]

                await fact_storage.handle_document_tracking(
-                    conn, bank_id, document_id, combined_content, is_first_batch, retain_params
+                    conn, bank_id, document_id, combined_content, is_first_batch, retain_params, document_tags
                )
                document_ids_added.append(document_id)
                doc_id_mapping[None] = document_id  # For backwards compatibility
@@ -269,7 +271,13 @@ async def retain_batch(
                                    retain_params["metadata"] = first_item["metadata"]

                            await fact_storage.handle_document_tracking(
-                                conn, bank_id, actual_doc_id, combined_content, is_first_batch, retain_params
+                                conn,
+                                bank_id,
+                                actual_doc_id,
+                                combined_content,
+                                is_first_batch,
+                                retain_params,
+                                document_tags,
                            )
                            document_ids_added.append(actual_doc_id)

@@ -396,27 +404,9 @@ async def retain_batch(
            causal_link_count = await link_creation.create_causal_links_batch(conn, unit_ids, non_duplicate_facts)
            log_buffer.append(f"[10] Causal links: {causal_link_count} links in {time.time() - step_start:.3f}s")

-            # Regenerate observations - sync (in transaction) or async (background task)
-            config = get_config()
-            if config.retain_observations_async:
-                # Queue for async processing after transaction commits
-                entity_ids_for_async = list(set(link.entity_id for link in entity_links)) if entity_links else []
-                log_buffer.append(
-                    f"[11] Observations: queued {len(entity_ids_for_async)} entities for async processing"
-                )
-            else:
-                # Run synchronously inside transaction for atomicity
-                await observation_regeneration.regenerate_observations_batch(
-                    conn, embeddings_model, llm_config, bank_id, entity_links, log_buffer
-                )
-                entity_ids_for_async = []
-
            # Map results back to original content items
            result_unit_ids = _map_results_to_contents(contents, extracted_facts, is_duplicate_flags, unit_ids)

-        # Trigger background tasks AFTER transaction commits
-        await _trigger_background_tasks(task_backend, bank_id, unit_ids, non_duplicate_facts, entity_ids_for_async)
-
        # Log final summary
        total_time = time.time() - start_time
        log_buffer.append(f"{'=' * 60}")
@@ -458,35 +448,3 @@ def _map_results_to_contents(
        result_unit_ids.append(content_unit_ids)

    return result_unit_ids
-
-
-async def _trigger_background_tasks(
-    task_backend,
-    bank_id: str,
-    unit_ids: list[str],
-    facts: list[ProcessedFact],
-    entity_ids_for_observations: list[str] | None = None,
-) -> None:
-    """Trigger background tasks after transaction commits."""
-    # Trigger opinion reinforcement if there are entities
-    fact_entities = [[e.name for e in fact.entities] for fact in facts]
-    if any(fact_entities):
-        await task_backend.submit_task(
-            {
-                "type": "reinforce_opinion",
-                "bank_id": bank_id,
-                "created_unit_ids": unit_ids,
-                "unit_texts": [fact.fact_text for fact in facts],
-                "unit_entities": fact_entities,
-            }
-        )
-
-    # Trigger observation regeneration if async mode is enabled
-    if entity_ids_for_observations:
-        await task_backend.submit_task(
-            {
-                "type": "regenerate_observations",
-                "bank_id": bank_id,
-                "entity_ids": entity_ids_for_observations,
-            }
-        )
@@ -21,6 +21,7 @@ class RetainContentDict(TypedDict, total=False):
        metadata: Custom key-value metadata (optional)
        document_id: Document ID for this content item (optional)
        entities: User-provided entities to merge with extracted entities (optional)
+        tags: Visibility scope tags for this content item (optional)
    """

    content: str  # Required
@@ -29,6 +30,7 @@ class RetainContentDict(TypedDict, total=False):
    metadata: dict[str, str]
    document_id: str
    entities: list[dict[str, str]]  # [{"text": "...", "type": "..."}]
+    tags: list[str]  # Visibility scope tags


 def _now_utc() -> datetime:
@@ -49,6 +51,7 @@ class RetainContent:
    event_date: datetime = field(default_factory=_now_utc)
    metadata: dict[str, str] = field(default_factory=dict)
    entities: list[dict[str, str]] = field(default_factory=list)  # User-provided entities
+    tags: list[str] = field(default_factory=list)  # Visibility scope tags


@dataclass
@@ -113,6 +116,7 @@ class ExtractedFact:
    context: str = ""
    mentioned_at: datetime | None = None
    metadata: dict[str, str] = field(default_factory=dict)
+    tags: list[str] = field(default_factory=list)  # Visibility scope tags


@dataclass
@@ -158,6 +162,9 @@ class ProcessedFact:
    # Track which content this fact came from (for user entity merging)
    content_index: int = 0

+    # Visibility scope tags
+    tags: list[str] = field(default_factory=list)
+
    @property
    def is_duplicate(self) -> bool:
        """Check if this fact was marked as a duplicate."""
@@ -201,6 +208,7 @@ class ProcessedFact:
            causal_relations=extracted_fact.causal_relations,
            chunk_id=chunk_id,
            content_index=extracted_fact.content_index,
+            tags=extracted_fact.tags,
        )


@@ -232,6 +240,7 @@ class RetainBatch:
    document_id: str | None = None
    fact_type_override: str | None = None
    confidence_score: float | None = None
+    document_tags: list[str] = field(default_factory=list)  # Tags applied to all items

    # Extracted data (populated during processing)
    extracted_facts: list[ExtractedFact] = field(default_factory=list)
@@ -11,6 +11,7 @@ from abc import ABC, abstractmethod

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

 logger = logging.getLogger(__name__)
@@ -43,6 +44,8 @@ class GraphRetriever(ABC):
        semantic_seeds: list[RetrievalResult] | None = None,
        temporal_seeds: list[RetrievalResult] | None = None,
        adjacency=None,  # TypedAdjacency, optional pre-loaded graph
+        tags: list[str] | None = None,  # Visibility scope tags for filtering
+        tags_match: TagsMatch = "any",  # How to match tags: 'any' (OR) or 'all' (AND)
    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
        """
        Retrieve relevant facts via graph traversal.
@@ -57,6 +60,7 @@ class GraphRetriever(ABC):
            semantic_seeds: Pre-computed semantic entry points (from semantic retrieval)
            temporal_seeds: Pre-computed temporal entry points (from temporal retrieval)
            adjacency: Pre-loaded typed adjacency graph (optional, for MPFP)
+            tags: Optional list of tags for visibility filtering (OR matching)

        Returns:
            Tuple of (List of RetrievalResult with activation scores, optional timing info)
@@ -114,6 +118,8 @@ class BFSGraphRetriever(GraphRetriever):
        semantic_seeds: list[RetrievalResult] | None = None,
        temporal_seeds: list[RetrievalResult] | None = None,
        adjacency=None,  # Not used by BFS
+        tags: list[str] | None = None,
+        tags_match: TagsMatch = "any",
    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
        """
        Retrieve facts using BFS spreading activation.
@@ -129,7 +135,9 @@ class BFSGraphRetriever(GraphRetriever):
        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)
+            results = await self._retrieve_with_conn(
+                conn, query_embedding_str, bank_id, fact_type, budget, tags=tags, tags_match=tags_match
+            )
            return results, None

    async def _retrieve_with_conn(
@@ -139,33 +147,46 @@ class BFSGraphRetriever(GraphRetriever):
        bank_id: str,
        fact_type: str,
        budget: int,
+        tags: list[str] | None = None,
+        tags_match: TagsMatch = "any",
    ) -> list[RetrievalResult]:
        """Internal implementation with connection."""
+        from .tags import build_tags_where_clause_simple
+
+        tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
+        params = [query_embedding_str, bank_id, fact_type, self.entry_point_threshold, self.entry_point_limit]
+        if tags:
+            params.append(tags)

        # Step 1: Find entry points
        entry_points = await conn.fetch(
            f"""
            SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+                   mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                   1 - (embedding <=> $1::vector) AS similarity
            FROM {fq_table("memory_units")}
            WHERE bank_id = $2
              AND embedding IS NOT NULL
              AND fact_type = $3
              AND (1 - (embedding <=> $1::vector)) >= $4
+              {tags_clause}
            ORDER BY embedding <=> $1::vector
            LIMIT $5
            """,
-            query_embedding_str,
-            bank_id,
-            fact_type,
-            self.entry_point_threshold,
-            self.entry_point_limit,
+            *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 = []
@@ -196,7 +217,7 @@ class BFSGraphRetriever(GraphRetriever):
                    f"""
                    SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.occurred_end,
                           mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type,
-                           mu.document_id, mu.chunk_id,
+                           mu.document_id, mu.chunk_id, mu.tags,
                           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
@@ -236,4 +257,8 @@ class BFSGraphRetriever(GraphRetriever):
                            neighbor_result = RetrievalResult.from_db_row(dict(n))
                            queue.append((neighbor_result, new_activation))

+        # Apply tags filtering (BFS may traverse into memories that don't match tags criteria)
+        if tags:
+            results = filter_results_by_tags(results, tags, match=tags_match)
+
        return results
@@ -18,6 +18,7 @@ import time
 from ..db_utils import acquire_with_retry
 from ..memory_engine import fq_table
 from .graph_retrieval import GraphRetriever
+from .tags import TagsMatch, filter_results_by_tags
 from .types import MPFPTimings, RetrievalResult

 logger = logging.getLogger(__name__)
@@ -30,26 +31,32 @@ async def _find_semantic_seeds(
    fact_type: str,
    limit: int = 20,
    threshold: float = 0.3,
+    tags: list[str] | None = None,
+    tags_match: TagsMatch = "any",
 ) -> list[RetrievalResult]:
    """Find semantic seeds via embedding search."""
+    from .tags import build_tags_where_clause_simple
+
+    tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
+    params = [query_embedding_str, bank_id, fact_type, threshold, limit]
+    if tags:
+        params.append(tags)
+
    rows = await conn.fetch(
        f"""
        SELECT id, text, context, event_date, occurred_start, occurred_end,
-               mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+               mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
               1 - (embedding <=> $1::vector) AS similarity
        FROM {fq_table("memory_units")}
        WHERE bank_id = $2
          AND embedding IS NOT NULL
          AND fact_type = $3
          AND (1 - (embedding <=> $1::vector)) >= $4
+          {tags_clause}
        ORDER BY embedding <=> $1::vector
        LIMIT $5
        """,
-        query_embedding_str,
-        bank_id,
-        fact_type,
-        threshold,
-        limit,
+        *params,
    )
    return [RetrievalResult.from_db_row(dict(r)) for r in rows]

@@ -95,6 +102,8 @@ class LinkExpansionRetriever(GraphRetriever):
        semantic_seeds: list[RetrievalResult] | None = None,
        temporal_seeds: list[RetrievalResult] | None = None,
        adjacency=None,
+        tags: list[str] | None = None,
+        tags_match: TagsMatch = "any",
    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
        """
        Retrieve facts by expanding links from seeds.
@@ -109,6 +118,7 @@ class LinkExpansionRetriever(GraphRetriever):
            semantic_seeds: Pre-computed semantic entry points
            temporal_seeds: Pre-computed temporal entry points
            adjacency: Unused, kept for interface compatibility
+            tags: Optional list of tags for visibility filtering (OR matching)

        Returns:
            Tuple of (results, timings)
@@ -125,15 +135,27 @@ class LinkExpansionRetriever(GraphRetriever):
            else:
                seeds_start = time.time()
                all_seeds = await _find_semantic_seeds(
-                    conn, query_embedding_str, bank_id, fact_type, limit=20, threshold=0.3
+                    conn,
+                    query_embedding_str,
+                    bank_id,
+                    fact_type,
+                    limit=20,
+                    threshold=0.3,
+                    tags=tags,
+                    tags_match=tags_match,
                )
                timings.seeds_time = time.time() - seeds_start
+                logger.debug(
+                    f"[LinkExpansion] Found {len(all_seeds)} semantic seeds for fact_type={fact_type} "
+                    f"(tags={tags}, tags_match={tags_match})"
+                )

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

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

            seed_ids = list({s.id for s in all_seeds})
@@ -147,7 +169,7 @@ class LinkExpansionRetriever(GraphRetriever):
                SELECT
                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
                    mu.occurred_end, mu.mentioned_at, mu.access_count, mu.embedding,
-                    mu.fact_type, mu.document_id, mu.chunk_id,
+                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
                    COUNT(*)::float AS score
                FROM {fq_table("unit_entities")} seed_ue
                JOIN {fq_table("entities")} e ON seed_ue.entity_id = e.id
@@ -172,7 +194,7 @@ 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.access_count, mu.embedding,
-                    mu.fact_type, mu.document_id, mu.chunk_id,
+                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
                    ml.weight + 1.0 AS score
                FROM {fq_table("memory_links")} ml
                JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
@@ -219,6 +241,10 @@ class LinkExpansionRetriever(GraphRetriever):
            result.activation = row["score"]
            results.append(result)

+        # Apply tags filtering (graph expansion may reach untagged memories)
+        if tags:
+            results = filter_results_by_tags(results, tags, match=tags_match)
+
        timings.result_count = len(results)
        timings.traverse = time.time() - start_time

@@ -23,6 +23,7 @@ from dataclasses import dataclass, field
 from ..db_utils import acquire_with_retry
 from ..memory_engine import fq_table
 from .graph_retrieval import GraphRetriever
+from .tags import TagsMatch
 from .types import MPFPTimings, RetrievalResult

 logger = logging.getLogger(__name__)
@@ -448,7 +449,7 @@ async def fetch_memory_units_by_ids(
        rows = await conn.fetch(
            f"""
            SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, access_count, embedding, fact_type, document_id, chunk_id
+                   mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags
            FROM {fq_table("memory_units")}
            WHERE id = ANY($1::uuid[])
              AND fact_type = $2
@@ -503,6 +504,8 @@ class MPFPGraphRetriever(GraphRetriever):
        semantic_seeds: list[RetrievalResult] | None = None,
        temporal_seeds: list[RetrievalResult] | None = None,
        adjacency=None,  # Ignored - kept for interface compatibility
+        tags: list[str] | None = None,
+        tags_match: TagsMatch = "any",
    ) -> tuple[list[RetrievalResult], MPFPTimings | None]:
        """
        Retrieve facts using MPFP algorithm with lazy edge loading.
@@ -517,6 +520,7 @@ class MPFPGraphRetriever(GraphRetriever):
            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)
@@ -532,8 +536,13 @@ class MPFPGraphRetriever(GraphRetriever):
        # If no semantic seeds provided, fall back to finding our own
        if not semantic_seed_nodes:
            seeds_start = time.time()
-            semantic_seed_nodes = await self._find_semantic_seeds(pool, query_embedding_str, bank_id, fact_type)
+            semantic_seed_nodes = await self._find_semantic_seeds(
+                pool, query_embedding_str, bank_id, fact_type, tags=tags, tags_match=tags_match
+            )
            timings.seeds_time = time.time() - seeds_start
+            logger.debug(
+                f"[MPFP] Found {len(semantic_seed_nodes)} semantic seeds for fact_type={fact_type} (tags={tags}, tags_match={tags_match})"
+            )

        # Collect all pattern jobs
        pattern_jobs = []
@@ -549,6 +558,9 @@ class MPFPGraphRetriever(GraphRetriever):
                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)
@@ -587,6 +599,7 @@ class MPFPGraphRetriever(GraphRetriever):
        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
@@ -596,6 +609,13 @@ class MPFPGraphRetriever(GraphRetriever):
        step_start = time.time()
        results = await fetch_memory_units_by_ids(pool, result_ids, fact_type)
        timings.fetch = time.time() - step_start
+
+        # Filter results by tags (graph traversal may have picked up unfiltered memories)
+        if tags:
+            from .tags import filter_results_by_tags
+
+            results = filter_results_by_tags(results, tags, match=tags_match)
+
        timings.result_count = len(results)

        # Add activation scores from fusion
@@ -634,8 +654,17 @@ class MPFPGraphRetriever(GraphRetriever):
        fact_type: str,
        limit: int = 20,
        threshold: float = 0.3,
+        tags: list[str] | None = None,
+        tags_match: TagsMatch = "any",
    ) -> list[SeedNode]:
        """Fallback: find semantic seeds via embedding search."""
+        from .tags import build_tags_where_clause_simple
+
+        tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
+        params = [query_embedding_str, bank_id, fact_type, threshold, limit]
+        if tags:
+            params.append(tags)
+
        async with acquire_with_retry(pool) as conn:
            rows = await conn.fetch(
                f"""
@@ -645,14 +674,11 @@ class MPFPGraphRetriever(GraphRetriever):
                  AND embedding IS NOT NULL
                  AND fact_type = $3
                  AND (1 - (embedding <=> $1::vector)) >= $4
+                  {tags_clause}
                ORDER BY embedding <=> $1::vector
                LIMIT $5
                """,
-                query_embedding_str,
-                bank_id,
-                fact_type,
-                threshold,
-                limit,
+                *params,
            )

        return [SeedNode(node_id=str(r["id"]), score=r["similarity"]) for r in rows]
@@ -1,125 +0,0 @@
-"""
-Observation utilities for generating entity observations from facts.
-
-Observations are objective facts synthesized from multiple memory facts
-about an entity, without personality influence.
-"""
-
-import logging
-
-from pydantic import BaseModel, Field
-
-from ..response_models import MemoryFact
-
-logger = logging.getLogger(__name__)
-
-
-class Observation(BaseModel):
-    """An observation about an entity."""
-
-    observation: str = Field(description="The observation text - a factual statement about the entity")
-
-
-class ObservationExtractionResponse(BaseModel):
-    """Response containing extracted observations."""
-
-    observations: list[Observation] = Field(default_factory=list, description="List of observations about the entity")
-
-
-def format_facts_for_observation_prompt(facts: list[MemoryFact]) -> str:
-    """Format facts as text for observation extraction prompt."""
-    import json
-
-    if not facts:
-        return "[]"
-    formatted = []
-    for fact in facts:
-        fact_obj = {"text": fact.text}
-
-        # Add context if available
-        if fact.context:
-            fact_obj["context"] = fact.context
-
-        # Add occurred_start if available
-        if fact.occurred_start:
-            fact_obj["occurred_at"] = fact.occurred_start
-
-        formatted.append(fact_obj)
-
-    return json.dumps(formatted, indent=2)
-
-
-def build_observation_prompt(
-    entity_name: str,
-    facts_text: str,
-) -> str:
-    """Build the observation extraction prompt for the LLM."""
-    return f"""Based on the following facts about "{entity_name}", generate a list of key observations.
-
-FACTS ABOUT {entity_name.upper()}:
-{facts_text}
-
-Your task: Synthesize the facts into clear, objective observations about {entity_name}.
-
-GUIDELINES:
-1. Each observation should be a factual statement about {entity_name}
-2. Combine related facts into single observations where appropriate
-3. Be objective - do not add opinions, judgments, or interpretations
-4. Focus on what we KNOW about {entity_name}, not what we assume
-5. Include observations about: identity, characteristics, roles, relationships, activities
-6. Write in third person (e.g., "John is..." not "I think John is...")
-7. If there are conflicting facts, note the most recent or most supported one
-
-EXAMPLES of good observations:
- "John works at Google as a software engineer"
- "John is detail-oriented and methodical in his approach"
- "John collaborates frequently with Sarah on the AI project"
- "John joined the company in 2023"
-
-EXAMPLES of bad observations (avoid these):
- "John seems like a good person" (opinion/judgment)
- "John probably likes his job" (assumption)
- "I believe John is reliable" (first-person opinion)
-
-Generate 3-7 observations based on the available facts. If there are very few facts, generate fewer observations."""
-
-
-def get_observation_system_message() -> str:
-    """Get the system message for observation extraction."""
-    return "You are an objective observer synthesizing facts about an entity. Generate clear, factual observations without opinions or personality influence. Be concise and accurate."
-
-
-async def extract_observations_from_facts(llm_config, entity_name: str, facts: list[MemoryFact]) -> list[str]:
-    """
-    Extract observations from facts about an entity using LLM.
-
-    Args:
-        llm_config: LLM configuration to use
-        entity_name: Name of the entity to generate observations about
-        facts: List of facts mentioning the entity
-
-    Returns:
-        List of observation strings
-    """
-    if not facts:
-        return []
-
-    facts_text = format_facts_for_observation_prompt(facts)
-    prompt = build_observation_prompt(entity_name, facts_text)
-
-    try:
-        result = await llm_config.call(
-            messages=[
-                {"role": "system", "content": get_observation_system_message()},
-                {"role": "user", "content": prompt},
-            ],
-            response_format=ObservationExtractionResponse,
-            scope="memory_extract_observation",
-        )
-
-        observations = [op.observation for op in result.observations]
-        return observations
-
-    except Exception as e:
-        logger.warning(f"Failed to extract observations for {entity_name}: {str(e)}")
-        return []
@@ -20,6 +20,7 @@ from ..memory_engine import fq_table
 from .graph_retrieval import BFSGraphRetriever, GraphRetriever
 from .link_expansion_retrieval import LinkExpansionRetriever
 from .mpfp_retrieval import MPFPGraphRetriever
+from .tags import TagsMatch, build_tags_where_clause_simple
 from .types import MPFPTimings, RetrievalResult

 logger = logging.getLogger(__name__)
@@ -85,7 +86,12 @@ def set_default_graph_retriever(retriever: GraphRetriever) -> None:


 async def retrieve_semantic(
-    conn, query_emb_str: str, bank_id: str, fact_type: str, limit: int
+    conn,
+    query_emb_str: str,
+    bank_id: str,
+    fact_type: str,
+    limit: int,
+    tags: list[str] | None = None,
 ) -> list[RetrievalResult]:
    """
    Semantic retrieval via vector similarity.
@@ -96,31 +102,44 @@ async def retrieve_semantic(
        agent_id: bank ID
        fact_type: Fact type to filter
        limit: Maximum results to return
+        tags: Optional list of tags for visibility filtering (OR matching)

    Returns:
        List of RetrievalResult objects
    """
+    from .tags import TagsMatch, build_tags_where_clause_simple
+
+    tags_clause = build_tags_where_clause_simple(tags, 5)
+    params = [query_emb_str, bank_id, fact_type, limit]
+    if tags:
+        params.append(tags)
+
    results = await conn.fetch(
        f"""
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
               1 - (embedding <=> $1::vector) AS similarity
        FROM {fq_table("memory_units")}
        WHERE bank_id = $2
          AND embedding IS NOT NULL
          AND fact_type = $3
          AND (1 - (embedding <=> $1::vector)) >= 0.3
+          {tags_clause}
        ORDER BY embedding <=> $1::vector
        LIMIT $4
        """,
-        query_emb_str,
-        bank_id,
-        fact_type,
-        limit,
+        *params,
    )
    return [RetrievalResult.from_db_row(dict(r)) for r in results]


-async def retrieve_bm25(conn, query_text: str, bank_id: str, fact_type: str, limit: int) -> list[RetrievalResult]:
+async def retrieve_bm25(
+    conn,
+    query_text: str,
+    bank_id: str,
+    fact_type: str,
+    limit: int,
+    tags: list[str] | None = None,
+) -> list[RetrievalResult]:
    """
    BM25 keyword retrieval via full-text search.

@@ -130,12 +149,15 @@ async def retrieve_bm25(conn, query_text: str, bank_id: str, fact_type: str, lim
        agent_id: bank ID
        fact_type: Fact type to filter
        limit: Maximum results to return
+        tags: Optional list of tags for visibility filtering (OR matching)

    Returns:
        List of RetrievalResult objects
    """
    import re

+    from .tags import TagsMatch, build_tags_where_clause_simple
+
    # Sanitize query text: remove special characters that have meaning in tsquery
    # Keep only alphanumeric characters and spaces
    sanitized_text = re.sub(r"[^\w\s]", " ", query_text.lower())
@@ -151,21 +173,24 @@ async def retrieve_bm25(conn, query_text: str, bank_id: str, fact_type: str, lim
    # This prevents empty results when some terms are missing
    query_tsquery = " | ".join(tokens)

+    tags_clause = build_tags_where_clause_simple(tags, 5)
+    params = [query_tsquery, bank_id, fact_type, limit]
+    if tags:
+        params.append(tags)
+
    results = await conn.fetch(
        f"""
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
               ts_rank_cd(search_vector, to_tsquery('english', $1)) AS bm25_score
        FROM {fq_table("memory_units")}
        WHERE bank_id = $2
          AND fact_type = $3
          AND search_vector @@ to_tsquery('english', $1)
+          {tags_clause}
        ORDER BY bm25_score DESC
        LIMIT $4
        """,
-        query_tsquery,
-        bank_id,
-        fact_type,
-        limit,
+        *params,
    )
    return [RetrievalResult.from_db_row(dict(r)) for r in results]

@@ -177,6 +202,8 @@ async def retrieve_semantic_bm25_combined(
    bank_id: str,
    fact_types: list[str],
    limit: int,
+    tags: list[str] | None = None,
+    tags_match: TagsMatch = "any",
 ) -> dict[str, tuple[list[RetrievalResult], list[RetrievalResult]]]:
    """
    Combined semantic + BM25 retrieval for multiple fact types in a single query.
@@ -203,10 +230,14 @@ async def retrieve_semantic_bm25_combined(

    # If no valid tokens for BM25, just run semantic
    if not tokens:
+        tags_clause = build_tags_where_clause_simple(tags, 5, match=tags_match)
+        params = [query_emb_str, bank_id, fact_types, limit]
+        if tags:
+            params.append(tags)
        results = await conn.fetch(
            f"""
            WITH semantic_ranked AS (
-                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                       1 - (embedding <=> $1::vector) AS similarity,
                       NULL::float AS bm25_score,
                       'semantic' AS source,
@@ -216,16 +247,14 @@ async def retrieve_semantic_bm25_combined(
                  AND embedding IS NOT NULL
                  AND fact_type = ANY($3)
                  AND (1 - (embedding <=> $1::vector)) >= 0.3
+                  {tags_clause}
            )
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                   similarity, bm25_score, source
            FROM semantic_ranked
            WHERE rn <= $4
            """,
-            query_emb_str,
-            bank_id,
-            fact_types,
-            limit,
+            *params,
        )
        # Group by fact_type
        result_dict: dict[str, tuple[list[RetrievalResult], list[RetrievalResult]]] = {
@@ -241,12 +270,18 @@ async def retrieve_semantic_bm25_combined(

    query_tsquery = " | ".join(tokens)

+    # Build tags clause - param 6 if tags provided
+    tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
+    params = [query_emb_str, bank_id, fact_types, limit, query_tsquery]
+    if tags:
+        params.append(tags)
+
    # Combined CTE query for both semantic and BM25 across all fact types
    # Uses window functions to limit per fact_type per method
    results = await conn.fetch(
        f"""
        WITH semantic_ranked AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                   1 - (embedding <=> $1::vector) AS similarity,
                   NULL::float AS bm25_score,
                   'semantic' AS source,
@@ -256,9 +291,10 @@ async def retrieve_semantic_bm25_combined(
              AND embedding IS NOT NULL
              AND fact_type = ANY($3)
              AND (1 - (embedding <=> $1::vector)) >= 0.3
+              {tags_clause}
        ),
        bm25_ranked AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                   NULL::float AS similarity,
                   ts_rank_cd(search_vector, to_tsquery('english', $5)) AS bm25_score,
                   'bm25' AS source,
@@ -267,14 +303,15 @@ async def retrieve_semantic_bm25_combined(
            WHERE bank_id = $2
              AND fact_type = ANY($3)
              AND search_vector @@ to_tsquery('english', $5)
+              {tags_clause}
        ),
        semantic AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                   similarity, bm25_score, source
            FROM semantic_ranked WHERE rn <= $4
        ),
        bm25 AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                   similarity, bm25_score, source
            FROM bm25_ranked WHERE rn <= $4
        )
@@ -282,11 +319,7 @@ async def retrieve_semantic_bm25_combined(
        UNION ALL
        SELECT * FROM bm25
        """,
-        query_emb_str,
-        bank_id,
-        fact_types,
-        limit,
-        query_tsquery,
+        *params,
    )

    # Group results by fact_type and source
@@ -313,6 +346,8 @@ async def retrieve_temporal_combined(
    end_date: datetime,
    budget: int,
    semantic_threshold: float = 0.1,
+    tags: list[str] | None = None,
+    tags_match: TagsMatch = "any",
 ) -> dict[str, list[RetrievalResult]]:
    """
    Temporal retrieval for multiple fact types in a single query.
@@ -341,11 +376,17 @@ async def retrieve_temporal_combined(
    if end_date.tzinfo is None:
        end_date = end_date.replace(tzinfo=UTC)

+    # Build tags clause
+    tags_clause = build_tags_where_clause_simple(tags, 7, match=tags_match)
+    params = [query_emb_str, bank_id, fact_types, start_date, end_date, semantic_threshold]
+    if tags:
+        params.append(tags)
+
    # Batch query: Get entry points for ALL fact types at once with window function
    entry_points = await conn.fetch(
        f"""
        WITH ranked_entries AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
                   1 - (embedding <=> $1::vector) AS similarity,
                   ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC, embedding <=> $1::vector) AS rn
            FROM {fq_table("memory_units")}
@@ -363,17 +404,13 @@ async def retrieve_temporal_combined(
                  (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
              )
              AND (1 - (embedding <=> $1::vector)) >= $6
+              {tags_clause}
        )
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, similarity
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags, similarity
        FROM ranked_entries
        WHERE rn <= 10
        """,
-        query_emb_str,
-        bank_id,
-        fact_types,
-        start_date,
-        end_date,
-        semantic_threshold,
+        *params,
    )

    if not entry_points:
@@ -436,13 +473,20 @@ async def retrieve_temporal_combined(
        budget_remaining = budget - len(ft_entry_points)
        batch_size = 20

+        # Build tags clause for spreading (use param 6 since 1-5 are used)
+        spreading_tags_clause = build_tags_where_clause_simple(tags, 6, table_alias="mu.", match=tags_match)
+
        while frontier and budget_remaining > 0:
            batch_ids = frontier[:batch_size]
            frontier = frontier[batch_size:]

+            spreading_params = [query_emb_str, batch_ids, ft, semantic_threshold, batch_size * 10]
+            if tags:
+                spreading_params.append(tags)
+
            neighbors = await conn.fetch(
                f"""
-                SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id,
+                SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
                       ml.weight, ml.link_type, ml.from_unit_id,
                       1 - (mu.embedding <=> $1::vector) AS similarity
                FROM {fq_table("memory_links")} ml
@@ -453,14 +497,11 @@ async def retrieve_temporal_combined(
                  AND mu.fact_type = $3
                  AND mu.embedding IS NOT NULL
                  AND (1 - (mu.embedding <=> $1::vector)) >= $4
+                  {spreading_tags_clause}
                ORDER BY ml.weight DESC
                LIMIT $5
                """,
-                query_emb_str,
-                batch_ids,
-                ft,
-                semantic_threshold,
-                batch_size * 10,
+                *spreading_params,
            )

            for n in neighbors:
@@ -529,6 +570,7 @@ async def retrieve_temporal(
    end_date: datetime,
    budget: int,
    semantic_threshold: float = 0.1,
+    tags: list[str] | None = None,
 ) -> list[RetrievalResult]:
    """
    Temporal retrieval with spreading activation.
@@ -547,6 +589,7 @@ async def retrieve_temporal(
        end_date: End of time range
        budget: Node budget for spreading
        semantic_threshold: Minimum semantic similarity to include
+        tags: Optional list of tags for visibility filtering (OR matching)

    Returns:
        List of RetrievalResult objects with temporal scores
@@ -558,9 +601,16 @@ async def retrieve_temporal(
    if end_date.tzinfo is None:
        end_date = end_date.replace(tzinfo=UTC)

+    from .tags import TagsMatch, build_tags_where_clause_simple
+
+    tags_clause = build_tags_where_clause_simple(tags, 7)
+    params = [query_emb_str, bank_id, fact_type, start_date, end_date, semantic_threshold]
+    if tags:
+        params.append(tags)
+
    entry_points = await conn.fetch(
        f"""
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
+        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
               1 - (embedding <=> $1::vector) AS similarity
        FROM {fq_table("memory_units")}
        WHERE bank_id = $2
@@ -580,15 +630,11 @@ async def retrieve_temporal(
              (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
          )
          AND (1 - (embedding <=> $1::vector)) >= $6
+          {tags_clause}
        ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC, (embedding <=> $1::vector) ASC
        LIMIT 10
        """,
-        query_emb_str,
-        bank_id,
-        fact_type,
-        start_date,
-        end_date,
-        semantic_threshold,
+        *params,
    )

    if not entry_points:
@@ -740,6 +786,7 @@ async def retrieve_parallel(
    query_analyzer: Optional["QueryAnalyzer"] = None,
    graph_retriever: GraphRetriever | None = None,
    temporal_constraint: tuple | None = None,  # Pre-extracted temporal constraint
+    tags: list[str] | None = None,  # Visibility scope tags for filtering
 ) -> ParallelRetrievalResult:
    """
    Run 3-way or 4-way parallel retrieval (adds temporal if detected).
@@ -755,6 +802,7 @@ async def retrieve_parallel(
        query_analyzer: Query analyzer to use (defaults to TransformerQueryAnalyzer)
        graph_retriever: Graph retrieval strategy (defaults to configured retriever)
        temporal_constraint: Pre-extracted temporal constraint (optional)
+        tags: Optional list of tags for visibility filtering (OR matching)

    Returns:
        ParallelRetrievalResult with semantic, bm25, graph, temporal results and timings
@@ -775,6 +823,7 @@ async def retrieve_parallel(
            retriever,
            question_date,
            query_analyzer,
+            tags=tags,
        )
    else:
        # For BFS, extract temporal constraint upfront (legacy path)
@@ -785,7 +834,15 @@ async def retrieve_parallel(
                query_text, reference_date=question_date, analyzer=query_analyzer
            )
        return await _retrieve_parallel_bfs(
-            pool, query_text, query_embedding_str, bank_id, fact_type, thinking_budget, temporal_constraint, retriever
+            pool,
+            query_text,
+            query_embedding_str,
+            bank_id,
+            fact_type,
+            thinking_budget,
+            temporal_constraint,
+            retriever,
+            tags=tags,
        )


@@ -809,6 +866,7 @@ async def _retrieve_parallel_mpfp(
    retriever: GraphRetriever,
    question_date: datetime | None = None,
    query_analyzer=None,
+    tags: list[str] | None = None,
 ) -> ParallelRetrievalResult:
    """
    MPFP retrieval with true parallelization.
@@ -830,7 +888,9 @@ async def _retrieve_parallel_mpfp(
        acquire_start = time.time()
        async with acquire_with_retry(pool) as conn:
            conn_wait = time.time() - acquire_start
-            results = await retrieve_semantic(conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget)
+            results = await retrieve_semantic(
+                conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget, tags=tags
+            )
        return _TimedResult(results, time.time() - start, conn_wait)

    async def run_bm25() -> _TimedResult:
@@ -839,7 +899,7 @@ async def _retrieve_parallel_mpfp(
        acquire_start = time.time()
        async with acquire_with_retry(pool) as conn:
            conn_wait = time.time() - acquire_start
-            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget)
+            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget, tags=tags)
        return _TimedResult(results, time.time() - start, conn_wait)

    async def run_graph() -> tuple[list[RetrievalResult], float, MPFPTimings | None]:
@@ -857,6 +917,7 @@ async def _retrieve_parallel_mpfp(
            query_text=query_text,
            semantic_seeds=None,  # Let MPFP find its own seeds
            temporal_seeds=None,  # Don't wait for temporal extraction
+            tags=tags,
        )
        return results, time.time() - start, mpfp_timing

@@ -1028,6 +1089,7 @@ async def _retrieve_parallel_bfs(
    thinking_budget: int,
    temporal_constraint: tuple | None,
    retriever: GraphRetriever,
+    tags: list[str] | None = None,
 ) -> ParallelRetrievalResult:
    """BFS retrieval: all methods run in parallel (original behavior)."""
    import time
@@ -1035,13 +1097,15 @@ async def _retrieve_parallel_bfs(
    async def run_semantic() -> _TimedResult:
        start = time.time()
        async with acquire_with_retry(pool) as conn:
-            results = await retrieve_semantic(conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget)
+            results = await retrieve_semantic(
+                conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget, tags=tags
+            )
        return _TimedResult(results, time.time() - start)

    async def run_bm25() -> _TimedResult:
        start = time.time()
        async with acquire_with_retry(pool) as conn:
-            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget)
+            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget, tags=tags)
        return _TimedResult(results, time.time() - start)

    async def run_graph() -> _TimedResult:
@@ -1053,6 +1117,7 @@ async def _retrieve_parallel_bfs(
            fact_type=fact_type,
            budget=thinking_budget,
            query_text=query_text,
+            tags=tags,
        )
        return _TimedResult(results, time.time() - start)

@@ -1068,6 +1133,7 @@ async def _retrieve_parallel_bfs(
                tc_end,
                budget=thinking_budget,
                semantic_threshold=0.1,
+                tags=tags,
            )
        return _TimedResult(results, time.time() - start)

@@ -1122,6 +1188,8 @@ async def retrieve_all_fact_types_parallel(
    question_date: datetime | None = None,
    query_analyzer: Optional["QueryAnalyzer"] = None,
    graph_retriever: GraphRetriever | None = None,
+    tags: list[str] | None = None,
+    tags_match: TagsMatch = "any",
 ) -> MultiFactTypeRetrievalResult:
    """
    Optimized retrieval for multiple fact types using batched queries.
@@ -1171,7 +1239,14 @@ async def retrieve_all_fact_types_parallel(

        # Semantic + BM25 combined
        semantic_bm25_results = await retrieve_semantic_bm25_combined(
-            conn, query_embedding_str, query_text, bank_id, fact_types, thinking_budget
+            conn,
+            query_embedding_str,
+            query_text,
+            bank_id,
+            fact_types,
+            thinking_budget,
+            tags=tags,
+            tags_match=tags_match,
        )
        semantic_bm25_time = time.time() - semantic_bm25_start

@@ -1188,6 +1263,8 @@ async def retrieve_all_fact_types_parallel(
                tc_end,
                budget=thinking_budget,
                semantic_threshold=0.1,
+                tags=tags,
+                tags_match=tags_match,
            )
            temporal_time = time.time() - temporal_start

@@ -1206,6 +1283,8 @@ async def retrieve_all_fact_types_parallel(
            query_text=query_text,
            semantic_seeds=None,
            temporal_seeds=None,
+            tags=tags,
+            tags_match=tags_match,
        )
        return ft, results, time.time() - graph_start, mpfp_timing

@@ -0,0 +1,172 @@
+"""
+Tags filtering utilities for retrieval.
+
+Provides SQL building functions for filtering memories by tags.
+Supports four matching modes via TagsMatch enum:
+- "any": OR matching, includes untagged memories (default, backward compatible)
+- "all": AND matching, includes untagged memories
+- "any_strict": OR matching, excludes untagged memories
+- "all_strict": AND matching, excludes untagged memories
+
+OR matching (any/any_strict): Memory matches if ANY of its tags overlap with request tags
+AND matching (all/all_strict): Memory matches if ALL request tags are present in its tags
+"""
+
+from typing import Literal
+
+TagsMatch = Literal["any", "all", "any_strict", "all_strict"]
+
+
+def _parse_tags_match(match: TagsMatch) -> tuple[str, bool]:
+    """
+    Parse TagsMatch into operator and include_untagged flag.
+
+    Returns:
+        Tuple of (operator, include_untagged)
+        - operator: "&&" for any/any_strict, "@>" for all/all_strict
+        - include_untagged: True for any/all, False for any_strict/all_strict
+    """
+    if match == "any":
+        return "&&", True
+    elif match == "all":
+        return "@>", True
+    elif match == "any_strict":
+        return "&&", False
+    elif match == "all_strict":
+        return "@>", False
+    else:
+        # Default to "any" behavior
+        return "&&", True
+
+
+def build_tags_where_clause(
+    tags: list[str] | None,
+    param_offset: int = 1,
+    table_alias: str = "",
+    match: TagsMatch = "any",
+) -> tuple[str, list, int]:
+    """
+    Build a SQL WHERE clause for filtering by tags.
+
+    Supports four matching modes:
+    - "any" (default): OR matching, includes untagged memories
+    - "all": AND matching, includes untagged memories
+    - "any_strict": OR matching, excludes untagged memories
+    - "all_strict": AND matching, excludes untagged memories
+
+    Args:
+        tags: List of tags to filter by. If None or empty, returns empty clause (no filtering).
+        param_offset: Starting parameter number for SQL placeholders (default 1).
+        table_alias: Optional table alias prefix (e.g., "mu." for "memory_units mu").
+        match: Matching mode. Defaults to "any".
+
+    Returns:
+        Tuple of (sql_clause, params, next_param_offset):
+        - sql_clause: SQL WHERE clause string
+        - params: List of parameter values to bind
+        - next_param_offset: Next available parameter number
+
+    Example:
+        >>> clause, params, next_offset = build_tags_where_clause(['user_a'], 3, 'mu.', 'any_strict')
+        >>> print(clause)  # "AND mu.tags IS NOT NULL AND mu.tags != '{}' AND mu.tags && $3"
+    """
+    if not tags:
+        return "", [], param_offset
+
+    column = f"{table_alias}tags" if table_alias else "tags"
+    operator, include_untagged = _parse_tags_match(match)
+
+    if include_untagged:
+        # Include untagged memories (NULL or empty array) OR matching tags
+        clause = f"AND ({column} IS NULL OR {column} = '{{}}' OR {column} {operator} ${param_offset})"
+    else:
+        # Strict: only memories with matching tags (exclude NULL and empty)
+        clause = f"AND {column} IS NOT NULL AND {column} != '{{}}' AND {column} {operator} ${param_offset}"
+
+    return clause, [tags], param_offset + 1
+
+
+def build_tags_where_clause_simple(
+    tags: list[str] | None,
+    param_num: int,
+    table_alias: str = "",
+    match: TagsMatch = "any",
+) -> str:
+    """
+    Build a simple SQL WHERE clause for tags filtering.
+
+    This is a convenience version that returns just the clause string,
+    assuming the caller will add the tags array to their params list.
+
+    Args:
+        tags: List of tags to filter by. If None or empty, returns empty string.
+        param_num: Parameter number to use in the clause.
+        table_alias: Optional table alias prefix.
+        match: Matching mode. Defaults to "any".
+
+    Returns:
+        SQL clause string or empty string.
+    """
+    if not tags:
+        return ""
+
+    column = f"{table_alias}tags" if table_alias else "tags"
+    operator, include_untagged = _parse_tags_match(match)
+
+    if include_untagged:
+        # Include untagged memories (NULL or empty array) OR matching tags
+        return f"AND ({column} IS NULL OR {column} = '{{}}' OR {column} {operator} ${param_num})"
+    else:
+        # Strict: only memories with matching tags (exclude NULL and empty)
+        return f"AND {column} IS NOT NULL AND {column} != '{{}}' AND {column} {operator} ${param_num}"
+
+
+def filter_results_by_tags(
+    results: list,
+    tags: list[str] | None,
+    match: TagsMatch = "any",
+) -> list:
+    """
+    Filter retrieval results by tags in Python (for post-processing).
+
+    Used when SQL filtering isn't possible (e.g., graph traversal results).
+
+    Args:
+        results: List of RetrievalResult objects with a 'tags' attribute.
+        tags: List of tags to filter by. If None or empty, returns all results.
+        match: Matching mode. Defaults to "any".
+
+    Returns:
+        Filtered list of results.
+    """
+    if not tags:
+        return results
+
+    _, include_untagged = _parse_tags_match(match)
+    is_any_match = match in ("any", "any_strict")
+
+    tags_set = set(tags)
+    filtered = []
+
+    for result in results:
+        result_tags = getattr(result, "tags", None)
+
+        # Check if untagged
+        is_untagged = result_tags is None or len(result_tags) == 0
+
+        if is_untagged:
+            if include_untagged:
+                filtered.append(result)
+            # else: skip untagged
+        else:
+            result_tags_set = set(result_tags)
+            if is_any_match:
+                # Any overlap
+                if result_tags_set & tags_set:
+                    filtered.append(result)
+            else:
+                # All tags must be present
+                if tags_set <= result_tags_set:
+                    filtered.append(result)
+
+    return filtered
@@ -3,31 +3,13 @@ Think operation utilities for formulating answers based on agent and world facts
 """

 import logging
-import re
 from datetime import datetime

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

 logger = logging.getLogger(__name__)


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


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

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

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

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

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

 QUESTION: {query}

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


 def get_system_message(disposition: DispositionTraits) -> str:
@@ -175,122 +191,11 @@ def get_system_message(disposition: DispositionTraits) -> str:
    return f"You are a person with your own thoughts, experiences, opinions, and disposition. When asked a question, respond naturally like a human would - share what you genuinely think based on what you know and have experienced. {disposition_instruction} Be direct, express your views confidently, and use 'I think', 'I believe', 'in my view', etc. Respond in plain text without markdown formatting. IMPORTANT: Detect the language of the question and respond in the SAME language. Do not translate to English if the question is in another language."


-async def extract_opinions_from_text(llm_config, text: str, query: str) -> list[Opinion]:
-    """
-    Extract opinions with reasons and confidence from text using LLM.
-
-    Args:
-        llm_config: LLM configuration to use
-        text: Text to extract opinions from
-        query: The original query that prompted this response
-
-    Returns:
-        List of Opinion objects with text and confidence
-    """
-    extraction_prompt = f"""Extract any NEW opinions or perspectives from the answer below and rewrite them in FIRST-PERSON as if YOU are stating the opinion directly.
-
-ORIGINAL QUESTION:
-{query}
-
-ANSWER PROVIDED:
-{text}
-
-Your task: Find opinions in the answer and rewrite them AS IF YOU ARE THE ONE SAYING THEM.
-
-An opinion is a judgment, viewpoint, or conclusion that goes beyond just stating facts.
-
-IMPORTANT: Do NOT extract statements like:
- "I don't have enough information"
- "The facts don't contain information about X"
- "I cannot answer because..."
-
-ONLY extract actual opinions about substantive topics.
-
-CRITICAL FORMAT REQUIREMENTS:
-1. **ALWAYS start with first-person phrases**: "I think...", "I believe...", "In my view...", "I've come to believe...", "Previously I thought... but now..."
-2. **NEVER use third-person**: Do NOT say "The speaker thinks..." or "They believe..." - always use "I"
-3. Include the reasoning naturally within the statement
-4. Provide a confidence score (0.0 to 1.0)
-
-CORRECT Examples (✓ FIRST-PERSON):
- "I think Alice is more reliable because she consistently delivers on time and writes clean code"
- "Previously I thought all engineers were equal, but now I feel that experience and track record really matter"
- "I believe reliability is best measured by consistent output over time"
- "I've come to believe that track records are more important than potential"
-
-WRONG Examples (✗ THIRD-PERSON - DO NOT USE):
- "The speaker thinks Alice is more reliable"
- "They believe reliability matters"
- "It is believed that Alice is better"
-
-If no genuine opinions are expressed (e.g., the response just says "I don't know"), return an empty list."""
-
-    try:
-        result = await llm_config.call(
-            messages=[
-                {
-                    "role": "system",
-                    "content": "You are converting opinions from text into first-person statements. Always use 'I think', 'I believe', 'I feel', etc. NEVER use third-person like 'The speaker' or 'They'.",
-                },
-                {"role": "user", "content": extraction_prompt},
-            ],
-            response_format=OpinionExtractionResponse,
-            scope="memory_extract_opinion",
-        )
-
-        # Format opinions with confidence score and convert to first-person
-        formatted_opinions = []
-        for op in result.opinions:
-            # Convert third-person to first-person if needed
-            opinion_text = op.opinion
-
-            # Replace common third-person patterns with first-person
-            def singularize_verb(verb):
-                if verb.endswith("es"):
-                    return verb[:-1]  # believes -> believe
-                elif verb.endswith("s"):
-                    return verb[:-1]  # thinks -> think
-                return verb
-
-            # Pattern: "The speaker/user [verb]..." -> "I [verb]..."
-            match = re.match(
-                r"^(The speaker|The user|They|It is believed) (believes?|thinks?|feels?|says|asserts?|considers?)(\s+that)?(.*)$",
-                opinion_text,
-                re.IGNORECASE,
-            )
-            if match:
-                verb = singularize_verb(match.group(2))
-                that_part = match.group(3) or ""  # Keep " that" if present
-                rest = match.group(4)
-                opinion_text = f"I {verb}{that_part}{rest}"
-
-            # If still doesn't start with first-person, prepend "I believe that "
-            first_person_starters = [
-                "I think",
-                "I believe",
-                "I feel",
-                "In my view",
-                "I've come to believe",
-                "Previously I",
-            ]
-            if not any(opinion_text.startswith(starter) for starter in first_person_starters):
-                opinion_text = "I believe that " + opinion_text[0].lower() + opinion_text[1:]
-
-            formatted_opinions.append(Opinion(opinion=opinion_text, confidence=op.confidence))
-
-        return formatted_opinions
-
-    except Exception as e:
-        logger.warning(f"Failed to extract opinions: {str(e)}")
-        return []
-
-
 async def reflect(
    llm_config,
    query: str,
    experience_facts: list[str] = None,
    world_facts: list[str] = None,
-    opinion_facts: list[str] = None,
    name: str = "Assistant",
    disposition: DispositionTraits = None,
    background: str = "",
@@ -307,7 +212,6 @@ async def reflect(
        query: Question to answer
        experience_facts: List of experience/agent fact strings
        world_facts: List of world fact strings
-        opinion_facts: List of opinion fact strings
        name: Name of the agent/persona
        disposition: Disposition traits (defaults to neutral)
        background: Background information
@@ -328,18 +232,15 @@ async def reflect(

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

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

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


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

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


 class EntryPoint(BaseModel):
@@ -22,6 +22,7 @@ from .trace import (
    SearchPhaseMetrics,
    SearchSummary,
    SearchTrace,
+    TemporalConstraint,
    WeightComponents,
 )

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

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

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

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

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

+    def record_temporal_constraint(self, start: datetime | None, end: datetime | None):
+        """Record the detected temporal constraint from query analysis."""
+        if start is not None or end is not None:
+            self.temporal_constraint = TemporalConstraint(start=start, end=end)
+
    def add_entry_point(self, node_id: str, text: str, similarity: float, rank: int):
        """
        Record an entry point.
@@ -428,6 +448,9 @@ class SearchTracer:
            timestamp=datetime.now(UTC),
            budget=self.budget,
            max_tokens=self.max_tokens,
+            tags=self.tags,
+            tags_match=self.tags_match,
+            temporal_constraint=self.temporal_constraint,
        )

        # Create summary
@@ -48,6 +48,7 @@ class RetrievalResult:
    chunk_id: str | None = None
    access_count: int = 0
    embedding: list[float] | None = None
+    tags: list[str] | None = None  # Visibility scope tags

    # Retrieval-specific scores (only one will be set depending on retrieval method)
    similarity: float | None = None  # Semantic retrieval
@@ -72,6 +73,7 @@ class RetrievalResult:
            chunk_id=row.get("chunk_id"),
            access_count=row.get("access_count", 0),
            embedding=row.get("embedding"),
+            tags=row.get("tags"),
            similarity=row.get("similarity"),
            bm25_score=row.get("bm25_score"),
            activation=row.get("activation"),
@@ -156,6 +158,7 @@ class ScoredResult:
            "chunk_id": self.retrieval.chunk_id,
            "access_count": self.retrieval.access_count,
            "embedding": self.retrieval.embedding,
+            "tags": self.retrieval.tags,
            "semantic_similarity": self.retrieval.similarity,
            "bm25_score": self.retrieval.bm25_score,
        }
@@ -1,31 +1,40 @@
 """
-Abstract task backend for running async tasks.
+Task backend for distributed task processing.

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

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

 logger = logging.getLogger(__name__)


+def fq_table(table: str, schema: str | None = None) -> str:
+    """Get fully-qualified table name with optional schema prefix."""
+    if schema:
+        return f'"{schema}".{table}'
+    return table
+
+
 class TaskBackend(ABC):
    """
    Abstract base class for task execution backends.

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

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

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

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

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

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

    async def initialize(self):
@@ -121,221 +129,123 @@ class SyncTaskBackend(TaskBackend):
        logger.debug("SyncTaskBackend shutdown")


-class NoopTaskBackend(TaskBackend):
+class BrokerTaskBackend(TaskBackend):
    """
-    No-op task backend that discards all tasks.
+    Task backend using PostgreSQL as broker.

-    This is useful for tests where background task execution is not needed
-    and would only slow down the test suite.
+    submit_task() stores task_payload in async_operations table.
+    Actual polling and execution is handled separately by WorkerPoller.
+
+    This backend is used by the API to store tasks. Workers poll
+    the database separately to claim and execute tasks.
    """

-    async def initialize(self):
-        """No-op."""
-        self._initialized = True
-        logger.debug("NoopTaskBackend initialized")
-
-    async def submit_task(self, task_dict: dict[str, Any]):
-        """Discard the task (do nothing)."""
-        pass
-
-    async def shutdown(self):
-        """No-op."""
-        self._initialized = False
-        logger.debug("NoopTaskBackend shutdown")
-
-
-class AsyncIOQueueBackend(TaskBackend):
-    """
-    Task backend implementation using asyncio queues.
-
-    This is the default implementation that uses in-process asyncio queues
-    and a periodic consumer worker.
-    """
-
-    def __init__(self, batch_size: int = 10, batch_interval: float = 1.0):
+    def __init__(
+        self,
+        pool_getter: Callable[[], "asyncpg.Pool"],
+        schema: str | None = None,
+    ):
        """
-        Initialize AsyncIO queue backend.
+        Initialize the broker task backend.

        Args:
-            batch_size: Maximum number of tasks to process in one batch
-            batch_interval: Maximum time (seconds) to wait before processing batch
+            pool_getter: Callable that returns the asyncpg connection pool
+            schema: Database schema for multi-tenant support (optional)
        """
        super().__init__()
-        self._queue: asyncio.Queue | None = None
-        self._worker_task: asyncio.Task | None = None
-        self._shutdown_event: asyncio.Event | None = None
-        self._batch_size = batch_size
-        self._batch_interval = batch_interval
-        self._in_flight_count = 0
-        self._in_flight_lock = asyncio.Lock()
+        self._pool_getter = pool_getter
+        self._schema = schema

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

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

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

-        await self._queue.put(task_dict)
+        pool = self._pool_getter()
+        operation_id = task_dict.get("operation_id")
+        task_type = task_dict.get("type", "unknown")
+        bank_id = task_dict.get("bank_id")
+        payload_json = json.dumps(task_dict)
+
+        table = fq_table("async_operations", self._schema)
+
+        if operation_id:
+            # Update existing operation with task payload
+            await pool.execute(
+                f"""
+                UPDATE {table}
+                SET task_payload = $1::jsonb, updated_at = now()
+                WHERE operation_id = $2
+                """,
+                payload_json,
+                operation_id,
+            )
+            logger.debug(f"Updated task payload for operation {operation_id}")
+        else:
+            # Insert new operation (for tasks without pre-created records)
+            # e.g., access_count_update tasks
+            import uuid
+
+            new_id = uuid.uuid4()
+            await pool.execute(
+                f"""
+                INSERT INTO {table} (operation_id, bank_id, operation_type, status, task_payload)
+                VALUES ($1, $2, $3, 'pending', $4::jsonb)
+                """,
+                new_id,
+                bank_id,
+                task_type,
+                payload_json,
+            )
+            logger.debug(f"Created new operation {new_id} for task type {task_type}")
+
+    async def shutdown(self):
+        """Shutdown the backend."""
+        self._initialized = False
+        logger.info("BrokerTaskBackend shutdown")

    async def wait_for_pending_tasks(self, timeout: float = 120.0):
        """
-        Wait for all pending tasks in the queue and in-flight tasks to complete.
+        Wait for pending tasks to be processed.

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

        Args:
-            timeout: Maximum time to wait in seconds (default 120s for long-running tasks)
+            timeout: Maximum time to wait in seconds
        """
-        if not self._initialized or self._queue is None:
-            return
+        import asyncio
+
+        pool = self._pool_getter()
+        table = fq_table("async_operations", self._schema)

-        # Wait for queue to be empty AND no in-flight tasks
        start_time = asyncio.get_event_loop().time()
        while asyncio.get_event_loop().time() - start_time < timeout:
-            async with self._in_flight_lock:
-                in_flight = self._in_flight_count
+            # Check if there are any pending tasks with payloads
+            count = await pool.fetchval(
+                f"""
+                SELECT COUNT(*) FROM {table}
+                WHERE status = 'pending' AND task_payload IS NOT NULL
+                """
+            )

-            if self._queue.empty() and in_flight == 0:
-                # Queue is empty and no tasks in flight, we're done
+            if count == 0:
                return

-            # Wait a bit before checking again
            await asyncio.sleep(0.5)

-    async def shutdown(self):
-        """Shutdown the worker and drain the queue."""
-        if not self._initialized:
-            return
-
-        logger.info("Shutting down AsyncIOQueueBackend...")
-
-        # Signal shutdown
-        self._shutdown_event.set()
-
-        # Cancel worker
-        if self._worker_task is not None:
-            self._worker_task.cancel()
-            try:
-                await self._worker_task
-            except asyncio.CancelledError:
-                pass  # Worker cancelled successfully
-
-        self._initialized = False
-        logger.info("AsyncIOQueueBackend shutdown complete")
-
-    async def _execute_task_with_tracking(self, task_dict: dict[str, Any]):
-        """Execute a task and track its in-flight status."""
-        async with self._in_flight_lock:
-            self._in_flight_count += 1
-        try:
-            await self._execute_task(task_dict)
-        finally:
-            async with self._in_flight_lock:
-                self._in_flight_count -= 1
-
-    async def _execute_task_no_tracking(self, task_dict: dict[str, Any]):
-        """Execute a task without in-flight tracking (tracking done at batch level)."""
-        await self._execute_task(task_dict)
-
-    def _get_queue_stats(self) -> tuple[int, dict[str, int]]:
-        """Get current queue size and bank_id distribution."""
-        queue_size = self._queue.qsize() if self._queue else 0
-        bank_distribution: dict[str, int] = {}
-
-        if queue_size > 0 and self._queue:
-            # Peek at queue items without removing them
-            # Note: This is a snapshot and may not be perfectly accurate due to concurrency
-            try:
-                # Access internal deque for logging purposes only
-                items = list(self._queue._queue)  # type: ignore[attr-defined]
-                for item in items:
-                    bank_id = item.get("bank_id", "unknown")
-                    bank_distribution[bank_id] = bank_distribution.get(bank_id, 0) + 1
-            except Exception:
-                pass  # Queue access failed, return empty distribution
-
-        return queue_size, bank_distribution
-
-    async def _worker(self):
-        """
-        Background worker that processes tasks in batches.
-
-        Collects tasks for up to batch_interval seconds or batch_size items,
-        then processes them.
-        """
-        while not self._shutdown_event.is_set():
-            try:
-                # Collect tasks for batching
-                tasks = []
-                deadline = asyncio.get_event_loop().time() + self._batch_interval
-
-                while len(tasks) < self._batch_size and asyncio.get_event_loop().time() < deadline:
-                    try:
-                        remaining_time = max(0.1, deadline - asyncio.get_event_loop().time())
-                        task_dict = await asyncio.wait_for(self._queue.get(), timeout=remaining_time)
-                        # Track task as in-flight immediately when picked up from queue
-                        # This prevents wait_for_pending_tasks from returning too early
-                        async with self._in_flight_lock:
-                            self._in_flight_count += 1
-                        tasks.append(task_dict)
-                    except TimeoutError:
-                        break
-
-                # Process batch
-                if tasks:
-                    # Log batch start with queue stats
-                    queue_size, bank_distribution = self._get_queue_stats()
-
-                    # Summarize batch by task type and bank
-                    batch_summary: dict[str, dict[str, int]] = {}
-                    for task_dict in tasks:
-                        task_type = task_dict.get("type", "unknown")
-                        bank_id = task_dict.get("bank_id", "unknown")
-                        if task_type not in batch_summary:
-                            batch_summary[task_type] = {}
-                        batch_summary[task_type][bank_id] = batch_summary[task_type].get(bank_id, 0) + 1
-
-                    # Build log message
-                    batch_parts = []
-                    for task_type, banks in sorted(batch_summary.items()):
-                        bank_str = ", ".join(f"{b}:{c}" for b, c in sorted(banks.items()))
-                        batch_parts.append(f"{task_type}[{bank_str}]")
-                    batch_str = ", ".join(batch_parts)
-
-                    if queue_size > 0:
-                        pending_str = ", ".join(f"{k}:{v}" for k, v in sorted(bank_distribution.items()))
-                        logger.info(
-                            f"Processing {len(tasks)} tasks: {batch_str} (pending={queue_size} [{pending_str}])"
-                        )
-                    else:
-                        logger.info(f"Processing {len(tasks)} tasks: {batch_str}")
-
-                    # Execute tasks concurrently (in_flight already tracked when picked up)
-                    await asyncio.gather(
-                        *[self._execute_task_no_tracking(task_dict) for task_dict in tasks], return_exceptions=True
-                    )
-
-                    # Decrement in_flight count after all tasks complete
-                    async with self._in_flight_lock:
-                        self._in_flight_count -= len(tasks)
-
-            except asyncio.CancelledError:
-                break
-            except Exception as e:
-                logger.error(f"Worker error: {e}")
-                await asyncio.sleep(1)  # Backoff on error
+        logger.warning(f"Timeout waiting for pending tasks after {timeout}s")
@@ -27,6 +27,8 @@ from hindsight_api.extensions.operation_validator import (
    RecallResult,
    ReflectContext,
    ReflectResultContext,
+    RefreshMentalModelContext,
+    RefreshMentalModelResult,
    RetainContext,
    RetainResult,
    ValidationResult,
@@ -54,6 +56,8 @@ __all__ = [
    "RecallResult",
    "ReflectContext",
    "ReflectResultContext",
+    "RefreshMentalModelContext",
+    "RefreshMentalModelResult",
    "RetainContext",
    "RetainResult",
    "ValidationResult",
@@ -97,6 +97,18 @@ class ReflectContext:
    context: str | None = None


+@dataclass
+class RefreshMentalModelContext:
+    """Context for a refresh mental model operation validation (pre-operation).
+
+    Contains ALL user-provided parameters for the refresh mental model operation.
+    """
+
+    bank_id: str
+    model_id: str
+    request_context: "RequestContext"
+
+
 # =============================================================================
 # Post-operation Contexts (includes results)
 # =============================================================================
@@ -164,6 +176,27 @@ class ReflectResultContext:
    error: str | None = None


+@dataclass
+class RefreshMentalModelResult:
+    """Result context for post-refresh-mental-model hook.
+
+    Contains the operation parameters and the result including token usage.
+    """
+
+    bank_id: str
+    model_id: str
+    request_context: "RequestContext"
+    # Result
+    model_name: str | None = None
+    observations_count: int = 0
+    input_tokens: int = 0
+    output_tokens: int = 0
+    total_tokens: int = 0
+    duration_ms: int = 0
+    success: bool = True
+    error: str | None = None
+
+
 class OperationValidatorExtension(Extension, ABC):
    """
    Validates and hooks into retain/recall/reflect operations.
@@ -265,6 +298,25 @@ class OperationValidatorExtension(Extension, ABC):
        """
        ...

+    @abstractmethod
+    async def validate_refresh_mental_model(self, ctx: RefreshMentalModelContext) -> ValidationResult:
+        """
+        Validate a refresh mental model operation before execution.
+
+        Called before the refresh mental model operation is processed.
+        Return ValidationResult.reject() to prevent the operation from executing.
+
+        Args:
+            ctx: Context containing all user-provided parameters:
+                - bank_id: Bank identifier
+                - model_id: Mental model ID to refresh
+                - request_context: Request context with auth info
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        ...
+
    # =========================================================================
    # Post-operation hooks (optional - override to implement)
    # =========================================================================
@@ -325,3 +377,28 @@ class OperationValidatorExtension(Extension, ABC):
                - error: Error message (if failed)
        """
        pass
+
+    async def on_refresh_mental_model_complete(self, result: RefreshMentalModelResult) -> None:
+        """
+        Called after a refresh mental model operation completes (success or failure).
+
+        Override this method to implement post-operation logic such as:
+        - Token usage tracking and billing
+        - Audit logging
+        - Metrics collection
+
+        Args:
+            result: Result context containing:
+                - bank_id: Bank identifier
+                - model_id: Mental model ID
+                - request_context: Request context with auth info
+                - model_name: Name of the mental model (if success)
+                - observations_count: Number of observations generated
+                - input_tokens: Number of input tokens used
+                - output_tokens: Number of output tokens used
+                - total_tokens: Total tokens used (input + output)
+                - duration_ms: Total operation duration in milliseconds
+                - success: Whether the operation succeeded
+                - error: Error message (if failed)
+        """
+        pass
@@ -187,15 +187,19 @@ def main():
            embeddings_provider=config.embeddings_provider,
            embeddings_local_model=config.embeddings_local_model,
            embeddings_tei_url=config.embeddings_tei_url,
+            embeddings_openai_base_url=config.embeddings_openai_base_url,
+            embeddings_cohere_base_url=config.embeddings_cohere_base_url,
            reranker_provider=config.reranker_provider,
            reranker_local_model=config.reranker_local_model,
            reranker_tei_url=config.reranker_tei_url,
            reranker_tei_batch_size=config.reranker_tei_batch_size,
            reranker_tei_max_concurrent=config.reranker_tei_max_concurrent,
            reranker_max_candidates=config.reranker_max_candidates,
+            reranker_cohere_base_url=config.reranker_cohere_base_url,
            host=args.host,
            port=args.port,
            log_level=args.log_level,
+            log_format=config.log_format,
            mcp_enabled=config.mcp_enabled,
            graph_retriever=config.graph_retriever,
            mpfp_top_k_neighbors=config.mpfp_top_k_neighbors,
@@ -215,9 +219,14 @@ def main():
            db_pool_max_size=config.db_pool_max_size,
            db_command_timeout=config.db_command_timeout,
            db_acquire_timeout=config.db_acquire_timeout,
-            task_backend=config.task_backend,
-            task_backend_memory_batch_size=config.task_backend_memory_batch_size,
-            task_backend_memory_batch_interval=config.task_backend_memory_batch_interval,
+            worker_enabled=config.worker_enabled,
+            worker_id=config.worker_id,
+            worker_poll_interval_ms=config.worker_poll_interval_ms,
+            worker_max_retries=config.worker_max_retries,
+            worker_batch_size=config.worker_batch_size,
+            worker_http_port=config.worker_http_port,
+            reflect_max_iterations=config.reflect_max_iterations,
+            mental_model_refresh_concurrency=config.mental_model_refresh_concurrency,
        )
    config.configure_logging()
    if not args.daemon:
@@ -28,6 +28,15 @@ from opentelemetry.sdk.resources import Resource
 if TYPE_CHECKING:
    import asyncpg

+
+def _get_tenant() -> str:
+    """Get current tenant (schema) from context for metrics labeling."""
+    # Import here to avoid circular imports
+    from hindsight_api.engine.memory_engine import get_current_schema
+
+    return get_current_schema()
+
+
 # Custom bucket boundaries for operation duration (in seconds)
 # Fine granularity in 0-30s range where most operations complete
 DURATION_BUCKETS = (0.1, 0.25, 0.5, 0.75, 1.0, 2.0, 3.0, 5.0, 7.5, 10.0, 15.0, 20.0, 30.0, 60.0, 120.0)
@@ -323,6 +332,7 @@ class MetricsCollector(MetricsCollectorBase):
            "operation": operation,
            "bank_id": bank_id,
            "source": source,
+            "tenant": _get_tenant(),
        }
        if budget:
            attributes["budget"] = budget
@@ -373,6 +383,7 @@ class MetricsCollector(MetricsCollectorBase):
            "model": model,
            "scope": scope,
            "success": str(success).lower(),
+            "tenant": _get_tenant(),
        }

        # Record duration
@@ -425,10 +436,14 @@ class MetricsCollector(MetricsCollectorBase):
            status_code = status_code_getter()
            status_class = f"{status_code // 100}xx"

+            # Get tenant from context (may be set during request processing)
+            tenant = _get_tenant()
+
            attributes = {
                **base_attributes,
                "status_code": str(status_code),
                "status_class": status_class,
+                "tenant": tenant,
            }

            # Record duration and count
@@ -7,6 +7,7 @@ This module provides the ASGI app for uvicorn import string usage:
 For CLI usage, use the hindsight-api command instead.
 """

+import logging
 import os
 import warnings

@@ -17,6 +18,12 @@ warnings.filterwarnings("ignore", message="websockets.server.WebSocketServerProt
 from hindsight_api import MemoryEngine
 from hindsight_api.api import create_app
 from hindsight_api.config import get_config
+from hindsight_api.extensions import (
+    DefaultExtensionContext,
+    OperationValidatorExtension,
+    TenantExtension,
+    load_extension,
+)

 # Disable tokenizers parallelism to avoid warnings
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -25,10 +32,33 @@ os.environ["TOKENIZERS_PARALLELISM"] = "false"
 config = get_config()
 config.configure_logging()

+# Load operation validator extension if configured
+operation_validator = load_extension("OPERATION_VALIDATOR", OperationValidatorExtension)
+if operation_validator:
+    logging.info(f"Loaded operation validator: {operation_validator.__class__.__name__}")
+
+# Load tenant extension if configured
+tenant_extension = load_extension("TENANT", TenantExtension)
+if tenant_extension:
+    logging.info(f"Loaded tenant extension: {tenant_extension.__class__.__name__}")
+
 # Create app at module level (required for uvicorn import string)
 # MemoryEngine reads configuration from environment variables automatically
 # Note: run_migrations=True by default, but migrations are idempotent so safe with workers
-_memory = MemoryEngine(run_migrations=config.run_migrations_on_startup)
+_memory = MemoryEngine(
+    operation_validator=operation_validator,
+    tenant_extension=tenant_extension,
+    run_migrations=config.run_migrations_on_startup,
+)
+
+# Set extension context on tenant extension (needed for schema provisioning)
+if tenant_extension:
+    extension_context = DefaultExtensionContext(
+        database_url=config.database_url,
+        memory_engine=_memory,
+    )
+    tenant_extension.set_context(extension_context)
+    logging.info("Extension context set on tenant extension")

 # Create unified app with both HTTP and optionally MCP
 app = create_app(
@@ -0,0 +1,11 @@
+"""
+Worker package for distributed task processing.
+
+This package provides:
+- WorkerPoller: Polls PostgreSQL for pending tasks and executes them
+- main: CLI entry point for hindsight-worker
+"""
+
+from .poller import WorkerPoller
+
+__all__ = ["WorkerPoller"]
@@ -0,0 +1,285 @@
+"""
+Command-line interface for Hindsight Worker.
+
+Run the worker with:
+    hindsight-worker
+
+Stop with Ctrl+C (graceful shutdown).
+"""
+
+import argparse
+import asyncio
+import atexit
+import logging
+import os
+import signal
+import socket
+import sys
+import warnings
+
+from ..config import get_config
+from ..engine.task_backend import SyncTaskBackend
+from .poller import WorkerPoller
+
+# Filter deprecation warnings from third-party libraries
+warnings.filterwarnings("ignore", message="websockets.legacy is deprecated")
+warnings.filterwarnings("ignore", message="websockets.server.WebSocketServerProtocol is deprecated")
+
+# Disable tokenizers parallelism to avoid warnings
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+logger = logging.getLogger(__name__)
+
+
+def create_worker_app(poller: WorkerPoller, memory):
+    """Create a minimal FastAPI app for worker metrics and health."""
+    from fastapi import FastAPI
+    from fastapi.responses import JSONResponse, Response
+    from prometheus_client import CONTENT_TYPE_LATEST, generate_latest
+
+    from ..metrics import create_metrics_collector, get_metrics_collector, initialize_metrics
+
+    app = FastAPI(
+        title="Hindsight Worker",
+        description="Worker process for distributed task execution",
+    )
+
+    # Initialize OpenTelemetry metrics
+    try:
+        prometheus_reader = initialize_metrics(service_name="hindsight-worker", service_version="1.0.0")
+        create_metrics_collector()
+        app.state.prometheus_reader = prometheus_reader
+        logger.info("Metrics initialized - available at /metrics endpoint")
+    except Exception as e:
+        logger.warning(f"Failed to initialize metrics: {e}. Metrics will be disabled.")
+        app.state.prometheus_reader = None
+
+    # Set up DB pool metrics if available
+    metrics_collector = get_metrics_collector()
+    if memory._pool is not None and hasattr(metrics_collector, "set_db_pool"):
+        metrics_collector.set_db_pool(memory._pool)
+        logger.info("DB pool metrics configured")
+
+    @app.get(
+        "/health",
+        summary="Health check endpoint",
+        description="Returns worker health status including database connectivity",
+        tags=["Monitoring"],
+    )
+    async def health_endpoint():
+        """Health check endpoint."""
+        health = await memory.health_check()
+        health["worker_id"] = poller.worker_id
+        health["is_shutdown"] = poller.is_shutdown
+        status_code = 200 if health.get("status") == "healthy" else 503
+        return JSONResponse(content=health, status_code=status_code)
+
+    @app.get(
+        "/metrics",
+        summary="Prometheus metrics endpoint",
+        description="Exports metrics in Prometheus format for scraping",
+        tags=["Monitoring"],
+    )
+    async def metrics_endpoint():
+        """Return Prometheus metrics."""
+        metrics_data = generate_latest()
+        return Response(content=metrics_data, media_type=CONTENT_TYPE_LATEST)
+
+    @app.get(
+        "/",
+        summary="Worker info",
+        description="Basic worker information",
+        tags=["Info"],
+    )
+    async def root():
+        """Return basic worker info."""
+        return {
+            "service": "hindsight-worker",
+            "worker_id": poller.worker_id,
+            "is_shutdown": poller.is_shutdown,
+        }
+
+    return app
+
+
+def main():
+    """Main entry point for the hindsight-worker CLI."""
+    # Load configuration from environment
+    config = get_config()
+
+    parser = argparse.ArgumentParser(
+        prog="hindsight-worker",
+        description="Hindsight Worker - distributed task processor",
+    )
+
+    # Worker options
+    parser.add_argument(
+        "--worker-id",
+        default=config.worker_id or socket.gethostname(),
+        help="Worker identifier (default: hostname, env: HINDSIGHT_API_WORKER_ID)",
+    )
+    parser.add_argument(
+        "--poll-interval",
+        type=int,
+        default=config.worker_poll_interval_ms,
+        help=f"Poll interval in milliseconds (default: {config.worker_poll_interval_ms}, env: HINDSIGHT_API_WORKER_POLL_INTERVAL_MS)",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=config.worker_batch_size,
+        help=f"Tasks to claim per poll (default: {config.worker_batch_size}, env: HINDSIGHT_API_WORKER_BATCH_SIZE)",
+    )
+    parser.add_argument(
+        "--max-retries",
+        type=int,
+        default=config.worker_max_retries,
+        help=f"Max retries before marking failed (default: {config.worker_max_retries}, env: HINDSIGHT_API_WORKER_MAX_RETRIES)",
+    )
+
+    # HTTP server options
+    parser.add_argument(
+        "--http-port",
+        type=int,
+        default=config.worker_http_port,
+        help=f"HTTP port for metrics/health endpoints (default: {config.worker_http_port}, env: HINDSIGHT_API_WORKER_HTTP_PORT)",
+    )
+    parser.add_argument(
+        "--http-host",
+        default="0.0.0.0",
+        help="HTTP host to bind (default: 0.0.0.0)",
+    )
+
+    # Logging options
+    parser.add_argument(
+        "--log-level",
+        default=config.log_level,
+        choices=["critical", "error", "warning", "info", "debug", "trace"],
+        help=f"Log level (default: {config.log_level}, env: HINDSIGHT_API_LOG_LEVEL)",
+    )
+
+    args = parser.parse_args()
+
+    # Configure logging
+    config.configure_logging()
+
+    # Import MemoryEngine here to avoid circular imports
+    from .. import MemoryEngine
+
+    print(f"Starting Hindsight Worker: {args.worker_id}")
+    print(f"  Poll interval: {args.poll_interval}ms")
+    print(f"  Batch size: {args.batch_size}")
+    print(f"  Max retries: {args.max_retries}")
+    print(f"  HTTP server: {args.http_host}:{args.http_port}")
+    print()
+
+    # Global references for cleanup
+    memory = None
+    poller = None
+
+    async def run():
+        nonlocal memory, poller
+        import uvicorn
+
+        # Initialize MemoryEngine
+        # Workers use SyncTaskBackend because they execute tasks directly,
+        # they don't need to store tasks (they poll from DB)
+        memory = MemoryEngine(
+            run_migrations=False,  # Workers don't run migrations
+            task_backend=SyncTaskBackend(),
+        )
+
+        await memory.initialize()
+
+        print(f"Database connected: {config.database_url}")
+
+        # Create and start the poller
+        poller = WorkerPoller(
+            pool=memory._pool,
+            worker_id=args.worker_id,
+            executor=memory.execute_task,
+            poll_interval_ms=args.poll_interval,
+            batch_size=args.batch_size,
+            max_retries=args.max_retries,
+        )
+
+        # Create the HTTP app for metrics/health
+        app = create_worker_app(poller, memory)
+
+        # Setup signal handlers for graceful shutdown
+        shutdown_requested = asyncio.Event()
+
+        def signal_handler(signum, frame):
+            print(f"\nReceived signal {signum}, initiating graceful shutdown...")
+            shutdown_requested.set()
+
+        signal.signal(signal.SIGINT, signal_handler)
+        signal.signal(signal.SIGTERM, signal_handler)
+
+        # Create uvicorn config and server
+        uvicorn_config = uvicorn.Config(
+            app,
+            host=args.http_host,
+            port=args.http_port,
+            log_level="info",  # Reduce uvicorn noise
+            access_log=False,
+        )
+        server = uvicorn.Server(uvicorn_config)
+
+        # Run the poller and HTTP server concurrently
+        poller_task = asyncio.create_task(poller.run())
+        http_task = asyncio.create_task(server.serve())
+
+        print(f"Worker started. Metrics available at http://{args.http_host}:{args.http_port}/metrics")
+
+        # Wait for shutdown signal
+        await shutdown_requested.wait()
+
+        # Graceful shutdown
+        print("Shutting down HTTP server...")
+        server.should_exit = True
+
+        print("Waiting for poller to finish...")
+        await poller.shutdown_graceful(timeout=30.0)
+        poller_task.cancel()
+        try:
+            await poller_task
+        except asyncio.CancelledError:
+            pass
+
+        # Wait for HTTP server to finish
+        try:
+            await asyncio.wait_for(http_task, timeout=5.0)
+        except asyncio.TimeoutError:
+            http_task.cancel()
+            try:
+                await http_task
+            except asyncio.CancelledError:
+                pass
+
+        # Close memory engine
+        await memory.close()
+        print("Worker shutdown complete")
+
+    def cleanup():
+        """Synchronous cleanup for atexit."""
+        if memory is not None and memory._pg0 is not None:
+            try:
+                loop = asyncio.new_event_loop()
+                loop.run_until_complete(memory._pg0.stop())
+                loop.close()
+                print("\npg0 stopped.")
+            except Exception as e:
+                print(f"\nError stopping pg0: {e}")
+
+    atexit.register(cleanup)
+
+    try:
+        asyncio.run(run())
+    except KeyboardInterrupt:
+        print("\nWorker interrupted")
+        sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,281 @@
+"""
+Worker poller for distributed task execution.
+
+Polls PostgreSQL for pending tasks and executes them using
+FOR UPDATE SKIP LOCKED for safe concurrent claiming.
+"""
+
+import asyncio
+import json
+import logging
+import traceback
+from collections.abc import Awaitable, Callable
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    import asyncpg
+
+logger = logging.getLogger(__name__)
+
+
+def fq_table(table: str, schema: str | None = None) -> str:
+    """Get fully-qualified table name with optional schema prefix."""
+    if schema:
+        return f'"{schema}".{table}'
+    return table
+
+
+class WorkerPoller:
+    """
+    Polls PostgreSQL for pending tasks and executes them.
+
+    Uses FOR UPDATE SKIP LOCKED for safe distributed claiming,
+    allowing multiple workers to process tasks without conflicts.
+    """
+
+    def __init__(
+        self,
+        pool: "asyncpg.Pool",
+        worker_id: str,
+        executor: Callable[[dict[str, Any]], Awaitable[None]],
+        poll_interval_ms: int = 500,
+        batch_size: int = 10,
+        max_retries: int = 3,
+        schema: str | None = None,
+    ):
+        """
+        Initialize the worker poller.
+
+        Args:
+            pool: asyncpg connection pool
+            worker_id: Unique identifier for this worker
+            executor: Async function to execute tasks (typically MemoryEngine.execute_task)
+            poll_interval_ms: Interval between polls when no tasks found (milliseconds)
+            batch_size: Maximum number of tasks to claim per poll cycle
+            max_retries: Maximum retry attempts before marking task as failed
+            schema: Database schema for multi-tenant support (optional)
+        """
+        self._pool = pool
+        self._worker_id = worker_id
+        self._executor = executor
+        self._poll_interval_ms = poll_interval_ms
+        self._batch_size = batch_size
+        self._max_retries = max_retries
+        self._schema = schema
+        self._shutdown = asyncio.Event()
+        self._current_tasks: set[asyncio.Task] = set()
+        self._in_flight_count = 0
+        self._in_flight_lock = asyncio.Lock()
+
+    async def claim_batch(self) -> list[tuple[str, dict[str, Any]]]:
+        """
+        Claim up to batch_size pending tasks atomically.
+
+        Uses FOR UPDATE SKIP LOCKED to ensure no conflicts with other workers.
+
+        Returns:
+            List of tuples (operation_id, task_dict)
+        """
+        table = fq_table("async_operations", self._schema)
+
+        async with self._pool.acquire() as conn:
+            async with conn.transaction():
+                # Select and lock pending tasks
+                rows = await conn.fetch(
+                    f"""
+                    SELECT operation_id, task_payload
+                    FROM {table}
+                    WHERE status = 'pending' AND task_payload IS NOT NULL
+                    ORDER BY created_at
+                    LIMIT $1
+                    FOR UPDATE SKIP LOCKED
+                    """,
+                    self._batch_size,
+                )
+
+                if not rows:
+                    return []
+
+                # Claim the tasks by updating status and worker_id
+                operation_ids = [row["operation_id"] for row in rows]
+                await conn.execute(
+                    f"""
+                    UPDATE {table}
+                    SET status = 'processing', worker_id = $1, claimed_at = now(), updated_at = now()
+                    WHERE operation_id = ANY($2)
+                    """,
+                    self._worker_id,
+                    operation_ids,
+                )
+
+                # Parse and return task payloads
+                return [(str(row["operation_id"]), json.loads(row["task_payload"])) for row in rows]
+
+    async def _mark_completed(self, operation_id: str):
+        """Mark a task as completed."""
+        table = fq_table("async_operations", self._schema)
+        await self._pool.execute(
+            f"""
+            UPDATE {table}
+            SET status = 'completed', completed_at = now(), updated_at = now()
+            WHERE operation_id = $1
+            """,
+            operation_id,
+        )
+
+    async def _mark_failed(self, operation_id: str, error_message: str):
+        """Mark a task as failed with error message."""
+        table = fq_table("async_operations", self._schema)
+        # Truncate error message if too long (max 5000 chars in schema)
+        error_message = error_message[:5000] if len(error_message) > 5000 else error_message
+        await self._pool.execute(
+            f"""
+            UPDATE {table}
+            SET status = 'failed', error_message = $2, completed_at = now(), updated_at = now()
+            WHERE operation_id = $1
+            """,
+            operation_id,
+            error_message,
+        )
+
+    async def _retry_or_fail(self, operation_id: str, error_message: str):
+        """Increment retry count or mark as failed if max retries exceeded."""
+        table = fq_table("async_operations", self._schema)
+
+        # Get current retry count
+        row = await self._pool.fetchrow(
+            f"SELECT retry_count FROM {table} WHERE operation_id = $1",
+            operation_id,
+        )
+
+        if row is None:
+            logger.warning(f"Operation {operation_id} not found, cannot retry")
+            return
+
+        retry_count = row["retry_count"]
+
+        if retry_count >= self._max_retries:
+            # Max retries exceeded, mark as failed
+            await self._mark_failed(
+                operation_id, f"Max retries ({self._max_retries}) exceeded. Last error: {error_message}"
+            )
+            logger.error(f"Task {operation_id} failed after {retry_count} retries")
+        else:
+            # Increment retry and reset to pending
+            await self._pool.execute(
+                f"""
+                UPDATE {table}
+                SET status = 'pending', worker_id = NULL, claimed_at = NULL,
+                    retry_count = retry_count + 1, updated_at = now()
+                WHERE operation_id = $1
+                """,
+                operation_id,
+            )
+            logger.warning(f"Task {operation_id} failed, will retry (attempt {retry_count + 1}/{self._max_retries})")
+
+    async def execute_task(self, operation_id: str, task_dict: dict[str, Any]):
+        """Execute a single task and update its status."""
+        task_type = task_dict.get("type", "unknown")
+        bank_id = task_dict.get("bank_id", "unknown")
+
+        try:
+            logger.debug(f"Executing task {operation_id} (type={task_type}, bank={bank_id})")
+            await self._executor(task_dict)
+            await self._mark_completed(operation_id)
+            logger.debug(f"Task {operation_id} completed successfully")
+        except Exception as e:
+            error_msg = f"{type(e).__name__}: {e}\n{traceback.format_exc()}"
+            logger.error(f"Task {operation_id} failed: {e}")
+            await self._retry_or_fail(operation_id, error_msg)
+
+    async def run(self):
+        """
+        Main polling loop.
+
+        Continuously polls for pending tasks, claims them, and executes them
+        until shutdown is signaled.
+        """
+        logger.info(f"Worker {self._worker_id} starting polling loop")
+
+        while not self._shutdown.is_set():
+            try:
+                # Claim a batch of tasks
+                tasks = await self.claim_batch()
+
+                if tasks:
+                    # Log batch info
+                    task_types = {}
+                    for _, task_dict in tasks:
+                        t = task_dict.get("type", "unknown")
+                        task_types[t] = task_types.get(t, 0) + 1
+                    types_str = ", ".join(f"{k}:{v}" for k, v in task_types.items())
+                    logger.info(f"Worker {self._worker_id} claimed {len(tasks)} tasks: {types_str}")
+
+                    # Track in-flight tasks
+                    async with self._in_flight_lock:
+                        self._in_flight_count += len(tasks)
+
+                    # Execute tasks concurrently
+                    try:
+                        await asyncio.gather(
+                            *[self.execute_task(op_id, task_dict) for op_id, task_dict in tasks],
+                            return_exceptions=True,
+                        )
+                    finally:
+                        async with self._in_flight_lock:
+                            self._in_flight_count -= len(tasks)
+                else:
+                    # No tasks found, wait before polling again
+                    try:
+                        await asyncio.wait_for(
+                            self._shutdown.wait(),
+                            timeout=self._poll_interval_ms / 1000,
+                        )
+                    except asyncio.TimeoutError:
+                        pass  # Normal timeout, continue polling
+
+            except asyncio.CancelledError:
+                logger.info(f"Worker {self._worker_id} polling loop cancelled")
+                break
+            except Exception as e:
+                logger.error(f"Worker {self._worker_id} error in polling loop: {e}")
+                traceback.print_exc()
+                # Backoff on error
+                await asyncio.sleep(1)
+
+        logger.info(f"Worker {self._worker_id} polling loop stopped")
+
+    async def shutdown_graceful(self, timeout: float = 30.0):
+        """
+        Signal shutdown and wait for current tasks to complete.
+
+        Args:
+            timeout: Maximum time to wait for in-flight tasks (seconds)
+        """
+        logger.info(f"Worker {self._worker_id} initiating graceful shutdown")
+        self._shutdown.set()
+
+        # Wait for in-flight tasks to complete
+        start_time = asyncio.get_event_loop().time()
+        while asyncio.get_event_loop().time() - start_time < timeout:
+            async with self._in_flight_lock:
+                in_flight = self._in_flight_count
+
+            if in_flight == 0:
+                logger.info(f"Worker {self._worker_id} graceful shutdown complete")
+                return
+
+            logger.info(f"Worker {self._worker_id} waiting for {in_flight} in-flight tasks")
+            await asyncio.sleep(0.5)
+
+        logger.warning(f"Worker {self._worker_id} shutdown timeout after {timeout}s")
+
+    @property
+    def worker_id(self) -> str:
+        """Get the worker ID."""
+        return self._worker_id
+
+    @property
+    def is_shutdown(self) -> bool:
+        """Check if shutdown has been signaled."""
+        return self._shutdown.is_set()
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"

 [project]
 name = "hindsight-api"
-version = "0.2.1"
+version = "0.3.0"
 description = "Hindsight: Agent Memory That Works Like Human Memory"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -25,7 +25,7 @@ dependencies = [
    "psycopg2-binary>=2.9.11",
    "tiktoken>=0.12.0",
    "httpx>=0.27.0",
-    "fastmcp>=2.3.0",
+    "fastmcp>=2.14.0",  # CVE-2025-66416
    "pg0-embedded>=0.11.0",
    "python-dateutil>=2.8.0",
    "opentelemetry-api>=1.20.0",
@@ -39,10 +39,17 @@ dependencies = [
    "cohere>=5.0.0",
    "flashrank>=0.2.0",
    # Local ML models for embeddings/reranking - can be excluded in Docker with INCLUDE_LOCAL_MODELS=false
-    "sentence-transformers>=3.0.0,<3.3.0",
-    "transformers>=4.30.0,<4.46.0",
-    "torch>=2.0.0",
+    "sentence-transformers>=3.3.0",
+    "transformers>=4.53.0",  # Security fixes for ReDoS vulnerabilities
+    "torch>=2.6.0",  # CVE fix for remote code execution
    "uvloop>=0.22.1",
+    # Transitive dependency security fixes
+    "pyasn1>=0.6.2",  # DoS vulnerability fix
+    "urllib3>=2.6.3",  # Decompression-bomb safeguards bypass fix
+    "langchain-core>=1.2.5",  # Serialization injection vulnerability fix
+    "filelock>=3.20.1",  # TOCTOU race condition fix
+    "authlib>=1.6.6",  # Account takeover vulnerability fix
+    "aiohttp>=3.13.3",  # Multiple DoS vulnerabilities
 ]

 [project.optional-dependencies]
@@ -51,11 +58,12 @@ test = [
    "pytest-asyncio>=0.21.0",
    "pytest-timeout>=2.4.0",
    "pytest-xdist>=3.0.0",
-    "filelock>=3.0.0",
+    "filelock>=3.20.1",  # TOCTOU race condition fix
 ]

 [project.scripts]
 hindsight-api = "hindsight_api.main:main"
+hindsight-worker = "hindsight_api.worker.main:main"
 hindsight-local-mcp = "hindsight_api.mcp_local:main"
 hindsight-admin = "hindsight_api.admin.cli:main"

@@ -97,7 +105,7 @@ dev = [
    "pytest-timeout>=2.4.0",
    "pytest-xdist>=3.8.0",
    "python-dotenv>=1.2.1",
-    "filelock>=3.0.0",
+    "filelock>=3.20.1",  # TOCTOU race condition fix
    "ruff>=0.8.0",
    "ty>=0.0.1",
 ]
@@ -12,6 +12,7 @@ from hindsight_api import MemoryEngine, LLMConfig, LocalSTEmbeddings, RequestCon

 from hindsight_api.engine.cross_encoder import LocalSTCrossEncoder
 from hindsight_api.engine.query_analyzer import DateparserQueryAnalyzer
+from hindsight_api.engine.task_backend import SyncTaskBackend
 from hindsight_api.pg0 import EmbeddedPostgres

 # Default pg0 instance configuration for tests
@@ -147,6 +148,7 @@ async def memory(pg0_db_url, embeddings, cross_encoder, query_analyzer):
    Uses pg0_db_url (a postgresql:// URL) directly, so MemoryEngine won't try to
    manage pg0 lifecycle - that's handled by the session-scoped pg0_db_url fixture.
    Migrations are disabled here since they're run once at session scope in pg0_db_url.
+    Uses SyncTaskBackend so async tasks execute immediately (no worker needed).
    """
    mem = MemoryEngine(
        db_url=pg0_db_url,  # Direct postgresql:// URL, not pg0://
@@ -160,6 +162,7 @@ async def memory(pg0_db_url, embeddings, cross_encoder, query_analyzer):
        pool_min_size=1,
        pool_max_size=5,
        run_migrations=False,  # Migrations already run at session scope
+        task_backend=SyncTaskBackend(),  # Execute tasks immediately in tests
    )
    await mem.initialize()
    yield mem
@@ -1,5 +1,5 @@
 """
-Tests for agent management API (profile, disposition, background).
+Tests for agent management API (profile, disposition).
 """
 import pytest
 import uuid
@@ -25,15 +25,12 @@ class TestAgentProfile:

        assert profile is not None
        assert "disposition" in profile
-        assert "background" in profile

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

-        assert profile["background"] == ""
-
    @pytest.mark.asyncio
    async def test_update_agent_disposition(self, memory: MemoryEngine, request_context):
        """Test updating agent disposition traits."""
@@ -76,63 +73,10 @@ class TestAgentProfile:
        for agent in agents:
            assert "bank_id" in agent
            assert "disposition" in agent
-            assert "background" in agent
            assert "created_at" in agent
            assert "updated_at" in agent


-class TestAgentBackground:
-    """Tests for agent background management."""
-
-    @pytest.mark.asyncio
-    async def test_merge_agent_background(self, memory: MemoryEngine, request_context):
-        """Test merging agent background information."""
-        bank_id = unique_agent_id("test_profile_merge")
-
-        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-        assert profile["background"] == ""
-
-        result1 = await memory.merge_bank_background(
-            bank_id,
-            "I was born in Texas",
-            update_disposition=False,
-            request_context=request_context,
-        )
-        assert "Texas" in result1["background"]
-
-        result2 = await memory.merge_bank_background(
-            bank_id,
-            "I have 10 years of startup experience",
-            update_disposition=False,
-            request_context=request_context,
-        )
-        assert "Texas" in result2["background"] or "startup" in result2["background"]
-
-        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-        assert final_profile["background"] != ""
-
-    @pytest.mark.asyncio
-    async def test_merge_background_handles_conflicts(self, memory: MemoryEngine, request_context):
-        """Test that merging background handles conflicts (new overwrites old)."""
-        bank_id = unique_agent_id("test_profile_conflict")
-
-        result1 = await memory.merge_bank_background(
-            bank_id,
-            "I was born in Colorado",
-            update_disposition=False,
-            request_context=request_context,
-        )
-        assert "Colorado" in result1["background"]
-
-        result2 = await memory.merge_bank_background(
-            bank_id,
-            "You were born in Texas",
-            update_disposition=False,
-            request_context=request_context,
-        )
-        assert "Texas" in result2["background"]
-
-
 class TestAgentEndpoint:
    """Tests for agent PUT endpoint logic."""

@@ -147,7 +91,6 @@ class TestAgentEndpoint:
                literalism=5,
                empathy=2
            ),
-            background="I am a creative software engineer"
        )

        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
@@ -159,55 +102,10 @@ class TestAgentEndpoint:
                request_context=request_context,
            )

-        if request.background is not None:
-            pool = await memory._get_pool()
-            async with pool.acquire() as conn:
-                await conn.execute(
-                    """
-                    UPDATE banks
-                    SET background = $2,
-                        updated_at = NOW()
-                    WHERE bank_id = $1
-                    """,
-                    bank_id,
-                    request.background
-                )
-
        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)

        assert final_profile["disposition"].skepticism == 4
        assert final_profile["disposition"].literalism == 5
-        assert final_profile["background"] == "I am a creative software engineer"
-
-    @pytest.mark.asyncio
-    async def test_put_agent_partial_update(self, memory: MemoryEngine, request_context):
-        """Test updating only background."""
-        bank_id = unique_agent_id("test_put_partial")
-
-        request = CreateBankRequest(
-            background="I am a data scientist"
-        )
-
-        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-
-        if request.background is not None:
-            pool = await memory._get_pool()
-            async with pool.acquire() as conn:
-                await conn.execute(
-                    """
-                    UPDATE banks
-                    SET background = $2,
-                        updated_at = NOW()
-                    WHERE bank_id = $1
-                    """,
-                    bank_id,
-                    request.background
-                )
-
-        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-
-        assert final_profile["disposition"].skepticism == 3  # Default
-        assert final_profile["background"] == "I am a data scientist"


 class TestAgentDispositionIntegration:
@@ -225,13 +123,6 @@ class TestAgentDispositionIntegration:
        }
        await memory.update_bank_disposition(bank_id, disposition, request_context=request_context)

-        await memory.merge_bank_background(
-            bank_id,
-            "I am a creative artist who values innovation over tradition",
-            update_disposition=False,
-            request_context=request_context,
-        )
-
        await memory.retain_batch_async(
            bank_id=bank_id,
            contents=[
@@ -17,6 +17,7 @@ from hindsight_api import MemoryEngine, RequestContext
 from hindsight_api.engine.embeddings import LocalSTEmbeddings, OpenAIEmbeddings, CohereEmbeddings
 from hindsight_api.engine.cross_encoder import LocalSTCrossEncoder, CohereCrossEncoder
 from hindsight_api.engine.query_analyzer import DateparserQueryAnalyzer
+from hindsight_api.engine.task_backend import SyncTaskBackend
 from hindsight_api.extensions import TenantExtension, TenantContext
 from hindsight_api.migrations import run_migrations, ensure_embedding_dimension

@@ -323,6 +324,7 @@ class TestOpenAIEmbeddings:
            pool_max_size=3,
            run_migrations=False,
            tenant_extension=SchemaTenantExtension(schema_name),
+            task_backend=SyncTaskBackend(),
        )

        try:
@@ -392,6 +394,7 @@ class TestOpenAIEmbeddings:
            pool_max_size=3,
            run_migrations=False,
            tenant_extension=SchemaTenantExtension(schema_name),
+            task_backend=SyncTaskBackend(),
        )

        try:
@@ -559,6 +562,7 @@ class TestCohereIntegration:
            pool_max_size=3,
            run_migrations=False,
            tenant_extension=SchemaTenantExtension(schema_name),
+            task_backend=SyncTaskBackend(),
        )

        try:
@@ -0,0 +1,516 @@
+"""Tests for emergent entity filtering."""
+
+import pytest
+from unittest.mock import AsyncMock, MagicMock
+
+from hindsight_api.engine.mental_models.emergent import (
+    build_mission_filter_prompt,
+    evaluate_emergent_models,
+    filter_candidates_by_mission,
+    MissionFilterResponse,
+    MissionFilterCandidate,
+)
+from hindsight_api.engine.mental_models.models import EmergentCandidate
+
+
+class TestBuildMissionFilterPrompt:
+    """Test prompt building for mission filtering."""
+
+    def test_prompt_contains_mission(self):
+        """Test that prompt includes the mission."""
+        candidates = [
+            EmergentCandidate(
+                name="Alice",
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            )
+        ]
+        prompt = build_mission_filter_prompt("Be a PM for engineering team", candidates)
+        assert "Be a PM for engineering team" in prompt
+
+    def test_prompt_contains_candidates(self):
+        """Test that prompt includes all candidates."""
+        candidates = [
+            EmergentCandidate(
+                name="Alice Chen",
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            ),
+            EmergentCandidate(
+                name="Project Phoenix",
+                detection_method="named_entity_extraction",
+                mention_count=5,
+            ),
+        ]
+        prompt = build_mission_filter_prompt("Track projects", candidates)
+        assert "Alice Chen" in prompt
+        assert "Project Phoenix" in prompt
+
+    def test_prompt_contains_rejection_guidance(self):
+        """Test that prompt contains guidance to reject generic entities."""
+        candidates = [
+            EmergentCandidate(
+                name="test",
+                detection_method="named_entity_extraction",
+                mention_count=1,
+            )
+        ]
+        prompt = build_mission_filter_prompt("Test mission", candidates)
+
+        # Should contain rejection guidance for generic terms
+        assert "promote=false" in prompt
+        assert "kids" in prompt  # Example of generic term to reject
+        assert "community" in prompt  # Example of abstract concept to reject
+        assert "motivation" in prompt  # Example of abstract concept to reject
+
+
+class TestFilterCandidatesByMission:
+    """Test the filter_candidates_by_mission function."""
+
+    @pytest.fixture
+    def mock_llm_config(self):
+        """Create a mock LLM config."""
+        config = MagicMock()
+        config.call = AsyncMock()
+        return config
+
+    async def test_empty_candidates(self, mock_llm_config):
+        """Test with empty candidate list."""
+        result = await filter_candidates_by_mission(
+            llm_config=mock_llm_config,
+            mission="Test mission",
+            candidates=[],
+        )
+        assert result == []
+        mock_llm_config.call.assert_not_called()
+
+    async def test_no_mission_keeps_all(self, mock_llm_config):
+        """Test that no mission keeps all candidates (skips filtering)."""
+        candidates = [
+            EmergentCandidate(
+                name="Alice",
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            )
+        ]
+        result = await filter_candidates_by_mission(
+            llm_config=mock_llm_config,
+            mission="",  # Empty mission
+            candidates=candidates,
+        )
+        assert len(result) == 1
+        assert result[0].name == "Alice"
+        mock_llm_config.call.assert_not_called()
+
+    async def test_filters_by_promote_flag(self, mock_llm_config):
+        """Test that candidates are filtered by promote flag."""
+        candidates = [
+            EmergentCandidate(
+                name="Alice Chen",
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            ),
+            EmergentCandidate(
+                name="community",
+                detection_method="named_entity_extraction",
+                mention_count=5,
+            ),
+        ]
+
+        # Mock LLM response - Alice is promoted, community is not
+        mock_llm_config.call.return_value = MissionFilterResponse(
+            candidates=[
+                MissionFilterCandidate(name="Alice Chen", promote=True, reason="Specific person"),
+                MissionFilterCandidate(name="community", promote=False, reason="Generic abstract concept"),
+            ]
+        )
+
+        result = await filter_candidates_by_mission(
+            llm_config=mock_llm_config,
+            mission="Be a PM for engineering team",
+            candidates=candidates,
+        )
+
+        assert len(result) == 1
+        assert result[0].name == "Alice Chen"
+
+    async def test_rejects_generic_entities(self, mock_llm_config):
+        """Test that generic entities are rejected."""
+        # These are all generic/abstract terms that should be rejected
+        generic_names = [
+            "user", "support", "community", "family", "motivation",
+            "photo", "gratitude", "difference", "volunteering",
+            "kids", "veterans", "impact", "kindness", "encouragement",
+            "education", "nature", "joy", "positivity", "inspiration",
+            "help", "commitment", "passion", "energy", "connection",
+        ]
+        candidates = [
+            EmergentCandidate(
+                name=name,
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            )
+            for name in generic_names
+        ]
+
+        # Add some valid candidates
+        valid_candidates = [
+            EmergentCandidate(
+                name="John",
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            ),
+            EmergentCandidate(
+                name="Maria",
+                detection_method="named_entity_extraction",
+                mention_count=8,
+            ),
+            EmergentCandidate(
+                name="Max",
+                detection_method="named_entity_extraction",
+                mention_count=6,
+            ),
+        ]
+        candidates.extend(valid_candidates)
+
+        # Mock LLM response - reject all generic, promote only specific names
+        response_candidates = [
+            MissionFilterCandidate(name=name, promote=False, reason="Generic/abstract term")
+            for name in generic_names
+        ]
+        response_candidates.extend([
+            MissionFilterCandidate(name=c.name, promote=True, reason="Specific person name")
+            for c in valid_candidates
+        ])
+
+        mock_llm_config.call.return_value = MissionFilterResponse(candidates=response_candidates)
+
+        result = await filter_candidates_by_mission(
+            llm_config=mock_llm_config,
+            mission="Be a health coach",
+            candidates=candidates,
+        )
+
+        # Should only have John, Maria, and Max
+        result_names = {c.name for c in result}
+        assert result_names == {"John", "Maria", "Max"}
+
+    async def test_accepts_specific_named_entities(self, mock_llm_config):
+        """Test that specific named entities are accepted."""
+        # These should all be accepted
+        valid_names = [
+            "Alice Chen",       # Full name
+            "Dr. Smith",        # Title + name
+            "John",             # First name (when it's clearly a person)
+            "Google",           # Organization
+            "Frontend Team",    # Named team
+            "Project Phoenix",  # Named project
+            "NYC Office",       # Named place
+            "Q4 Planning",      # Named event
+            "Sprint 23 Review", # Named meeting
+        ]
+        candidates = [
+            EmergentCandidate(
+                name=name,
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            )
+            for name in valid_names
+        ]
+
+        # Mock LLM response - promote all
+        response_candidates = [
+            MissionFilterCandidate(name=name, promote=True, reason="Specific named entity")
+            for name in valid_names
+        ]
+        mock_llm_config.call.return_value = MissionFilterResponse(candidates=response_candidates)
+
+        result = await filter_candidates_by_mission(
+            llm_config=mock_llm_config,
+            mission="Be a PM for engineering team",
+            candidates=candidates,
+        )
+
+        # Should have all valid names
+        result_names = {c.name for c in result}
+        assert result_names == set(valid_names)
+
+    async def test_llm_error_rejects_all_candidates(self, mock_llm_config):
+        """Test that LLM errors result in rejecting all candidates (fail-safe)."""
+        candidates = [
+            EmergentCandidate(
+                name="Alice",
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            )
+        ]
+
+        mock_llm_config.call.side_effect = Exception("LLM error")
+
+        result = await filter_candidates_by_mission(
+            llm_config=mock_llm_config,
+            mission="Test mission",
+            candidates=candidates,
+        )
+
+        # Should reject all candidates on error (fail-safe)
+        assert len(result) == 0
+
+    async def test_missing_candidate_in_response_is_rejected(self, mock_llm_config):
+        """Test that candidates not in LLM response are rejected by default."""
+        candidates = [
+            EmergentCandidate(
+                name="Alice",
+                detection_method="named_entity_extraction",
+                mention_count=10,
+            ),
+            EmergentCandidate(
+                name="Bob",
+                detection_method="named_entity_extraction",
+                mention_count=5,
+            ),
+        ]
+
+        # Mock LLM response - only includes Alice, not Bob
+        mock_llm_config.call.return_value = MissionFilterResponse(
+            candidates=[
+                MissionFilterCandidate(name="Alice", promote=True, reason="Specific person"),
+            ]
+        )
+
+        result = await filter_candidates_by_mission(
+            llm_config=mock_llm_config,
+            mission="Test mission",
+            candidates=candidates,
+        )
+
+        # Only Alice should be in result (Bob was missing from response, so rejected)
+        assert len(result) == 1
+        assert result[0].name == "Alice"
+
+
+class TestEvaluateEmergentModels:
+    """Test the evaluate_emergent_models function for cleanup of existing models."""
+
+    @pytest.fixture
+    def mock_llm_config(self):
+        """Create a mock LLM config."""
+        config = MagicMock()
+        config.call = AsyncMock()
+        return config
+
+    async def test_empty_models(self, mock_llm_config):
+        """Test with empty model list."""
+        result = await evaluate_emergent_models(
+            llm_config=mock_llm_config,
+            models=[],
+        )
+        assert result == []
+        mock_llm_config.call.assert_not_called()
+
+    async def test_removes_generic_models(self, mock_llm_config):
+        """Test that generic/abstract models are marked for removal."""
+        models = [
+            {"id": "id-kids", "name": "kids"},
+            {"id": "id-community", "name": "community"},
+            {"id": "id-motivation", "name": "motivation"},
+            {"id": "id-john", "name": "John"},
+            {"id": "id-maria", "name": "Maria"},
+        ]
+
+        # Mock LLM response - reject generic, keep specific names
+        mock_llm_config.call.return_value = MissionFilterResponse(
+            candidates=[
+                MissionFilterCandidate(name="kids", promote=False, reason="Generic category"),
+                MissionFilterCandidate(name="community", promote=False, reason="Abstract concept"),
+                MissionFilterCandidate(name="motivation", promote=False, reason="Abstract concept"),
+                MissionFilterCandidate(name="John", promote=True, reason="Person name"),
+                MissionFilterCandidate(name="Maria", promote=True, reason="Person name"),
+            ]
+        )
+
+        result = await evaluate_emergent_models(
+            llm_config=mock_llm_config,
+            models=models,
+        )
+
+        # Should return IDs of generic models to remove
+        assert set(result) == {"id-kids", "id-community", "id-motivation"}
+
+    async def test_keeps_specific_named_models(self, mock_llm_config):
+        """Test that specific named models are kept."""
+        models = [
+            {"id": "id-john", "name": "John"},
+            {"id": "id-google", "name": "Google"},
+            {"id": "id-project", "name": "Project Phoenix"},
+        ]
+
+        # Mock LLM response - keep all
+        mock_llm_config.call.return_value = MissionFilterResponse(
+            candidates=[
+                MissionFilterCandidate(name="John", promote=True, reason="Person name"),
+                MissionFilterCandidate(name="Google", promote=True, reason="Organization"),
+                MissionFilterCandidate(name="Project Phoenix", promote=True, reason="Named project"),
+            ]
+        )
+
+        result = await evaluate_emergent_models(
+            llm_config=mock_llm_config,
+            models=models,
+        )
+
+        # No models should be removed
+        assert result == []
+
+    async def test_llm_error_keeps_all_models(self, mock_llm_config):
+        """Test that LLM errors result in keeping all models (safe default)."""
+        models = [
+            {"id": "id-kids", "name": "kids"},
+            {"id": "id-john", "name": "John"},
+        ]
+
+        mock_llm_config.call.side_effect = Exception("LLM error")
+
+        result = await evaluate_emergent_models(
+            llm_config=mock_llm_config,
+            models=models,
+        )
+
+        # Should keep all models on error (return empty removal list)
+        assert result == []
+
+    async def test_missing_model_in_response_is_removed(self, mock_llm_config):
+        """Test that models not in LLM response are marked for removal."""
+        models = [
+            {"id": "id-alice", "name": "Alice"},
+            {"id": "id-bob", "name": "Bob"},
+        ]
+
+        # Mock LLM response - only includes Alice
+        mock_llm_config.call.return_value = MissionFilterResponse(
+            candidates=[
+                MissionFilterCandidate(name="Alice", promote=True, reason="Person name"),
+            ]
+        )
+
+        result = await evaluate_emergent_models(
+            llm_config=mock_llm_config,
+            models=models,
+        )
+
+        # Bob should be marked for removal (missing from response)
+        assert result == ["id-bob"]
+
+
+class TestRemovedEntitiesNotRepromoted:
+    """Test that entities removed by evaluation are not re-promoted.
+
+    This tests the fix for a bug where:
+    1. evaluate_emergent_models returns model IDs to remove (e.g., 'entity-maya')
+    2. We delete those models
+    3. detect_entity_candidates finds the same entities (now eligible since model was deleted)
+    4. filter_candidates_by_goal approves them (different LLM call)
+    5. BUG: We were re-promoting the same entities we just removed
+
+    The fix tracks removed entity_ids and excludes them from promotion.
+    """
+
+    async def test_removed_entity_ids_excluded_from_promotion(self):
+        """Test that entities whose models were removed are not re-promoted."""
+        from hindsight_api.engine.mental_models.models import EmergentCandidate
+
+        # Simulate the scenario from the bug:
+        # - existing_emergent has model 'entity-maya' with entity_id='uuid-maya'
+        # - evaluate_emergent_models says to remove 'entity-maya'
+        # - detect_entity_candidates returns 'Maya' with entity_id='uuid-maya' (now eligible)
+        # - filter_candidates_by_goal says to promote 'Maya'
+        # - But we should NOT promote because we just removed it
+
+        existing_emergent = [
+            {"id": "entity-maya", "name": "Maya", "entity_id": "uuid-maya"},
+            {"id": "entity-alex", "name": "Alex", "entity_id": "uuid-alex"},
+            {"id": "entity-john", "name": "John", "entity_id": "uuid-john"},  # This one will be kept
+        ]
+
+        # Models to remove (evaluate_emergent_models would return these)
+        models_to_remove = ["entity-maya", "entity-alex"]
+
+        # Build model_id -> entity_id mapping (this is what the fix does)
+        model_to_entity = {m["id"]: m.get("entity_id") for m in existing_emergent}
+
+        # Track removed entity_ids
+        removed_entity_ids: set[str] = set()
+        for model_id in models_to_remove:
+            entity_id = model_to_entity.get(model_id)
+            if entity_id:
+                removed_entity_ids.add(str(entity_id))
+
+        # Verify we tracked the right entity_ids
+        assert removed_entity_ids == {"uuid-maya", "uuid-alex"}
+
+        # Now simulate candidates that were detected (includes removed entities)
+        candidates = [
+            EmergentCandidate(
+                name="Maya", entity_id="uuid-maya", detection_method="named_entity", mention_count=10
+            ),
+            EmergentCandidate(
+                name="Alex", entity_id="uuid-alex", detection_method="named_entity", mention_count=8
+            ),
+            EmergentCandidate(
+                name="NewPerson", entity_id="uuid-new", detection_method="named_entity", mention_count=5
+            ),
+        ]
+
+        # Filter out candidates whose entity was just removed (the fix)
+        filtered_candidates = [c for c in candidates if c.entity_id not in removed_entity_ids]
+
+        # Only NewPerson should remain - Maya and Alex were removed and should not be re-promoted
+        assert len(filtered_candidates) == 1
+        assert filtered_candidates[0].name == "NewPerson"
+        assert filtered_candidates[0].entity_id == "uuid-new"
+
+    async def test_candidates_without_matching_removal_are_kept(self):
+        """Test that candidates not in the removed set are still promoted."""
+        from hindsight_api.engine.mental_models.models import EmergentCandidate
+
+        # No models removed
+        removed_entity_ids: set[str] = set()
+
+        candidates = [
+            EmergentCandidate(
+                name="Alice", entity_id="uuid-alice", detection_method="named_entity", mention_count=10
+            ),
+            EmergentCandidate(
+                name="Bob", entity_id="uuid-bob", detection_method="named_entity", mention_count=8
+            ),
+        ]
+
+        # Filter (should keep all since nothing was removed)
+        filtered_candidates = [c for c in candidates if c.entity_id not in removed_entity_ids]
+
+        assert len(filtered_candidates) == 2
+        assert {c.name for c in filtered_candidates} == {"Alice", "Bob"}
+
+    async def test_partial_removal_keeps_other_candidates(self):
+        """Test that only removed entities are excluded, others pass through."""
+        from hindsight_api.engine.mental_models.models import EmergentCandidate
+
+        # Only one entity removed
+        removed_entity_ids = {"uuid-removed"}
+
+        candidates = [
+            EmergentCandidate(
+                name="Removed", entity_id="uuid-removed", detection_method="named_entity", mention_count=10
+            ),
+            EmergentCandidate(
+                name="Kept1", entity_id="uuid-kept1", detection_method="named_entity", mention_count=8
+            ),
+            EmergentCandidate(
+                name="Kept2", entity_id="uuid-kept2", detection_method="named_entity", mention_count=5
+            ),
+        ]
+
+        filtered_candidates = [c for c in candidates if c.entity_id not in removed_entity_ids]
+
+        assert len(filtered_candidates) == 2
+        assert {c.name for c in filtered_candidates} == {"Kept1", "Kept2"}
@@ -17,6 +17,8 @@ from hindsight_api.extensions import (
    RecallResult,
    ReflectContext,
    ReflectResultContext,
+    RefreshMentalModelContext,
+    RefreshMentalModelResult,
    RequestContext,
    RetainContext,
    RetainResult,
@@ -93,6 +95,7 @@ class RateLimitingValidator(OperationValidatorExtension):
        self.retain_counts: dict[str, int] = defaultdict(int)
        self.recall_counts: dict[str, int] = defaultdict(int)
        self.reflect_counts: dict[str, int] = defaultdict(int)
+        self.refresh_mental_model_counts: dict[str, int] = defaultdict(int)

    async def validate_retain(self, ctx: RetainContext) -> ValidationResult:
        self.retain_counts[ctx.bank_id] += 1
@@ -118,6 +121,16 @@ class RateLimitingValidator(OperationValidatorExtension):
            )
        return ValidationResult.accept()

+    async def validate_refresh_mental_model(
+        self, ctx: RefreshMentalModelContext
+    ) -> ValidationResult:
+        self.refresh_mental_model_counts[ctx.bank_id] += 1
+        if self.refresh_mental_model_counts[ctx.bank_id] > self.max_attempts:
+            return ValidationResult.reject(
+                f"Refresh mental model limit exceeded for bank {ctx.bank_id}"
+            )
+        return ValidationResult.accept()
+

 class TrackingValidator(OperationValidatorExtension):
    """
@@ -132,10 +145,12 @@ class TrackingValidator(OperationValidatorExtension):
        self.pre_retain_calls: list[RetainContext] = []
        self.pre_recall_calls: list[RecallContext] = []
        self.pre_reflect_calls: list[ReflectContext] = []
+        self.pre_refresh_mental_model_calls: list[RefreshMentalModelContext] = []
        # Post-hook tracking
        self.post_retain_calls: list[RetainResult] = []
        self.post_recall_calls: list[RecallResult] = []
        self.post_reflect_calls: list[ReflectResultContext] = []
+        self.post_refresh_mental_model_calls: list[RefreshMentalModelResult] = []

    async def validate_retain(self, ctx: RetainContext) -> ValidationResult:
        self.pre_retain_calls.append(ctx)
@@ -149,6 +164,12 @@ class TrackingValidator(OperationValidatorExtension):
        self.pre_reflect_calls.append(ctx)
        return ValidationResult.accept()

+    async def validate_refresh_mental_model(
+        self, ctx: RefreshMentalModelContext
+    ) -> ValidationResult:
+        self.pre_refresh_mental_model_calls.append(ctx)
+        return ValidationResult.accept()
+
    async def on_retain_complete(self, result: RetainResult) -> None:
        self.post_retain_calls.append(result)

@@ -158,6 +179,11 @@ class TrackingValidator(OperationValidatorExtension):
    async def on_reflect_complete(self, result: ReflectResultContext) -> None:
        self.post_reflect_calls.append(result)

+    async def on_refresh_mental_model_complete(
+        self, result: RefreshMentalModelResult
+    ) -> None:
+        self.post_refresh_mental_model_calls.append(result)
+

 class TestMemoryEngineValidation:
    """Tests for validation integration with MemoryEngine.
@@ -515,6 +541,105 @@ class TestOperationHooksParameters:
        assert len(validator.pre_recall_calls) == 1
        assert len(validator.post_recall_calls) == 1

+    @pytest.mark.asyncio
+    async def test_refresh_mental_model_pre_hook_receives_all_parameters(
+        self, memory_with_tracking_validator
+    ):
+        """Pre-refresh-mental-model hook receives all user-provided parameters."""
+        import uuid
+
+        memory, validator = memory_with_tracking_validator
+        bank_id = f"test-refresh-mm-params-{uuid.uuid4().hex[:8]}"
+        ctx = RequestContext(api_key="test-key")
+
+        # Create bank first (get_bank_profile auto-creates if needed)
+        await memory.get_bank_profile(bank_id, request_context=ctx)
+
+        # Create a pinned mental model
+        model = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Test Model",
+            description="Test description",
+            subtype="pinned",
+            request_context=ctx,
+        )
+
+        assert model is not None
+        model_id = model["id"]
+
+        # Attempt to refresh (may not actually refresh if no data, but hook should be called)
+        try:
+            await memory.refresh_mental_model(
+                bank_id=bank_id,
+                model_id=model_id,
+                request_context=ctx,
+            )
+        except Exception:
+            pass  # May fail if no data
+
+        # Check pre-hook was called
+        assert len(validator.pre_refresh_mental_model_calls) == 1
+        pre_ctx = validator.pre_refresh_mental_model_calls[0]
+        assert pre_ctx.bank_id == bank_id
+        assert pre_ctx.model_id == model_id
+        assert pre_ctx.request_context == ctx
+
+    @pytest.mark.asyncio
+    async def test_refresh_mental_model_post_hook_receives_token_usage(
+        self, memory_with_tracking_validator
+    ):
+        """Post-refresh-mental-model hook receives token usage information."""
+        import uuid
+
+        memory, validator = memory_with_tracking_validator
+        bank_id = f"test-refresh-mm-tokens-{uuid.uuid4().hex[:8]}"
+        ctx = RequestContext(api_key="test-key")
+
+        # Store some content first
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[
+                {"content": "Alice is a software engineer who works on machine learning."},
+                {"content": "Alice enjoys hiking and outdoor activities on weekends."},
+                {"content": "Alice has been working at the company for 5 years."},
+            ],
+            request_context=ctx,
+        )
+
+        # Create a pinned mental model
+        model = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Alice Profile",
+            description="Profile of Alice including work and hobbies",
+            subtype="pinned",
+            request_context=ctx,
+        )
+
+        if model:
+            model_id = model["id"]
+
+            # Refresh the mental model
+            result = await memory.refresh_mental_model(
+                bank_id=bank_id,
+                model_id=model_id,
+                request_context=ctx,
+            )
+
+            # Check post-hook was called with token usage
+            if validator.post_refresh_mental_model_calls:
+                post_result = validator.post_refresh_mental_model_calls[0]
+                assert post_result.bank_id == bank_id
+                assert post_result.model_id == model_id
+                assert post_result.request_context == ctx
+                assert post_result.success is True
+                assert post_result.error is None
+
+                # Token usage should be populated (may be 0 if refresh was skipped)
+                assert post_result.total_tokens >= 0
+                assert post_result.input_tokens >= 0
+                assert post_result.output_tokens >= 0
+                assert post_result.duration_ms >= 0
+

 class TestTenantExtension:
    """Tests for TenantExtension and ApiKeyTenantExtension."""
@@ -947,172 +947,3 @@ so the algorithm learns to box out. See you next week!
                    raise e


-# =============================================================================
-# DISPOSITION INFERENCE TESTS
-# =============================================================================
-
-class TestDispositionInference:
-    """Tests for LLM-based disposition trait inference from background."""
-
-    @pytest.mark.asyncio
-    async def test_background_merge_with_disposition_inference(self, memory, request_context):
-        """Test that background merge infers disposition traits by default."""
-        import uuid
-        bank_id = f"test_infer_{uuid.uuid4().hex[:8]}"
-
-        result = await memory.merge_bank_background(
-            bank_id,
-            "I am a creative software engineer who loves innovation and trying new technologies",
-            update_disposition=True,
-            request_context=request_context,
-        )
-
-        assert "background" in result
-        assert "disposition" in result
-
-        background = result["background"]
-        disposition = result["disposition"]
-
-        assert "creative" in background.lower() or "innovation" in background.lower()
-
-        # Check that new traits are present with valid values (1-5)
-        required_traits = ["skepticism", "literalism", "empathy"]
-        for trait in required_traits:
-            assert trait in disposition
-            assert 1 <= disposition[trait] <= 5
-
-    @pytest.mark.asyncio
-    async def test_background_merge_without_disposition_inference(self, memory, request_context):
-        """Test that background merge skips disposition inference when disabled."""
-        import uuid
-        bank_id = f"test_no_infer_{uuid.uuid4().hex[:8]}"
-
-        initial_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-        initial_disposition = initial_profile["disposition"]
-
-        result = await memory.merge_bank_background(
-            bank_id,
-            "I am a data scientist",
-            update_disposition=False,
-            request_context=request_context,
-        )
-
-        assert "background" in result
-        assert "disposition" not in result
-
-        final_profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-        final_disposition = final_profile["disposition"]
-
-        assert initial_disposition == final_disposition
-
-    @pytest.mark.asyncio
-    async def test_disposition_inference_for_lawyer(self, memory, request_context):
-        """Test disposition inference for lawyer profile (high skepticism, high literalism)."""
-        import uuid
-        bank_id = f"test_lawyer_{uuid.uuid4().hex[:8]}"
-
-        result = await memory.merge_bank_background(
-            bank_id,
-            "I am a lawyer who focuses on contract details and never takes claims at face value",
-            update_disposition=True,
-            request_context=request_context,
-        )
-
-        disposition = result["disposition"]
-
-        # Lawyers should have higher skepticism and literalism
-        assert disposition["skepticism"] >= 3
-        assert disposition["literalism"] >= 3
-
-    @pytest.mark.asyncio
-    async def test_disposition_inference_for_therapist(self, memory, request_context):
-        """Test disposition inference for therapist profile (high empathy)."""
-        import uuid
-        bank_id = f"test_therapist_{uuid.uuid4().hex[:8]}"
-
-        result = await memory.merge_bank_background(
-            bank_id,
-            "I am a therapist who deeply understands and connects with people's emotional struggles",
-            update_disposition=True,
-            request_context=request_context,
-        )
-
-        disposition = result["disposition"]
-
-        # Therapists should have higher empathy
-        assert disposition["empathy"] >= 3
-
-    @pytest.mark.asyncio
-    async def test_disposition_updates_in_database(self, memory, request_context):
-        """Test that inferred disposition is actually stored in database."""
-        import uuid
-        bank_id = f"test_db_update_{uuid.uuid4().hex[:8]}"
-
-        result = await memory.merge_bank_background(
-            bank_id,
-            "I am an innovative designer",
-            update_disposition=True,
-            request_context=request_context,
-        )
-
-        inferred_disposition = result["disposition"]
-
-        profile = await memory.get_bank_profile(bank_id, request_context=request_context)
-        db_disposition = profile["disposition"]
-
-        # Compare values (db_disposition is a Pydantic model)
-        assert db_disposition.skepticism == inferred_disposition["skepticism"]
-        assert db_disposition.literalism == inferred_disposition["literalism"]
-        assert db_disposition.empathy == inferred_disposition["empathy"]
-
-    @pytest.mark.asyncio
-    async def test_multiple_background_merges_update_disposition(self, memory, request_context):
-        """Test that each background merge can update disposition."""
-        import uuid
-        bank_id = f"test_multi_merge_{uuid.uuid4().hex[:8]}"
-
-        result1 = await memory.merge_bank_background(
-            bank_id,
-            "I am a software engineer",
-            update_disposition=True,
-            request_context=request_context,
-        )
-        disposition1 = result1["disposition"]
-
-        result2 = await memory.merge_bank_background(
-            bank_id,
-            "I love creative problem solving and innovation",
-            update_disposition=True,
-            request_context=request_context,
-        )
-        disposition2 = result2["disposition"]
-
-        assert "engineer" in result2["background"].lower() or "software" in result2["background"].lower()
-        assert "creative" in result2["background"].lower() or "innovation" in result2["background"].lower()
-
-    @pytest.mark.asyncio
-    async def test_background_merge_conflict_resolution_with_disposition(self, memory, request_context):
-        """Test that conflicts are resolved and disposition reflects final background."""
-        import uuid
-        bank_id = f"test_conflict_{uuid.uuid4().hex[:8]}"
-
-        await memory.merge_bank_background(
-            bank_id,
-            "I was born in Colorado and prefer stability",
-            update_disposition=True,
-            request_context=request_context,
-        )
-
-        result = await memory.merge_bank_background(
-            bank_id,
-            "You were born in Texas and are very skeptical of people",
-            update_disposition=True,
-            request_context=request_context,
-        )
-
-        background = result["background"]
-        disposition = result["disposition"]
-
-        assert "texas" in background.lower()
-        # Higher skepticism expected from "very skeptical of people"
-        assert disposition["skepticism"] >= 3
@@ -51,7 +51,7 @@ Marcus: Yeah, I realized I was being too optimistic about their defense.
    results = await memory.recall_async(
        bank_id=bank_id,
        query="Marcus prediction Rams",
-        fact_type=['opinion', 'experience', 'world'],
+        fact_type=['experience', 'world'],
        budget=Budget.LOW,
        max_tokens=8192,
        request_context=request_context,
@@ -61,8 +61,8 @@ Marcus: Yeah, I realized I was being too optimistic about their defense.
    for i, result in enumerate(results.results):
        print(f"{i+1}. [{result.mentioned_at}] {result.text[:100]}")

-    # Get all opinion facts (Marcus's predictions/statements)
-    agent_facts = [r for r in results.results if r.fact_type == 'opinion']
+    # Get all facts (Marcus's predictions/statements)
+    agent_facts = results.results

    print(f"\n=== Agent facts (Marcus's statements) ===")
    for i, fact in enumerate(agent_facts):
@@ -70,6 +70,7 @@ Marcus: Yeah, I realized I was being too optimistic about their defense.

    # Check that agent facts have different timestamps
    if len(agent_facts) >= 2:
+        # Parse timestamps
        timestamps = [datetime.fromisoformat(f.mentioned_at.replace('Z', '+00:00')) for f in agent_facts]

        # Verify timestamps are different (have time offsets)
@@ -77,42 +78,40 @@ Marcus: Yeah, I realized I was being too optimistic about their defense.
        assert len(unique_timestamps) == len(timestamps), \
            f"Expected unique timestamps for each fact, but got duplicates: {timestamps}"

-        # Verify timestamps are in order (ascending)
-        for i in range(len(timestamps) - 1):
-            assert timestamps[i] < timestamps[i + 1], \
-                f"Facts should be ordered by time. Fact {i} ({timestamps[i]}) >= Fact {i+1} ({timestamps[i+1]})"
+        # Sort facts by timestamp for ordering check
+        # Note: recall returns by relevance, not time order
+        sorted_facts = sorted(agent_facts, key=lambda f: datetime.fromisoformat(f.mentioned_at.replace('Z', '+00:00')))
+        sorted_timestamps = [datetime.fromisoformat(f.mentioned_at.replace('Z', '+00:00')) for f in sorted_facts]
+
+        # Verify sorted timestamps are in ascending order
+        for i in range(len(sorted_timestamps) - 1):
+            assert sorted_timestamps[i] < sorted_timestamps[i + 1], \
+                f"Facts should have sequential timestamps. Fact {i} ({sorted_timestamps[i]}) >= Fact {i+1} ({sorted_timestamps[i+1]})"

        # Verify reasonable time spacing (should be ~10 seconds apart)
-        time_diffs = [(timestamps[i+1] - timestamps[i]).total_seconds() for i in range(len(timestamps) - 1)]
+        time_diffs = [(sorted_timestamps[i+1] - sorted_timestamps[i]).total_seconds() for i in range(len(sorted_timestamps) - 1)]
        print(f"\n=== Time differences between facts: {time_diffs} seconds ===")

        # Each fact should be 10+ seconds apart (allowing for some flexibility)
        for diff in time_diffs:
            assert diff >= 5, f"Expected at least 5 seconds between facts, got {diff}"

+        # Update agent_facts to be sorted for subsequent checks
+        agent_facts = sorted_facts
+        timestamps = sorted_timestamps
+
        print(f"\n✅ All {len(agent_facts)} agent facts have properly ordered timestamps")

-    # Verify that retrieval returns facts in chronological order
-    # The first prediction should come before the changed prediction
+    # Verify that facts capture the key information
+    # Note: LLM may merge related predictions into single facts
    agent_texts = [f.text.lower() for f in agent_facts]
+    all_text = " ".join(agent_texts)

-    # Look for evidence of the sequence
-    has_first_prediction = any('27' in text and '24' in text for text in agent_texts)
-    has_changed_prediction = any('chang' in text or 'by 3' in text or 'realized' in text for text in agent_texts)
+    # Look for evidence of the predictions being captured (may be merged or separate)
+    has_prediction_info = '27' in all_text or 'rams' in all_text or 'prediction' in all_text

-    if has_first_prediction and has_changed_prediction:
-        # Find indices
-        first_idx = next(i for i, text in enumerate(agent_texts) if '27' in text and '24' in text)
-        changed_idx = next(i for i, text in enumerate(agent_texts) if 'chang' in text or 'by 3' in text or 'realized' in text)
-
-        print(f"\nFirst prediction at index {first_idx}: {agent_facts[first_idx].text[:100]}")
-        print(f"Changed prediction at index {changed_idx}: {agent_facts[changed_idx].text[:100]}")
-
-        # The original prediction should come before the changed one
-        assert timestamps[first_idx] < timestamps[changed_idx], \
-            "Original prediction should have earlier timestamp than changed prediction"
-
-        print(f"\n✅ Temporal ordering preserved: First prediction came before changed prediction")
+    assert has_prediction_info, "Facts should contain information about Marcus's predictions"
+    print(f"\n✅ Facts capture prediction information")

    # Cleanup
    await memory.delete_bank(bank_id, request_context=request_context)
@@ -156,14 +155,14 @@ Alice: I reconsidered the team's experience level.
    results = await memory.recall_async(
        bank_id=bank_id,
        query="Alice preference React Vue",
-        fact_type=['opinion', 'experience'],
+        fact_type=['experience', 'world'],
        budget=Budget.LOW,
        max_tokens=8192,
        request_context=request_context,
    )

    print(f"\n=== Retrieved {len(results.results)} agent facts ===")
-    agent_facts = [r for r in results.results if r.fact_type in ('opinion', 'experience')]
+    agent_facts = results.results

    for i, fact in enumerate(agent_facts):
        print(f"{i+1}. [{fact.mentioned_at}] {fact.text[:80]}")
@@ -60,17 +60,6 @@ async def test_full_api_workflow(api_client, test_bank_id):
    assert response.status_code == 200
    profile = response.json()
    assert "disposition" in profile
-    assert "background" in profile
-
-    # Add background
-    response = await api_client.post(
-        f"/v1/default/banks/{test_bank_id}/background",
-        json={
-            "content": "A software engineer passionate about AI and memory systems."
-        }
-    )
-    assert response.status_code == 200
-    assert "software engineer" in response.json()["background"].lower()

    # ================================================================
    # 2. Memory Storage
@@ -244,7 +233,9 @@ async def test_full_api_workflow(api_client, test_bank_id):
    response = await api_client.get(f"/v1/default/banks/{test_bank_id}/profile")
    assert response.status_code == 200
    updated_profile = response.json()
-    assert "software engineer" in updated_profile["background"].lower()
+    assert updated_profile["disposition"]["skepticism"] == 4
+    assert updated_profile["disposition"]["literalism"] == 3
+    assert updated_profile["disposition"]["empathy"] == 4

    # ================================================================
    # 8. Test Entity Endpoints
@@ -289,11 +280,11 @@ async def test_full_api_workflow(api_client, test_bank_id):
        entity_detail = response.json()
        assert "id" in entity_detail

-        # Test regenerate observations
+        # Test regenerate observations (deprecated - returns 410 Gone)
        response = await api_client.post(
            f"/v1/default/banks/{test_bank_id}/entities/{entity_id}/regenerate"
        )
-        assert response.status_code == 200
+        assert response.status_code == 410  # Deprecated endpoint

    # ================================================================
    # 9. List All Banks (should include our test bank)
@@ -845,9 +836,8 @@ async def test_reflect_structured_output(api_client):
    assert response.status_code == 200
    result = response.json()

-    # Verify text field exists (empty when using structured output)
+    # Verify text field exists (may contain text even with structured output)
    assert "text" in result
-    assert result["text"] == ""

    # Verify structured output exists and has expected structure
    assert "structured_output" in result
@@ -979,20 +969,24 @@ async def test_reflect_returns_token_usage(api_client):
    assert "text" in result
    assert len(result["text"]) > 0

-    # Verify usage field exists and has expected structure
+    # Verify usage field exists (may be None for agentic reflect which makes multiple LLM calls)
    assert "usage" in result, "Response should include 'usage' field"
    usage = result["usage"]
-    assert usage is not None, "Usage should not be None for reflect"
-    assert "input_tokens" in usage, "Usage should have 'input_tokens'"
-    assert "output_tokens" in usage, "Usage should have 'output_tokens'"
-    assert "total_tokens" in usage, "Usage should have 'total_tokens'"

-    # Verify token counts are valid
-    assert usage["input_tokens"] > 0, f"Expected input_tokens > 0, got {usage['input_tokens']}"
-    assert usage["output_tokens"] >= 0, f"Expected output_tokens >= 0, got {usage['output_tokens']}"
-    assert usage["total_tokens"] == usage["input_tokens"] + usage["output_tokens"]
+    # Usage is optional - agentic reflect doesn't aggregate multiple LLM call usages
+    if usage is not None:
+        assert "input_tokens" in usage, "Usage should have 'input_tokens'"
+        assert "output_tokens" in usage, "Usage should have 'output_tokens'"
+        assert "total_tokens" in usage, "Usage should have 'total_tokens'"

-    print(f"Reflect token usage: input={usage['input_tokens']}, output={usage['output_tokens']}, total={usage['total_tokens']}")
+        # Verify token counts are valid
+        assert usage["input_tokens"] > 0, f"Expected input_tokens > 0, got {usage['input_tokens']}"
+        assert usage["output_tokens"] >= 0, f"Expected output_tokens >= 0, got {usage['output_tokens']}"
+        assert usage["total_tokens"] == usage["input_tokens"] + usage["output_tokens"]
+
+        print(f"Reflect token usage: input={usage['input_tokens']}, output={usage['output_tokens']}, total={usage['total_tokens']}")
+    else:
+        print("Reflect usage is None (expected for agentic reflect)")


@pytest.mark.asyncio
@@ -0,0 +1,325 @@
+"""
+Tests for LLM tool calling functionality.
+"""
+
+import pytest
+
+from hindsight_api.engine.llm_wrapper import LLMProvider
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult
+
+
+# Sample tools for testing
+SAMPLE_TOOLS = [
+    {
+        "type": "function",
+        "function": {
+            "name": "get_weather",
+            "description": "Get weather for a location",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "location": {"type": "string", "description": "City name"},
+                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
+                },
+                "required": ["location"],
+            },
+        },
+    },
+    {
+        "type": "function",
+        "function": {
+            "name": "search",
+            "description": "Search for information",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "query": {"type": "string", "description": "Search query"},
+                },
+                "required": ["query"],
+            },
+        },
+    },
+]
+
+
+class TestMockToolCalling:
+    """Test tool calling with mock provider."""
+
+    @pytest.mark.asyncio
+    async def test_call_with_tools_returns_tool_calls(self):
+        """Test that mock provider can return tool calls."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+
+        # Set mock response to return tool calls
+        llm.set_mock_response([
+            {"name": "get_weather", "arguments": {"location": "Paris", "unit": "celsius"}},
+        ])
+
+        result = await llm.call_with_tools(
+            messages=[{"role": "user", "content": "What's the weather in Paris?"}],
+            tools=SAMPLE_TOOLS,
+        )
+
+        assert isinstance(result, LLMToolCallResult)
+        assert len(result.tool_calls) == 1
+        assert result.tool_calls[0].name == "get_weather"
+        assert result.tool_calls[0].arguments == {"location": "Paris", "unit": "celsius"}
+        assert result.finish_reason == "tool_calls"
+
+    @pytest.mark.asyncio
+    async def test_call_with_tools_returns_content(self):
+        """Test that mock provider can return plain content."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+
+        # Default mock response is plain content
+        result = await llm.call_with_tools(
+            messages=[{"role": "user", "content": "Hello"}],
+            tools=SAMPLE_TOOLS,
+        )
+
+        assert isinstance(result, LLMToolCallResult)
+        assert result.content == "mock response"
+        assert len(result.tool_calls) == 0
+        assert result.finish_reason == "stop"
+
+    @pytest.mark.asyncio
+    async def test_call_with_tools_records_calls(self):
+        """Test that mock calls are recorded."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+        llm.clear_mock_calls()
+
+        await llm.call_with_tools(
+            messages=[{"role": "user", "content": "Test message"}],
+            tools=SAMPLE_TOOLS,
+            scope="test_scope",
+        )
+
+        calls = llm.get_mock_calls()
+        assert len(calls) == 1
+        assert calls[0]["scope"] == "test_scope"
+        assert "get_weather" in calls[0]["tools"]
+        assert "search" in calls[0]["tools"]
+
+    @pytest.mark.asyncio
+    async def test_call_with_tools_multiple_tool_calls(self):
+        """Test handling multiple tool calls in one response."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+
+        llm.set_mock_response([
+            {"name": "get_weather", "arguments": {"location": "Paris"}},
+            {"name": "search", "arguments": {"query": "weather forecast"}},
+        ])
+
+        result = await llm.call_with_tools(
+            messages=[{"role": "user", "content": "Weather in Paris and search for forecasts"}],
+            tools=SAMPLE_TOOLS,
+        )
+
+        assert len(result.tool_calls) == 2
+        assert result.tool_calls[0].name == "get_weather"
+        assert result.tool_calls[1].name == "search"
+
+    @pytest.mark.asyncio
+    async def test_call_with_tools_accepts_llm_tool_call_result(self):
+        """Test that mock can accept LLMToolCallResult directly."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+
+        expected_result = LLMToolCallResult(
+            content="Here's the info",
+            tool_calls=[LLMToolCall(id="call_123", name="search", arguments={"query": "test"})],
+            finish_reason="tool_calls",
+        )
+        llm.set_mock_response(expected_result)
+
+        result = await llm.call_with_tools(
+            messages=[{"role": "user", "content": "Search for test"}],
+            tools=SAMPLE_TOOLS,
+        )
+
+        assert result == expected_result
+
+
+class TestToolCallConversation:
+    """Test tool call conversation flow."""
+
+    @pytest.mark.asyncio
+    async def test_tool_result_message_format(self):
+        """Test that tool result messages can be passed in subsequent calls."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+
+        # First call returns tool call
+        llm.set_mock_response([{"name": "get_weather", "arguments": {"location": "Paris"}}])
+
+        result1 = await llm.call_with_tools(
+            messages=[{"role": "user", "content": "What's the weather?"}],
+            tools=SAMPLE_TOOLS,
+        )
+
+        # Build conversation with tool result
+        messages = [
+            {"role": "user", "content": "What's the weather?"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {
+                        "id": result1.tool_calls[0].id,
+                        "type": "function",
+                        "function": {
+                            "name": result1.tool_calls[0].name,
+                            "arguments": '{"location": "Paris"}',
+                        },
+                    }
+                ],
+            },
+            {
+                "role": "tool",
+                "tool_call_id": result1.tool_calls[0].id,
+                "content": '{"temperature": 20, "conditions": "sunny"}',
+            },
+        ]
+
+        # Second call should work with tool result in history
+        llm.set_mock_response(None)  # Reset to default
+        result2 = await llm.call_with_tools(
+            messages=messages,
+            tools=SAMPLE_TOOLS,
+        )
+
+        assert result2.content == "mock response"
+
+
+class TestToolSchemas:
+    """Test tool schema handling."""
+
+    @pytest.mark.asyncio
+    async def test_empty_tools_list(self):
+        """Test calling with empty tools list."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+
+        result = await llm.call_with_tools(
+            messages=[{"role": "user", "content": "Hello"}],
+            tools=[],
+        )
+
+        assert result.content == "mock response"
+
+    @pytest.mark.asyncio
+    async def test_tool_with_no_required_params(self):
+        """Test tool with no required parameters."""
+        llm = LLMProvider(provider="mock", api_key="", base_url="", model="mock")
+
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "list_items",
+                    "description": "List all items",
+                    "parameters": {"type": "object", "properties": {}, "required": []},
+                },
+            }
+        ]
+
+        llm.set_mock_response([{"name": "list_items", "arguments": {}}])
+
+        result = await llm.call_with_tools(
+            messages=[{"role": "user", "content": "List items"}],
+            tools=tools,
+        )
+
+        assert len(result.tool_calls) == 1
+        assert result.tool_calls[0].name == "list_items"
+        assert result.tool_calls[0].arguments == {}
+
+
+class TestReflectToolSchemas:
+    """Test reflect agent tool schemas."""
+
+    def test_get_reflect_tools_default(self):
+        """Test getting default reflect tools."""
+        from hindsight_api.engine.reflect.tools_schema import get_reflect_tools
+
+        tools = get_reflect_tools()
+
+        tool_names = [t["function"]["name"] for t in tools]
+        assert "list_mental_models" in tool_names
+        assert "get_mental_model" in tool_names
+        assert "recall" in tool_names
+        assert "learn" in tool_names
+        assert "expand" in tool_names
+        assert "done" in tool_names
+
+    def test_get_reflect_tools_without_learn(self):
+        """Test getting reflect tools without learn."""
+        from hindsight_api.engine.reflect.tools_schema import get_reflect_tools
+
+        tools = get_reflect_tools(enable_learn=False)
+
+        tool_names = [t["function"]["name"] for t in tools]
+        assert "learn" not in tool_names
+        assert "recall" in tool_names
+        assert "done" in tool_names
+
+    def test_get_reflect_tools_answer_mode(self):
+        """Test getting reflect tools with answer output mode."""
+        from hindsight_api.engine.reflect.tools_schema import get_reflect_tools
+
+        tools = get_reflect_tools()
+
+        done_tool = next(t for t in tools if t["function"]["name"] == "done")
+        params = done_tool["function"]["parameters"]["properties"]
+
+        assert "answer" in params
+        assert "memory_ids" in params
+        assert "model_ids" in params
+
+
+class TestLLMToolCallResult:
+    """Test LLMToolCallResult model."""
+
+    def test_tool_call_result_defaults(self):
+        """Test default values for LLMToolCallResult."""
+        result = LLMToolCallResult()
+
+        assert result.content is None
+        assert result.tool_calls == []
+        assert result.finish_reason is None
+
+    def test_tool_call_result_with_content(self):
+        """Test LLMToolCallResult with content."""
+        result = LLMToolCallResult(content="Hello", finish_reason="stop")
+
+        assert result.content == "Hello"
+        assert result.tool_calls == []
+        assert result.finish_reason == "stop"
+
+    def test_tool_call_result_with_tool_calls(self):
+        """Test LLMToolCallResult with tool calls."""
+        result = LLMToolCallResult(
+            tool_calls=[
+                LLMToolCall(id="call_1", name="test_tool", arguments={"arg": "value"}),
+            ],
+            finish_reason="tool_calls",
+        )
+
+        assert result.content is None
+        assert len(result.tool_calls) == 1
+        assert result.tool_calls[0].name == "test_tool"
+        assert result.finish_reason == "tool_calls"
+
+
+class TestLLMToolCall:
+    """Test LLMToolCall model."""
+
+    def test_tool_call_basic(self):
+        """Test basic LLMToolCall creation."""
+        call = LLMToolCall(id="call_123", name="get_weather", arguments={"location": "Paris"})
+
+        assert call.id == "call_123"
+        assert call.name == "get_weather"
+        assert call.arguments == {"location": "Paris"}
+
+    def test_tool_call_empty_arguments(self):
+        """Test LLMToolCall with empty arguments."""
+        call = LLMToolCall(id="call_456", name="list_items", arguments={})
+
+        assert call.arguments == {}
@@ -19,6 +19,7 @@ import pytest_asyncio
 from hindsight_api import MemoryEngine, LLMConfig, LocalSTEmbeddings, RequestContext
 from hindsight_api.engine.cross_encoder import LocalSTCrossEncoder
 from hindsight_api.engine.query_analyzer import DateparserQueryAnalyzer
+from hindsight_api.engine.task_backend import SyncTaskBackend
 from hindsight_api.engine.retain.fact_extraction import FactExtractionResponse, ExtractedFact
 from hindsight_api.engine.llm_wrapper import TokenUsage

@@ -106,6 +107,7 @@ class TestLargeBatchRetain:
            pool_max_size=10,
            run_migrations=False,
            skip_llm_verification=True,  # Skip LLM verification since we're mocking
+            task_backend=SyncTaskBackend(),  # Execute tasks immediately in tests
        )
        await mem.initialize()
        yield mem
@@ -0,0 +1,400 @@
+"""
+Tests for hindsight_api.main module (single-worker code path).
+
+The main.py module is used when running with a single worker:
+    hindsight-api  (or hindsight-api --workers 1)
+
+When workers=1, main.py creates the app directly and passes it to uvicorn.
+These tests ensure that extensions are properly loaded in this code path.
+
+Compare with test_server_module.py which tests the multi-worker path (workers > 1).
+"""
+
+import sys
+from unittest.mock import MagicMock, patch
+
+
+class TestMainModuleExtensionLoading:
+    """Tests that main.py correctly loads extensions when configured via environment."""
+
+    def test_main_loads_tenant_extension_when_configured(self, monkeypatch):
+        """
+        Verify that main.py loads tenant extension from HINDSIGHT_API_TENANT_EXTENSION.
+
+        This ensures extension loading works in the single-worker code path.
+        """
+        # Set up environment to configure a tenant extension
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TENANT_EXTENSION",
+            "tests.test_main_module:MockTenantExtension",
+        )
+        # Ensure single worker mode
+        monkeypatch.setenv("HINDSIGHT_API_WORKERS", "1")
+
+        # Track what extensions were loaded via load_extension
+        loaded_extensions = {}
+
+        # Get the real load_extension function
+        from hindsight_api.extensions.loader import load_extension as real_load_extension
+
+        def tracking_load_extension(name, base_class):
+            """Track calls to load_extension and delegate to original."""
+            result = real_load_extension(name, base_class)
+            loaded_extensions[name] = result
+            return result
+
+        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
+             patch("hindsight_api.main.create_app") as mock_create_app, \
+             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main.load_extension", side_effect=tracking_load_extension), \
+             patch("hindsight_api.main.DefaultExtensionContext"), \
+             patch("hindsight_api.main.print_banner"), \
+             patch("uvicorn.run"):  # Don't actually start uvicorn
+
+            mock_config = MagicMock()
+            mock_config.host = "0.0.0.0"
+            mock_config.port = 8888
+            mock_config.log_level = "info"
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_engine.return_value = MagicMock()
+            mock_create_app.return_value = MagicMock()
+
+            # Mock sys.argv to simulate CLI invocation
+            with patch.object(sys, 'argv', ['hindsight-api']):
+                from hindsight_api.main import main
+                main()
+
+        # Verify TENANT extension was loaded
+        assert "TENANT" in loaded_extensions, \
+            "main.py did not call load_extension('TENANT', ...) - extensions not loaded!"
+        assert loaded_extensions["TENANT"] is not None, \
+            "load_extension('TENANT', ...) returned None despite env var being set"
+        assert isinstance(loaded_extensions["TENANT"], MockTenantExtension), \
+            f"Expected MockTenantExtension, got {type(loaded_extensions['TENANT'])}"
+
+    def test_main_loads_operation_validator_when_configured(self, monkeypatch):
+        """
+        Verify that main.py loads operation validator from HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION.
+        """
+        monkeypatch.setenv(
+            "HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION",
+            "tests.test_main_module:MockOperationValidator",
+        )
+        monkeypatch.setenv("HINDSIGHT_API_WORKERS", "1")
+
+        loaded_extensions = {}
+
+        from hindsight_api.extensions.loader import load_extension as real_load_extension
+
+        def tracking_load_extension(name, base_class):
+            result = real_load_extension(name, base_class)
+            loaded_extensions[name] = result
+            return result
+
+        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
+             patch("hindsight_api.main.create_app") as mock_create_app, \
+             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main.load_extension", side_effect=tracking_load_extension), \
+             patch("hindsight_api.main.DefaultExtensionContext"), \
+             patch("hindsight_api.main.print_banner"), \
+             patch("uvicorn.run"):
+
+            mock_config = MagicMock()
+            mock_config.host = "0.0.0.0"
+            mock_config.port = 8888
+            mock_config.log_level = "info"
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_engine.return_value = MagicMock()
+            mock_create_app.return_value = MagicMock()
+
+            with patch.object(sys, 'argv', ['hindsight-api']):
+                from hindsight_api.main import main
+                main()
+
+        assert "OPERATION_VALIDATOR" in loaded_extensions, \
+            "main.py did not call load_extension('OPERATION_VALIDATOR', ...)"
+        assert loaded_extensions["OPERATION_VALIDATOR"] is not None
+        assert isinstance(loaded_extensions["OPERATION_VALIDATOR"], MockOperationValidator)
+
+    def test_main_passes_extensions_to_memory_engine(self, monkeypatch):
+        """
+        Verify that main.py passes loaded extensions to MemoryEngine constructor.
+
+        This is the critical test - even if extensions are loaded, they must be
+        passed to MemoryEngine for authentication to work.
+        """
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TENANT_EXTENSION",
+            "tests.test_main_module:MockTenantExtension",
+        )
+        monkeypatch.setenv("HINDSIGHT_API_WORKERS", "1")
+
+        memory_engine_calls = []
+
+        def capture_memory_engine(*args, **kwargs):
+            memory_engine_calls.append({"args": args, "kwargs": kwargs})
+            return MagicMock()
+
+        with patch("hindsight_api.main.MemoryEngine", side_effect=capture_memory_engine), \
+             patch("hindsight_api.main.create_app") as mock_create_app, \
+             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main.DefaultExtensionContext"), \
+             patch("hindsight_api.main.print_banner"), \
+             patch("uvicorn.run"):
+
+            mock_config = MagicMock()
+            mock_config.host = "0.0.0.0"
+            mock_config.port = 8888
+            mock_config.log_level = "info"
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_create_app.return_value = MagicMock()
+
+            with patch.object(sys, 'argv', ['hindsight-api']):
+                from hindsight_api.main import main
+                main()
+
+        # Verify MemoryEngine was called
+        assert len(memory_engine_calls) == 1, "MemoryEngine should be called exactly once"
+
+        call_kwargs = memory_engine_calls[0]["kwargs"]
+
+        # THE CRITICAL ASSERTION: tenant_extension must be passed and not None
+        assert "tenant_extension" in call_kwargs, \
+            "MemoryEngine was not called with tenant_extension parameter!"
+        assert call_kwargs["tenant_extension"] is not None, \
+            "tenant_extension was None - main.py did not pass loaded extension to MemoryEngine!"
+
+    def test_main_sets_extension_context_on_tenant_extension(self, monkeypatch):
+        """
+        Verify that main.py sets the extension context on tenant extension.
+
+        This is required for tenant extensions that need to provision schemas.
+        """
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TENANT_EXTENSION",
+            "tests.test_main_module:MockTenantExtension",
+        )
+        monkeypatch.setenv("HINDSIGHT_API_WORKERS", "1")
+
+        captured_tenant_ext = [None]
+
+        def capture_memory_engine(*args, **kwargs):
+            captured_tenant_ext[0] = kwargs.get("tenant_extension")
+            return MagicMock()
+
+        context_created = []
+
+        def capture_context(*args, **kwargs):
+            ctx = MagicMock()
+            context_created.append(ctx)
+            return ctx
+
+        with patch("hindsight_api.main.MemoryEngine", side_effect=capture_memory_engine), \
+             patch("hindsight_api.main.create_app") as mock_create_app, \
+             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main.DefaultExtensionContext", side_effect=capture_context), \
+             patch("hindsight_api.main.print_banner"), \
+             patch("uvicorn.run"):
+
+            mock_config = MagicMock()
+            mock_config.host = "0.0.0.0"
+            mock_config.port = 8888
+            mock_config.log_level = "info"
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_create_app.return_value = MagicMock()
+
+            with patch.object(sys, 'argv', ['hindsight-api']):
+                from hindsight_api.main import main
+                main()
+
+        # Verify context was created and set
+        assert len(context_created) == 1, "DefaultExtensionContext should be created"
+        assert captured_tenant_ext[0] is not None, "Tenant extension should be captured"
+        assert captured_tenant_ext[0]._context_set, \
+            "set_context was not called on tenant extension"
+
+    def test_main_works_without_extensions(self, monkeypatch):
+        """
+        Verify that main.py works correctly when no extensions are configured.
+        """
+        # Ensure no extension env vars are set
+        monkeypatch.delenv("HINDSIGHT_API_TENANT_EXTENSION", raising=False)
+        monkeypatch.delenv("HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION", raising=False)
+        monkeypatch.setenv("HINDSIGHT_API_WORKERS", "1")
+
+        memory_engine_calls = []
+
+        def capture_memory_engine(*args, **kwargs):
+            memory_engine_calls.append({"args": args, "kwargs": kwargs})
+            return MagicMock()
+
+        with patch("hindsight_api.main.MemoryEngine", side_effect=capture_memory_engine), \
+             patch("hindsight_api.main.create_app") as mock_create_app, \
+             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main.print_banner"), \
+             patch("uvicorn.run"):
+
+            mock_config = MagicMock()
+            mock_config.host = "0.0.0.0"
+            mock_config.port = 8888
+            mock_config.log_level = "info"
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_create_app.return_value = MagicMock()
+
+            with patch.object(sys, 'argv', ['hindsight-api']):
+                from hindsight_api.main import main
+                main()
+
+        # Should work without extensions
+        assert len(memory_engine_calls) == 1
+        call_kwargs = memory_engine_calls[0]["kwargs"]
+
+        # Extensions should be None when not configured
+        assert call_kwargs.get("tenant_extension") is None
+        assert call_kwargs.get("operation_validator") is None
+
+    def test_main_uses_app_object_for_single_worker(self, monkeypatch):
+        """
+        Verify that main.py passes the app object (not import string) when workers=1.
+
+        This is important because it means single-worker mode uses the app created
+        in main.py (with extensions loaded), not server.py.
+        """
+        monkeypatch.setenv("HINDSIGHT_API_WORKERS", "1")
+        monkeypatch.delenv("HINDSIGHT_API_TENANT_EXTENSION", raising=False)
+
+        uvicorn_calls = []
+
+        def capture_uvicorn_run(**kwargs):
+            uvicorn_calls.append(kwargs)
+
+        mock_app = MagicMock()
+
+        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
+             patch("hindsight_api.main.create_app", return_value=mock_app), \
+             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main.print_banner"), \
+             patch("uvicorn.run", side_effect=capture_uvicorn_run):
+
+            mock_config = MagicMock()
+            mock_config.host = "0.0.0.0"
+            mock_config.port = 8888
+            mock_config.log_level = "info"
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_engine.return_value = MagicMock()
+
+            with patch.object(sys, 'argv', ['hindsight-api', '--workers', '1']):
+                from hindsight_api.main import main
+                main()
+
+        assert len(uvicorn_calls) == 1
+        # With workers=1, should pass app object, not import string
+        assert uvicorn_calls[0]["app"] is mock_app, \
+            "main.py should pass app object (not import string) when workers=1"
+
+    def test_main_uses_import_string_for_multiple_workers(self, monkeypatch):
+        """
+        Verify that main.py uses import string when workers > 1.
+
+        This is important because multi-worker mode requires server.py to be imported
+        by each worker process.
+        """
+        monkeypatch.setenv("HINDSIGHT_API_WORKERS", "2")
+        monkeypatch.delenv("HINDSIGHT_API_TENANT_EXTENSION", raising=False)
+
+        uvicorn_calls = []
+
+        def capture_uvicorn_run(**kwargs):
+            uvicorn_calls.append(kwargs)
+
+        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
+             patch("hindsight_api.main.create_app") as mock_create_app, \
+             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main.print_banner"), \
+             patch("uvicorn.run", side_effect=capture_uvicorn_run):
+
+            mock_config = MagicMock()
+            mock_config.host = "0.0.0.0"
+            mock_config.port = 8888
+            mock_config.log_level = "info"
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_engine.return_value = MagicMock()
+            mock_create_app.return_value = MagicMock()
+
+            with patch.object(sys, 'argv', ['hindsight-api', '--workers', '2']):
+                from hindsight_api.main import main
+                main()
+
+        assert len(uvicorn_calls) == 1
+        # With workers > 1, should use import string
+        assert uvicorn_calls[0]["app"] == "hindsight_api.server:app", \
+            "main.py should use import string when workers > 1"
+        assert uvicorn_calls[0]["workers"] == 2
+
+
+# Mock extensions for testing
+from hindsight_api.extensions import (
+    TenantExtension,
+    TenantContext,
+    RequestContext,
+    OperationValidatorExtension,
+    ValidationResult,
+    RetainContext,
+    RecallContext,
+    ReflectContext,
+    RefreshMentalModelContext,
+)
+
+
+class MockTenantExtension(TenantExtension):
+    """Mock tenant extension for testing main.py extension loading."""
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+        self._context_set = False
+
+    async def authenticate(self, request_context: RequestContext) -> TenantContext:
+        return TenantContext(schema_name="public")
+
+    def set_context(self, context) -> None:
+        self._context_set = True
+
+
+class MockOperationValidator(OperationValidatorExtension):
+    """Mock operation validator for testing main.py extension loading."""
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+
+    async def validate_retain(self, ctx: RetainContext) -> ValidationResult:
+        return ValidationResult.accept()
+
+    async def validate_recall(self, ctx: RecallContext) -> ValidationResult:
+        return ValidationResult.accept()
+
+    async def validate_reflect(self, ctx: ReflectContext) -> ValidationResult:
+        return ValidationResult.accept()
+
+    async def validate_refresh_mental_model(self, ctx: RefreshMentalModelContext) -> ValidationResult:
+        return ValidationResult.accept()
@@ -0,0 +1,405 @@
+"""Tests for observation trend computation and evidence-grounded models."""
+
+from datetime import datetime, timedelta, timezone
+
+import pytest
+
+from hindsight_api.engine.reflect.observations import (
+    CandidateObservation,
+    Observation,
+    ObservationEvidence,
+    Trend,
+    compute_trend,
+    verify_evidence_quotes,
+)
+
+
+class TestComputeTrend:
+    """Tests for the compute_trend function."""
+
+    def test_empty_evidence_returns_stale(self):
+        """No evidence should return STALE trend."""
+        trend = compute_trend([])
+        assert trend == Trend.STALE
+
+    def test_all_recent_evidence_returns_new(self):
+        """All evidence within recent window (30 days) should return NEW trend.
+
+        Scenario: User just started using the app and mentioned they like coffee twice.
+        Both mentions are within the last 2 weeks, so this is a NEW observation.
+        """
+        now = datetime.now(timezone.utc)
+        evidence = [
+            ObservationEvidence(
+                memory_id="mem-coffee-morning",
+                quote="I always start my day with a large black coffee",
+                relevance="Shows preference for coffee and morning routine",
+                timestamp=now - timedelta(days=5),
+            ),
+            ObservationEvidence(
+                memory_id="mem-coffee-meeting",
+                quote="grabbed coffee before the standup meeting",
+                relevance="Confirms regular coffee consumption",
+                timestamp=now - timedelta(days=10),
+            ),
+        ]
+
+        trend = compute_trend(evidence, now=now)
+        assert trend == Trend.NEW
+
+    def test_no_recent_evidence_returns_stale(self):
+        """No evidence in recent window should return STALE trend.
+
+        Scenario: User mentioned running 3 months ago but hasn't mentioned it since.
+        The observation about running as a hobby may no longer be accurate.
+        """
+        now = datetime.now(timezone.utc)
+        evidence = [
+            ObservationEvidence(
+                memory_id="mem-running-march",
+                quote="training for a half marathon in the spring",
+                relevance="Shows interest in running",
+                timestamp=now - timedelta(days=60),
+            ),
+            ObservationEvidence(
+                memory_id="mem-running-feb",
+                quote="went for a 10k run this morning",
+                relevance="Active runner",
+                timestamp=now - timedelta(days=100),
+            ),
+        ]
+
+        trend = compute_trend(evidence, now=now)
+        assert trend == Trend.STALE
+
+    def test_stable_evidence_distribution(self):
+        """Evidence spread evenly across time should return STABLE trend.
+
+        Scenario: User has consistently mentioned working remotely over 4 months.
+        Evidence is well-distributed, indicating a stable, ongoing preference.
+        """
+        now = datetime.now(timezone.utc)
+        evidence = [
+            # Recent (within 30 days)
+            ObservationEvidence(
+                memory_id="mem-remote-jan",
+                quote="working from my home office today",
+                relevance="Current remote work",
+                timestamp=now - timedelta(days=5),
+            ),
+            ObservationEvidence(
+                memory_id="mem-remote-dec",
+                quote="the flexibility of remote work is great",
+                relevance="Values remote work",
+                timestamp=now - timedelta(days=15),
+            ),
+            # Middle period (30-90 days)
+            ObservationEvidence(
+                memory_id="mem-remote-nov",
+                quote="set up a standing desk at home",
+                relevance="Invested in home office",
+                timestamp=now - timedelta(days=45),
+            ),
+            ObservationEvidence(
+                memory_id="mem-remote-oct",
+                quote="prefer async communication over meetings",
+                relevance="Remote work style preference",
+                timestamp=now - timedelta(days=60),
+            ),
+            # Older (90+ days)
+            ObservationEvidence(
+                memory_id="mem-remote-sep",
+                quote="switched to fully remote last quarter",
+                relevance="Original transition to remote",
+                timestamp=now - timedelta(days=100),
+            ),
+            ObservationEvidence(
+                memory_id="mem-remote-aug",
+                quote="negotiated remote work in my new contract",
+                relevance="Intentional choice for remote",
+                timestamp=now - timedelta(days=120),
+            ),
+        ]
+
+        trend = compute_trend(evidence, now=now)
+        assert trend == Trend.STABLE
+
+    def test_strengthening_trend(self):
+        """Much more recent evidence than older should return STRENGTHENING trend.
+
+        Scenario: User has been increasingly talking about learning Python recently
+        after mentioning it once months ago. Interest appears to be growing.
+        """
+        now = datetime.now(timezone.utc)
+        evidence = [
+            # Lots of recent evidence - actively learning
+            ObservationEvidence(
+                memory_id="mem-python-project",
+                quote="finished my first Python project - a web scraper",
+                relevance="Completed Python project",
+                timestamp=now - timedelta(days=2),
+            ),
+            ObservationEvidence(
+                memory_id="mem-python-course",
+                quote="halfway through the Python bootcamp",
+                relevance="Active learning",
+                timestamp=now - timedelta(days=5),
+            ),
+            ObservationEvidence(
+                memory_id="mem-python-book",
+                quote="reading Fluent Python, it's excellent",
+                relevance="Deepening knowledge",
+                timestamp=now - timedelta(days=10),
+            ),
+            ObservationEvidence(
+                memory_id="mem-python-practice",
+                quote="solved 50 LeetCode problems in Python",
+                relevance="Practicing skills",
+                timestamp=now - timedelta(days=15),
+            ),
+            ObservationEvidence(
+                memory_id="mem-python-ide",
+                quote="set up VS Code with all the Python extensions",
+                relevance="Setting up environment",
+                timestamp=now - timedelta(days=20),
+            ),
+            # Only one old mention - initial interest
+            ObservationEvidence(
+                memory_id="mem-python-start",
+                quote="thinking about learning Python someday",
+                relevance="Initial interest",
+                timestamp=now - timedelta(days=100),
+            ),
+        ]
+
+        trend = compute_trend(evidence, now=now)
+        assert trend == Trend.STRENGTHENING
+
+    def test_weakening_trend(self):
+        """Much less recent evidence than older should return WEAKENING trend.
+
+        Scenario: User was very active in a book club last year but mentions
+        have tapered off. The observation about being a book club member
+        may be becoming less relevant.
+        """
+        now = datetime.now(timezone.utc)
+        evidence = [
+            # Only one recent mention
+            ObservationEvidence(
+                memory_id="mem-book-recent",
+                quote="haven't had time for book club lately",
+                relevance="Reduced participation",
+                timestamp=now - timedelta(days=10),
+            ),
+            # Lots of older evidence - was very active
+            ObservationEvidence(
+                memory_id="mem-book-aug",
+                quote="hosting book club at my place next week",
+                relevance="Active organizer",
+                timestamp=now - timedelta(days=40),
+            ),
+            ObservationEvidence(
+                memory_id="mem-book-july",
+                quote="leading the discussion on 1984",
+                relevance="Active participant",
+                timestamp=now - timedelta(days=50),
+            ),
+            ObservationEvidence(
+                memory_id="mem-book-june",
+                quote="we picked The Midnight Library for June",
+                relevance="Regular member",
+                timestamp=now - timedelta(days=60),
+            ),
+            ObservationEvidence(
+                memory_id="mem-book-may",
+                quote="book club was amazing tonight",
+                relevance="Enthusiastic member",
+                timestamp=now - timedelta(days=100),
+            ),
+            ObservationEvidence(
+                memory_id="mem-book-april",
+                quote="joined a new book club in my neighborhood",
+                relevance="Started participation",
+                timestamp=now - timedelta(days=110),
+            ),
+            ObservationEvidence(
+                memory_id="mem-book-march",
+                quote="excited to finally join a book club",
+                relevance="Initial enthusiasm",
+                timestamp=now - timedelta(days=120),
+            ),
+        ]
+
+        trend = compute_trend(evidence, now=now)
+        assert trend == Trend.WEAKENING
+
+
+class TestObservationModel:
+    """Tests for the Observation model."""
+
+    def test_observation_computed_trend(self):
+        """Observation should have computed trend property based on evidence."""
+        now = datetime.now(timezone.utc)
+        obs = Observation(
+            title="Morning meeting preference",
+            content="Prefers morning meetings over afternoon ones",
+            evidence=[
+                ObservationEvidence(
+                    memory_id="mem-morning-standup",
+                    quote="I'm most productive in morning meetings",
+                    relevance="Direct preference statement",
+                    timestamp=now - timedelta(days=5),
+                ),
+            ],
+            created_at=now,
+        )
+
+        assert obs.trend == Trend.NEW
+        assert obs.evidence_count == 1
+
+    def test_observation_evidence_span(self):
+        """Observation should compute evidence span correctly.
+
+        The span shows the date range of supporting evidence, helping
+        understand how long this pattern has been observed.
+        """
+        now = datetime.now(timezone.utc)
+        old_time = now - timedelta(days=100)
+        recent_time = now - timedelta(days=5)
+
+        obs = Observation(
+            title="Values work-life balance",
+            content="Values work-life balance highly",
+            evidence=[
+                ObservationEvidence(
+                    memory_id="mem-balance-old",
+                    quote="turned down a promotion because of the hours",
+                    relevance="Prioritized balance over advancement",
+                    timestamp=old_time,
+                ),
+                ObservationEvidence(
+                    memory_id="mem-balance-recent",
+                    quote="always log off by 6pm no matter what",
+                    relevance="Maintains boundaries",
+                    timestamp=recent_time,
+                ),
+            ],
+            created_at=now,
+        )
+
+        evidence_span = obs.evidence_span
+        assert evidence_span["from"] == old_time.isoformat()
+        assert evidence_span["to"] == recent_time.isoformat()
+
+    def test_observation_empty_evidence_span(self):
+        """Observation with no evidence should have null span."""
+        obs = Observation(
+            title="Test observation",
+            content="Test observation without evidence",
+            evidence=[],
+        )
+
+        evidence_span = obs.evidence_span
+        assert evidence_span["from"] is None
+        assert evidence_span["to"] is None
+
+
+class TestVerifyEvidenceQuotes:
+    """Tests for evidence quote verification.
+
+    This ensures the LLM isn't hallucinating quotes - every quote
+    must actually appear in the source memory.
+    """
+
+    def test_valid_quotes(self):
+        """Should return True when quotes exist in their source memories."""
+        obs = Observation(
+            title="Enjoys hiking",
+            content="Enjoys hiking on weekends",
+            evidence=[
+                ObservationEvidence(
+                    memory_id="mem-hiking-trip",
+                    quote="went hiking at Mount Tam",
+                    relevance="Shows hiking activity",
+                    timestamp=datetime.now(timezone.utc),
+                ),
+            ],
+        )
+
+        memories = {
+            "mem-hiking-trip": "Had a great Saturday - went hiking at Mount Tam with friends and saw amazing views."
+        }
+        is_valid, errors = verify_evidence_quotes(obs, memories)
+
+        assert is_valid is True
+        assert len(errors) == 0
+
+    def test_invalid_quote(self):
+        """Should return False when quote doesn't exist in memory.
+
+        This catches LLM hallucinations where it fabricates quotes.
+        """
+        obs = Observation(
+            title="Loves spicy food",
+            content="Loves spicy food",
+            evidence=[
+                ObservationEvidence(
+                    memory_id="mem-dinner",
+                    quote="I love extra hot salsa",
+                    relevance="Shows spicy food preference",
+                    timestamp=datetime.now(timezone.utc),
+                ),
+            ],
+        )
+
+        memories = {"mem-dinner": "Had tacos for dinner. The guacamole was really fresh."}
+        is_valid, errors = verify_evidence_quotes(obs, memories)
+
+        assert is_valid is False
+        assert len(errors) == 1
+        assert "Quote not found" in errors[0]
+
+    def test_missing_memory(self):
+        """Should return False when referenced memory doesn't exist.
+
+        This catches cases where the LLM references a memory ID that
+        was never actually retrieved.
+        """
+        obs = Observation(
+            title="Has a dog named Max",
+            content="Has a dog named Max",
+            evidence=[
+                ObservationEvidence(
+                    memory_id="mem-pet-story",
+                    quote="took Max to the vet",
+                    relevance="Shows pet ownership",
+                    timestamp=datetime.now(timezone.utc),
+                ),
+            ],
+        )
+
+        memories = {"mem-different-id": "Some unrelated memory content"}
+        is_valid, errors = verify_evidence_quotes(obs, memories)
+
+        assert is_valid is False
+        assert len(errors) == 1
+        assert "not found" in errors[0]
+
+
+class TestCandidateObservation:
+    """Tests for candidate observation model.
+
+    Candidates are generated in the SEED phase and validated
+    before becoming full observations.
+    """
+
+    def test_create_candidate(self):
+        """Should create candidate with content and seed memories."""
+        candidate = CandidateObservation(
+            content="User prefers async communication over meetings",
+            seed_memory_ids=["mem-slack-pref", "mem-meeting-decline"],
+        )
+
+        assert candidate.content == "User prefers async communication over meetings"
+        assert len(candidate.seed_memory_ids) == 2
+        assert "mem-slack-pref" in candidate.seed_memory_ids
@@ -1,5 +1,9 @@
 """
 Test observation generation and entity state functionality.
+
+NOTE: Observations are now stored as summaries on the entities table,
+not as separate memory_units. The observations list in EntityState is
+populated from the summary for backwards compatibility.
 """
 import pytest
 from hindsight_api.engine.memory_engine import Budget
@@ -8,21 +12,16 @@ from datetime import datetime, timezone


@pytest.mark.asyncio
-async def test_observation_generation_on_put(memory, request_context):
+async def test_entity_extraction_on_retain(memory, request_context):
    """
-    Test that observations are generated SYNCHRONOUSLY when new facts are added.
+    Test that entities are extracted when new facts are added.

-    Observations are generated during retain when:
-    - Entity has >= 5 facts (MIN_FACTS_THRESHOLD)
-    - Entity is in top 5 by mention count
-
-    This test stores enough facts to trigger automatic observation generation.
+    This test stores multiple facts and verifies entities are extracted.
    """
-    bank_id = f"test_obs_{datetime.now(timezone.utc).timestamp()}"
+    bank_id = f"test_entity_extraction_{datetime.now(timezone.utc).timestamp()}"

    try:
-        # Store multiple facts about John to reach the MIN_FACTS_THRESHOLD (5)
-        # Each retain call should extract at least one fact about John
+        # Store multiple facts about John
        contents = [
            "John is a software engineer at Google.",
            "John is detail-oriented and methodical in his work.",
@@ -41,9 +40,8 @@ async def test_observation_generation_on_put(memory, request_context):
                request_context=request_context,
            )

-        # Observations are generated SYNCHRONOUSLY during retain,
-        # so they should be available immediately after retain completes.
-        # No need to wait for background tasks for observations.
+        # Wait for background tasks
+        await memory.wait_for_background_tasks()

        # Find the John entity
        pool = await memory._get_pool()
@@ -58,7 +56,7 @@ async def test_observation_generation_on_put(memory, request_context):
                bank_id
            )

-            # Also check the fact count for this entity
+            # Check the fact count for this entity
            if entity_row:
                fact_count = await conn.fetchval(
                    """
@@ -70,30 +68,9 @@ async def test_observation_generation_on_put(memory, request_context):
                print(f"Entity: {entity_row['canonical_name']} has {fact_count} linked facts")

        assert entity_row is not None, "John entity should have been extracted"
-
-        entity_id = str(entity_row['id'])
-        entity_name = entity_row['canonical_name']
        print(f"\n=== Found Entity ===")
-        print(f"Entity: {entity_name} (id: {entity_id})")
-
-        # Get observations for the entity - should be available immediately
-        observations = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
-
-        print(f"\n=== Observations for {entity_name} ===")
-        print(f"Total observations: {len(observations)}")
-        for obs in observations:
-            print(f"  - {obs.text}")
-
-        # Verify observations were created (requires >= 5 facts)
-        assert len(observations) > 0, \
-            f"Observations should have been generated synchronously during retain (entity has {fact_count} facts, threshold is 5)"
-
-        # Check that observations mention relevant content
-        obs_texts = " ".join([o.text.lower() for o in observations])
-        assert any(keyword in obs_texts for keyword in ["google", "engineer", "ai", "machine learning", "detail"]), \
-            "Observations should contain relevant information about John"
-
-        print(f"✓ Observations were successfully generated synchronously during retain")
+        print(f"Entity: {entity_row['canonical_name']} (id: {entity_row['id']})")
+        print(f"Entity was successfully extracted")

    finally:
        # Cleanup
@@ -106,7 +83,7 @@ async def test_observation_generation_on_put(memory, request_context):
@pytest.mark.asyncio
 async def test_regenerate_entity_observations(memory, request_context):
    """
-    Test explicit regeneration of observations for an entity.
+    Test explicit regeneration of summary for an entity.
    """
    bank_id = f"test_regen_obs_{datetime.now(timezone.utc).timestamp()}"

@@ -139,7 +116,7 @@ async def test_regenerate_entity_observations(memory, request_context):
            entity_id = str(entity_row['id'])
            entity_name = entity_row['canonical_name']

-            # Manually regenerate observations
+            # Manually regenerate summary (via observations API for backwards compat)
            created_ids = await memory.regenerate_entity_observations(
                bank_id=bank_id,
                entity_id=entity_id,
@@ -147,23 +124,25 @@ async def test_regenerate_entity_observations(memory, request_context):
                request_context=request_context,
            )

-            print(f"\n=== Regenerated Observations ===")
-            print(f"Created {len(created_ids)} observations for {entity_name}")
+            print(f"\n=== Regenerated Summary ===")
+            print(f"Created {len(created_ids)} summary for {entity_name}")

-            # Get the observations
-            observations = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
-            for obs in observations:
+            # Get entity state
+            state = await memory.get_entity_state(
+                bank_id, entity_id, entity_name, request_context=request_context
+            )
+            for obs in state.observations:
                print(f"  - {obs.text}")

-            # Verify observations were created
+            # Verify summary was created
            if len(created_ids) > 0:
-                assert len(observations) == len(created_ids), "Should have same number of observations as created IDs"
-                print(f"✓ Observations regenerated successfully")
+                assert len(state.observations) == 1, "Should have exactly 1 observation (the summary)"
+                print(f"Summary regenerated successfully")
            else:
-                print(f"⚠ Note: No observations were regenerated")
+                print(f"Note: No summary was regenerated")

        else:
-            print(f"⚠ Note: No 'Sarah' entity was extracted")
+            print(f"Note: No 'Sarah' entity was extracted")

    finally:
        # Cleanup
@@ -174,19 +153,14 @@ async def test_regenerate_entity_observations(memory, request_context):


@pytest.mark.asyncio
-async def test_manual_regenerate_with_few_facts(memory, request_context):
+async def test_entity_state_retrieval(memory, request_context):
    """
-    Test that manual regeneration works even with fewer than 5 facts.
-
-    This is important because:
-    - Automatic generation during retain requires MIN_FACTS_THRESHOLD (5)
-    - But manual regeneration via API should work with any number of facts
-    - The UI triggers manual regeneration, so it should work regardless of fact count
+    Test retrieving entity state with facts.
    """
-    bank_id = f"test_manual_regen_{datetime.now(timezone.utc).timestamp()}"
+    bank_id = f"test_entity_state_{datetime.now(timezone.utc).timestamp()}"

    try:
-        # Store only 2 facts - below the automatic threshold
+        # Store facts
        await memory.retain_async(
            bank_id=bank_id,
            content="Alice works at Google as a senior software engineer.",
@@ -220,51 +194,25 @@ async def test_manual_regenerate_with_few_facts(memory, request_context):
        entity_id = str(entity_row['id'])
        entity_name = entity_row['canonical_name']

-        # Check fact count - should be < 5
+        # Check fact count
        async with pool.acquire() as conn:
            fact_count = await conn.fetchval(
                "SELECT COUNT(*) FROM unit_entities WHERE entity_id = $1",
                entity_row['id']
            )

-        print(f"\n=== Manual Regeneration Test ===")
+        print(f"\n=== Entity State Test ===")
        print(f"Entity: {entity_name} (id: {entity_id})")
        print(f"Linked facts: {fact_count}")

-        # Verify we're testing with fewer than the automatic threshold
-        assert fact_count < 5, f"Test requires < 5 facts, but entity has {fact_count}"
-
-        # Before regeneration - should have no observations (auto threshold not met)
-        obs_before = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
-        print(f"Observations before manual regenerate: {len(obs_before)}")
-
-        # Manually regenerate observations - this should work regardless of fact count
-        created_ids = await memory.regenerate_entity_observations(
-            bank_id=bank_id,
-            entity_id=entity_id,
-            entity_name=entity_name,
-            request_context=request_context,
+        # Get entity state
+        state = await memory.get_entity_state(
+            bank_id, entity_id, entity_name, request_context=request_context
        )

-        print(f"Observations created by manual regenerate: {len(created_ids)}")
-
-        # Get observations after regeneration
-        observations = await memory.get_entity_observations(bank_id, entity_id, limit=10, request_context=request_context)
-        print(f"Observations after manual regenerate: {len(observations)}")
-        for obs in observations:
-            print(f"  - {obs.text}")
-
-        # Manual regeneration should create observations even with < 5 facts
-        assert len(observations) > 0, \
-            f"Manual regeneration should create observations even with only {fact_count} facts. " \
-            f"The LLM should synthesize at least 1 observation from the available facts."
-
-        # Verify observations contain relevant content
-        obs_texts = " ".join([o.text.lower() for o in observations])
-        assert any(keyword in obs_texts for keyword in ["google", "engineer", "hiking", "photography", "alice"]), \
-            "Observations should contain relevant information about Alice"
-
-        print(f"✓ Manual regeneration works with {fact_count} facts (below automatic threshold of 5)")
+        assert state.entity_id == entity_id
+        assert state.canonical_name == entity_name
+        print(f"Entity state retrieved successfully")

    finally:
        # Cleanup
@@ -277,16 +225,16 @@ async def test_manual_regenerate_with_few_facts(memory, request_context):
@pytest.mark.asyncio
 async def test_search_with_include_entities(memory, request_context):
    """
-    Test that search with include_entities=True returns entity observations.
+    Test that search with include_entities=True returns entity information.

    This test verifies that:
-    1. Observations are generated during retain (when entity has >= 5 facts)
-    2. Observations are returned in recall results with include_entities=True
+    1. Entities are extracted after retain
+    2. Entity info is returned in recall results with include_entities=True
    """
    bank_id = f"test_search_ent_{datetime.now(timezone.utc).timestamp()}"

    try:
-        # Store enough facts about Alice to trigger observation generation (>= 5 facts)
+        # Store facts about Alice
        contents = [
            "Alice is a data scientist who works on recommendation systems at Netflix.",
            "Alice presented her research at the ML conference last month.",
@@ -305,7 +253,8 @@ async def test_search_with_include_entities(memory, request_context):
                request_context=request_context,
            )

-        # Observations are generated synchronously during retain, no need to wait
+        # Wait for background tasks
+        await memory.wait_for_background_tasks()

        # Search with include_entities=True
        result = await memory.recall_async(
@@ -315,7 +264,7 @@ async def test_search_with_include_entities(memory, request_context):
            budget=Budget.LOW,
            max_tokens=2000,
            include_entities=True,
-            max_entity_tokens=500,
+            max_entity_tokens=5000,
            request_context=request_context,
        )

@@ -326,40 +275,28 @@ async def test_search_with_include_entities(memory, request_context):
            if fact.entities:
                print(f"    Entities: {', '.join(fact.entities)}")

-        print(f"\n=== Entity Observations in Recall ===")
-        if result.entities:
-            for name, state in result.entities.items():
-                print(f"\n{name}:")
-                for obs in state.observations:
-                    print(f"  - {obs.text}")
-        else:
-            print("No entity observations returned")
-
        # Verify results
        assert len(result.results) > 0, "Should find some facts"

        # Check if entities are included in facts
        facts_with_entities = [f for f in result.results if f.entities]
        assert len(facts_with_entities) > 0, "Some facts should have entity information"
-        print(f"✓ {len(facts_with_entities)} facts have entity information")
+        print(f"{len(facts_with_entities)} facts have entity information")

-        # Check if entity observations are included in recall
-        assert result.entities is not None and len(result.entities) > 0, \
-            "Entity observations should be included in recall results"
-        print(f"✓ Entity observations included for {len(result.entities)} entities")
+        # Check if entity info is returned
+        if result.entities:
+            print(f"Entity info included for {len(result.entities)} entities")

-        # Verify Alice entity has observations
-        alice_found = False
-        for name, state in result.entities.items():
-            assert state.canonical_name == name, "Entity canonical_name should match key"
-            assert state.entity_id, "Entity should have an ID"
-            if "alice" in name.lower():
-                alice_found = True
-                assert len(state.observations) > 0, \
-                    "Alice should have observations (generated during retain)"
-                print(f"✓ Alice has {len(state.observations)} observations in recall result")
+            # Verify Alice entity is in results
+            alice_found = False
+            for name, state in result.entities.items():
+                assert state.canonical_name == name, "Entity canonical_name should match key"
+                assert state.entity_id, "Entity should have an ID"
+                if "alice" in name.lower():
+                    alice_found = True
+                    print(f"Alice entity found: {name}")

-        assert alice_found, "Alice entity should be in recall results"
+            assert alice_found, "Alice entity should be in recall results"

    finally:
        # Cleanup
@@ -435,7 +372,10 @@ async def test_get_entity_state(memory, request_context):
@pytest.mark.asyncio
 async def test_observation_fact_type_in_database(memory, request_context):
    """
-    Test that observations are stored with correct fact_type in database.
+    Test that observations are NOT stored as memory_units with fact_type='observation'.
+
+    NOTE: Observations are now handled via mental models, not as memory_units
+    or entity summaries.
    """
    bank_id = f"test_obs_db_{datetime.now(timezone.utc).timestamp()}"

@@ -451,7 +391,7 @@ async def test_observation_fact_type_in_database(memory, request_context):

        await memory.wait_for_background_tasks()

-        # Check that observations have correct fact_type
+        # Check that NO observations exist in memory_units
        pool = await memory._get_pool()
        async with pool.acquire() as conn:
            observations = await conn.fetch(
@@ -463,17 +403,11 @@ async def test_observation_fact_type_in_database(memory, request_context):
                bank_id
            )

-        print(f"\n=== Observation Records in Database ===")
-        print(f"Found {len(observations)} observation records")
-        for obs in observations:
-            print(f"  - fact_type: {obs['fact_type']}")
-            print(f"    text: {obs['text']}")
-            print(f"    context: {obs['context']}")
+        print(f"\n=== Observation Records in memory_units ===")
+        print(f"Found {len(observations)} observation records (should be 0)")

-        if len(observations) > 0:
-            for obs in observations:
-                assert obs['fact_type'] == 'observation', "All observation records should have fact_type='observation'"
-            print(f"✓ All observations have correct fact_type")
+        # Observations are no longer stored as memory_units
+        assert len(observations) == 0, "Observations should NOT be stored as memory_units"

    finally:
        # Cleanup
@@ -484,23 +418,183 @@ async def test_observation_fact_type_in_database(memory, request_context):


@pytest.mark.asyncio
-async def test_user_entity_prioritized_for_observations(memory, request_context):
+async def test_entity_mention_counts(memory, request_context):
    """
-    Test that the 'user' entity gets observations even when many other entities exist.
+    Test that entity mention counts are tracked correctly.

-    The retain pipeline only regenerates observations for TOP_N_ENTITIES (5) entities,
-    sorted by mention count. This test verifies that the most mentioned entity ('user')
-    gets prioritized and receives observations.
-
-    This is critical because 'user' is often the most important entity in personal memory.
+    This test creates entities with varying mention counts and verifies
+    that the counts are accurate.
    """
-    bank_id = f"test_user_priority_{datetime.now(timezone.utc).timestamp()}"
+    bank_id = f"test_mention_counts_{datetime.now(timezone.utc).timestamp()}"

    try:
-        # Create content where 'user' (the user) is mentioned many times
-        # along with several other entities
+        # Create content with varying entity mention counts:
+        # - "HighMention Corp" mentioned 10+ times
+        # - "LowMention Ltd" mentioned 1 time
+        contents = [
+            # High mentions - HighMention Corp
+            "HighMention Corp is a tech company based in San Francisco.",
+            "HighMention Corp was founded in 2010 by experienced entrepreneurs.",
+            "HighMention Corp has over 500 employees worldwide.",
+            "HighMention Corp specializes in cloud computing solutions.",
+            "HighMention Corp recently raised $50 million in Series C funding.",
+            "HighMention Corp has partnerships with major tech companies.",
+            "HighMention Corp is known for its innovative culture.",
+            "HighMention Corp offers competitive salaries and benefits.",
+            "HighMention Corp has offices in 5 countries.",
+            "HighMention Corp won the best workplace award last year.",
+            # Low mentions - LowMention Ltd
+            "LowMention Ltd is a small consulting firm.",
+        ]
+
+        for i, content in enumerate(contents):
+            await memory.retain_async(
+                bank_id=bank_id,
+                content=content,
+                context="company info",
+                event_date=datetime(2024, 1, 15 + i, tzinfo=timezone.utc),
+                request_context=request_context,
+            )
+
+        # Wait for background tasks
+        await memory.wait_for_background_tasks()
+
+        # Check entity mention counts
+        pool = await memory._get_pool()
+        async with pool.acquire() as conn:
+            entities = await conn.fetch(
+                """
+                SELECT e.id, e.canonical_name, e.mention_count
+                FROM entities e
+                WHERE e.bank_id = $1
+                ORDER BY e.mention_count DESC
+                """,
+                bank_id
+            )
+
+        print(f"\n=== Entity Mention Counts Test ===")
+        print(f"Total entities: {len(entities)}")
+
+        high_mention_entity = None
+        low_mention_entity = None
+
+        for entity in entities:
+            name = entity['canonical_name'].lower()
+            mention_count = entity['mention_count']
+
+            print(f"  {entity['canonical_name']}: mentions={mention_count}")
+
+            if "highmention" in name:
+                high_mention_entity = entity
+            elif "lowmention" in name:
+                low_mention_entity = entity
+
+        # Verify HighMention Corp has higher mention count
+        if high_mention_entity and low_mention_entity:
+            assert high_mention_entity['mention_count'] > low_mention_entity['mention_count'], \
+                "HighMention Corp should have more mentions than LowMention Ltd"
+            print("PASS: Entity mention counts are tracked correctly")
+
+    finally:
+        # Cleanup
+        pool = await memory._get_pool()
+        async with pool.acquire() as conn:
+            await conn.execute("DELETE FROM memory_units WHERE bank_id = $1", bank_id)
+            await conn.execute("DELETE FROM entities WHERE bank_id = $1", bank_id)
+
+
+@pytest.mark.asyncio
+async def test_entity_mention_ranking(memory, request_context):
+    """
+    Test that entity mention counts correctly rank entities.
+
+    This test:
+    1. Creates an entity with 6 mentions
+    2. Adds more entities with higher mention counts
+    3. Verifies entities are ranked correctly by mention count
+    """
+    bank_id = f"test_ranking_{datetime.now(timezone.utc).timestamp()}"
+
+    try:
+        # Phase 1: Create "OriginalEntity" with 6 mentions
+        print("\n=== Phase 1: Create OriginalEntity with 6 mentions ===")
+        for i in range(6):
+            await memory.retain_async(
+                bank_id=bank_id,
+                content=f"OriginalEntity is mentioned here in fact {i+1}.",
+                context="test",
+                event_date=datetime(2024, 1, 1 + i, tzinfo=timezone.utc),
+                request_context=request_context,
+            )
+
+        await memory.wait_for_background_tasks()
+
+        # Phase 2: Add more entities with MORE mentions
+        print("\n=== Phase 2: Add entities with 10+ mentions each ===")
+        for entity_num in range(3):  # Reduced from 10 to 3 to speed up test
+            entity_name = f"NewEntity{entity_num}"
+            for mention in range(10):
+                await memory.retain_async(
+                    bank_id=bank_id,
+                    content=f"{entity_name} is a very important entity, mention {mention+1}.",
+                    context="test",
+                    event_date=datetime(2024, 2, 1 + mention, tzinfo=timezone.utc),
+                    request_context=request_context,
+                )
+
+        await memory.wait_for_background_tasks()
+
+        # Phase 3: Verify entities are ranked by mention count
+        print("\n=== Phase 3: Check entity ranking ===")
+        pool = await memory._get_pool()
+        async with pool.acquire() as conn:
+            all_entities = await conn.fetch(
+                """
+                SELECT canonical_name, mention_count
+                FROM entities
+                WHERE bank_id = $1
+                ORDER BY mention_count DESC
+                """,
+                bank_id
+            )
+
+        print(f"\nAll entities by mention count:")
+        for e in all_entities:
+            print(f"  {e['canonical_name']}: mentions={e['mention_count']}")
+
+        # Verify new entities have higher counts than OriginalEntity
+        original = next((e for e in all_entities if 'originalentity' in e['canonical_name'].lower()), None)
+        new_entities = [e for e in all_entities if 'newentity' in e['canonical_name'].lower()]
+
+        assert original is not None, "OriginalEntity should exist"
+        assert len(new_entities) > 0, "NewEntity entities should exist"
+
+        # Verify entities are created and have mention counts
+        # Note: LLM may merge mentions, so we just check that new entities exist
+        print(f"OriginalEntity mentions: {original['mention_count']}")
+        for new_entity in new_entities:
+            print(f"{new_entity['canonical_name']} mentions: {new_entity['mention_count']}")
+
+        print("PASS: Entities are created with mention counts tracked")
+
+    finally:
+        # Cleanup
+        pool = await memory._get_pool()
+        async with pool.acquire() as conn:
+            await conn.execute("DELETE FROM memory_units WHERE bank_id = $1", bank_id)
+            await conn.execute("DELETE FROM entities WHERE bank_id = $1", bank_id)
+
+
+@pytest.mark.asyncio
+async def test_user_entity_extraction(memory, request_context):
+    """
+    Test that the 'user' entity is correctly extracted when mentioned frequently.
+    """
+    bank_id = f"test_user_entity_{datetime.now(timezone.utc).timestamp()}"
+
+    try:
+        # Create content where 'user' is mentioned many times
        contents = [
-            # User mentioned frequently
            "The user loves hiking in the mountains during summer.",
            "The user works as a software engineer at Microsoft.",
            "The user has a dog named Max who is a golden retriever.",
@@ -510,11 +604,8 @@ async def test_user_entity_prioritized_for_observations(memory, request_context)
            # Other entities mentioned fewer times
            "Sarah is a friend who works at Google.",
            "Bob is a colleague from the data science team.",
-            "Tokyo is a city the user visited last year.",
-            "Python is the user's favorite programming language.",
        ]

-        # Retain all content in a single batch for efficiency
        for i, content in enumerate(contents):
            await memory.retain_async(
                bank_id=bank_id,
@@ -524,12 +615,12 @@ async def test_user_entity_prioritized_for_observations(memory, request_context)
                request_context=request_context,
            )

-        # Observations are generated synchronously during retain
+        # Wait for background tasks
+        await memory.wait_for_background_tasks()

        # Find the 'user' entity
        pool = await memory._get_pool()
        async with pool.acquire() as conn:
-            # Find user entity (may be named "user", "the user", etc.)
            user_entity = await conn.fetchrow(
                """
                SELECT e.id, e.canonical_name,
@@ -544,7 +635,7 @@ async def test_user_entity_prioritized_for_observations(memory, request_context)
                bank_id
            )

-            # Get all entities with their fact counts to verify prioritization
+            # Get all entities with their fact counts
            all_entities = await conn.fetch(
                """
                SELECT e.id, e.canonical_name,
@@ -564,41 +655,10 @@ async def test_user_entity_prioritized_for_observations(memory, request_context)

        # Verify user entity exists
        assert user_entity is not None, "User entity should have been extracted"
-        user_entity_id = str(user_entity['id'])
-        user_entity_name = user_entity['canonical_name']
-        user_fact_count = user_entity['fact_count']
-
        print(f"\n=== User Entity ===")
-        print(f"Entity: {user_entity_name} (id: {user_entity_id})")
-        print(f"Fact count: {user_fact_count}")
-
-        # Verify user has enough facts for observations (>= MIN_FACTS_THRESHOLD of 5)
-        assert user_fact_count >= 5, \
-            f"User entity should have at least 5 facts, but has {user_fact_count}"
-
-        # Get observations for user entity
-        observations = await memory.get_entity_observations(bank_id, user_entity_id, limit=10, request_context=request_context)
-
-        print(f"\n=== User Entity Observations ===")
-        print(f"Total observations: {len(observations)}")
-        for obs in observations:
-            print(f"  - {obs.text}")
-
-        # Verify observations were generated for user (critical assertion)
-        assert len(observations) > 0, \
-            f"User entity should have observations (has {user_fact_count} facts, threshold is 5). " \
-            f"This may indicate that 'user' is not being prioritized in the top 5 entities by mention count."
-
-        # Verify observations mention relevant content about the user
-        obs_texts = " ".join([o.text.lower() for o in observations])
-        user_keywords = ["hiking", "software", "engineer", "dog", "max", "cooking",
-                        "italian", "mit", "dune", "microsoft"]
-        matching_keywords = [k for k in user_keywords if k in obs_texts]
-        assert len(matching_keywords) > 0, \
-            f"Observations should contain relevant information about the user. Keywords found: {matching_keywords}"
-
-        print(f"✓ User entity was prioritized and received {len(observations)} observations")
-        print(f"✓ Observations contain relevant keywords: {matching_keywords}")
+        print(f"Entity: {user_entity['canonical_name']} (id: {user_entity['id']})")
+        print(f"Fact count: {user_entity['fact_count']}")
+        print(f"User entity was successfully extracted")

    finally:
        # Cleanup
@@ -465,7 +465,7 @@ async def test_occurred_dates_not_defaulted(memory, request_context):
            query="Tell me about Alice",
            budget=Budget.LOW,
            max_tokens=500,
-            fact_type=["world", "opinion"],
+            fact_type=["world", "experience"],
            request_context=request_context,
        )

@@ -2058,3 +2058,26 @@ async def test_user_provided_entities(memory, request_context):

    finally:
        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+def test_recall_result_model_empty_construction():
+    """
+    Test that RecallResultModel can be constructed with empty results.
+
+    This is a regression test for the bug where constructing an empty RecallResultModel
+    would cause an UnboundLocalError because RecallResult was imported as RecallResultModel
+    but the code mistakenly used the wrong name.
+
+    The fix ensures RecallResultModel is used consistently throughout memory_engine.py.
+    """
+    from hindsight_api.engine.response_models import RecallResult
+
+    # This should not raise any errors
+    result = RecallResult(results=[], entities={}, chunks={})
+
+    assert result is not None, "Should create a result object"
+    assert result.results == [], "Should have empty results"
+    assert result.entities == {}, "Should have empty entities"
+    assert result.chunks == {}, "Should have empty chunks"
+
+    logger.info("✓ RecallResult empty construction works correctly")
@@ -0,0 +1,294 @@
+"""
+Tests for hindsight_api.server module (multi-worker code path).
+
+The server.py module is used when running with multiple workers:
+    uvicorn hindsight_api.server:app --workers 2
+
+This module executes code at import time, creating the app at module level.
+These tests ensure that extensions are properly loaded in this code path,
+which was previously a regression that caused authentication bypass in production.
+"""
+
+import importlib
+import sys
+from unittest.mock import MagicMock, patch
+
+
+def _clean_server_module():
+    """Remove hindsight_api.server from sys.modules for fresh import."""
+    modules_to_remove = [k for k in sys.modules.keys() if k.startswith("hindsight_api.server")]
+    for mod in modules_to_remove:
+        del sys.modules[mod]
+
+
+class TestServerModuleExtensionLoading:
+    """Tests that server.py correctly loads extensions when configured via environment."""
+
+    def test_server_loads_tenant_extension_when_configured(self, monkeypatch):
+        """
+        Verify that server.py loads tenant extension from HINDSIGHT_API_TENANT_EXTENSION.
+
+        This test catches the regression where server.py didn't call load_extension(),
+        causing authentication to be bypassed in multi-worker deployments.
+        """
+        # Set up environment to configure a tenant extension
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TENANT_EXTENSION",
+            "tests.test_server_module:MockTenantExtension",
+        )
+
+        _clean_server_module()
+
+        # Track what extensions were loaded via load_extension
+        loaded_extensions = {}
+
+        # Get the real load_extension function
+        from hindsight_api.extensions.loader import load_extension as real_load_extension
+
+        def tracking_load_extension(name, base_class):
+            """Track calls to load_extension and delegate to original."""
+            result = real_load_extension(name, base_class)
+            loaded_extensions[name] = result
+            return result
+
+        # Patch at source level BEFORE importing server
+        # Note: We patch the entire hindsight_api module namespace
+        with patch("hindsight_api.MemoryEngine") as mock_engine, \
+             patch("hindsight_api.api.create_app") as mock_create_app, \
+             patch("hindsight_api.config.get_config") as mock_get_config, \
+             patch("hindsight_api.extensions.load_extension", side_effect=tracking_load_extension), \
+             patch("hindsight_api.extensions.DefaultExtensionContext"):
+
+            mock_config = MagicMock()
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_engine.return_value = MagicMock()
+            mock_create_app.return_value = MagicMock()
+
+            # Now import server - this triggers module-level code
+            import hindsight_api.server
+
+        # Verify TENANT extension was loaded
+        assert "TENANT" in loaded_extensions, \
+            "server.py did not call load_extension('TENANT', ...) - extensions not loaded!"
+        assert loaded_extensions["TENANT"] is not None, \
+            "load_extension('TENANT', ...) returned None despite env var being set"
+        assert isinstance(loaded_extensions["TENANT"], MockTenantExtension), \
+            f"Expected MockTenantExtension, got {type(loaded_extensions['TENANT'])}"
+
+    def test_server_loads_operation_validator_when_configured(self, monkeypatch):
+        """
+        Verify that server.py loads operation validator from HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION.
+        """
+        monkeypatch.setenv(
+            "HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION",
+            "tests.test_server_module:MockOperationValidator",
+        )
+
+        _clean_server_module()
+
+        loaded_extensions = {}
+
+        from hindsight_api.extensions.loader import load_extension as real_load_extension
+
+        def tracking_load_extension(name, base_class):
+            result = real_load_extension(name, base_class)
+            loaded_extensions[name] = result
+            return result
+
+        with patch("hindsight_api.MemoryEngine") as mock_engine, \
+             patch("hindsight_api.api.create_app") as mock_create_app, \
+             patch("hindsight_api.config.get_config") as mock_get_config, \
+             patch("hindsight_api.extensions.load_extension", side_effect=tracking_load_extension), \
+             patch("hindsight_api.extensions.DefaultExtensionContext"):
+
+            mock_config = MagicMock()
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_engine.return_value = MagicMock()
+            mock_create_app.return_value = MagicMock()
+
+            import hindsight_api.server
+
+        assert "OPERATION_VALIDATOR" in loaded_extensions, \
+            "server.py did not call load_extension('OPERATION_VALIDATOR', ...)"
+        assert loaded_extensions["OPERATION_VALIDATOR"] is not None
+        assert isinstance(loaded_extensions["OPERATION_VALIDATOR"], MockOperationValidator)
+
+    def test_server_passes_extensions_to_memory_engine(self, monkeypatch):
+        """
+        Verify that server.py passes loaded extensions to MemoryEngine constructor.
+
+        This is the critical test - even if extensions are loaded, they must be
+        passed to MemoryEngine for authentication to work.
+        """
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TENANT_EXTENSION",
+            "tests.test_server_module:MockTenantExtension",
+        )
+
+        _clean_server_module()
+
+        memory_engine_calls = []
+
+        def capture_memory_engine(*args, **kwargs):
+            memory_engine_calls.append({"args": args, "kwargs": kwargs})
+            return MagicMock()
+
+        with patch("hindsight_api.MemoryEngine", side_effect=capture_memory_engine), \
+             patch("hindsight_api.api.create_app") as mock_create_app, \
+             patch("hindsight_api.config.get_config") as mock_get_config, \
+             patch("hindsight_api.extensions.DefaultExtensionContext"):
+
+            mock_config = MagicMock()
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_create_app.return_value = MagicMock()
+
+            import hindsight_api.server
+
+        # Verify MemoryEngine was called
+        assert len(memory_engine_calls) == 1, "MemoryEngine should be called exactly once"
+
+        call_kwargs = memory_engine_calls[0]["kwargs"]
+
+        # THE CRITICAL ASSERTION: tenant_extension must be passed and not None
+        assert "tenant_extension" in call_kwargs, \
+            "MemoryEngine was not called with tenant_extension parameter!"
+        assert call_kwargs["tenant_extension"] is not None, \
+            "tenant_extension was None - server.py did not pass loaded extension to MemoryEngine!"
+
+    def test_server_sets_extension_context_on_tenant_extension(self, monkeypatch):
+        """
+        Verify that server.py sets the extension context on tenant extension.
+
+        This is required for tenant extensions that need to provision schemas.
+        """
+        monkeypatch.setenv(
+            "HINDSIGHT_API_TENANT_EXTENSION",
+            "tests.test_server_module:MockTenantExtension",
+        )
+
+        _clean_server_module()
+
+        context_set_calls = []
+        captured_tenant_ext = [None]
+
+        def capture_memory_engine(*args, **kwargs):
+            captured_tenant_ext[0] = kwargs.get("tenant_extension")
+            return MagicMock()
+
+        def capture_context(*args, **kwargs):
+            ctx = MagicMock()
+            context_set_calls.append(ctx)
+            return ctx
+
+        with patch("hindsight_api.MemoryEngine", side_effect=capture_memory_engine), \
+             patch("hindsight_api.api.create_app") as mock_create_app, \
+             patch("hindsight_api.config.get_config") as mock_get_config, \
+             patch("hindsight_api.extensions.DefaultExtensionContext", side_effect=capture_context):
+
+            mock_config = MagicMock()
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_create_app.return_value = MagicMock()
+
+            import hindsight_api.server
+
+        # Verify context was created and set
+        assert len(context_set_calls) == 1, "DefaultExtensionContext should be created"
+        assert captured_tenant_ext[0] is not None, "Tenant extension should be captured"
+        assert captured_tenant_ext[0]._context_set, \
+            "set_context was not called on tenant extension"
+
+    def test_server_works_without_extensions(self, monkeypatch):
+        """
+        Verify that server.py works correctly when no extensions are configured.
+        """
+        # Ensure no extension env vars are set
+        monkeypatch.delenv("HINDSIGHT_API_TENANT_EXTENSION", raising=False)
+        monkeypatch.delenv("HINDSIGHT_API_OPERATION_VALIDATOR_EXTENSION", raising=False)
+
+        _clean_server_module()
+
+        memory_engine_calls = []
+
+        def capture_memory_engine(*args, **kwargs):
+            memory_engine_calls.append({"args": args, "kwargs": kwargs})
+            return MagicMock()
+
+        with patch("hindsight_api.MemoryEngine", side_effect=capture_memory_engine), \
+             patch("hindsight_api.api.create_app") as mock_create_app, \
+             patch("hindsight_api.config.get_config") as mock_get_config:
+
+            mock_config = MagicMock()
+            mock_config.mcp_enabled = False
+            mock_config.run_migrations_on_startup = False
+            mock_config.database_url = "postgresql://test:test@localhost/test"
+            mock_get_config.return_value = mock_config
+            mock_create_app.return_value = MagicMock()
+
+            import hindsight_api.server
+
+        # Should work without extensions
+        assert len(memory_engine_calls) == 1
+        call_kwargs = memory_engine_calls[0]["kwargs"]
+
+        # Extensions should be None when not configured
+        assert call_kwargs.get("tenant_extension") is None
+        assert call_kwargs.get("operation_validator") is None
+
+
+# Mock extensions for testing
+from hindsight_api.extensions import (
+    TenantExtension,
+    TenantContext,
+    RequestContext,
+    OperationValidatorExtension,
+    ValidationResult,
+    RetainContext,
+    RecallContext,
+    ReflectContext,
+    RefreshMentalModelContext,
+)
+
+
+class MockTenantExtension(TenantExtension):
+    """Mock tenant extension for testing server.py extension loading."""
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+        self._context_set = False
+
+    async def authenticate(self, request_context: RequestContext) -> TenantContext:
+        return TenantContext(schema_name="public")
+
+    def set_context(self, context) -> None:
+        self._context_set = True
+
+
+class MockOperationValidator(OperationValidatorExtension):
+    """Mock operation validator for testing server.py extension loading."""
+
+    def __init__(self, config: dict):
+        super().__init__(config)
+
+    async def validate_retain(self, ctx: RetainContext) -> ValidationResult:
+        return ValidationResult.accept()
+
+    async def validate_recall(self, ctx: RecallContext) -> ValidationResult:
+        return ValidationResult.accept()
+
+    async def validate_reflect(self, ctx: ReflectContext) -> ValidationResult:
+        return ValidationResult.accept()
+
+    async def validate_refresh_mental_model(self, ctx: RefreshMentalModelContext) -> ValidationResult:
+        return ValidationResult.accept()
@@ -0,0 +1,886 @@
+"""
+Tests for tags-based visibility scoping.
+
+This module tests the tags feature which allows filtering memories by visibility tags.
+Use cases:
+- Multi-user agent: Agent has a single memory bank, users should only see memories from
+  conversations they participated in
+- Student tracking: Teacher tracks students, students should only see their own data
+
+The tags use OR-based matching: a memory matches if ANY of its tags overlap with the request tags.
+"""
+from datetime import datetime
+
+import httpx
+import pytest
+import pytest_asyncio
+
+from hindsight_api.api import create_app
+from hindsight_api.engine.search.tags import build_tags_where_clause_simple, filter_results_by_tags
+
+# ============================================================================
+# Unit Tests for tags SQL builder
+# ============================================================================
+
+
+class TestTagsWhereClauseBuilder:
+    """Unit tests for the tags WHERE clause SQL builder."""
+
+    def test_no_tags_returns_empty_string(self):
+        """When tags is None, should return empty string (no filtering)."""
+        result = build_tags_where_clause_simple(None, 5)
+        assert result == ""
+
+    def test_empty_tags_list_returns_empty_string(self):
+        """When tags is an empty list, should return empty string (no filtering)."""
+        result = build_tags_where_clause_simple([], 5)
+        assert result == ""
+
+    def test_tags_with_different_param_num(self):
+        """Should use the provided parameter number."""
+        result = build_tags_where_clause_simple(["user_a", "user_b"], 3)
+        # Default is "any" which includes untagged
+        assert "$3" in result
+
+    def test_tags_with_table_alias(self):
+        """Should include table alias when provided."""
+        result = build_tags_where_clause_simple(["user_a"], 5, table_alias="mu.")
+        assert "mu.tags" in result
+
+    # ---- Test "any" mode (OR, includes untagged - default) ----
+
+    def test_tags_match_any_includes_untagged(self):
+        """When match='any', should include untagged memories (NULL or empty)."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="any")
+        # Should use OR with NULL/empty check
+        assert "IS NULL" in result
+        assert "= '{}'" in result
+        assert "&&" in result  # overlap operator
+
+    def test_tags_match_any_uses_overlap(self):
+        """When match='any', should use overlap operator (&&)."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="any")
+        assert "&&" in result
+
+    # ---- Test "all" mode (AND, includes untagged) ----
+
+    def test_tags_match_all_includes_untagged(self):
+        """When match='all', should include untagged memories (NULL or empty)."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="all")
+        # Should use OR with NULL/empty check
+        assert "IS NULL" in result
+        assert "= '{}'" in result
+        assert "@>" in result  # contains operator
+
+    def test_tags_match_all_uses_contains(self):
+        """When match='all', should use contains operator (@>)."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="all")
+        assert "@>" in result
+
+    # ---- Test "any_strict" mode (OR, excludes untagged) ----
+
+    def test_tags_match_any_strict_excludes_untagged(self):
+        """When match='any_strict', should exclude untagged memories."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="any_strict")
+        # Should require tags to be NOT NULL and not empty
+        assert "IS NOT NULL" in result
+        assert "!= '{}'" in result
+        assert "&&" in result  # overlap operator
+
+    def test_tags_match_any_strict_uses_overlap(self):
+        """When match='any_strict', should use overlap operator (&&)."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="any_strict")
+        assert "&&" in result
+        # Should NOT include untagged
+        assert "IS NULL" not in result or "IS NOT NULL" in result
+
+    # ---- Test "all_strict" mode (AND, excludes untagged) ----
+
+    def test_tags_match_all_strict_excludes_untagged(self):
+        """When match='all_strict', should exclude untagged memories."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="all_strict")
+        # Should require tags to be NOT NULL and not empty
+        assert "IS NOT NULL" in result
+        assert "!= '{}'" in result
+        assert "@>" in result  # contains operator
+
+    def test_tags_match_all_strict_uses_contains(self):
+        """When match='all_strict', should use contains operator (@>)."""
+        result = build_tags_where_clause_simple(["user_a"], 5, match="all_strict")
+        assert "@>" in result
+
+    # ---- Test table alias with all modes ----
+
+    def test_tags_match_any_with_table_alias(self):
+        """Should include table alias with any mode."""
+        result = build_tags_where_clause_simple(["user_a"], 3, table_alias="mu.", match="any")
+        assert "mu.tags" in result
+
+    def test_tags_match_all_strict_with_table_alias(self):
+        """Should include table alias with all_strict mode."""
+        result = build_tags_where_clause_simple(["user_a", "user_b"], 3, table_alias="mu.", match="all_strict")
+        assert "mu.tags" in result
+        assert "@>" in result
+        assert "IS NOT NULL" in result
+
+
+# ============================================================================
+# Unit Tests for filter_results_by_tags (Python-side filtering)
+# ============================================================================
+
+
+class MockResult:
+    """Mock result object for testing filter_results_by_tags."""
+
+    def __init__(self, tags):
+        self.tags = tags
+
+
+class TestFilterResultsByTags:
+    """Unit tests for the Python-side tags filter function."""
+
+    def test_no_tags_returns_all(self):
+        """When tags is None, should return all results."""
+        results = [MockResult(["a"]), MockResult(["b"]), MockResult(None)]
+        filtered = filter_results_by_tags(results, None)
+        assert len(filtered) == 3
+
+    def test_empty_tags_returns_all(self):
+        """When tags is empty list, should return all results."""
+        results = [MockResult(["a"]), MockResult(["b"]), MockResult(None)]
+        filtered = filter_results_by_tags(results, [])
+        assert len(filtered) == 3
+
+    # ---- Test "any" mode (OR, includes untagged) ----
+
+    def test_any_mode_includes_matching_tags(self):
+        """'any' mode should include results with matching tags."""
+        results = [MockResult(["a"]), MockResult(["b"]), MockResult(["c"])]
+        filtered = filter_results_by_tags(results, ["a", "b"], match="any")
+        # "a" and "b" match, "c" doesn't match and isn't untagged, so excluded
+        assert len(filtered) == 2
+        tags_found = [r.tags[0] for r in filtered if r.tags]
+        assert "a" in tags_found
+        assert "b" in tags_found
+        assert "c" not in tags_found
+
+    def test_any_mode_includes_untagged(self):
+        """'any' mode should include untagged results."""
+        results = [MockResult(["a"]), MockResult(None), MockResult([])]
+        filtered = filter_results_by_tags(results, ["a"], match="any")
+        assert len(filtered) == 3  # a matches, None is untagged, [] is untagged
+
+    def test_any_mode_includes_partial_overlap(self):
+        """'any' mode should include results with ANY overlapping tag."""
+        results = [MockResult(["a", "x"]), MockResult(["b", "y"])]
+        filtered = filter_results_by_tags(results, ["a"], match="any")
+        # ["a", "x"] matches, ["b", "y"] doesn't, but untagged would be included
+        tags_found = [r.tags for r in filtered]
+        assert ["a", "x"] in tags_found
+
+    # ---- Test "any_strict" mode (OR, excludes untagged) ----
+
+    def test_any_strict_excludes_untagged(self):
+        """'any_strict' mode should exclude untagged results."""
+        results = [MockResult(["a"]), MockResult(None), MockResult([])]
+        filtered = filter_results_by_tags(results, ["a"], match="any_strict")
+        assert len(filtered) == 1  # Only ["a"] matches
+        assert filtered[0].tags == ["a"]
+
+    def test_any_strict_excludes_non_matching(self):
+        """'any_strict' mode should exclude non-matching tagged results."""
+        results = [MockResult(["a"]), MockResult(["b"]), MockResult(["c"])]
+        filtered = filter_results_by_tags(results, ["a"], match="any_strict")
+        assert len(filtered) == 1
+        assert filtered[0].tags == ["a"]
+
+    # ---- Test "all" mode (AND, includes untagged) ----
+
+    def test_all_mode_requires_all_tags(self):
+        """'all' mode should require ALL requested tags to be present."""
+        results = [MockResult(["a", "b"]), MockResult(["a"]), MockResult(["b"])]
+        filtered = filter_results_by_tags(results, ["a", "b"], match="all")
+        # Only ["a", "b"] has both tags, but untagged would also be included
+        tags_found = [r.tags for r in filtered]
+        assert ["a", "b"] in tags_found
+
+    def test_all_mode_includes_untagged(self):
+        """'all' mode should include untagged results."""
+        results = [MockResult(["a", "b"]), MockResult(None), MockResult([])]
+        filtered = filter_results_by_tags(results, ["a", "b"], match="all")
+        assert len(filtered) == 3  # ["a", "b"] matches, None is untagged, [] is untagged
+
+    # ---- Test "all_strict" mode (AND, excludes untagged) ----
+
+    def test_all_strict_requires_all_tags(self):
+        """'all_strict' mode should require ALL requested tags."""
+        results = [MockResult(["a", "b"]), MockResult(["a"]), MockResult(["b"])]
+        filtered = filter_results_by_tags(results, ["a", "b"], match="all_strict")
+        assert len(filtered) == 1
+        assert filtered[0].tags == ["a", "b"]
+
+    def test_all_strict_excludes_untagged(self):
+        """'all_strict' mode should exclude untagged results."""
+        results = [MockResult(["a", "b"]), MockResult(None), MockResult([])]
+        filtered = filter_results_by_tags(results, ["a", "b"], match="all_strict")
+        assert len(filtered) == 1
+        assert filtered[0].tags == ["a", "b"]
+
+    def test_all_strict_allows_superset(self):
+        """'all_strict' mode should allow results with MORE tags than requested."""
+        results = [MockResult(["a", "b", "c"]), MockResult(["a"])]
+        filtered = filter_results_by_tags(results, ["a", "b"], match="all_strict")
+        assert len(filtered) == 1
+        assert filtered[0].tags == ["a", "b", "c"]  # Has a, b, AND c
+
+
+# ============================================================================
+# Integration Tests for tags in retain/recall/reflect
+# ============================================================================
+
+
+@pytest_asyncio.fixture
+async def api_client(memory):
+    """Create an async test client for the FastAPI app."""
+    app = create_app(memory, initialize_memory=False)
+    transport = httpx.ASGITransport(app=app)
+    async with httpx.AsyncClient(transport=transport, base_url="http://test") as client:
+        yield client
+
+
+@pytest.fixture
+def test_bank_id():
+    """Provide a unique bank ID for this test run."""
+    return f"tags_test_{datetime.now().timestamp()}"
+
+
+@pytest.mark.asyncio
+async def test_retain_with_tags(api_client, test_bank_id):
+    """Test that memories can be stored with tags."""
+    # Store memory with tags
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "Alice loves hiking in the mountains.",
+                    "tags": ["user_alice"]
+                }
+            ]
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+    assert result["success"] is True
+    assert result["items_count"] == 1
+
+
+@pytest.mark.asyncio
+async def test_retain_with_document_tags(api_client, test_bank_id):
+    """Test that document-level tags are applied to all items."""
+    # Store memories with document-level tags
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "document_tags": ["session_123"],
+            "items": [
+                {"content": "Bob discussed the quarterly report."},
+                {"content": "Charlie mentioned the new product launch."}
+            ]
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+    assert result["success"] is True
+    assert result["items_count"] == 2
+
+
+@pytest.mark.asyncio
+async def test_retain_merges_document_and_item_tags(api_client, test_bank_id):
+    """Test that document tags and item tags are merged."""
+    # Store memory with both document and item tags
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "document_tags": ["session_abc"],
+            "items": [
+                {
+                    "content": "Dave talked about machine learning.",
+                    "tags": ["user_dave"]
+                }
+            ]
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+    assert result["success"] is True
+
+
+@pytest.mark.asyncio
+async def test_recall_without_tags_returns_all_memories(api_client, test_bank_id):
+    """Test that recall without tags returns all memories (no filtering)."""
+    # Store memories for different users
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Eve works on natural language processing.", "tags": ["user_eve"]},
+                {"content": "Frank specializes in computer vision.", "tags": ["user_frank"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Recall without tags - should return all
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories/recall",
+        json={"query": "Who works on what?", "budget": "low"}
+    )
+    assert response.status_code == 200
+    results = response.json()["results"]
+
+    # Should find both Eve and Frank
+    texts = [r["text"] for r in results]
+    assert any("Eve" in t for t in texts), "Should find Eve"
+    assert any("Frank" in t for t in texts), "Should find Frank"
+
+
+@pytest.mark.asyncio
+async def test_recall_with_tags_filters_memories(api_client, test_bank_id):
+    """Test that recall with tags only returns matching memories."""
+    # Store memories for different users
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Grace is a data scientist at Google.", "tags": ["user_grace"]},
+                {"content": "Henry is a software engineer at Meta.", "tags": ["user_henry"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Recall with user_grace tag - should only return Grace's memory
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories/recall",
+        json={"query": "Who works at which company?", "budget": "low", "tags": ["user_grace"]}
+    )
+    assert response.status_code == 200
+    results = response.json()["results"]
+
+    # Should find Grace but not Henry
+    texts = [r["text"] for r in results]
+    assert any("Grace" in t for t in texts), "Should find Grace with user_grace tag"
+    # Henry should NOT be found since he has user_henry tag
+    assert not any("Henry" in t for t in texts), "Should NOT find Henry (different tag)"
+
+
+@pytest.mark.asyncio
+async def test_recall_with_multiple_tags_uses_or_matching(api_client, test_bank_id):
+    """Test that multiple tags use OR matching (any match returns the memory)."""
+    # Store memories for different users
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Ivan leads the security team.", "tags": ["user_ivan"]},
+                {"content": "Julia manages the design team.", "tags": ["user_julia"]},
+                {"content": "Karl oversees the marketing team.", "tags": ["user_karl"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Recall with user_ivan OR user_julia - should return both Ivan and Julia, but not Karl
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories/recall",
+        json={"query": "Who leads which team?", "budget": "low", "tags": ["user_ivan", "user_julia"]}
+    )
+    assert response.status_code == 200
+    results = response.json()["results"]
+
+    texts = [r["text"] for r in results]
+    assert any("Ivan" in t for t in texts), "Should find Ivan (tag matches)"
+    assert any("Julia" in t for t in texts), "Should find Julia (tag matches)"
+    assert not any("Karl" in t for t in texts), "Should NOT find Karl (tag doesn't match)"
+
+
+@pytest.mark.asyncio
+async def test_recall_returns_memories_with_any_overlapping_tag(api_client, test_bank_id):
+    """Test that memories with multiple tags are returned if ANY tag matches."""
+    # Store memory with multiple tags
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "Lisa and Mike discussed the budget in a group chat.",
+                    "tags": ["user_lisa", "user_mike"]  # Memory visible to both
+                },
+                {"content": "Nancy reviewed the budget alone.", "tags": ["user_nancy"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Recall with user_lisa - should return the group chat memory
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories/recall",
+        json={"query": "What was discussed about the budget?", "budget": "low", "tags": ["user_lisa"]}
+    )
+    assert response.status_code == 200
+    results = response.json()["results"]
+
+    texts = [r["text"] for r in results]
+    assert any("Lisa" in t and "Mike" in t for t in texts), "Should find group chat (Lisa is in tags)"
+    assert not any("Nancy" in t for t in texts), "Should NOT find Nancy's memory"
+
+
+@pytest.mark.asyncio
+async def test_reflect_with_tags_filters_memories(api_client, test_bank_id):
+    """Test that reflect with tags only uses matching memories for reasoning."""
+    # Store different memories for different users
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Oscar's favorite color is blue.", "tags": ["user_oscar"]},
+                {"content": "Peter's favorite color is red.", "tags": ["user_peter"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Reflect with user_oscar tag - should only use Oscar's memories
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/reflect",
+        json={
+            "query": "What is the favorite color?",
+            "budget": "low",
+            "tags": ["user_oscar"],
+            "include": {"facts": {}}  # Request facts to verify what was used
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+
+    # The response should mention Oscar's color (blue), not Peter's (red)
+    # Note: We can check based_on facts if they're returned
+    if result.get("based_on"):
+        based_on = result["based_on"]
+        memories = based_on.get("memories", []) if isinstance(based_on, dict) else []
+        fact_texts = [f["text"] for f in memories]
+        # Should use Oscar's memory (if facts are included)
+        if fact_texts:
+            assert any("Oscar" in t or "blue" in t for t in fact_texts), "Should use Oscar's memory"
+
+
+@pytest.mark.asyncio
+async def test_recall_with_empty_tags_returns_all(api_client, test_bank_id):
+    """Test that empty tags list behaves same as no tags (returns all)."""
+    # Store memories
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Quinn studies mathematics.", "tags": ["user_quinn"]},
+                {"content": "Rachel studies physics.", "tags": ["user_rachel"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Recall with empty tags list - should return all
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories/recall",
+        json={"query": "Who studies what?", "budget": "low", "tags": []}
+    )
+    assert response.status_code == 200
+    results = response.json()["results"]
+
+    texts = [r["text"] for r in results]
+    assert any("Quinn" in t for t in texts), "Should find Quinn"
+    assert any("Rachel" in t for t in texts), "Should find Rachel"
+
+
+@pytest.mark.asyncio
+async def test_multi_user_agent_visibility(api_client):
+    """
+    Test multi-user agent visibility scoping.
+
+    Scenario:
+    - Agent has one memory bank
+    - Agent chats with User A (room 1) and User B (room 2) separately
+    - Agent also hosts a group chat with both users (room 3)
+    - User A should only see memories from rooms 1 and 3
+    - User B should only see memories from rooms 2 and 3
+    - Agent (no filter) should see all memories
+    """
+    bank_id = f"multi_user_test_{datetime.now().timestamp()}"
+
+    # Store memories from different chat rooms
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                # Room 1: Agent + User A private chat
+                {"content": "User A said they prefer morning meetings.", "tags": ["user_a"]},
+                # Room 2: Agent + User B private chat
+                {"content": "User B mentioned they like afternoon meetings.", "tags": ["user_b"]},
+                # Room 3: Group chat with both users
+                {"content": "In the group meeting, they agreed to meet at noon.", "tags": ["user_a", "user_b"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # User A queries - should see their private chat and group chat
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories/recall",
+        json={"query": "What meeting time preferences were discussed?", "budget": "low", "tags": ["user_a"]}
+    )
+    assert response.status_code == 200
+    user_a_results = response.json()["results"]
+    user_a_texts = [r["text"] for r in user_a_results]
+
+    assert any("morning" in t for t in user_a_texts), "User A should see their own preference (morning)"
+    assert any("noon" in t for t in user_a_texts), "User A should see group chat (noon)"
+    assert not any("afternoon" in t for t in user_a_texts), "User A should NOT see User B's private preference"
+
+    # User B queries - should see their private chat and group chat
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories/recall",
+        json={"query": "What meeting time preferences were discussed?", "budget": "low", "tags": ["user_b"]}
+    )
+    assert response.status_code == 200
+    user_b_results = response.json()["results"]
+    user_b_texts = [r["text"] for r in user_b_results]
+
+    assert any("afternoon" in t for t in user_b_texts), "User B should see their own preference (afternoon)"
+    assert any("noon" in t for t in user_b_texts), "User B should see group chat (noon)"
+    assert not any("morning" in t for t in user_b_texts), "User B should NOT see User A's private preference"
+
+    # Agent queries (no filter) - should see everything
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories/recall",
+        json={"query": "What meeting time preferences were discussed?", "budget": "low"}  # No tags
+    )
+    assert response.status_code == 200
+    agent_results = response.json()["results"]
+    agent_texts = [r["text"] for r in agent_results]
+
+    assert any("morning" in t for t in agent_texts), "Agent should see User A's preference"
+    assert any("afternoon" in t for t in agent_texts), "Agent should see User B's preference"
+    assert any("noon" in t for t in agent_texts), "Agent should see group chat"
+
+
+@pytest.mark.asyncio
+async def test_student_tracking_visibility(api_client):
+    """
+    Test student tracking visibility scoping.
+
+    Scenario:
+    - Teacher bot has one memory bank
+    - Teacher records observations for Student A, Student B
+    - Student A should only see their own data
+    - Teacher (no filter) should see all student data
+    """
+    bank_id = f"student_test_{datetime.now().timestamp()}"
+
+    # Store memories for different students
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Student A showed improvement in algebra today.", "tags": ["student_a"]},
+                {"content": "Student B struggled with geometry concepts.", "tags": ["student_b"]},
+                {"content": "Student A participated actively in class discussion.", "tags": ["student_a"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # Student A queries - should only see their own data
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories/recall",
+        json={"query": "How am I doing in class?", "budget": "low", "tags": ["student_a"]}
+    )
+    assert response.status_code == 200
+    student_a_results = response.json()["results"]
+    student_a_texts = [r["text"] for r in student_a_results]
+
+    assert any("algebra" in t for t in student_a_texts), "Student A should see their algebra progress"
+    assert any("participated" in t for t in student_a_texts), "Student A should see their participation"
+    assert not any("Student B" in t or "geometry" in t for t in student_a_texts), "Student A should NOT see Student B's data"
+
+    # Teacher queries (no filter) - should see all students
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories/recall",
+        json={"query": "Which students need help?", "budget": "low"}  # No tags
+    )
+    assert response.status_code == 200
+    teacher_results = response.json()["results"]
+    teacher_texts = [r["text"] for r in teacher_results]
+
+    assert any("Student A" in t for t in teacher_texts), "Teacher should see Student A's data"
+    assert any("Student B" in t for t in teacher_texts), "Teacher should see Student B's data"
+
+
+# ============================================================================
+# Tests for list_tags API endpoint
+# ============================================================================
+
+
+@pytest.mark.asyncio
+async def test_list_tags_returns_all_tags(api_client):
+    """Test that list_tags returns all unique tags with counts."""
+    bank_id = f"list_tags_test_{datetime.now().timestamp()}"
+
+    # Store memories with various tags
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Memory 1 for user alice.", "tags": ["user:alice"]},
+                {"content": "Memory 2 for user alice.", "tags": ["user:alice"]},
+                {"content": "Memory 3 for user bob.", "tags": ["user:bob"]},
+                {"content": "Memory 4 in session 123.", "tags": ["session:123"]},
+                {"content": "Memory 5 for alice in session 456.", "tags": ["user:alice", "session:456"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # List all tags
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags")
+    assert response.status_code == 200
+    result = response.json()
+
+    # Verify structure
+    assert "items" in result
+    assert "total" in result
+    assert "limit" in result
+    assert "offset" in result
+
+    # Verify tags and counts
+    tags_map = {item["tag"]: item["count"] for item in result["items"]}
+    assert "user:alice" in tags_map
+    assert tags_map["user:alice"] == 3  # 3 memories have this tag
+    assert "user:bob" in tags_map
+    assert tags_map["user:bob"] == 1
+    assert "session:123" in tags_map
+    assert tags_map["session:123"] == 1
+    assert "session:456" in tags_map
+    assert tags_map["session:456"] == 1
+
+    assert result["total"] == 4  # 4 unique tags
+
+
+@pytest.mark.asyncio
+async def test_list_tags_with_wildcard_prefix(api_client):
+    """Test that list_tags filters with prefix wildcard pattern (user:*)."""
+    bank_id = f"list_tags_wildcard_test_{datetime.now().timestamp()}"
+
+    # Store memories with various tags
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Memory for alice who works at tech.", "tags": ["user:alice"]},
+                {"content": "Memory for bob who is an engineer.", "tags": ["user:bob"]},
+                {"content": "Memory for charlie the designer.", "tags": ["user:charlie"]},
+                {"content": "Session memory about the meeting.", "tags": ["session:abc"]},
+                {"content": "Room memory for conference room.", "tags": ["room:123"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # List tags with 'user:*' wildcard pattern
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags", params={"q": "user:*"})
+    assert response.status_code == 200
+    result = response.json()
+
+    # Should only return user:* tags
+    tags = [item["tag"] for item in result["items"]]
+    assert "user:alice" in tags
+    assert "user:bob" in tags
+    assert "user:charlie" in tags
+    assert "session:abc" not in tags
+    assert "room:123" not in tags
+    assert result["total"] == 3
+
+
+@pytest.mark.asyncio
+async def test_list_tags_with_wildcard_suffix(api_client):
+    """Test that list_tags filters with suffix wildcard pattern (*-admin)."""
+    bank_id = f"list_tags_suffix_test_{datetime.now().timestamp()}"
+
+    # Store memories with various tags - use meaningful content for reliable fact extraction
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {"content": "John has the role-admin permission and can manage user accounts.", "tags": ["role-admin"]},
+                {"content": "Sarah has super-admin access and can modify system settings.", "tags": ["super-admin"]},
+                {"content": "Mike is a standard role-user who can only view content.", "tags": ["role-user"]},
+                {"content": "Alice is a role-guest visitor with limited read access.", "tags": ["role-guest"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # List tags with '*-admin' wildcard pattern (suffix match)
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags", params={"q": "*-admin"})
+    assert response.status_code == 200
+    result = response.json()
+
+    # Should only return *-admin tags
+    tags = [item["tag"] for item in result["items"]]
+    assert "role-admin" in tags
+    assert "super-admin" in tags
+    assert "role-user" not in tags
+    assert "role-guest" not in tags
+    assert result["total"] == 2
+
+
+@pytest.mark.asyncio
+async def test_list_tags_with_wildcard_middle(api_client):
+    """Test that list_tags filters with middle wildcard pattern (env*-prod)."""
+    bank_id = f"list_tags_middle_test_{datetime.now().timestamp()}"
+
+    # Store memories with various tags - use meaningful content for fact extraction
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {"content": "The production environment is configured with high availability and uses AWS infrastructure.", "tags": ["env-prod"]},
+                {"content": "The enterprise environment for production runs on dedicated servers with 24/7 monitoring.", "tags": ["environment-prod"]},
+                {"content": "The staging environment mirrors production but uses smaller instance sizes.", "tags": ["env-staging"]},
+                {"content": "The development environment allows developers to test their code locally.", "tags": ["env-dev"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # List tags with 'env*-prod' wildcard pattern (middle match)
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags", params={"q": "env*-prod"})
+    assert response.status_code == 200
+    result = response.json()
+
+    # Should only return env*-prod tags
+    tags = [item["tag"] for item in result["items"]]
+    assert "env-prod" in tags
+    assert "environment-prod" in tags
+    assert "env-staging" not in tags
+    assert "env-dev" not in tags
+    assert result["total"] == 2
+
+
+@pytest.mark.asyncio
+async def test_list_tags_case_insensitive(api_client):
+    """Test that list_tags wildcard matching is case-insensitive."""
+    bank_id = f"list_tags_case_test_{datetime.now().timestamp()}"
+
+    # Store memories with mixed case tags - use meaningful content
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Alice is a software engineer who specializes in machine learning algorithms.", "tags": ["User:Alice"]},
+                {"content": "Bob works as a data scientist at a large technology company.", "tags": ["user:bob"]},
+                {"content": "Charlie is the lead designer responsible for the user interface.", "tags": ["USER:CHARLIE"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # List tags with lowercase pattern - should match all cases
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags", params={"q": "user:*"})
+    assert response.status_code == 200
+    result = response.json()
+
+    # Should match all user tags regardless of case
+    tags = [item["tag"] for item in result["items"]]
+    assert len(tags) == 3
+    assert result["total"] == 3
+
+
+@pytest.mark.asyncio
+async def test_list_tags_pagination(api_client):
+    """Test that list_tags supports pagination."""
+    bank_id = f"list_tags_pagination_test_{datetime.now().timestamp()}"
+
+    # Store memories with many tags - use meaningful content for fact extraction
+    names = ["Alice", "Bob", "Charlie", "Diana", "Eve", "Frank", "Grace", "Henry", "Ivan", "Julia"]
+    items = [
+        {"content": f"{name} works as a software engineer at company {i}.", "tags": [f"tag:{i:03d}"]}
+        for i, name in enumerate(names)
+    ]
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={"items": items}
+    )
+    assert response.status_code == 200
+
+    # Get first page (limit 3)
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags", params={"limit": 3, "offset": 0})
+    assert response.status_code == 200
+    result = response.json()
+    assert len(result["items"]) == 3
+    assert result["total"] == 10
+    assert result["limit"] == 3
+    assert result["offset"] == 0
+
+    # Get second page
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags", params={"limit": 3, "offset": 3})
+    assert response.status_code == 200
+    result = response.json()
+    assert len(result["items"]) == 3
+    assert result["offset"] == 3
+
+
+@pytest.mark.asyncio
+async def test_list_tags_empty_bank(api_client):
+    """Test that list_tags returns empty for bank with no tags."""
+    bank_id = f"list_tags_empty_test_{datetime.now().timestamp()}"
+
+    # List tags without storing anything
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags")
+    assert response.status_code == 200
+    result = response.json()
+
+    assert result["items"] == []
+    assert result["total"] == 0
+
+
+@pytest.mark.asyncio
+async def test_list_tags_ordered_by_count(api_client):
+    """Test that list_tags returns tags ordered by frequency (most used first)."""
+    bank_id = f"list_tags_order_test_{datetime.now().timestamp()}"
+
+    # Store memories with tags having different frequencies - use meaningful content
+    response = await api_client.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {"content": "Alice works at a startup company as a developer.", "tags": ["rare"]},
+                {"content": "Bob is a senior engineer at Google.", "tags": ["common"]},
+                {"content": "Charlie manages the marketing team at Microsoft.", "tags": ["common"]},
+                {"content": "Diana leads the design department at Apple.", "tags": ["common"]},
+                {"content": "Eve is a data scientist at Amazon.", "tags": ["medium"]},
+                {"content": "Frank handles customer support at Meta.", "tags": ["medium"]},
+            ]
+        }
+    )
+    assert response.status_code == 200
+
+    # List tags - should be ordered by count descending
+    response = await api_client.get(f"/v1/default/banks/{bank_id}/tags")
+    assert response.status_code == 200
+    result = response.json()
+
+    tags = [item["tag"] for item in result["items"]]
+    # common (3) should come before medium (2) which should come before rare (1)
+    assert tags.index("common") < tags.index("medium")
+    assert tags.index("medium") < tags.index("rare")
@@ -0,0 +1,592 @@
+"""
+Tests for the distributed worker system.
+
+Tests cover:
+- BrokerTaskBackend task submission and storage
+- WorkerPoller task claiming with FOR UPDATE SKIP LOCKED
+- Concurrent workers claiming different tasks (no duplicates)
+- Task completion and failure handling
+- Retry mechanism
+- Worker decommissioning
+"""
+
+import asyncio
+import json
+import uuid
+
+import pytest
+import pytest_asyncio
+
+from hindsight_api.engine.task_backend import BrokerTaskBackend, SyncTaskBackend
+
+
+# Use loadgroup to ensure these tests run in the same worker
+# since they share database state
+pytestmark = pytest.mark.xdist_group("worker_tests")
+
+
+@pytest_asyncio.fixture
+async def pool(pg0_db_url):
+    """Create a dedicated connection pool for worker tests."""
+    import asyncpg
+
+    from hindsight_api.pg0 import resolve_database_url
+
+    # Resolve pg0:// URL to postgresql:// URL if needed
+    resolved_url = await resolve_database_url(pg0_db_url)
+
+    pool = await asyncpg.create_pool(
+        resolved_url,
+        min_size=2,
+        max_size=10,
+        command_timeout=30,
+    )
+    yield pool
+    await pool.close()
+
+
+@pytest_asyncio.fixture
+async def clean_operations(pool):
+    """Clean up async_operations table before and after tests."""
+    # Clean before test
+    await pool.execute("DELETE FROM async_operations WHERE bank_id LIKE 'test-worker-%'")
+    yield
+    # Clean after test
+    await pool.execute("DELETE FROM async_operations WHERE bank_id LIKE 'test-worker-%'")
+
+
+class TestBrokerTaskBackend:
+    """Tests for BrokerTaskBackend task storage."""
+
+    @pytest.mark.asyncio
+    async def test_submit_task_updates_existing_operation(self, pool, clean_operations):
+        """Test that submit_task updates task_payload for existing operations."""
+        # Create an operation record first
+        operation_id = uuid.uuid4()
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+
+        await pool.execute(
+            """
+            INSERT INTO async_operations (operation_id, bank_id, operation_type, status)
+            VALUES ($1, $2, 'test_operation', 'pending')
+            """,
+            operation_id,
+            bank_id,
+        )
+
+        # Submit task with same operation_id
+        backend = BrokerTaskBackend(pool_getter=lambda: pool)
+        await backend.initialize()
+
+        task_dict = {
+            "operation_id": str(operation_id),
+            "type": "test_task",
+            "bank_id": bank_id,
+            "data": {"key": "value"},
+        }
+        await backend.submit_task(task_dict)
+
+        # Verify task_payload was stored
+        row = await pool.fetchrow(
+            "SELECT task_payload, status FROM async_operations WHERE operation_id = $1",
+            operation_id,
+        )
+        assert row is not None
+        assert row["status"] == "pending"
+        payload = json.loads(row["task_payload"])
+        assert payload["type"] == "test_task"
+        assert payload["data"] == {"key": "value"}
+
+    @pytest.mark.asyncio
+    async def test_submit_task_creates_new_operation(self, pool, clean_operations):
+        """Test that submit_task creates new operation when no operation_id provided."""
+        backend = BrokerTaskBackend(pool_getter=lambda: pool)
+        await backend.initialize()
+
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        task_dict = {
+            "type": "access_count_update",
+            "bank_id": bank_id,
+            "node_ids": ["node1", "node2"],
+        }
+        await backend.submit_task(task_dict)
+
+        # Verify new operation was created
+        row = await pool.fetchrow(
+            "SELECT operation_type, status, task_payload FROM async_operations WHERE bank_id = $1",
+            bank_id,
+        )
+        assert row is not None
+        assert row["operation_type"] == "access_count_update"
+        assert row["status"] == "pending"
+        payload = json.loads(row["task_payload"])
+        assert payload["node_ids"] == ["node1", "node2"]
+
+
+class TestWorkerPoller:
+    """Tests for WorkerPoller task claiming and execution."""
+
+    @pytest.mark.asyncio
+    async def test_claim_batch_claims_pending_tasks(self, pool, clean_operations):
+        """Test that claim_batch claims pending tasks with task_payload."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create some pending tasks
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        for i in range(3):
+            op_id = uuid.uuid4()
+            payload = json.dumps({"type": "test_task", "index": i, "bank_id": bank_id})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload)
+                VALUES ($1, $2, 'test', 'pending', $3::jsonb)
+                """,
+                op_id,
+                bank_id,
+                payload,
+            )
+
+        # Create poller and claim tasks
+        executed_tasks = []
+
+        async def mock_executor(task_dict):
+            executed_tasks.append(task_dict)
+
+        poller = WorkerPoller(
+            pool=pool,
+            worker_id="test-worker-1",
+            executor=mock_executor,
+            batch_size=10,
+        )
+
+        claimed = await poller.claim_batch()
+        assert len(claimed) == 3
+
+        # Verify tasks are marked as processing with worker_id
+        rows = await pool.fetch(
+            "SELECT status, worker_id FROM async_operations WHERE bank_id = $1",
+            bank_id,
+        )
+        for row in rows:
+            assert row["status"] == "processing"
+            assert row["worker_id"] == "test-worker-1"
+
+    @pytest.mark.asyncio
+    async def test_claim_batch_respects_batch_size(self, pool, clean_operations):
+        """Test that claim_batch respects the batch_size limit."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create 10 pending tasks
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        for i in range(10):
+            op_id = uuid.uuid4()
+            payload = json.dumps({"type": "test_task", "index": i, "bank_id": bank_id})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload)
+                VALUES ($1, $2, 'test', 'pending', $3::jsonb)
+                """,
+                op_id,
+                bank_id,
+                payload,
+            )
+
+        # Claim with batch_size=3
+        poller = WorkerPoller(
+            pool=pool,
+            worker_id="test-worker-1",
+            executor=lambda x: None,
+            batch_size=3,
+        )
+
+        claimed = await poller.claim_batch()
+        assert len(claimed) == 3
+
+    @pytest.mark.asyncio
+    async def test_execute_task_marks_completed(self, pool, clean_operations):
+        """Test that successful task execution marks task as completed."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create a pending task
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        op_id = uuid.uuid4()
+        payload = json.dumps({"type": "test_task", "operation_id": str(op_id), "bank_id": bank_id})
+        await pool.execute(
+            """
+            INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload, worker_id)
+            VALUES ($1, $2, 'test', 'processing', $3::jsonb, 'test-worker-1')
+            """,
+            op_id,
+            bank_id,
+            payload,
+        )
+
+        executed = []
+
+        async def mock_executor(task_dict):
+            executed.append(task_dict)
+
+        poller = WorkerPoller(
+            pool=pool,
+            worker_id="test-worker-1",
+            executor=mock_executor,
+        )
+
+        # Execute the task
+        task_dict = json.loads(payload)
+        await poller.execute_task(str(op_id), task_dict)
+
+        assert len(executed) == 1
+
+        # Verify task is marked as completed
+        row = await pool.fetchrow(
+            "SELECT status, completed_at FROM async_operations WHERE operation_id = $1",
+            op_id,
+        )
+        assert row["status"] == "completed"
+        assert row["completed_at"] is not None
+
+    @pytest.mark.asyncio
+    async def test_execute_task_retries_on_failure(self, pool, clean_operations):
+        """Test that failed task execution triggers retry mechanism."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create a pending task with retry_count=0
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        op_id = uuid.uuid4()
+        payload = json.dumps({"type": "test_task", "operation_id": str(op_id), "bank_id": bank_id})
+        await pool.execute(
+            """
+            INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload, worker_id, retry_count)
+            VALUES ($1, $2, 'test', 'processing', $3::jsonb, 'test-worker-1', 0)
+            """,
+            op_id,
+            bank_id,
+            payload,
+        )
+
+        async def failing_executor(task_dict):
+            raise ValueError("Simulated failure")
+
+        poller = WorkerPoller(
+            pool=pool,
+            worker_id="test-worker-1",
+            executor=failing_executor,
+            max_retries=3,
+        )
+
+        # Execute (should fail and retry)
+        task_dict = json.loads(payload)
+        await poller.execute_task(str(op_id), task_dict)
+
+        # Verify task is back to pending with incremented retry_count
+        row = await pool.fetchrow(
+            "SELECT status, retry_count, worker_id FROM async_operations WHERE operation_id = $1",
+            op_id,
+        )
+        assert row["status"] == "pending"
+        assert row["retry_count"] == 1
+        assert row["worker_id"] is None  # Worker ID cleared for retry
+
+    @pytest.mark.asyncio
+    async def test_execute_task_fails_after_max_retries(self, pool, clean_operations):
+        """Test that task is marked failed after exceeding max retries."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create a task that has already used all retries
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        op_id = uuid.uuid4()
+        payload = json.dumps({"type": "test_task", "operation_id": str(op_id), "bank_id": bank_id})
+        await pool.execute(
+            """
+            INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload, worker_id, retry_count)
+            VALUES ($1, $2, 'test', 'processing', $3::jsonb, 'test-worker-1', 3)
+            """,
+            op_id,
+            bank_id,
+            payload,
+        )
+
+        async def failing_executor(task_dict):
+            raise ValueError("Simulated failure")
+
+        poller = WorkerPoller(
+            pool=pool,
+            worker_id="test-worker-1",
+            executor=failing_executor,
+            max_retries=3,
+        )
+
+        # Execute (should fail permanently)
+        task_dict = json.loads(payload)
+        await poller.execute_task(str(op_id), task_dict)
+
+        # Verify task is marked as failed
+        row = await pool.fetchrow(
+            "SELECT status, error_message FROM async_operations WHERE operation_id = $1",
+            op_id,
+        )
+        assert row["status"] == "failed"
+        assert "Max retries" in row["error_message"]
+
+
+class TestConcurrentWorkers:
+    """Tests for concurrent worker task claiming (FOR UPDATE SKIP LOCKED)."""
+
+    @pytest.mark.asyncio
+    async def test_concurrent_workers_claim_different_tasks(self, pool, clean_operations):
+        """Test that multiple workers claim different tasks (no duplicates)."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create 10 pending tasks
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        task_ids = []
+        for i in range(10):
+            op_id = uuid.uuid4()
+            task_ids.append(op_id)
+            payload = json.dumps({"type": "test_task", "index": i, "bank_id": bank_id, "operation_id": str(op_id)})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload)
+                VALUES ($1, $2, 'test', 'pending', $3::jsonb)
+                """,
+                op_id,
+                bank_id,
+                payload,
+            )
+
+        # Create 3 workers that will claim tasks concurrently
+        workers_claimed: dict[str, list[str]] = {"worker-1": [], "worker-2": [], "worker-3": []}
+
+        async def claim_for_worker(worker_id: str):
+            poller = WorkerPoller(
+                pool=pool,
+                worker_id=worker_id,
+                executor=lambda x: None,
+                batch_size=5,  # Each worker tries to claim 5
+            )
+            claimed = await poller.claim_batch()
+            workers_claimed[worker_id] = [op_id for op_id, _ in claimed]
+
+        # Run all workers concurrently
+        await asyncio.gather(
+            claim_for_worker("worker-1"),
+            claim_for_worker("worker-2"),
+            claim_for_worker("worker-3"),
+        )
+
+        # Verify no duplicates - each task claimed by exactly one worker
+        all_claimed = workers_claimed["worker-1"] + workers_claimed["worker-2"] + workers_claimed["worker-3"]
+        assert len(all_claimed) == len(set(all_claimed)), "Duplicate task claimed by multiple workers!"
+
+        # Verify total claimed equals available tasks (10)
+        assert len(all_claimed) == 10, f"Expected 10 tasks claimed, got {len(all_claimed)}"
+
+        # Verify each task is assigned to exactly one worker in DB
+        rows = await pool.fetch(
+            "SELECT operation_id, worker_id FROM async_operations WHERE bank_id = $1",
+            bank_id,
+        )
+        worker_assignments = {str(row["operation_id"]): row["worker_id"] for row in rows}
+
+        # With FOR UPDATE SKIP LOCKED, it's a race condition which workers get tasks.
+        # The important invariant is no duplicates and all tasks claimed, which we verified above.
+        # Just verify that at least 1 worker got tasks and all tasks have a worker assigned.
+        assert len(set(worker_assignments.values())) >= 1, "At least one worker should have claimed tasks"
+        assert all(w is not None for w in worker_assignments.values()), "All tasks should have a worker assigned"
+
+    @pytest.mark.asyncio
+    async def test_workers_do_not_claim_already_processing_tasks(self, pool, clean_operations):
+        """Test that workers skip tasks already being processed by another worker."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create tasks - some pending, some already processing
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+
+        # Create 3 pending tasks
+        for i in range(3):
+            op_id = uuid.uuid4()
+            payload = json.dumps({"type": "test_task", "index": i, "bank_id": bank_id})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload)
+                VALUES ($1, $2, 'test', 'pending', $3::jsonb)
+                """,
+                op_id,
+                bank_id,
+                payload,
+            )
+
+        # Create 2 already-processing tasks owned by another worker
+        for i in range(2):
+            op_id = uuid.uuid4()
+            payload = json.dumps({"type": "test_task", "index": i + 10, "bank_id": bank_id})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload, worker_id)
+                VALUES ($1, $2, 'test', 'processing', $3::jsonb, 'other-worker')
+                """,
+                op_id,
+                bank_id,
+                payload,
+            )
+
+        # New worker should only claim the 3 pending tasks
+        poller = WorkerPoller(
+            pool=pool,
+            worker_id="new-worker",
+            executor=lambda x: None,
+            batch_size=10,
+        )
+
+        claimed = await poller.claim_batch()
+        assert len(claimed) == 3, "Worker should only claim pending tasks"
+
+        # Verify other worker's tasks are still owned by them
+        row = await pool.fetchrow(
+            "SELECT COUNT(*) as count FROM async_operations WHERE bank_id = $1 AND worker_id = 'other-worker'",
+            bank_id,
+        )
+        assert row["count"] == 2
+
+
+class TestWorkerDecommission:
+    """Tests for worker decommissioning functionality."""
+
+    @pytest.mark.asyncio
+    async def test_decommission_releases_worker_tasks(self, pool, clean_operations):
+        """Test that decommissioning a worker releases all its processing tasks."""
+        # Create tasks being processed by a worker
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+        worker_id = "worker-to-decommission"
+
+        for i in range(5):
+            op_id = uuid.uuid4()
+            payload = json.dumps({"type": "test_task", "index": i, "bank_id": bank_id})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload, worker_id, claimed_at)
+                VALUES ($1, $2, 'test', 'processing', $3::jsonb, $4, now())
+                """,
+                op_id,
+                bank_id,
+                payload,
+                worker_id,
+            )
+
+        # Run decommission
+        result = await pool.fetch(
+            """
+            UPDATE async_operations
+            SET status = 'pending', worker_id = NULL, claimed_at = NULL, updated_at = now()
+            WHERE worker_id = $1 AND status = 'processing'
+            RETURNING operation_id
+            """,
+            worker_id,
+        )
+
+        assert len(result) == 5
+
+        # Verify all tasks are back to pending
+        rows = await pool.fetch(
+            "SELECT status, worker_id, claimed_at FROM async_operations WHERE bank_id = $1",
+            bank_id,
+        )
+        for row in rows:
+            assert row["status"] == "pending"
+            assert row["worker_id"] is None
+            assert row["claimed_at"] is None
+
+    @pytest.mark.asyncio
+    async def test_decommission_does_not_affect_other_workers(self, pool, clean_operations):
+        """Test that decommissioning one worker doesn't affect another worker's tasks."""
+        bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+
+        # Create tasks for worker-1
+        for i in range(3):
+            op_id = uuid.uuid4()
+            payload = json.dumps({"type": "test_task", "index": i, "bank_id": bank_id})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload, worker_id)
+                VALUES ($1, $2, 'test', 'processing', $3::jsonb, 'worker-1')
+                """,
+                op_id,
+                bank_id,
+                payload,
+            )
+
+        # Create tasks for worker-2
+        for i in range(3):
+            op_id = uuid.uuid4()
+            payload = json.dumps({"type": "test_task", "index": i + 10, "bank_id": bank_id})
+            await pool.execute(
+                """
+                INSERT INTO async_operations (operation_id, bank_id, operation_type, status, task_payload, worker_id)
+                VALUES ($1, $2, 'test', 'processing', $3::jsonb, 'worker-2')
+                """,
+                op_id,
+                bank_id,
+                payload,
+            )
+
+        # Decommission worker-1 only
+        await pool.execute(
+            """
+            UPDATE async_operations
+            SET status = 'pending', worker_id = NULL, claimed_at = NULL
+            WHERE worker_id = 'worker-1' AND status = 'processing'
+            """
+        )
+
+        # Verify worker-1 tasks are released
+        worker1_rows = await pool.fetch(
+            "SELECT status, worker_id FROM async_operations WHERE bank_id = $1 AND worker_id IS NULL",
+            bank_id,
+        )
+        assert len(worker1_rows) == 3
+
+        # Verify worker-2 tasks are unaffected
+        worker2_rows = await pool.fetch(
+            "SELECT status, worker_id FROM async_operations WHERE bank_id = $1 AND worker_id = 'worker-2'",
+            bank_id,
+        )
+        assert len(worker2_rows) == 3
+        for row in worker2_rows:
+            assert row["status"] == "processing"
+
+
+class TestSyncTaskBackend:
+    """Tests for SyncTaskBackend (used in tests and embedded mode)."""
+
+    @pytest.mark.asyncio
+    async def test_sync_backend_executes_immediately(self):
+        """Test that SyncTaskBackend executes tasks immediately."""
+        executed = []
+
+        async def mock_executor(task_dict):
+            executed.append(task_dict)
+
+        backend = SyncTaskBackend()
+        backend.set_executor(mock_executor)
+        await backend.initialize()
+
+        task_dict = {"type": "test", "data": "value"}
+        await backend.submit_task(task_dict)
+
+        assert len(executed) == 1
+        assert executed[0] == task_dict
+
+    @pytest.mark.asyncio
+    async def test_sync_backend_handles_errors(self):
+        """Test that SyncTaskBackend handles executor errors gracefully."""
+
+        async def failing_executor(task_dict):
+            raise ValueError("Test error")
+
+        backend = SyncTaskBackend()
+        backend.set_executor(failing_executor)
+        await backend.initialize()
+
+        # Should not raise, error is logged
+        await backend.submit_task({"type": "test"})
@@ -1,6 +1,6 @@
 [package]
 name = "hindsight-cli"
-version = "0.2.1"
+version = "0.3.0"
 edition = "2021"
 authors = ["Hindsight Team"]
 description = "A beautiful CLI for Hindsight - semantic memory system"
@@ -55,6 +55,7 @@ pub struct MemoryPutResult {
    pub items_count: i64,
    pub message: String,
    pub is_async: bool,
+    pub operation_id: Option<String>,
 }

 #[derive(Clone)]
@@ -116,6 +117,7 @@ impl ApiClient {
        self.runtime.block_on(async {
            let request = types::CreateBankRequest {
                name: Some(name.to_string()),
+                mission: None,
                background: None,
                disposition: None,
            };
@@ -161,10 +163,54 @@ impl ApiClient {
                items_count: result.items_count,
                message: format!("Stored {} memory units", result.items_count),
                is_async: result.async_,
+                operation_id: result.operation_id,
            })
        })
    }

+    /// Poll an operation until it completes or fails.
+    /// Returns Ok(true) if completed successfully, Ok(false) if failed, Err if polling error.
+    pub fn poll_operation(&self, agent_id: &str, operation_id: &str, verbose: bool) -> Result<(bool, Option<String>)> {
+        self.runtime.block_on(async {
+            loop {
+                let response = self.client.list_operations(agent_id, None).await?;
+                let ops = response.into_inner();
+
+                // Find our operation
+                let op = ops.operations.iter().find(|o| o.id == operation_id);
+
+                match op {
+                    Some(operation) => {
+                        if verbose {
+                            eprintln!("Operation {} status: {}", operation_id, operation.status);
+                        }
+                        match operation.status.as_str() {
+                            "pending" => {
+                                // Still running, wait and poll again
+                                tokio::time::sleep(std::time::Duration::from_millis(500)).await;
+                            }
+                            "completed" => {
+                                // Operation completed successfully
+                                return Ok((true, None));
+                            }
+                            "failed" => {
+                                return Ok((false, operation.error_message.clone()));
+                            }
+                            _ => {
+                                // Unknown status, treat as failed
+                                return Ok((false, Some(format!("Unknown status: {}", operation.status))));
+                            }
+                        }
+                    }
+                    None => {
+                        // Operation not in list means it completed successfully (removed from pending/failed)
+                        return Ok((true, None));
+                    }
+                }
+            }
+        })
+    }
+
    pub fn delete_memory(&self, _agent_id: &str, _unit_id: &str, _verbose: bool) -> Result<types::DeleteResponse> {
        // Note: Individual memory deletion is no longer supported in the API
        anyhow::bail!("Individual memory deletion is no longer supported. Use 'memory clear' to clear all memories.")
@@ -281,3 +327,105 @@ pub use types::{
    ReflectResponse,
    RetainRequest,
 };
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_operation_deserialize() {
+        let json = r#"{
+            "id": "test-op-123",
+            "task_type": "retain",
+            "items_count": 5,
+            "document_id": "doc-456",
+            "created_at": "2024-01-15T10:00:00Z",
+            "status": "pending",
+            "error_message": null
+        }"#;
+        let op: Operation = serde_json::from_str(json).unwrap();
+        assert_eq!(op.id, "test-op-123");
+        assert_eq!(op.task_type, "retain");
+        assert_eq!(op.items_count, 5);
+        assert_eq!(op.document_id, Some("doc-456".to_string()));
+        assert_eq!(op.status, "pending");
+        assert!(op.error_message.is_none());
+    }
+
+    #[test]
+    fn test_operation_deserialize_with_error() {
+        let json = r#"{
+            "id": "test-op-456",
+            "task_type": "retain",
+            "items_count": 3,
+            "document_id": null,
+            "created_at": "2024-01-15T10:00:00Z",
+            "status": "failed",
+            "error_message": "Something went wrong"
+        }"#;
+        let op: Operation = serde_json::from_str(json).unwrap();
+        assert_eq!(op.status, "failed");
+        assert_eq!(op.error_message, Some("Something went wrong".to_string()));
+    }
+
+    #[test]
+    fn test_memory_put_result_serialize() {
+        let result = MemoryPutResult {
+            success: true,
+            items_count: 10,
+            message: "Stored 10 memory units".to_string(),
+            is_async: true,
+            operation_id: Some("op-789".to_string()),
+        };
+        let json = serde_json::to_string(&result).unwrap();
+        assert!(json.contains("\"success\":true"));
+        assert!(json.contains("\"items_count\":10"));
+        assert!(json.contains("\"is_async\":true"));
+        assert!(json.contains("\"operation_id\":\"op-789\""));
+    }
+
+    #[test]
+    fn test_memory_put_result_without_operation_id() {
+        let result = MemoryPutResult {
+            success: true,
+            items_count: 5,
+            message: "Stored 5 memory units".to_string(),
+            is_async: false,
+            operation_id: None,
+        };
+        let json = serde_json::to_string(&result).unwrap();
+        assert!(json.contains("\"operation_id\":null"));
+    }
+
+    #[test]
+    fn test_operations_response_deserialize() {
+        let json = r#"{
+            "bank_id": "test-bank",
+            "operations": [
+                {
+                    "id": "op-1",
+                    "task_type": "retain",
+                    "items_count": 2,
+                    "document_id": null,
+                    "created_at": "2024-01-15T10:00:00Z",
+                    "status": "pending",
+                    "error_message": null
+                },
+                {
+                    "id": "op-2",
+                    "task_type": "retain",
+                    "items_count": 3,
+                    "document_id": "doc-123",
+                    "created_at": "2024-01-15T11:00:00Z",
+                    "status": "completed",
+                    "error_message": null
+                }
+            ]
+        }"#;
+        let ops: OperationsResponse = serde_json::from_str(json).unwrap();
+        assert_eq!(ops.bank_id, "test-bank");
+        assert_eq!(ops.operations.len(), 2);
+        assert_eq!(ops.operations[0].status, "pending");
+        assert_eq!(ops.operations[1].status, "completed");
+    }
+}
@@ -201,7 +201,7 @@ pub fn update_background(
        Ok(profile) => {
            if output_format == OutputFormat::Pretty {
                ui::print_success("Background updated successfully");
-                println!("\n{}", profile.background);
+                println!("\n{}", profile.mission);

                if !no_update_disposition {
                    if let (Some(old_p), Some(new_p)) =
@@ -5,7 +5,7 @@ use crossterm::{
    execute,
    terminal::{disable_raw_mode, enable_raw_mode, EnterAlternateScreen, LeaveAlternateScreen},
 };
-use hindsight_client::types::{BankListItem, RecallResult, EntityListItem, Budget};
+use hindsight_client::types::{BankListItem, RecallResult, EntityListItem, Budget, TagsMatch};
 use serde_json::{Map, Value};
 use ratatui::{
    backend::{Backend, CrosstermBackend},
@@ -341,6 +341,8 @@ impl App {
                            trace: false,
                            query_timestamp: None,
                            include: None,
+                            tags: None,
+                            tags_match: TagsMatch::Any,
                        };

                        let result = client.recall(&bank_id, &request, false)
@@ -357,6 +359,8 @@ impl App {
                            max_tokens: 4096,
                            include: None,
                            response_schema: None,
+                            tags: None,
+                            tags_match: TagsMatch::Any,
                        };

                        let result = client.reflect(&bank_id, &request, false)
@@ -9,7 +9,7 @@ use crate::output::{self, OutputFormat};
 use crate::ui;

 // Import types from generated client
-use hindsight_client::types::{Budget, ChunkIncludeOptions, IncludeOptions};
+use hindsight_client::types::{Budget, ChunkIncludeOptions, IncludeOptions, TagsMatch};
 use serde_json;

 // Helper function to parse budget string to Budget enum
@@ -21,6 +21,17 @@ fn parse_budget(budget: &str) -> Budget {
    }
 }

+// Helper function to check if a file has a text-based extension
+fn is_text_file(path: &std::path::Path) -> bool {
+    const TEXT_EXTENSIONS: &[&str] = &[
+        "txt", "md", "json", "yaml", "yml", "toml", "xml", "csv", "log", "rst", "adoc",
+    ];
+    path.extension()
+        .and_then(|ext| ext.to_str())
+        .map(|ext| TEXT_EXTENSIONS.contains(&ext.to_lowercase().as_str()))
+        .unwrap_or(false)
+}
+
 pub fn recall(
    client: &ApiClient,
    agent_id: &str,
@@ -60,6 +71,8 @@ pub fn recall(
        trace,
        query_timestamp: None,
        include,
+        tags: None,
+        tags_match: TagsMatch::Any,
    };

    let response = client.recall(agent_id, &request, verbose);
@@ -116,6 +129,8 @@ pub fn reflect(
        max_tokens: max_tokens.unwrap_or(4096),
        include: None,
        response_schema,
+        tags: None,
+        tags_match: TagsMatch::Any,
    };

    let response = client.reflect(agent_id, &request, verbose);
@@ -162,11 +177,13 @@ pub fn retain(
        timestamp: None,
        document_id: Some(doc_id.clone()),
        entities: None,
+        tags: None,
    };

    let request = RetainRequest {
        items: vec![item],
        async_: r#async,
+        document_tags: None,
    };

    let response = client.retain(agent_id, &request, r#async, verbose);
@@ -223,29 +240,23 @@ pub fn retain_files(
                .filter(|e| e.file_type().is_file())
            {
                let path = entry.path();
-                if let Some(ext) = path.extension() {
-                    if ext == "txt" || ext == "md" {
-                        files.push(path.to_path_buf());
-                    }
+                if is_text_file(&path) {
+                    files.push(path.to_path_buf());
                }
            }
        } else {
            for entry in fs::read_dir(&path)? {
                let entry = entry?;
                let path = entry.path();
-                if path.is_file() {
-                    if let Some(ext) = path.extension() {
-                        if ext == "txt" || ext == "md" {
-                            files.push(path);
-                        }
-                    }
+                if path.is_file() && is_text_file(&path) {
+                    files.push(path);
                }
            }
        }
    }

    if files.is_empty() {
-        ui::print_warning("No .txt or .md files found");
+        ui::print_warning("No text files found (supported: txt, md, json, yaml, yml, toml, xml, csv, log, rst, adoc)");
        return Ok(());
    }

@@ -272,6 +283,7 @@ pub fn retain_files(
            timestamp: None,
            document_id: Some(doc_id),
            entities: None,
+            tags: None,
        });

        pb.inc(1);
@@ -279,18 +291,20 @@ pub fn retain_files(

    pb.finish_with_message("Files processed");

+    // Always use async mode for the API call
+    let request = RetainRequest {
+        items,
+        async_: true,
+        document_tags: None,
+    };
+
    let spinner = if output_format == OutputFormat::Pretty {
-        Some(ui::create_spinner("Retaining memories..."))
+        Some(ui::create_spinner("Submitting retain request..."))
    } else {
        None
    };

-    let request = RetainRequest {
-        items,
-        async_: r#async,
-    };
-
-    let response = client.retain(agent_id, &request, r#async, verbose);
+    let response = client.retain(agent_id, &request, true, verbose);

    if let Some(mut sp) = spinner {
        sp.finish();
@@ -298,16 +312,55 @@ pub fn retain_files(

    match response {
        Ok(result) => {
-            if output_format == OutputFormat::Pretty {
-                ui::print_success("Files retained successfully");
-                if result.is_async {
-                    println!("  Status: queued for background processing");
+            if r#async {
+                // User requested async mode - return immediately
+                if output_format == OutputFormat::Pretty {
+                    ui::print_success("Files queued for processing");
                    println!("  Items: {}", result.items_count);
+                    if let Some(op_id) = &result.operation_id {
+                        println!("  Operation ID: {}", op_id);
+                    }
                } else {
-                    println!("  Total units created: {}", result.items_count);
+                    output::print_output(&result, output_format)?;
                }
            } else {
-                output::print_output(&result, output_format)?;
+                // Poll until completion
+                if let Some(operation_id) = &result.operation_id {
+                    let poll_spinner = if output_format == OutputFormat::Pretty {
+                        Some(ui::create_spinner("Processing memories..."))
+                    } else {
+                        None
+                    };
+
+                    let (success, error_msg) = client.poll_operation(agent_id, operation_id, verbose)?;
+
+                    if let Some(mut sp) = poll_spinner {
+                        sp.finish();
+                    }
+
+                    if success {
+                        if output_format == OutputFormat::Pretty {
+                            ui::print_success("Files retained successfully");
+                            println!("  Items processed: {}", result.items_count);
+                        } else {
+                            output::print_output(&result, output_format)?;
+                        }
+                    } else {
+                        let msg = error_msg.unwrap_or_else(|| "Unknown error".to_string());
+                        if output_format == OutputFormat::Pretty {
+                            ui::print_error(&format!("Retain operation failed: {}", msg));
+                        }
+                        anyhow::bail!("Retain operation failed: {}", msg);
+                    }
+                } else {
+                    // No operation ID returned, shouldn't happen with async=true
+                    if output_format == OutputFormat::Pretty {
+                        ui::print_success("Files retained successfully");
+                        println!("  Items processed: {}", result.items_count);
+                    } else {
+                        output::print_output(&result, output_format)?;
+                    }
+                }
            }
            Ok(())
        }
@@ -420,3 +473,84 @@ pub fn clear(
        Err(e) => Err(e)
    }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::path::Path;
+
+    #[test]
+    fn test_is_text_file_supported_extensions() {
+        let supported = [
+            "file.txt", "file.md", "file.json", "file.yaml", "file.yml",
+            "file.toml", "file.xml", "file.csv", "file.log", "file.rst", "file.adoc",
+        ];
+        for filename in supported {
+            assert!(
+                is_text_file(Path::new(filename)),
+                "{} should be recognized as a text file",
+                filename
+            );
+        }
+    }
+
+    #[test]
+    fn test_is_text_file_case_insensitive() {
+        assert!(is_text_file(Path::new("file.JSON")));
+        assert!(is_text_file(Path::new("file.TXT")));
+        assert!(is_text_file(Path::new("file.Md")));
+        assert!(is_text_file(Path::new("file.YAML")));
+    }
+
+    #[test]
+    fn test_is_text_file_unsupported_extensions() {
+        let unsupported = [
+            "file.pdf", "file.doc", "file.docx", "file.png", "file.jpg",
+            "file.exe", "file.bin", "file.zip", "file.tar", "file.gz",
+        ];
+        for filename in unsupported {
+            assert!(
+                !is_text_file(Path::new(filename)),
+                "{} should NOT be recognized as a text file",
+                filename
+            );
+        }
+    }
+
+    #[test]
+    fn test_is_text_file_no_extension() {
+        assert!(!is_text_file(Path::new("README")));
+        assert!(!is_text_file(Path::new("Makefile")));
+        assert!(!is_text_file(Path::new(".gitignore")));
+    }
+
+    #[test]
+    fn test_is_text_file_with_path() {
+        assert!(is_text_file(Path::new("/some/path/to/file.json")));
+        assert!(is_text_file(Path::new("../relative/path/file.md")));
+        assert!(!is_text_file(Path::new("/path/to/image.png")));
+    }
+
+    #[test]
+    fn test_parse_budget_valid_values() {
+        assert!(matches!(parse_budget("low"), Budget::Low));
+        assert!(matches!(parse_budget("mid"), Budget::Mid));
+        assert!(matches!(parse_budget("high"), Budget::High));
+    }
+
+    #[test]
+    fn test_parse_budget_case_insensitive() {
+        assert!(matches!(parse_budget("LOW"), Budget::Low));
+        assert!(matches!(parse_budget("MID"), Budget::Mid));
+        assert!(matches!(parse_budget("HIGH"), Budget::High));
+        assert!(matches!(parse_budget("Low"), Budget::Low));
+        assert!(matches!(parse_budget("High"), Budget::High));
+    }
+
+    #[test]
+    fn test_parse_budget_defaults_to_mid() {
+        assert!(matches!(parse_budget("invalid"), Budget::Mid));
+        assert!(matches!(parse_budget(""), Budget::Mid));
+        assert!(matches!(parse_budget("unknown"), Budget::Mid));
+    }
+}
@@ -8,6 +8,7 @@ const DEFAULT_API_URL: &str = "http://localhost:8888";
 const CONFIG_FILE_NAME: &str = "config";
 const CONFIG_DIR_NAME: &str = ".hindsight";

+#[derive(Debug)]
 pub struct Config {
    pub api_url: String,
    pub api_key: Option<String>,
@@ -174,3 +175,156 @@ pub fn generate_doc_id() -> String {
    let now = chrono::Local::now();
    format!("cli_put_{}", now.format("%Y%m%d_%H%M%S"))
 }
+
+/// Parse a simple TOML-like config line and extract value.
+/// Handles both quoted and unquoted values.
+pub fn parse_config_value(line: &str, key: &str) -> Option<String> {
+    let line = line.trim();
+    if !line.starts_with(key) {
+        return None;
+    }
+    line.split('=').nth(1).map(|value| {
+        value.trim().trim_matches('"').trim_matches('\'').to_string()
+    }).filter(|v| !v.is_empty())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_config_source_display() {
+        assert_eq!(format!("{}", ConfigSource::LocalFile), "config file");
+        assert_eq!(format!("{}", ConfigSource::Environment), "environment variable");
+        assert_eq!(format!("{}", ConfigSource::Default), "default");
+    }
+
+    #[test]
+    fn test_validate_and_create_valid_http() {
+        let config = Config::validate_and_create(
+            "http://localhost:8888".to_string(),
+            None,
+            ConfigSource::Default,
+        );
+        assert!(config.is_ok());
+        let config = config.unwrap();
+        assert_eq!(config.api_url, "http://localhost:8888");
+        assert_eq!(config.source, ConfigSource::Default);
+    }
+
+    #[test]
+    fn test_validate_and_create_valid_https() {
+        let config = Config::validate_and_create(
+            "https://api.example.com".to_string(),
+            Some("secret-key".to_string()),
+            ConfigSource::Environment,
+        );
+        assert!(config.is_ok());
+        let config = config.unwrap();
+        assert_eq!(config.api_url, "https://api.example.com");
+        assert_eq!(config.api_key, Some("secret-key".to_string()));
+        assert_eq!(config.source, ConfigSource::Environment);
+    }
+
+    #[test]
+    fn test_validate_and_create_invalid_url() {
+        let config = Config::validate_and_create(
+            "localhost:8888".to_string(),
+            None,
+            ConfigSource::Default,
+        );
+        assert!(config.is_err());
+        let err = config.unwrap_err().to_string();
+        assert!(err.contains("Invalid API URL"));
+        assert!(err.contains("Must start with http:// or https://"));
+    }
+
+    #[test]
+    fn test_validate_and_create_ftp_url() {
+        let config = Config::validate_and_create(
+            "ftp://example.com".to_string(),
+            None,
+            ConfigSource::Default,
+        );
+        assert!(config.is_err());
+    }
+
+    #[test]
+    fn test_generate_doc_id_format() {
+        let doc_id = generate_doc_id();
+        assert!(doc_id.starts_with("cli_put_"));
+        // Should be cli_put_YYYYMMDD_HHMMSS format
+        assert!(doc_id.len() > 20); // cli_put_ (8) + date (8) + _ (1) + time (6) = 23
+    }
+
+    #[test]
+    fn test_generate_doc_id_uniqueness() {
+        let id1 = generate_doc_id();
+        std::thread::sleep(std::time::Duration::from_secs(1));
+        let id2 = generate_doc_id();
+        // IDs generated at different times should be different
+        assert_ne!(id1, id2);
+    }
+
+    #[test]
+    fn test_parse_config_value_quoted() {
+        assert_eq!(
+            parse_config_value(r#"api_url = "http://localhost:8888""#, "api_url"),
+            Some("http://localhost:8888".to_string())
+        );
+    }
+
+    #[test]
+    fn test_parse_config_value_single_quoted() {
+        assert_eq!(
+            parse_config_value("api_url = 'http://localhost:8888'", "api_url"),
+            Some("http://localhost:8888".to_string())
+        );
+    }
+
+    #[test]
+    fn test_parse_config_value_unquoted() {
+        assert_eq!(
+            parse_config_value("api_url = http://localhost:8888", "api_url"),
+            Some("http://localhost:8888".to_string())
+        );
+    }
+
+    #[test]
+    fn test_parse_config_value_with_spaces() {
+        assert_eq!(
+            parse_config_value("  api_url  =  \"http://localhost:8888\"  ", "api_url"),
+            Some("http://localhost:8888".to_string())
+        );
+    }
+
+    #[test]
+    fn test_parse_config_value_wrong_key() {
+        assert_eq!(
+            parse_config_value("api_key = secret", "api_url"),
+            None
+        );
+    }
+
+    #[test]
+    fn test_parse_config_value_empty() {
+        assert_eq!(
+            parse_config_value("api_url = ", "api_url"),
+            None
+        );
+        assert_eq!(
+            parse_config_value("api_url = \"\"", "api_url"),
+            None
+        );
+    }
+
+    #[test]
+    fn test_config_api_url_accessor() {
+        let config = Config {
+            api_url: "http://test:8080".to_string(),
+            api_key: None,
+            source: ConfigSource::Default,
+        };
+        assert_eq!(config.api_url(), "http://test:8080");
+    }
+}
@@ -8,13 +8,35 @@ pub enum OutputFormat {
    Yaml,
 }

+impl OutputFormat {
+    /// Parse output format from string
+    pub fn from_str(s: &str) -> Option<Self> {
+        match s.to_lowercase().as_str() {
+            "json" => Some(OutputFormat::Json),
+            "yaml" | "yml" => Some(OutputFormat::Yaml),
+            "pretty" | "text" => Some(OutputFormat::Pretty),
+            _ => None,
+        }
+    }
+}
+
+/// Format data as JSON string
+pub fn to_json<T: Serialize>(data: &T) -> Result<String> {
+    Ok(serde_json::to_string_pretty(data)?)
+}
+
+/// Format data as YAML string
+pub fn to_yaml<T: Serialize>(data: &T) -> Result<String> {
+    Ok(serde_yaml::to_string(data)?)
+}
+
 pub fn print_output<T: Serialize>(data: &T, format: OutputFormat) -> Result<()> {
    match format {
        OutputFormat::Json => {
-            println!("{}", serde_json::to_string_pretty(data)?);
+            println!("{}", to_json(data)?);
        }
        OutputFormat::Yaml => {
-            println!("{}", serde_yaml::to_string(data)?);
+            println!("{}", to_yaml(data)?);
        }
        OutputFormat::Pretty => {
            // This should not be called - pretty printing is handled in ui.rs
@@ -23,3 +45,121 @@ pub fn print_output<T: Serialize>(data: &T, format: OutputFormat) -> Result<()>
    }
    Ok(())
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde::{Deserialize, Serialize};
+
+    #[derive(Debug, Serialize, Deserialize, PartialEq)]
+    struct TestData {
+        name: String,
+        count: i32,
+        active: bool,
+    }
+
+    #[test]
+    fn test_output_format_from_str_json() {
+        assert_eq!(OutputFormat::from_str("json"), Some(OutputFormat::Json));
+        assert_eq!(OutputFormat::from_str("JSON"), Some(OutputFormat::Json));
+        assert_eq!(OutputFormat::from_str("Json"), Some(OutputFormat::Json));
+    }
+
+    #[test]
+    fn test_output_format_from_str_yaml() {
+        assert_eq!(OutputFormat::from_str("yaml"), Some(OutputFormat::Yaml));
+        assert_eq!(OutputFormat::from_str("YAML"), Some(OutputFormat::Yaml));
+        assert_eq!(OutputFormat::from_str("yml"), Some(OutputFormat::Yaml));
+        assert_eq!(OutputFormat::from_str("YML"), Some(OutputFormat::Yaml));
+    }
+
+    #[test]
+    fn test_output_format_from_str_pretty() {
+        assert_eq!(OutputFormat::from_str("pretty"), Some(OutputFormat::Pretty));
+        assert_eq!(OutputFormat::from_str("PRETTY"), Some(OutputFormat::Pretty));
+        assert_eq!(OutputFormat::from_str("text"), Some(OutputFormat::Pretty));
+    }
+
+    #[test]
+    fn test_output_format_from_str_invalid() {
+        assert_eq!(OutputFormat::from_str("xml"), None);
+        assert_eq!(OutputFormat::from_str("csv"), None);
+        assert_eq!(OutputFormat::from_str(""), None);
+    }
+
+    #[test]
+    fn test_to_json() {
+        let data = TestData {
+            name: "test".to_string(),
+            count: 42,
+            active: true,
+        };
+        let json = to_json(&data).unwrap();
+        assert!(json.contains("\"name\": \"test\""));
+        assert!(json.contains("\"count\": 42"));
+        assert!(json.contains("\"active\": true"));
+    }
+
+    #[test]
+    fn test_to_yaml() {
+        let data = TestData {
+            name: "test".to_string(),
+            count: 42,
+            active: true,
+        };
+        let yaml = to_yaml(&data).unwrap();
+        assert!(yaml.contains("name: test"));
+        assert!(yaml.contains("count: 42"));
+        assert!(yaml.contains("active: true"));
+    }
+
+    #[test]
+    fn test_to_json_array() {
+        let data = vec![
+            TestData { name: "a".to_string(), count: 1, active: true },
+            TestData { name: "b".to_string(), count: 2, active: false },
+        ];
+        let json = to_json(&data).unwrap();
+        assert!(json.contains("\"name\": \"a\""));
+        assert!(json.contains("\"name\": \"b\""));
+    }
+
+    #[test]
+    fn test_to_yaml_array() {
+        let data = vec![
+            TestData { name: "a".to_string(), count: 1, active: true },
+            TestData { name: "b".to_string(), count: 2, active: false },
+        ];
+        let yaml = to_yaml(&data).unwrap();
+        assert!(yaml.contains("name: a"));
+        assert!(yaml.contains("name: b"));
+    }
+
+    #[test]
+    fn test_output_format_equality() {
+        assert_eq!(OutputFormat::Json, OutputFormat::Json);
+        assert_ne!(OutputFormat::Json, OutputFormat::Yaml);
+        assert_ne!(OutputFormat::Yaml, OutputFormat::Pretty);
+    }
+
+    #[test]
+    fn test_output_format_clone() {
+        let format = OutputFormat::Json;
+        let cloned = format.clone();
+        assert_eq!(format, cloned);
+    }
+
+    #[test]
+    fn test_to_json_special_chars() {
+        let data = TestData {
+            name: "test\"with\\special\nchars".to_string(),
+            count: 0,
+            active: false,
+        };
+        let json = to_json(&data).unwrap();
+        // JSON should properly escape special characters
+        assert!(json.contains("\\\""));
+        assert!(json.contains("\\\\"));
+        assert!(json.contains("\\n"));
+    }
+}
@@ -172,8 +172,11 @@ pub fn print_think_response(response: &ReflectResponse) {
    println!("{}", response.text);
    println!();

-    if !response.based_on.is_empty() {
-        println!("{}", dim(&format!("Based on {} memory units", response.based_on.len())));
+    if let Some(based_on) = &response.based_on {
+        let count = based_on.memories.len() + based_on.mental_models.len();
+        if count > 0 {
+            println!("{}", dim(&format!("Based on {} memory units", count)));
+        }
    }

    // Display structured output if present
@@ -322,10 +325,10 @@ pub fn print_disposition(profile: &BankProfileResponse) {
    println!("{} {}", dim("Name:"), gradient_start(&profile.name));
    println!();

-    // Print background if available
-    if !profile.background.is_empty() {
-        println!("{}", gradient_mid("Background:"));
-        for line in profile.background.lines() {
+    // Print mission if available
+    if !profile.mission.is_empty() {
+        println!("{}", gradient_mid("Mission:"));
+        for line in profile.mission.lines() {
            println!("{}", line);
        }
        println!();
@@ -4,6 +4,7 @@ hindsight_client_api/api/banks_api.py
 hindsight_client_api/api/documents_api.py
 hindsight_client_api/api/entities_api.py
 hindsight_client_api/api/memory_api.py
+hindsight_client_api/api/mental_models_api.py
 hindsight_client_api/api/monitoring_api.py
 hindsight_client_api/api/operations_api.py
 hindsight_client_api/api_client.py
@@ -12,6 +13,7 @@ hindsight_client_api/configuration.py
 hindsight_client_api/exceptions.py
 hindsight_client_api/models/__init__.py
 hindsight_client_api/models/add_background_request.py
+hindsight_client_api/models/async_operation_submit_response.py
 hindsight_client_api/models/background_response.py
 hindsight_client_api/models/bank_list_item.py
 hindsight_client_api/models/bank_list_response.py
@@ -23,6 +25,8 @@ hindsight_client_api/models/chunk_data.py
 hindsight_client_api/models/chunk_include_options.py
 hindsight_client_api/models/chunk_response.py
 hindsight_client_api/models/create_bank_request.py
+hindsight_client_api/models/create_mental_model_request.py
+hindsight_client_api/models/created_mental_model.py
 hindsight_client_api/models/delete_document_response.py
 hindsight_client_api/models/delete_response.py
 hindsight_client_api/models/disposition_traits.py
@@ -39,20 +43,37 @@ hindsight_client_api/models/http_validation_error.py
 hindsight_client_api/models/include_options.py
 hindsight_client_api/models/list_documents_response.py
 hindsight_client_api/models/list_memory_units_response.py
+hindsight_client_api/models/list_tags_response.py
 hindsight_client_api/models/memory_item.py
+hindsight_client_api/models/mental_model_freshness_response.py
+hindsight_client_api/models/mental_model_list_response.py
+hindsight_client_api/models/mental_model_observation_response.py
+hindsight_client_api/models/mental_model_response.py
+hindsight_client_api/models/observation_evidence_response.py
+hindsight_client_api/models/observation_input.py
 hindsight_client_api/models/operation_response.py
+hindsight_client_api/models/operation_status_response.py
 hindsight_client_api/models/operations_list_response.py
 hindsight_client_api/models/recall_request.py
 hindsight_client_api/models/recall_response.py
 hindsight_client_api/models/recall_result.py
+hindsight_client_api/models/reflect_based_on.py
 hindsight_client_api/models/reflect_fact.py
 hindsight_client_api/models/reflect_include_options.py
+hindsight_client_api/models/reflect_llm_call.py
+hindsight_client_api/models/reflect_mental_model.py
 hindsight_client_api/models/reflect_request.py
 hindsight_client_api/models/reflect_response.py
+hindsight_client_api/models/reflect_tool_call.py
+hindsight_client_api/models/reflect_trace.py
+hindsight_client_api/models/refresh_mental_models_request.py
 hindsight_client_api/models/retain_request.py
 hindsight_client_api/models/retain_response.py
+hindsight_client_api/models/tag_item.py
 hindsight_client_api/models/token_usage.py
+hindsight_client_api/models/tool_calls_include_options.py
 hindsight_client_api/models/update_disposition_request.py
+hindsight_client_api/models/update_mental_model_request.py
 hindsight_client_api/models/validation_error.py
 hindsight_client_api/models/validation_error_loc_inner.py
 hindsight_client_api/rest.py
@@ -115,6 +115,7 @@ class Hindsight:
        document_id: Optional[str] = None,
        metadata: Optional[Dict[str, str]] = None,
        entities: Optional[List[Dict[str, str]]] = None,
+        tags: Optional[List[str]] = None,
    ) -> RetainResponse:
        """
        Store a single memory (simplified interface).
@@ -127,13 +128,14 @@ class Hindsight:
            document_id: Optional document ID for grouping
            metadata: Optional user-defined metadata
            entities: Optional list of entities [{"text": "...", "type": "..."}]
+            tags: Optional list of tags for this memory

        Returns:
            RetainResponse with success status
        """
        return self.retain_batch(
            bank_id=bank_id,
-            items=[{"content": content, "timestamp": timestamp, "context": context, "metadata": metadata, "entities": entities}],
+            items=[{"content": content, "timestamp": timestamp, "context": context, "metadata": metadata, "entities": entities, "tags": tags}],
            document_id=document_id,
        )

@@ -143,15 +145,17 @@ class Hindsight:
        items: List[Dict[str, Any]],
        document_id: Optional[str] = None,
        retain_async: bool = False,
+        document_tags: Optional[List[str]] = None,
    ) -> RetainResponse:
        """
        Store multiple memories in batch.

        Args:
            bank_id: The memory bank ID
-            items: List of memory items with 'content' and optional 'timestamp', 'context', 'metadata', 'document_id', 'entities'
+            items: List of memory items with 'content' and optional 'timestamp', 'context', 'metadata', 'document_id', 'entities', 'tags'
            document_id: Optional document ID for grouping memories (applied to items that don't have their own)
            retain_async: If True, process asynchronously in background (default: False)
+            document_tags: Optional list of tags to apply to all memories in this batch

        Returns:
            RetainResponse with success status and item count
@@ -175,12 +179,14 @@ class Hindsight:
                    # Use item's document_id if provided, otherwise fall back to batch-level document_id
                    document_id=item.get("document_id") or document_id,
                    entities=entities,
+                    tags=item.get("tags"),
                )
            )

        request_obj = retain_request.RetainRequest(
            items=memory_items,
            async_=retain_async,
+            document_tags=document_tags,
        )

        return _run_async(self._memory_api.retain_memories(bank_id, request_obj))
@@ -198,6 +204,8 @@ class Hindsight:
        max_entity_tokens: int = 500,
        include_chunks: bool = False,
        max_chunk_tokens: int = 8192,
+        tags: Optional[List[str]] = None,
+        tags_match: str = "any",
    ) -> RecallResponse:
        """
        Recall memories using semantic similarity.
@@ -214,6 +222,9 @@ class Hindsight:
            max_entity_tokens: Maximum tokens for entity observations (default: 500)
            include_chunks: Include raw text chunks in results (default: False)
            max_chunk_tokens: Maximum tokens for chunks (default: 8192)
+            tags: Optional list of tags to filter memories by
+            tags_match: How to match tags: 'any' (OR, includes untagged), 'all' (AND, includes untagged),
+                'any_strict' (OR, excludes untagged), 'all_strict' (AND, excludes untagged). Default: 'any'

        Returns:
            RecallResponse with results, optional entities, optional chunks, and optional trace
@@ -233,6 +244,8 @@ class Hindsight:
            trace=trace,
            query_timestamp=query_timestamp,
            include=include_opts,
+            tags=tags,
+            tags_match=tags_match,
        )

        return _run_async(self._memory_api.recall_memories(bank_id, request_obj))
@@ -245,6 +258,8 @@ class Hindsight:
        context: Optional[str] = None,
        max_tokens: Optional[int] = None,
        response_schema: Optional[Dict[str, Any]] = None,
+        tags: Optional[List[str]] = None,
+        tags_match: str = "any",
    ) -> ReflectResponse:
        """
        Generate a contextual answer based on bank identity and memories.
@@ -258,6 +273,9 @@ class Hindsight:
            response_schema: Optional JSON Schema for structured output. When provided,
                the response will include a 'structured_output' field with the LLM
                response parsed according to this schema.
+            tags: Optional list of tags to filter memories by
+            tags_match: How to match tags: 'any' (OR, includes untagged), 'all' (AND, includes untagged),
+                'any_strict' (OR, excludes untagged), 'all_strict' (AND, excludes untagged). Default: 'any'

        Returns:
            ReflectResponse with answer text, optionally facts used, and optionally
@@ -269,6 +287,8 @@ class Hindsight:
            context=context,
            max_tokens=max_tokens,
            response_schema=response_schema,
+            tags=tags,
+            tags_match=tags_match,
        )

        return _run_async(self._memory_api.reflect(bank_id, request_obj))
@@ -21,6 +21,7 @@ from hindsight_client_api.api.banks_api import BanksApi
 from hindsight_client_api.api.documents_api import DocumentsApi
 from hindsight_client_api.api.entities_api import EntitiesApi
 from hindsight_client_api.api.memory_api import MemoryApi
+from hindsight_client_api.api.mental_models_api import MentalModelsApi
 from hindsight_client_api.api.monitoring_api import MonitoringApi
 from hindsight_client_api.api.operations_api import OperationsApi

@@ -37,6 +38,7 @@ from hindsight_client_api.exceptions import ApiException

 # import models into sdk package
 from hindsight_client_api.models.add_background_request import AddBackgroundRequest
+from hindsight_client_api.models.async_operation_submit_response import AsyncOperationSubmitResponse
 from hindsight_client_api.models.background_response import BackgroundResponse
 from hindsight_client_api.models.bank_list_item import BankListItem
 from hindsight_client_api.models.bank_list_response import BankListResponse
@@ -48,6 +50,8 @@ from hindsight_client_api.models.chunk_data import ChunkData
 from hindsight_client_api.models.chunk_include_options import ChunkIncludeOptions
 from hindsight_client_api.models.chunk_response import ChunkResponse
 from hindsight_client_api.models.create_bank_request import CreateBankRequest
+from hindsight_client_api.models.create_mental_model_request import CreateMentalModelRequest
+from hindsight_client_api.models.created_mental_model import CreatedMentalModel
 from hindsight_client_api.models.delete_document_response import DeleteDocumentResponse
 from hindsight_client_api.models.delete_response import DeleteResponse
 from hindsight_client_api.models.disposition_traits import DispositionTraits
@@ -64,19 +68,36 @@ from hindsight_client_api.models.http_validation_error import HTTPValidationErro
 from hindsight_client_api.models.include_options import IncludeOptions
 from hindsight_client_api.models.list_documents_response import ListDocumentsResponse
 from hindsight_client_api.models.list_memory_units_response import ListMemoryUnitsResponse
+from hindsight_client_api.models.list_tags_response import ListTagsResponse
 from hindsight_client_api.models.memory_item import MemoryItem
+from hindsight_client_api.models.mental_model_freshness_response import MentalModelFreshnessResponse
+from hindsight_client_api.models.mental_model_list_response import MentalModelListResponse
+from hindsight_client_api.models.mental_model_observation_response import MentalModelObservationResponse
+from hindsight_client_api.models.mental_model_response import MentalModelResponse
+from hindsight_client_api.models.observation_evidence_response import ObservationEvidenceResponse
+from hindsight_client_api.models.observation_input import ObservationInput
 from hindsight_client_api.models.operation_response import OperationResponse
+from hindsight_client_api.models.operation_status_response import OperationStatusResponse
 from hindsight_client_api.models.operations_list_response import OperationsListResponse
 from hindsight_client_api.models.recall_request import RecallRequest
 from hindsight_client_api.models.recall_response import RecallResponse
 from hindsight_client_api.models.recall_result import RecallResult
+from hindsight_client_api.models.reflect_based_on import ReflectBasedOn
 from hindsight_client_api.models.reflect_fact import ReflectFact
 from hindsight_client_api.models.reflect_include_options import ReflectIncludeOptions
+from hindsight_client_api.models.reflect_llm_call import ReflectLLMCall
+from hindsight_client_api.models.reflect_mental_model import ReflectMentalModel
 from hindsight_client_api.models.reflect_request import ReflectRequest
 from hindsight_client_api.models.reflect_response import ReflectResponse
+from hindsight_client_api.models.reflect_tool_call import ReflectToolCall
+from hindsight_client_api.models.reflect_trace import ReflectTrace
+from hindsight_client_api.models.refresh_mental_models_request import RefreshMentalModelsRequest
 from hindsight_client_api.models.retain_request import RetainRequest
 from hindsight_client_api.models.retain_response import RetainResponse
+from hindsight_client_api.models.tag_item import TagItem
 from hindsight_client_api.models.token_usage import TokenUsage
+from hindsight_client_api.models.tool_calls_include_options import ToolCallsIncludeOptions
 from hindsight_client_api.models.update_disposition_request import UpdateDispositionRequest
+from hindsight_client_api.models.update_mental_model_request import UpdateMentalModelRequest
 from hindsight_client_api.models.validation_error import ValidationError
 from hindsight_client_api.models.validation_error_loc_inner import ValidationErrorLocInner
@@ -5,6 +5,7 @@ from hindsight_client_api.api.banks_api import BanksApi
 from hindsight_client_api.api.documents_api import DocumentsApi
 from hindsight_client_api.api.entities_api import EntitiesApi
 from hindsight_client_api.api.memory_api import MemoryApi
+from hindsight_client_api.api.mental_models_api import MentalModelsApi
 from hindsight_client_api.api.monitoring_api import MonitoringApi
 from hindsight_client_api.api.operations_api import OperationsApi

@@ -64,9 +64,9 @@ class BanksApi:
        _headers: Optional[Dict[StrictStr, Any]] = None,
        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
    ) -> BackgroundResponse:
-        """Add/merge memory bank background
+        """(Deprecated) Add/merge memory bank background (deprecated)

-        Add new background information or merge with existing. LLM intelligently resolves conflicts, normalizes to first person, and optionally infers disposition traits.
+        Deprecated: Use PUT /mission instead. This endpoint now updates the mission field.

        :param bank_id: (required)
        :type bank_id: str
@@ -95,6 +95,7 @@ class BanksApi:
        :type _host_index: int, optional
        :return: Returns the result object.
        """ # noqa: E501
+        warnings.warn("POST /v1/default/banks/{bank_id}/background is deprecated.", DeprecationWarning)

        _param = self._add_bank_background_serialize(
            bank_id=bank_id,
@@ -140,9 +141,9 @@ class BanksApi:
        _headers: Optional[Dict[StrictStr, Any]] = None,
        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
    ) -> ApiResponse[BackgroundResponse]:
-        """Add/merge memory bank background
+        """(Deprecated) Add/merge memory bank background (deprecated)

-        Add new background information or merge with existing. LLM intelligently resolves conflicts, normalizes to first person, and optionally infers disposition traits.
+        Deprecated: Use PUT /mission instead. This endpoint now updates the mission field.

        :param bank_id: (required)
        :type bank_id: str
@@ -171,6 +172,7 @@ class BanksApi:
        :type _host_index: int, optional
        :return: Returns the result object.
        """ # noqa: E501
+        warnings.warn("POST /v1/default/banks/{bank_id}/background is deprecated.", DeprecationWarning)

        _param = self._add_bank_background_serialize(
            bank_id=bank_id,
@@ -216,9 +218,9 @@ class BanksApi:
        _headers: Optional[Dict[StrictStr, Any]] = None,
        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
    ) -> RESTResponseType:
-        """Add/merge memory bank background
+        """(Deprecated) Add/merge memory bank background (deprecated)

-        Add new background information or merge with existing. LLM intelligently resolves conflicts, normalizes to first person, and optionally infers disposition traits.
+        Deprecated: Use PUT /mission instead. This endpoint now updates the mission field.

        :param bank_id: (required)
        :type bank_id: str
@@ -247,6 +249,7 @@ class BanksApi:
        :type _host_index: int, optional
        :return: Returns the result object.
        """ # noqa: E501
+        warnings.warn("POST /v1/default/banks/{bank_id}/background is deprecated.", DeprecationWarning)

        _param = self._add_bank_background_serialize(
            bank_id=bank_id,
@@ -372,7 +375,7 @@ class BanksApi:
    ) -> BankProfileResponse:
        """Create or update memory bank

-        Create a new agent or update existing agent with disposition and background. Auto-fills missing fields with defaults.
+        Create a new agent or update existing agent with disposition and mission. Auto-fills missing fields with defaults.

        :param bank_id: (required)
        :type bank_id: str
@@ -448,7 +451,7 @@ class BanksApi:
    ) -> ApiResponse[BankProfileResponse]:
        """Create or update memory bank

-        Create a new agent or update existing agent with disposition and background. Auto-fills missing fields with defaults.
+        Create a new agent or update existing agent with disposition and mission. Auto-fills missing fields with defaults.

        :param bank_id: (required)
        :type bank_id: str
@@ -524,7 +527,7 @@ class BanksApi:
    ) -> RESTResponseType:
        """Create or update memory bank

-        Create a new agent or update existing agent with disposition and background. Auto-fills missing fields with defaults.
+        Create a new agent or update existing agent with disposition and mission. Auto-fills missing fields with defaults.

        :param bank_id: (required)
        :type bank_id: str
@@ -1233,7 +1236,7 @@ class BanksApi:
    ) -> BankProfileResponse:
        """Get memory bank profile

-        Get disposition traits and background for a memory bank. Auto-creates agent with defaults if not exists.
+        Get disposition traits and mission for a memory bank. Auto-creates agent with defaults if not exists.

        :param bank_id: (required)
        :type bank_id: str
@@ -1305,7 +1308,7 @@ class BanksApi:
    ) -> ApiResponse[BankProfileResponse]:
        """Get memory bank profile

-        Get disposition traits and background for a memory bank. Auto-creates agent with defaults if not exists.
+        Get disposition traits and mission for a memory bank. Auto-creates agent with defaults if not exists.

        :param bank_id: (required)
        :type bank_id: str
@@ -1377,7 +1380,7 @@ class BanksApi:
    ) -> RESTResponseType:
        """Get memory bank profile

-        Get disposition traits and background for a memory bank. Auto-creates agent with defaults if not exists.
+        Get disposition traits and mission for a memory bank. Auto-creates agent with defaults if not exists.

        :param bank_id: (required)
        :type bank_id: str
@@ -1754,6 +1757,312 @@ class BanksApi:



+    @validate_call
+    async def update_bank(
+        self,
+        bank_id: StrictStr,
+        create_bank_request: CreateBankRequest,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> BankProfileResponse:
+        """Partial update memory bank
+
+        Partially update an agent's profile. Only provided fields will be updated.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param create_bank_request: (required)
+        :type create_bank_request: CreateBankRequest
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._update_bank_serialize(
+            bank_id=bank_id,
+            create_bank_request=create_bank_request,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "BankProfileResponse",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        await response_data.read()
+        return self.api_client.response_deserialize(
+            response_data=response_data,
+            response_types_map=_response_types_map,
+        ).data
+
+
+    @validate_call
+    async def update_bank_with_http_info(
+        self,
+        bank_id: StrictStr,
+        create_bank_request: CreateBankRequest,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> ApiResponse[BankProfileResponse]:
+        """Partial update memory bank
+
+        Partially update an agent's profile. Only provided fields will be updated.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param create_bank_request: (required)
+        :type create_bank_request: CreateBankRequest
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._update_bank_serialize(
+            bank_id=bank_id,
+            create_bank_request=create_bank_request,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "BankProfileResponse",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        await response_data.read()
+        return self.api_client.response_deserialize(
+            response_data=response_data,
+            response_types_map=_response_types_map,
+        )
+
+
+    @validate_call
+    async def update_bank_without_preload_content(
+        self,
+        bank_id: StrictStr,
+        create_bank_request: CreateBankRequest,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> RESTResponseType:
+        """Partial update memory bank
+
+        Partially update an agent's profile. Only provided fields will be updated.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param create_bank_request: (required)
+        :type create_bank_request: CreateBankRequest
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._update_bank_serialize(
+            bank_id=bank_id,
+            create_bank_request=create_bank_request,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "BankProfileResponse",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        return response_data.response
+
+
+    def _update_bank_serialize(
+        self,
+        bank_id,
+        create_bank_request,
+        authorization,
+        _request_auth,
+        _content_type,
+        _headers,
+        _host_index,
+    ) -> RequestSerialized:
+
+        _host = None
+
+        _collection_formats: Dict[str, str] = {
+        }
+
+        _path_params: Dict[str, str] = {}
+        _query_params: List[Tuple[str, str]] = []
+        _header_params: Dict[str, Optional[str]] = _headers or {}
+        _form_params: List[Tuple[str, str]] = []
+        _files: Dict[
+            str, Union[str, bytes, List[str], List[bytes], List[Tuple[str, bytes]]]
+        ] = {}
+        _body_params: Optional[bytes] = None
+
+        # process the path parameters
+        if bank_id is not None:
+            _path_params['bank_id'] = bank_id
+        # process the query parameters
+        # process the header parameters
+        if authorization is not None:
+            _header_params['authorization'] = authorization
+        # process the form parameters
+        # process the body parameter
+        if create_bank_request is not None:
+            _body_params = create_bank_request
+
+
+        # set the HTTP header `Accept`
+        if 'Accept' not in _header_params:
+            _header_params['Accept'] = self.api_client.select_header_accept(
+                [
+                    'application/json'
+                ]
+            )
+
+        # set the HTTP header `Content-Type`
+        if _content_type:
+            _header_params['Content-Type'] = _content_type
+        else:
+            _default_content_type = (
+                self.api_client.select_header_content_type(
+                    [
+                        'application/json'
+                    ]
+                )
+            )
+            if _default_content_type is not None:
+                _header_params['Content-Type'] = _default_content_type
+
+        # authentication setting
+        _auth_settings: List[str] = [
+        ]
+
+        return self.api_client.param_serialize(
+            method='PATCH',
+            resource_path='/v1/default/banks/{bank_id}',
+            path_params=_path_params,
+            query_params=_query_params,
+            header_params=_header_params,
+            body=_body_params,
+            post_params=_form_params,
+            files=_files,
+            auth_settings=_auth_settings,
+            collection_formats=_collection_formats,
+            _host=_host,
+            _request_auth=_request_auth
+        )
+
+
+
+
    @validate_call
    async def update_bank_disposition(
        self,
@@ -664,9 +664,9 @@ class EntitiesApi:
        _headers: Optional[Dict[StrictStr, Any]] = None,
        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
    ) -> EntityDetailResponse:
-        """Regenerate entity observations
+        """(Deprecated) Regenerate entity observations (deprecated)

-        Regenerate observations for an entity based on all facts mentioning it.
+        This endpoint is deprecated. Entity observations have been replaced by mental models.

        :param bank_id: (required)
        :type bank_id: str
@@ -695,6 +695,7 @@ class EntitiesApi:
        :type _host_index: int, optional
        :return: Returns the result object.
        """ # noqa: E501
+        warnings.warn("POST /v1/default/banks/{bank_id}/entities/{entity_id}/regenerate is deprecated.", DeprecationWarning)

        _param = self._regenerate_entity_observations_serialize(
            bank_id=bank_id,
@@ -740,9 +741,9 @@ class EntitiesApi:
        _headers: Optional[Dict[StrictStr, Any]] = None,
        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
    ) -> ApiResponse[EntityDetailResponse]:
-        """Regenerate entity observations
+        """(Deprecated) Regenerate entity observations (deprecated)

-        Regenerate observations for an entity based on all facts mentioning it.
+        This endpoint is deprecated. Entity observations have been replaced by mental models.

        :param bank_id: (required)
        :type bank_id: str
@@ -771,6 +772,7 @@ class EntitiesApi:
        :type _host_index: int, optional
        :return: Returns the result object.
        """ # noqa: E501
+        warnings.warn("POST /v1/default/banks/{bank_id}/entities/{entity_id}/regenerate is deprecated.", DeprecationWarning)

        _param = self._regenerate_entity_observations_serialize(
            bank_id=bank_id,
@@ -816,9 +818,9 @@ class EntitiesApi:
        _headers: Optional[Dict[StrictStr, Any]] = None,
        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
    ) -> RESTResponseType:
-        """Regenerate entity observations
+        """(Deprecated) Regenerate entity observations (deprecated)

-        Regenerate observations for an entity based on all facts mentioning it.
+        This endpoint is deprecated. Entity observations have been replaced by mental models.

        :param bank_id: (required)
        :type bank_id: str
@@ -847,6 +849,7 @@ class EntitiesApi:
        :type _host_index: int, optional
        :return: Returns the result object.
        """ # noqa: E501
+        warnings.warn("POST /v1/default/banks/{bank_id}/entities/{entity_id}/regenerate is deprecated.", DeprecationWarning)

        _param = self._regenerate_entity_observations_serialize(
            bank_id=bank_id,
@@ -17,11 +17,12 @@ from typing import Any, Dict, List, Optional, Tuple, Union
 from typing_extensions import Annotated

 from pydantic import Field, StrictInt, StrictStr
-from typing import Optional
+from typing import Any, Optional
 from typing_extensions import Annotated
 from hindsight_client_api.models.delete_response import DeleteResponse
 from hindsight_client_api.models.graph_data_response import GraphDataResponse
 from hindsight_client_api.models.list_memory_units_response import ListMemoryUnitsResponse
+from hindsight_client_api.models.list_tags_response import ListTagsResponse
 from hindsight_client_api.models.recall_request import RecallRequest
 from hindsight_client_api.models.recall_response import RecallResponse
 from hindsight_client_api.models.reflect_request import ReflectRequest
@@ -654,6 +655,299 @@ class MemoryApi:



+    @validate_call
+    async def get_memory(
+        self,
+        bank_id: StrictStr,
+        memory_id: StrictStr,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> object:
+        """Get memory unit
+
+        Get a single memory unit by ID with all its metadata including entities and tags.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param memory_id: (required)
+        :type memory_id: str
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._get_memory_serialize(
+            bank_id=bank_id,
+            memory_id=memory_id,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "object",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        await response_data.read()
+        return self.api_client.response_deserialize(
+            response_data=response_data,
+            response_types_map=_response_types_map,
+        ).data
+
+
+    @validate_call
+    async def get_memory_with_http_info(
+        self,
+        bank_id: StrictStr,
+        memory_id: StrictStr,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> ApiResponse[object]:
+        """Get memory unit
+
+        Get a single memory unit by ID with all its metadata including entities and tags.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param memory_id: (required)
+        :type memory_id: str
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._get_memory_serialize(
+            bank_id=bank_id,
+            memory_id=memory_id,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "object",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        await response_data.read()
+        return self.api_client.response_deserialize(
+            response_data=response_data,
+            response_types_map=_response_types_map,
+        )
+
+
+    @validate_call
+    async def get_memory_without_preload_content(
+        self,
+        bank_id: StrictStr,
+        memory_id: StrictStr,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> RESTResponseType:
+        """Get memory unit
+
+        Get a single memory unit by ID with all its metadata including entities and tags.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param memory_id: (required)
+        :type memory_id: str
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._get_memory_serialize(
+            bank_id=bank_id,
+            memory_id=memory_id,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "object",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        return response_data.response
+
+
+    def _get_memory_serialize(
+        self,
+        bank_id,
+        memory_id,
+        authorization,
+        _request_auth,
+        _content_type,
+        _headers,
+        _host_index,
+    ) -> RequestSerialized:
+
+        _host = None
+
+        _collection_formats: Dict[str, str] = {
+        }
+
+        _path_params: Dict[str, str] = {}
+        _query_params: List[Tuple[str, str]] = []
+        _header_params: Dict[str, Optional[str]] = _headers or {}
+        _form_params: List[Tuple[str, str]] = []
+        _files: Dict[
+            str, Union[str, bytes, List[str], List[bytes], List[Tuple[str, bytes]]]
+        ] = {}
+        _body_params: Optional[bytes] = None
+
+        # process the path parameters
+        if bank_id is not None:
+            _path_params['bank_id'] = bank_id
+        if memory_id is not None:
+            _path_params['memory_id'] = memory_id
+        # process the query parameters
+        # process the header parameters
+        if authorization is not None:
+            _header_params['authorization'] = authorization
+        # process the form parameters
+        # process the body parameter
+
+
+        # set the HTTP header `Accept`
+        if 'Accept' not in _header_params:
+            _header_params['Accept'] = self.api_client.select_header_accept(
+                [
+                    'application/json'
+                ]
+            )
+
+
+        # authentication setting
+        _auth_settings: List[str] = [
+        ]
+
+        return self.api_client.param_serialize(
+            method='GET',
+            resource_path='/v1/default/banks/{bank_id}/memories/{memory_id}',
+            path_params=_path_params,
+            query_params=_query_params,
+            header_params=_header_params,
+            body=_body_params,
+            post_params=_form_params,
+            files=_files,
+            auth_settings=_auth_settings,
+            collection_formats=_collection_formats,
+            _host=_host,
+            _request_auth=_request_auth
+        )
+
+
+
+
    @validate_call
    async def list_memories(
        self,
@@ -1000,6 +1294,335 @@ class MemoryApi:



+    @validate_call
+    async def list_tags(
+        self,
+        bank_id: StrictStr,
+        q: Annotated[Optional[StrictStr], Field(description="Wildcard pattern to filter tags (e.g., 'user:*' for user:alice, '*-admin' for role-admin). Use '*' as wildcard. Case-insensitive.")] = None,
+        limit: Annotated[Optional[StrictInt], Field(description="Maximum number of tags to return")] = None,
+        offset: Annotated[Optional[StrictInt], Field(description="Offset for pagination")] = None,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> ListTagsResponse:
+        """List tags
+
+        List all unique tags in a memory bank with usage counts. Supports wildcard search using '*' (e.g., 'user:*', '*-fred', 'tag*-2'). Case-insensitive.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param q: Wildcard pattern to filter tags (e.g., 'user:*' for user:alice, '*-admin' for role-admin). Use '*' as wildcard. Case-insensitive.
+        :type q: str
+        :param limit: Maximum number of tags to return
+        :type limit: int
+        :param offset: Offset for pagination
+        :type offset: int
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._list_tags_serialize(
+            bank_id=bank_id,
+            q=q,
+            limit=limit,
+            offset=offset,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "ListTagsResponse",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        await response_data.read()
+        return self.api_client.response_deserialize(
+            response_data=response_data,
+            response_types_map=_response_types_map,
+        ).data
+
+
+    @validate_call
+    async def list_tags_with_http_info(
+        self,
+        bank_id: StrictStr,
+        q: Annotated[Optional[StrictStr], Field(description="Wildcard pattern to filter tags (e.g., 'user:*' for user:alice, '*-admin' for role-admin). Use '*' as wildcard. Case-insensitive.")] = None,
+        limit: Annotated[Optional[StrictInt], Field(description="Maximum number of tags to return")] = None,
+        offset: Annotated[Optional[StrictInt], Field(description="Offset for pagination")] = None,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> ApiResponse[ListTagsResponse]:
+        """List tags
+
+        List all unique tags in a memory bank with usage counts. Supports wildcard search using '*' (e.g., 'user:*', '*-fred', 'tag*-2'). Case-insensitive.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param q: Wildcard pattern to filter tags (e.g., 'user:*' for user:alice, '*-admin' for role-admin). Use '*' as wildcard. Case-insensitive.
+        :type q: str
+        :param limit: Maximum number of tags to return
+        :type limit: int
+        :param offset: Offset for pagination
+        :type offset: int
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._list_tags_serialize(
+            bank_id=bank_id,
+            q=q,
+            limit=limit,
+            offset=offset,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "ListTagsResponse",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        await response_data.read()
+        return self.api_client.response_deserialize(
+            response_data=response_data,
+            response_types_map=_response_types_map,
+        )
+
+
+    @validate_call
+    async def list_tags_without_preload_content(
+        self,
+        bank_id: StrictStr,
+        q: Annotated[Optional[StrictStr], Field(description="Wildcard pattern to filter tags (e.g., 'user:*' for user:alice, '*-admin' for role-admin). Use '*' as wildcard. Case-insensitive.")] = None,
+        limit: Annotated[Optional[StrictInt], Field(description="Maximum number of tags to return")] = None,
+        offset: Annotated[Optional[StrictInt], Field(description="Offset for pagination")] = None,
+        authorization: Optional[StrictStr] = None,
+        _request_timeout: Union[
+            None,
+            Annotated[StrictFloat, Field(gt=0)],
+            Tuple[
+                Annotated[StrictFloat, Field(gt=0)],
+                Annotated[StrictFloat, Field(gt=0)]
+            ]
+        ] = None,
+        _request_auth: Optional[Dict[StrictStr, Any]] = None,
+        _content_type: Optional[StrictStr] = None,
+        _headers: Optional[Dict[StrictStr, Any]] = None,
+        _host_index: Annotated[StrictInt, Field(ge=0, le=0)] = 0,
+    ) -> RESTResponseType:
+        """List tags
+
+        List all unique tags in a memory bank with usage counts. Supports wildcard search using '*' (e.g., 'user:*', '*-fred', 'tag*-2'). Case-insensitive.
+
+        :param bank_id: (required)
+        :type bank_id: str
+        :param q: Wildcard pattern to filter tags (e.g., 'user:*' for user:alice, '*-admin' for role-admin). Use '*' as wildcard. Case-insensitive.
+        :type q: str
+        :param limit: Maximum number of tags to return
+        :type limit: int
+        :param offset: Offset for pagination
+        :type offset: int
+        :param authorization:
+        :type authorization: str
+        :param _request_timeout: timeout setting for this request. If one
+                                 number provided, it will be total request
+                                 timeout. It can also be a pair (tuple) of
+                                 (connection, read) timeouts.
+        :type _request_timeout: int, tuple(int, int), optional
+        :param _request_auth: set to override the auth_settings for an a single
+                              request; this effectively ignores the
+                              authentication in the spec for a single request.
+        :type _request_auth: dict, optional
+        :param _content_type: force content-type for the request.
+        :type _content_type: str, Optional
+        :param _headers: set to override the headers for a single
+                         request; this effectively ignores the headers
+                         in the spec for a single request.
+        :type _headers: dict, optional
+        :param _host_index: set to override the host_index for a single
+                            request; this effectively ignores the host_index
+                            in the spec for a single request.
+        :type _host_index: int, optional
+        :return: Returns the result object.
+        """ # noqa: E501
+
+        _param = self._list_tags_serialize(
+            bank_id=bank_id,
+            q=q,
+            limit=limit,
+            offset=offset,
+            authorization=authorization,
+            _request_auth=_request_auth,
+            _content_type=_content_type,
+            _headers=_headers,
+            _host_index=_host_index
+        )
+
+        _response_types_map: Dict[str, Optional[str]] = {
+            '200': "ListTagsResponse",
+            '422': "HTTPValidationError",
+        }
+        response_data = await self.api_client.call_api(
+            *_param,
+            _request_timeout=_request_timeout
+        )
+        return response_data.response
+
+
+    def _list_tags_serialize(
+        self,
+        bank_id,
+        q,
+        limit,
+        offset,
+        authorization,
+        _request_auth,
+        _content_type,
+        _headers,
+        _host_index,
+    ) -> RequestSerialized:
+
+        _host = None
+
+        _collection_formats: Dict[str, str] = {
+        }
+
+        _path_params: Dict[str, str] = {}
+        _query_params: List[Tuple[str, str]] = []
+        _header_params: Dict[str, Optional[str]] = _headers or {}
+        _form_params: List[Tuple[str, str]] = []
+        _files: Dict[
+            str, Union[str, bytes, List[str], List[bytes], List[Tuple[str, bytes]]]
+        ] = {}
+        _body_params: Optional[bytes] = None
+
+        # process the path parameters
+        if bank_id is not None:
+            _path_params['bank_id'] = bank_id
+        # process the query parameters
+        if q is not None:
+            
+            _query_params.append(('q', q))
+            
+        if limit is not None:
+            
+            _query_params.append(('limit', limit))
+            
+        if offset is not None:
+            
+            _query_params.append(('offset', offset))
+            
+        # process the header parameters
+        if authorization is not None:
+            _header_params['authorization'] = authorization
+        # process the form parameters
+        # process the body parameter
+
+
+        # set the HTTP header `Accept`
+        if 'Accept' not in _header_params:
+            _header_params['Accept'] = self.api_client.select_header_accept(
+                [
+                    'application/json'
+                ]
+            )
+
+
+        # authentication setting
+        _auth_settings: List[str] = [
+        ]
+
+        return self.api_client.param_serialize(
+            method='GET',
+            resource_path='/v1/default/banks/{bank_id}/tags',
+            path_params=_path_params,
+            query_params=_query_params,
+            header_params=_header_params,
+            body=_body_params,
+            post_params=_form_params,
+            files=_files,
+            auth_settings=_auth_settings,
+            collection_formats=_collection_formats,
+            _host=_host,
+            _request_auth=_request_auth
+        )
+
+
+
+
    @validate_call
    async def recall_memories(
        self,
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Nicolò Boschi	85d5b3bfd9	tests	2026-01-19 18:34:32 +01:00
Nicolò Boschi	f15b76fba8	docs	2026-01-19 17:12:55 +01:00
Nicolò Boschi	61f457736e	doc	2026-01-19 16:44:45 +01:00
Nicolò Boschi	5de1447ff5	feat: new 'worker' service	2026-01-19 16:32:18 +01:00
Nicolò Boschi	bac4b24e30	fix(sec): upgrade vulnerable deps (#174 )	2026-01-19 14:27:31 +01:00
Nicolò Boschi	3290f4bfff	chore: unify agents.md and claude.md (#173 )	2026-01-19 14:18:33 +01:00
Nicolò Boschi	63a65d0723	feat: improve mental model refresh and add directives (#166 ) * feat: improve mental model refresh and add directives * feat: improve mental model refresh and add directives * tags * ui * fix * fix * update * update	2026-01-19 11:38:35 +01:00
Chris Bartholomew	870cfccabb	Add structured JSON logging support (#170 ) * Add structured JSON logging support Add HINDSIGHT_API_LOG_FORMAT environment variable to configure log output format. Options are "text" (default, human-readable) and "json" (structured). JSON format outputs logs with a "severity" field that cloud logging systems can parse for proper log level categorization. Also writes to stdout instead of stderr so log levels are correctly interpreted. * Rename GCPJsonFormatter to JsonFormatter	2026-01-19 09:01:24 +01:00
Nicolò Boschi	4476a10aa3	doc: refinement for 0.3.0 new features (#159 ) * doc: refinement for 0.3.0 new features * fix * fix * fixes	2026-01-16 11:16:52 +01:00
Nicolò Boschi	4f2833873c	feat: introduce mental models (#132 ) * mental models * DRAFT: refactor entity observations * fix db patch * agentic * agentic * reflect agent * new style * more * fix ci * fix * fix	2026-01-16 11:16:41 +01:00
Nicolò Boschi	1eeced3116	feat(cli): accept more file types on retain-files (#163 ) * feat(cli): accept more file types on retain-files * feat(cli): accept more file types on retain-files	2026-01-15 18:34:44 +01:00
Chris Bartholomew	55c216e069	Fix skill installer test examples to use meaningful content (#160 ) The "Test memory" example is too short for the LLM to extract meaningful facts from, causing the test to silently fail (0 memories created). Replace with "Alice works at Google as a software engineer" which has enough context for fact extraction. Fixes test examples in: - get-skill installer (local and cloud modes) - hindsight-embed configure output - skills.md documentation	2026-01-14 18:41:04 +01:00
Chris Bartholomew	e64d3634a9	feat: add cloud mode to skill installer for team memory sharing (#158 ) * doc: update expired Slack invite link * feat: add cloud mode to skill installer for team memory sharing Adds support for Hindsight Cloud in the skill installer, enabling teams to share memories about a codebase. Changes include: - Add `--mode cloud` option to get-skill installer - Install hindsight CLI binary for cloud mode (via get-cli) - Configure ~/.hindsight/config with API URL and key - Generate cloud-specific SKILL.md with team-aware guidance - Distinguish between project conventions and individual preferences - Update skills.md documentation with cloud setup instructions Cloud mode workflow: 1. Team admin creates a bank in Hindsight Cloud 2. Each developer runs: curl ... \| bash -s -- --mode cloud 3. All team members share the same memory bank 4. Knowledge retained by one member benefits everyone	2026-01-14 09:05:40 +01:00
Chris Bartholomew	70ce979fbe	doc: update expired Slack invite link (#157 )	2026-01-13 16:57:23 -05:00
Nicolò Boschi	de132501c6	doc: changelog for 0.3.0 (#156 )	2026-01-13 19:09:13 +01:00
Nicolò Boschi	a75dcfebf5	Release v0.3.0 - Update version to 0.3.0 in all components - Python packages: hindsight-api, hindsight-dev, hindsight-all, hindsight-litellm, hindsight-embed - Python client: hindsight-clients/python - TypeScript client: hindsight-clients/typescript - Rust CLI: hindsight-cli - Control Plane: hindsight-control-plane - Helm chart	2026-01-13 18:43:33 +01:00
Nicolò Boschi	20c8f8b06a	feat: add memory tags (#152 ) * feat: add memory tags * feat: add memory tags * support tags * support tags	2026-01-13 18:28:40 +01:00
Chris Bartholomew	f5f3fca4ad	Fix: Load extensions in server.py for multi-worker deployments (#155 ) * Fix: Load extensions in server.py for multi-worker deployments When running with multiple workers (--workers 2), uvicorn uses `hindsight_api.server:app` import string instead of passing an app object. The server.py module was not loading tenant/operation validator extensions, causing authentication bypass in production. This fix: - Adds extension loading to server.py matching main.py behavior - Sets extension context on tenant extension for schema provisioning - Adds comprehensive unit tests for server.py extension loading The tests specifically verify: - TENANT extension is loaded when HINDSIGHT_API_TENANT_EXTENSION is set - OPERATION_VALIDATOR is loaded when configured - Extensions are passed to MemoryEngine constructor - Extension context is set on tenant extension - Server works correctly without extensions configured * Add unit tests for main.py extension loading (single-worker path)	2026-01-13 17:55:33 +01:00
Nicolò Boschi	d47c8a28cc	feat: support litellm gateway (#154 )	2026-01-13 16:55:28 +01:00
Nicolò Boschi	1ffc2a418c	feat: add tenant to metrics labels (#151 )	2026-01-13 15:31:58 +01:00
Nicolò Boschi	fa53917c63	feat: support custom url for openai embeddings & cohere (#150 ) * feat: support custom url for openai embeddings & cohere * feat: support custom url for openai embeddings & cohere	2026-01-13 14:01:44 +01:00