---
phase: 02-agent-features
plan: 01
subsystem: memory
tags: [redis, pgvector, sentence-transformers, celery, memory, multi-tenancy]
dependency_graph:
  requires: [01-foundation (all plans)]
  provides: [conversational-memory, semantic-recall, embedding-backfill]
  affects: [orchestrator/tasks.py, orchestrator/agents/builder.py, orchestrator/agents/runner.py]
tech_stack:
  added: [pgvector>=0.4.2, sentence-transformers>=3.0.0 (all-MiniLM-L6-v2), fakeredis (tests)]
  patterns: [Redis RPUSH/LTRIM sliding window, pgvector HNSW cosine search, fire-and-forget Celery task, lazy singleton model loading]
key_files:
  created:
    - packages/shared/shared/models/memory.py
    - packages/orchestrator/orchestrator/memory/__init__.py
    - packages/orchestrator/orchestrator/memory/short_term.py
    - packages/orchestrator/orchestrator/memory/long_term.py
    - packages/orchestrator/orchestrator/memory/embedder.py
    - migrations/versions/002_phase2_memory.py
    - tests/unit/test_memory_short_term.py
    - tests/integration/test_memory_long_term.py
  modified:
    - packages/shared/shared/redis_keys.py
    - packages/shared/pyproject.toml
    - packages/orchestrator/pyproject.toml
    - packages/orchestrator/orchestrator/agents/builder.py
    - packages/orchestrator/orchestrator/agents/runner.py
    - packages/orchestrator/orchestrator/tasks.py
    - docker-compose.yml
decisions:
  - "pgvector/pgvector:pg16 Docker image used instead of postgres:16-alpine — postgres:16-alpine does not include the pgvector extension; switched to official pgvector image"
  - "Lazy singleton pattern for SentenceTransformer — model loaded once on first use, not at import time, to avoid 2s load penalty when module imported but not used"
  - "embed_and_store is a fire-and-forget Celery task (ignore_result=True) — embedding never blocks the LLM response path"
  - "All queries pre-filter by (tenant_id, agent_id, user_id) BEFORE ANN operator in pgvector — defense-in-depth with RLS as secondary backstop"
  - "No TTL on Redis memory keys — indefinite retention per plan spec; caller controls window size"
  - "pgvector context injected as system message BEFORE sliding window — provides LLM background context without polluting conversation flow"
metrics:
  duration: 9m 22s
  completed_date: "2026-03-23"
  tasks_completed: 2
  files_created: 8
  files_modified: 7
  tests_added: 16
  tests_total: 202
---

# Phase 2 Plan 01: Two-Layer Conversational Memory Summary

**One-liner:** Redis sliding window (last 20 msgs) + pgvector HNSW semantic recall (all-MiniLM-L6-v2, 384d) with per-user per-agent per-tenant isolation and async Celery embedding backfill.

## What Was Built

Transforms the stateless Phase 1 agent into one with persistent conversational memory across sessions. Every LLM call now receives:

1. **Short-term context**: Last 20 messages from Redis (zero-latency in-session history)
2. **Long-term context**: Up to 3 semantically relevant past exchanges from pgvector (cross-session recall)

The embedding backfill runs asynchronously — the LLM response is never blocked.

## Task Execution

### Task 1: DB models, migration, and memory modules with tests

**Status:** Complete

**TDD Cycle:**
- RED: Wrote 10 unit tests (fakeredis, short_term) + 6 integration tests (pgvector, long_term) → ImportError confirmed failure
- GREEN: Implemented all modules → 16/16 tests pass

**Files created:**
- `packages/shared/shared/models/memory.py` — ConversationEmbedding ORM with Vector(384) column
- `packages/orchestrator/orchestrator/memory/__init__.py` — Package init with architecture doc
- `packages/orchestrator/orchestrator/memory/short_term.py` — RPUSH/LTRIM sliding window
- `packages/orchestrator/orchestrator/memory/long_term.py` — pgvector HNSW cosine search
- `migrations/versions/002_phase2_memory.py` — Alembic migration with HNSW index and RLS

**Files modified:**
- `packages/shared/shared/redis_keys.py` — Added memory_short_key, escalation_status_key, pending_tool_confirm_key
- `packages/shared/pyproject.toml` — Added pgvector>=0.4.2
- `packages/orchestrator/pyproject.toml` — Added sentence-transformers>=3.0.0

**Commit:** `28a5ee9`

### Task 2: Wire memory into orchestrator pipeline

**Status:** Complete

**Files created:**
- `packages/orchestrator/orchestrator/memory/embedder.py` — Lazy singleton SentenceTransformer with embed_text() / embed_texts()

**Files modified:**
- `packages/orchestrator/orchestrator/agents/builder.py` — Added build_messages_with_memory()
- `packages/orchestrator/orchestrator/agents/runner.py` — Added optional messages parameter (backward compat)
- `packages/orchestrator/orchestrator/tasks.py` — Wired memory pipeline + new embed_and_store Celery task

**Commit:** `45b9573`

## Memory Pipeline Flow

```
handle_message (Celery task)
    │
    ├── BEFORE LLM CALL:
    │     1. get_recent_messages(redis)         → last 20 turns
    │     2. embed_text(user_text)              → 384-dim query vector
    │     3. retrieve_relevant(session, ...)    → top-3 past exchanges (cosine >= 0.75)
    │     4. build_messages_with_memory(...)    → enriched messages array
    │
    ├── run_agent(msg, agent, messages=enriched_messages)
    │     └── POST /complete to llm-pool
    │
    └── AFTER LLM RESPONSE:
          5. append_message(redis, user_turn)   → update sliding window
          6. append_message(redis, asst_turn)   → update sliding window
          7. embed_and_store.delay(...)          → fire-and-forget pgvector backfill
```

## Security Properties

- **Cross-tenant isolation**: pgvector queries pre-filter by tenant_id BEFORE ANN operator; RLS enforces at DB level as secondary backstop
- **Cross-agent isolation**: agent_id pre-filter ensures agent A cannot recall agent B's memory
- **Cross-user isolation**: user_id pre-filter ensures user A cannot see user B's memory
- **Redis isolation**: memory_short_key format `{tenant_id}:memory:short:{agent_id}:{user_id}` ensures namespace separation

## Deviations from Plan

### Auto-fixed Issues

**1. [Rule 3 - Blocking] Updated Docker postgres image to support pgvector extension**

- **Found during:** Task 1 integration test run
- **Issue:** `postgres:16-alpine` does not include the pgvector extension. Migration `002_phase2_memory.py` fails with: `extension "vector" is not available — Could not open extension control file`
- **Fix:** Changed `docker-compose.yml` postgres image from `postgres:16-alpine` to `pgvector/pgvector:pg16`. Restarted container. Migration ran cleanly.
- **Files modified:** `docker-compose.yml`
- **Commit:** `28a5ee9` (included in Task 1 commit)

## Self-Check: PASSED

All 8 created files exist on disk. Both commits (28a5ee9, 45b9573) confirmed present in git log. 202 tests pass.