File size: 12,079 Bytes
0e3d4b8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 | """Persistent Memory — long-term memory that survives restarts.
Three layers:
- Working memory: current conversation context (in-memory, per session)
- Episodic memory: past conversations and events (SQLite, persistent)
- Semantic memory: extracted facts and knowledge (SQLite, persistent)
All layers are persisted to SQLite so the LLM remembers everything
across restarts. Semantic memory is auto-extracted from conversations.
Linked to the recursive link graph for knowledge graph traversal.
"""
from __future__ import annotations
import hashlib
import json
import logging
import os
import sqlite3
import time
from dataclasses import dataclass, field
from typing import Any
logger = logging.getLogger(__name__)
@dataclass
class EpisodicMemory:
"""A single episodic memory (event/conversation)."""
id: str
session_id: str
role: str # "user", "assistant", "system", "event"
content: str
channel: str = "cli"
timestamp: float = field(default_factory=time.time)
importance: float = 0.5 # 0-1, higher = more important
tags: list[str] = field(default_factory=list)
@dataclass
class SemanticMemory:
"""A extracted fact or piece of knowledge."""
id: str
fact: str
source: str = "" # what conversation/event it came from
confidence: float = 0.5
timestamp: float = field(default_factory=time.time)
access_count: int = 0
last_accessed: float = field(default_factory=time.time)
tags: list[str] = field(default_factory=list)
class PersistentMemory:
"""Persistent long-term memory backed by SQLite.
Survives restarts. Three layers:
- Working: current session context (in-memory)
- Episodic: past conversations (SQLite)
- Semantic: extracted facts (SQLite)
Auto-extracts semantic memories from conversations.
Provides context injection for inference.
"""
IMPORTANCE_DECAY = 0.0001 # per second
MAX_WORKING_MEMORY = 20 # max items in working memory
def __init__(self, db_path: str) -> None:
self.db_path = db_path
os.makedirs(os.path.dirname(db_path) or ".", exist_ok=True)
self._working: list[dict[str, Any]] = []
self._session_id: str = ""
self._stats = {
"episodic_stored": 0,
"semantic_extracted": 0,
"memories_recalled": 0,
"context_injections": 0,
}
self._init_db()
self._load_stats()
def _init_db(self) -> None:
"""Initialize SQLite tables."""
with sqlite3.connect(self.db_path) as conn:
conn.executescript("""
CREATE TABLE IF NOT EXISTS episodic (
id TEXT PRIMARY KEY,
session_id TEXT,
role TEXT,
content TEXT,
channel TEXT,
timestamp REAL,
importance REAL,
tags TEXT
);
CREATE INDEX IF NOT EXISTS idx_episodic_session ON episodic(session_id);
CREATE INDEX IF NOT EXISTS idx_episodic_importance ON episodic(importance);
CREATE INDEX IF NOT EXISTS idx_episodic_timestamp ON episodic(timestamp);
CREATE TABLE IF NOT EXISTS semantic (
id TEXT PRIMARY KEY,
fact TEXT,
source TEXT,
confidence REAL,
timestamp REAL,
access_count INTEGER DEFAULT 0,
last_accessed REAL,
tags TEXT
);
CREATE INDEX IF NOT EXISTS idx_semantic_confidence ON semantic(confidence);
CREATE INDEX IF NOT EXISTS idx_semantic_tags ON semantic(tags);
""")
def _load_stats(self) -> None:
"""Load counts from DB."""
with sqlite3.connect(self.db_path) as conn:
self._stats["episodic_stored"] = conn.execute("SELECT COUNT(*) FROM episodic").fetchone()[0]
self._stats["semantic_extracted"] = conn.execute("SELECT COUNT(*) FROM semantic").fetchone()[0]
def set_session(self, session_id: str) -> None:
"""Set the current session ID."""
self._session_id = session_id
self._working.clear()
def add_episodic(self, role: str, content: str, channel: str = "cli",
importance: float = 0.5, tags: list[str] | None = None) -> str:
"""Store an episodic memory (conversation turn or event)."""
mem_id = hashlib.sha256(f"{role}:{content}:{time.time()}".encode()).hexdigest()[:16]
mem = EpisodicMemory(
id=mem_id, session_id=self._session_id, role=role,
content=content, channel=channel, importance=importance,
tags=tags or [],
)
with sqlite3.connect(self.db_path) as conn:
conn.execute(
"INSERT OR REPLACE INTO episodic VALUES (?,?,?,?,?,?,?,?)",
(mem.id, mem.session_id, mem.role, mem.content, mem.channel,
mem.timestamp, mem.importance, json.dumps(mem.tags))
)
# Also add to working memory
self._working.append({"role": role, "content": content, "timestamp": time.time()})
if len(self._working) > self.MAX_WORKING_MEMORY:
self._working = self._working[-self.MAX_WORKING_MEMORY:]
self._stats["episodic_stored"] += 1
return mem_id
def add_semantic(self, fact: str, source: str = "", confidence: float = 0.5,
tags: list[str] | None = None) -> str:
"""Store a semantic memory (extracted fact)."""
fact_id = hashlib.sha256(f"{fact}:{time.time()}".encode()).hexdigest()[:16]
mem = SemanticMemory(
id=fact_id, fact=fact, source=source, confidence=confidence,
tags=tags or [],
)
with sqlite3.connect(self.db_path) as conn:
conn.execute(
"INSERT OR REPLACE INTO semantic VALUES (?,?,?,?,?,?,?,?)",
(mem.id, mem.fact, mem.source, mem.confidence, mem.timestamp,
mem.access_count, mem.last_accessed, json.dumps(mem.tags))
)
self._stats["semantic_extracted"] += 1
return fact_id
def recall_episodic(self, query: str, max_results: int = 5) -> list[EpisodicMemory]:
"""Recall episodic memories related to a query."""
# Simple keyword search
keywords = query.lower().split()
with sqlite3.connect(self.db_path) as conn:
rows = conn.execute(
"SELECT * FROM episodic ORDER BY importance DESC, timestamp DESC LIMIT ?",
(max_results * 3,)
).fetchall()
results = []
for row in rows:
mem = self._row_to_episodic(row)
# Score by keyword overlap
content_lower = mem.content.lower()
score = sum(1 for kw in keywords if kw in content_lower)
if score > 0:
# Apply time decay
age = time.time() - mem.timestamp
mem.importance = max(0.01, mem.importance - age * self.IMPORTANCE_DECAY)
results.append((score + mem.importance, mem))
results.sort(key=lambda x: -x[0])
self._stats["memories_recalled"] += len(results[:max_results])
return [mem for _, mem in results[:max_results]]
def recall_semantic(self, query: str, max_results: int = 5) -> list[SemanticMemory]:
"""Recall semantic memories (facts) related to a query."""
keywords = query.lower().split()
with sqlite3.connect(self.db_path) as conn:
rows = conn.execute(
"SELECT * FROM semantic ORDER BY confidence DESC, last_accessed DESC LIMIT ?",
(max_results * 3,)
).fetchall()
results = []
for row in rows:
mem = self._row_to_semantic(row)
fact_lower = mem.fact.lower()
score = sum(1 for kw in keywords if kw in fact_lower)
if score > 0:
results.append((score + mem.confidence, mem))
# Update access count
with sqlite3.connect(self.db_path) as conn:
conn.execute(
"UPDATE semantic SET access_count = access_count + 1, last_accessed = ? WHERE id = ?",
(time.time(), mem.id)
)
results.sort(key=lambda x: -x[0])
return [mem for _, mem in results[:max_results]]
def extract_semantic(self, user_message: str, assistant_response: str) -> list[str]:
"""Auto-extract semantic memories (facts) from a conversation.
Simple extraction: look for statements that contain facts.
"""
facts: list[str] = []
# Simple heuristics for fact extraction
sentences = assistant_response.replace("!", ".").replace("?", ".").split(".")
for sentence in sentences:
s = sentence.strip()
if len(s) < 10 or len(s) > 200:
continue
# Skip questions and commands
if s.endswith("?") or s.startswith("You ") or s.startswith("I "):
continue
# Look for factual statements (contains "is", "are", "was", "has", etc.)
fact_indicators = [" is ", " are ", " was ", " has ", " have ", " can ", " cannot ",
" means ", " refers to ", " defined as ", " consists of "]
if any(ind in s.lower() for ind in fact_indicators):
fact_id = self.add_semantic(s, source=self._session_id, confidence=0.6)
facts.append(fact_id)
return facts
def get_context(self, query: str, max_episodic: int = 3, max_semantic: int = 3) -> str:
"""Get memory context to inject into the prompt for inference."""
parts: list[str] = []
# Working memory (current session)
if self._working:
recent = self._working[-5:]
working_text = " | ".join(f"{m['role']}: {m['content'][:80]}" for m in recent)
parts.append(f"Recent: {working_text}")
# Episodic memory
episodic = self.recall_episodic(query, max_results=max_episodic)
if episodic:
ep_text = " | ".join(f"{m.role}: {m.content[:80]}" for m in episodic)
parts.append(f"Past: {ep_text}")
# Semantic memory
semantic = self.recall_semantic(query, max_results=max_semantic)
if semantic:
sem_text = " | ".join(m.fact[:80] for m in semantic)
parts.append(f"Facts: {sem_text}")
if parts:
self._stats["context_injections"] += 1
return " | ".join(parts)
def get_working_memory(self) -> list[dict[str, Any]]:
"""Get current working memory (this session)."""
return self._working.copy()
def clear_working(self) -> None:
"""Clear working memory."""
self._working.clear()
def _row_to_episodic(self, row: tuple) -> EpisodicMemory:
return EpisodicMemory(
id=row[0], session_id=row[1], role=row[2], content=row[3],
channel=row[4], timestamp=row[5], importance=row[6],
tags=json.loads(row[7]) if row[7] else [],
)
def _row_to_semantic(self, row: tuple) -> SemanticMemory:
return SemanticMemory(
id=row[0], fact=row[1], source=row[2], confidence=row[3],
timestamp=row[4], access_count=row[5], last_accessed=row[6],
tags=json.loads(row[7]) if row[7] else [],
)
def get_stats(self) -> dict[str, Any]:
with sqlite3.connect(self.db_path) as conn:
ep_count = conn.execute("SELECT COUNT(*) FROM episodic").fetchone()[0]
sem_count = conn.execute("SELECT COUNT(*) FROM semantic").fetchone()[0]
return {
**self._stats,
"episodic_total": ep_count,
"semantic_total": sem_count,
"working_size": len(self._working),
"session_id": self._session_id,
}
|