Files
project-lyra/lyra/memory.py
Claude 0ee5a9ce47 feat: SQLite-backed memory with brute-force cosine recall
- lyra.memory.remember(session_id, role, content) embeds and stores
- lyra.memory.recent(session_id, n) returns the last N from a session
- lyra.memory.recall(query, k, session_id=None) returns top-k by cosine
  similarity across the chosen scope (all sessions by default)
- Embeddings live in the exchanges.embedding BLOB column as float32 bytes
- Connection reopens automatically if LYRA_DB_PATH changes (test-friendly)
2026-05-16 06:35:52 +00:00

134 lines
4.0 KiB
Python

"""Persistent memory: SQLite storage + brute-force cosine recall over embeddings.
Each exchange is stored with its OpenAI embedding as a float32 BLOB. Recall
loads all embeddings (optionally scoped to a session) into a matrix and
returns the top-k by cosine similarity. Brute force is fine up to tens of
thousands of rows; swap in a vector index when that stops being true.
"""
from __future__ import annotations
import sqlite3
from dataclasses import dataclass
from datetime import datetime, timezone
from pathlib import Path
import numpy as np
from lyra import llm
from lyra.config import load
SCHEMA = """
CREATE TABLE IF NOT EXISTS exchanges (
id INTEGER PRIMARY KEY AUTOINCREMENT,
session_id TEXT NOT NULL,
role TEXT NOT NULL,
content TEXT NOT NULL,
embedding BLOB NOT NULL,
created_at TEXT NOT NULL
);
CREATE INDEX IF NOT EXISTS idx_session_created ON exchanges(session_id, created_at);
"""
_conn: sqlite3.Connection | None = None
_conn_path: Path | None = None
def _connection() -> sqlite3.Connection:
"""Lazily open the SQLite connection. Reopens if LYRA_DB_PATH changed (for tests)."""
global _conn, _conn_path
cfg = load()
if _conn is None or _conn_path != cfg.db_path:
if _conn is not None:
_conn.close()
cfg.db_path.parent.mkdir(parents=True, exist_ok=True)
_conn = sqlite3.connect(cfg.db_path)
_conn.row_factory = sqlite3.Row
_conn.executescript(SCHEMA)
_conn_path = cfg.db_path
return _conn
@dataclass
class Exchange:
id: int
session_id: str
role: str
content: str
created_at: str
score: float | None = None
def _to_blob(vec: list[float]) -> bytes:
return np.asarray(vec, dtype=np.float32).tobytes()
def _from_blob(blob: bytes) -> np.ndarray:
return np.frombuffer(blob, dtype=np.float32)
def remember(session_id: str, role: str, content: str) -> int:
"""Embed and persist a single exchange. Returns the new row id."""
[embedding] = llm.embed([content])
now = datetime.now(timezone.utc).isoformat()
conn = _connection()
with conn:
cur = conn.execute(
"INSERT INTO exchanges (session_id, role, content, embedding, created_at) "
"VALUES (?, ?, ?, ?, ?)",
(session_id, role, content, _to_blob(embedding), now),
)
return int(cur.lastrowid)
def recent(session_id: str, n: int = 10) -> list[Exchange]:
"""Last `n` exchanges from a session, oldest first."""
conn = _connection()
rows = conn.execute(
"SELECT id, session_id, role, content, created_at FROM exchanges "
"WHERE session_id = ? ORDER BY id DESC LIMIT ?",
(session_id, n),
).fetchall()
return [
Exchange(
id=r["id"],
session_id=r["session_id"],
role=r["role"],
content=r["content"],
created_at=r["created_at"],
)
for r in reversed(rows)
]
def recall(query: str, k: int = 5, session_id: str | None = None) -> list[Exchange]:
"""Top-k exchanges semantically similar to `query`, optionally scoped to a session."""
[q_vec] = llm.embed([query])
q = np.asarray(q_vec, dtype=np.float32)
conn = _connection()
sql = "SELECT id, session_id, role, content, embedding, created_at FROM exchanges"
params: tuple = ()
if session_id is not None:
sql += " WHERE session_id = ?"
params = (session_id,)
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 [
Exchange(
id=rows[i]["id"],
session_id=rows[i]["session_id"],
role=rows[i]["role"],
content=rows[i]["content"],
created_at=rows[i]["created_at"],
score=float(scores[i]),
)
for i in top_idx
]