feat: tiered, compacting memory (phase 1.5)

Older sessions fade to a general idea; details stay retrievable.

- memory: summaries table (one compacted gist per session, embedded), plus
  store_summary/get_summary/recall_summaries and unsummarized_count (tracks
  exchanges newer than the current summary)
- lyra/summary.py: summarize_session compacts a session's raw turns into a
  third-person gist (default SUMMARY_BACKEND=local, so compaction is free);
  maybe_summarize re-summarizes once SUMMARIZE_AFTER new turns accumulate
- chat.build_messages now layers context in tiers: persona -> gists of other
  sessions -> a few sharp raw cross-session details -> current session raw
  turns -> new message; respond() compacts the session after each turn
- web: POST /sessions/{id}/summarize to compact on demand
- summarization activity surfaces in the live log

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

lyra/memory.py

@@ -33,6 +33,16 @@ CREATE TABLE IF NOT EXISTS sessions (
     name TEXT,
     created_at TEXT NOT NULL
 );
+
+-- One compacted "gist" per session. last_exchange_id marks how far the summary
+-- covers, so we know when enough new turns have accumulated to re-summarize.
+CREATE TABLE IF NOT EXISTS summaries (
+    session_id TEXT PRIMARY KEY,
+    content TEXT NOT NULL,
+    embedding BLOB NOT NULL,
+    last_exchange_id INTEGER NOT NULL,
+    created_at TEXT NOT NULL
+);
 """
 
 _conn: sqlite3.Connection | None = None
@@ -67,6 +77,15 @@ class Exchange:
     score: float | None = None
 
 
+@dataclass
+class Summary:
+    session_id: str
+    content: str
+    last_exchange_id: int
+    created_at: str
+    score: float | None = None
+
+
 def _to_blob(vec: list[float]) -> bytes:
     return np.asarray(vec, dtype=np.float32).tobytes()
 
@@ -171,6 +190,7 @@ def delete_session(session_id: str) -> None:
     with conn:
         conn.execute("DELETE FROM exchanges WHERE session_id = ?", (session_id,))
         conn.execute("DELETE FROM sessions WHERE id = ?", (session_id,))
+        conn.execute("DELETE FROM summaries WHERE session_id = ?", (session_id,))
 
 
 def recall(query: str, k: int = 5, session_id: str | None = None) -> list[Exchange]:
@@ -204,3 +224,83 @@ def recall(query: str, k: int = 5, session_id: str | None = None) -> list[Exchan
         )
         for i in top_idx
     ]
+
+
+# --- Summary tier (compacted per-session gists) ---
+
+
+def store_summary(session_id: str, content: str, last_exchange_id: int) -> None:
+    """Embed and persist the gist of a session, replacing any prior summary."""
+    [embedding] = llm.embed([content])
+    now = datetime.now(timezone.utc).isoformat()
+    conn = _connection()
+    with conn:
+        conn.execute(
+            "INSERT INTO summaries (session_id, content, embedding, last_exchange_id, created_at) "
+            "VALUES (?, ?, ?, ?, ?) "
+            "ON CONFLICT(session_id) DO UPDATE SET "
+            "content=excluded.content, embedding=excluded.embedding, "
+            "last_exchange_id=excluded.last_exchange_id, created_at=excluded.created_at",
+            (session_id, content, _to_blob(embedding), last_exchange_id, now),
+        )
+
+
+def get_summary(session_id: str) -> Summary | None:
+    conn = _connection()
+    r = conn.execute(
+        "SELECT session_id, content, last_exchange_id, created_at FROM summaries "
+        "WHERE session_id = ?",
+        (session_id,),
+    ).fetchone()
+    if r is None:
+        return None
+    return Summary(
+        session_id=r["session_id"],
+        content=r["content"],
+        last_exchange_id=r["last_exchange_id"],
+        created_at=r["created_at"],
+    )
+
+
+def unsummarized_count(session_id: str) -> int:
+    """How many exchanges in this session are newer than its current summary."""
+    conn = _connection()
+    summary = get_summary(session_id)
+    cutoff = summary.last_exchange_id if summary else 0
+    r = conn.execute(
+        "SELECT COUNT(*) AS n FROM exchanges WHERE session_id = ? AND id > ?",
+        (session_id, cutoff),
+    ).fetchone()
+    return int(r["n"])
+
+
+def recall_summaries(query: str, k: int = 3, exclude_session: str | None = None) -> list[Summary]:
+    """Top-k session summaries most similar to `query` (the long-term gist tier)."""
+    [q_vec] = llm.embed([query])
+    q = np.asarray(q_vec, dtype=np.float32)
+
+    conn = _connection()
+    sql = "SELECT session_id, content, embedding, last_exchange_id, created_at FROM summaries"
+    params: tuple = ()
+    if exclude_session is not None:
+        sql += " WHERE session_id != ?"
+        params = (exclude_session,)
+    rows = conn.execute(sql, params).fetchall()
+    if not rows:
+        return []
+
+    matrix = np.stack([_from_blob(r["embedding"]) for r in rows])
+    norms = np.linalg.norm(matrix, axis=1)
+    scores = (matrix @ q) / (norms * np.linalg.norm(q) + 1e-9)
+
+    top_idx = np.argsort(scores)[::-1][:k]
+    return [
+        Summary(
+            session_id=rows[i]["session_id"],
+            content=rows[i]["content"],
+            last_exchange_id=rows[i]["last_exchange_id"],
+            created_at=rows[i]["created_at"],
+            score=float(scores[i]),
+        )
+        for i in top_idx
+    ]