feat(02-01): add two-layer memory system — Redis sliding window + pgvector long-term

- ConversationEmbedding ORM model with Vector(384) column (pgvector)
- memory_short_key, escalation_status_key, pending_tool_confirm_key in redis_keys.py
- orchestrator/memory/short_term.py: RPUSH/LTRIM sliding window (get_recent_messages, append_message)
- orchestrator/memory/long_term.py: pgvector HNSW cosine search (retrieve_relevant, store_embedding)
- Migration 002: conversation_embeddings table, HNSW index, RLS with FORCE, SELECT/INSERT only
- 10 unit tests (fakeredis), 6 integration tests (pgvector) — all passing
- Auto-fix [Rule 3]: postgres image updated to pgvector/pgvector:pg16 (extension required)

packages/orchestrator/orchestrator/memory/__init__.py

@@ -0,0 +1,22 @@
+"""
+Konstruct Agent Memory Layer.
+
+Two-layer conversational memory system:
+
+1. Short-term (Redis sliding window):
+   - Stores the last N messages verbatim
+   - Zero latency — Redis is always available
+   - Provides immediate in-session context continuity
+   - See: short_term.py
+
+2. Long-term (pgvector HNSW similarity search):
+   - Stores all messages as semantic embeddings
+   - Retrieves top-K semantically relevant past exchanges
+   - Provides cross-session recall (user preferences, past issues, etc.)
+   - Embedding model: all-MiniLM-L6-v2 (384 dimensions)
+   - See: long_term.py
+
+Memory scoping: All operations are keyed by (tenant_id, agent_id, user_id).
+This ensures complete isolation — no cross-tenant, cross-agent, or cross-user
+contamination is possible.
+"""

packages/orchestrator/orchestrator/memory/long_term.py

@@ -0,0 +1,164 @@
+"""
+pgvector-backed long-term conversational memory.
+
+Stores conversation turns as 384-dimensional embeddings (all-MiniLM-L6-v2)
+and retrieves semantically relevant past exchanges using HNSW cosine similarity
+search.
+
+CRITICAL SECURITY CONSTRAINTS:
+1. ALL queries MUST pre-filter by (tenant_id, agent_id, user_id) BEFORE the
+   ANN operator. This prevents cross-tenant, cross-agent, or cross-user data
+   leakage even in the face of embedding collisions.
+2. Cosine similarity threshold filters out low-relevance results — only content
+   genuinely related to the query should be injected into the LLM prompt.
+3. RLS (Row Level Security) is the DB-level backstop — the application-level
+   filters above are the primary guard; RLS is the safety net.
+
+pgvector cosine operations:
+- <=> operator: cosine DISTANCE (0 = identical, 2 = opposite)
+- cosine similarity = 1 - cosine distance
+- A threshold of 0.75 means: only return results where
+  1 - (embedding <=> query) >= 0.75  →  distance <= 0.25
+"""
+
+from __future__ import annotations
+
+import logging
+import uuid
+
+from sqlalchemy import text
+from sqlalchemy.ext.asyncio import AsyncSession
+
+logger = logging.getLogger(__name__)
+
+
+async def retrieve_relevant(
+    session: AsyncSession,
+    tenant_id: uuid.UUID,
+    agent_id: uuid.UUID,
+    user_id: str,
+    query_embedding: list[float],
+    top_k: int = 3,
+    threshold: float = 0.75,
+) -> list[str]:
+    """
+    Retrieve semantically relevant past conversation content.
+
+    Performs an HNSW approximate nearest neighbor search scoped strictly to
+    (tenant_id, agent_id, user_id). Results below the cosine similarity
+    threshold are discarded.
+
+    IMPORTANT: The (tenant_id, agent_id, user_id) pre-filter is applied
+    BEFORE the ANN operator to guarantee isolation. This is not optional —
+    removing these WHERE clauses would allow cross-tenant data leakage.
+
+    Args:
+        session:         Async SQLAlchemy session (must have RLS configured).
+        tenant_id:       Tenant UUID — mandatory for isolation.
+        agent_id:        Agent UUID — mandatory for isolation.
+        user_id:         End-user identifier — mandatory for isolation.
+        query_embedding: 384-dimensional query vector (all-MiniLM-L6-v2).
+        top_k:           Maximum number of results to return. Default 3.
+        threshold:       Minimum cosine similarity (0.0–1.0). Default 0.75.
+                         Set lower for broader recall, higher for precision.
+
+    Returns:
+        List of content strings (original message text), most relevant first.
+        Returns empty list if no results meet the threshold.
+    """
+    # Convert embedding list to pgvector string format: '[0.1, 0.2, ...]'
+    vec_str = "[" + ",".join(str(float(v)) for v in query_embedding) + "]"
+
+    # CRITICAL: pre-filter by all three isolation columns BEFORE ANN search.
+    # The ORDER BY uses <=> (cosine distance) — lower is more similar.
+    # We convert to similarity (1 - distance) to apply the threshold filter.
+    stmt = text("""
+        SELECT content, 1 - (embedding <=> CAST(:query AS vector)) AS similarity
+        FROM conversation_embeddings
+        WHERE tenant_id = :tenant_id
+          AND agent_id = :agent_id
+          AND user_id = :user_id
+          AND 1 - (embedding <=> CAST(:query AS vector)) >= :threshold
+        ORDER BY embedding <=> CAST(:query AS vector)
+        LIMIT :top_k
+    """)
+
+    try:
+        result = await session.execute(
+            stmt,
+            {
+                "query": vec_str,
+                "tenant_id": str(tenant_id),
+                "agent_id": str(agent_id),
+                "user_id": user_id,
+                "threshold": threshold,
+                "top_k": top_k,
+            },
+        )
+        rows = result.fetchall()
+    except Exception:
+        logger.exception(
+            "pgvector retrieve_relevant failed for tenant=%s agent=%s user=%s",
+            tenant_id,
+            agent_id,
+            user_id,
+        )
+        return []
+
+    return [row.content for row in rows]
+
+
+async def store_embedding(
+    session: AsyncSession,
+    tenant_id: uuid.UUID,
+    agent_id: uuid.UUID,
+    user_id: str,
+    content: str,
+    role: str,
+    embedding: list[float],
+) -> None:
+    """
+    Store a conversation turn embedding in the database.
+
+    Inserts a new row into conversation_embeddings. Embeddings are immutable
+    once stored — there is no UPDATE path. This matches the audit-log-like
+    nature of conversation history.
+
+    Args:
+        session:   Async SQLAlchemy session (must have RLS configured).
+        tenant_id: Tenant UUID for isolation.
+        agent_id:  Agent UUID for isolation.
+        user_id:   End-user identifier for isolation.
+        content:   Original message text.
+        role:      "user" or "assistant".
+        embedding: 384-dimensional float list (all-MiniLM-L6-v2).
+    """
+    vec_str = "[" + ",".join(str(float(v)) for v in embedding) + "]"
+
+    stmt = text("""
+        INSERT INTO conversation_embeddings
+            (id, tenant_id, agent_id, user_id, content, role, embedding)
+        VALUES
+            (gen_random_uuid(), :tenant_id, :agent_id, :user_id, :content, :role, CAST(:embedding AS vector))
+    """)
+
+    try:
+        await session.execute(
+            stmt,
+            {
+                "tenant_id": str(tenant_id),
+                "agent_id": str(agent_id),
+                "user_id": user_id,
+                "content": content,
+                "role": role,
+                "embedding": vec_str,
+            },
+        )
+    except Exception:
+        logger.exception(
+            "pgvector store_embedding failed for tenant=%s agent=%s user=%s",
+            tenant_id,
+            agent_id,
+            user_id,
+        )
+        raise

packages/orchestrator/orchestrator/memory/short_term.py

@@ -0,0 +1,111 @@
+"""
+Redis sliding window for short-term conversational memory.
+
+Implements a RPUSH + LTRIM pattern:
+- RPUSH appends new messages to the right (tail) of the list
+- LTRIM trims the list to the last `window` entries
+- LRANGE retrieves all current entries
+
+This gives O(1) append + O(1) trim + O(N) read where N <= window size.
+
+Key format: {tenant_id}:memory:short:{agent_id}:{user_id}
+
+Messages are stored as JSON objects with "role" and "content" keys,
+matching the OpenAI chat messages format for direct injection into
+the LLM messages array.
+
+Design decisions:
+- No TTL: message retention is indefinite per user preference. If TTL-based
+  expiry is needed in the future, add it via a separate expiry policy.
+- No compression: messages are stored as plain JSON. At 20 messages * ~200
+  bytes average, storage per user/agent is ~4KB — negligible.
+- Parameterized window: callers control the window size, defaulting to 20.
+  This allows future policy changes without code modification.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+
+from shared.redis_keys import memory_short_key
+
+logger = logging.getLogger(__name__)
+
+
+async def get_recent_messages(
+    redis: object,
+    tenant_id: str,
+    agent_id: str,
+    user_id: str,
+    n: int = 20,
+) -> list[dict[str, str]]:
+    """
+    Retrieve the most recent N messages from the sliding window.
+
+    Returns messages in insertion order (oldest first) — this matches the
+    expected LLM message array format where conversation flows chronologically.
+
+    Args:
+        redis:     Redis async client (redis.asyncio.Redis or compatible).
+        tenant_id: Konstruct tenant identifier.
+        agent_id:  Agent UUID string.
+        user_id:   End-user identifier (channel-native).
+        n:         Maximum number of messages to retrieve. Default 20.
+                   Pass a larger value than the window size to get all messages.
+
+    Returns:
+        List of message dicts with "role" and "content" keys, oldest first.
+        Returns empty list if no messages exist for this key.
+    """
+    key = memory_short_key(tenant_id, agent_id, user_id)
+
+    # LRANGE -n -1 returns the last n items in insertion order
+    raw_messages = await redis.lrange(key, -n, -1)  # type: ignore[union-attr]
+
+    messages: list[dict[str, str]] = []
+    for raw in raw_messages:
+        try:
+            msg = json.loads(raw)
+            messages.append({"role": str(msg["role"]), "content": str(msg["content"])})
+        except (json.JSONDecodeError, KeyError):
+            logger.warning("Malformed message in Redis key %s — skipping", key)
+
+    return messages
+
+
+async def append_message(
+    redis: object,
+    tenant_id: str,
+    agent_id: str,
+    user_id: str,
+    role: str,
+    content: str,
+    window: int = 20,
+) -> None:
+    """
+    Append a message to the sliding window and trim to window size.
+
+    Uses a pipeline to make RPUSH + LTRIM atomic — no race condition
+    between append and trim even under concurrent writes.
+
+    Args:
+        redis:     Redis async client.
+        tenant_id: Konstruct tenant identifier.
+        agent_id:  Agent UUID string.
+        user_id:   End-user identifier (channel-native).
+        role:      Message role: "user" or "assistant".
+        content:   Message text content.
+        window:    Maximum number of messages to retain. Default 20.
+                   After this operation the list will contain at most `window`
+                   entries (the most recent ones).
+    """
+    key = memory_short_key(tenant_id, agent_id, user_id)
+    serialized = json.dumps({"role": role, "content": content})
+
+    # Pipeline ensures RPUSH + LTRIM are sent atomically
+    pipe = redis.pipeline()  # type: ignore[union-attr]
+    pipe.rpush(key, serialized)
+    # LTRIM to last `window` entries: keep index -(window) through -1
+    pipe.ltrim(key, -window, -1)
+    await pipe.execute()

packages/orchestrator/pyproject.toml

@@ -12,6 +12,7 @@ dependencies = [
     "fastapi[standard]>=0.115.0",
     "celery[redis]>=5.4.0",
     "httpx>=0.28.0",
+    "sentence-transformers>=3.0.0",
 ]
 
 [tool.uv.sources]