""" Auto Dream — 夜间记忆整合服务。 参考 Claude Code 的 Auto Dream 机制： - 每天凌晨 3:00 触发 - 扫描过去 24 小时的向量记忆 - 合并相似条目（余弦相似度 > 0.85） - 生成整合摘要并写入向量记忆池 """ from __future__ import annotations import asyncio import logging from datetime import datetime, timedelta, timezone from typing import Any, Dict, List, Optional, Tuple from sqlalchemy.orm import Session from app.core.database import SessionLocal from app.services.embedding_service import embedding_service logger = logging.getLogger(__name__) # 相似度阈值：高于此值视为可合并 MERGE_SIMILARITY_THRESHOLD = 0.85 # 每天凌晨 3:00 触发（UTC+8） DREAM_HOUR = 3 DREAM_MINUTE = 0 # 上次整合日期（模块级，进程重启后重置） _last_dream_date: Optional[str] = None _dream_lock = asyncio.Lock() def _should_dream_today() -> bool: """检查是否到了今天的整合时间且尚未执行。""" global _last_dream_date now = datetime.now(timezone.utc).astimezone( timezone(timedelta(hours=8)) ) today_str = now.strftime("%Y-%m-%d") # 已执行过 if _last_dream_date == today_str: return False # 在凌晨 3:00-4:00 之间触发 if now.hour != DREAM_HOUR: return False return True async def run_auto_dream() -> dict: """ 执行一次记忆整合。 返回: {"merged": int, "deleted": int, "dreams": int, "elapsed_s": float} """ import time start = time.time() if not _dream_lock.locked(): async with _dream_lock: return await _do_consolidate() logger.info("Auto Dream：上一次整合仍在进行中，跳过") return {"merged": 0, "deleted": 0, "dreams": 0, "elapsed_s": 0, "skipped": True} async def _do_consolidate() -> dict: """实际的整合逻辑。""" global _last_dream_date from app.models.agent_vector_memory import AgentVectorMemory now = datetime.now(timezone.utc).astimezone( timezone(timedelta(hours=8)) ) today_str = now.strftime("%Y-%m-%d") cutoff = now - timedelta(hours=24) db: Optional[Session] = None merged = 0 deleted = 0 dreams = 0 try: db = SessionLocal() # 1. 获取过去 24h 的所有向量记忆 rows = ( db.query(AgentVectorMemory) .filter(AgentVectorMemory.created_at >= cutoff) .order_by(AgentVectorMemory.created_at.desc()) .limit(200) .all() ) if len(rows) < 3: logger.info("Auto Dream：最近 24h 记忆不足（%d 条），跳过", len(rows)) _last_dream_date = today_str return {"merged": 0, "deleted": 0, "dreams": 0, "elapsed_s": 0} logger.info("Auto Dream：开始整合 %d 条最近记忆", len(rows)) # 2. 构建带 embedding 的条目列表 entries: List[Tuple[Any, List[float]]] = [] for row in rows: if not row.embedding: continue emb = embedding_service.deserialize_embedding(row.embedding) if emb and len(emb) > 0: entries.append((row, emb)) if len(entries) < 3: _last_dream_date = today_str return {"merged": 0, "deleted": 0, "dreams": 0, "elapsed_s": 0} # 3. 两两计算相似度，找出可合并的对 to_delete_ids: set = set() to_merge_pairs: List[Tuple[Any, Any]] = [] for i in range(len(entries)): if entries[i][0].id in to_delete_ids: continue for j in range(i + 1, len(entries)): if entries[j][0].id in to_delete_ids: continue sim = embedding_service.cosine_similarity( entries[i][1], entries[j][1] ) if sim >= MERGE_SIMILARITY_THRESHOLD: # 保留较新的，删除较旧的 newer = entries[i][0] if entries[i][0].created_at >= entries[j][0].created_at else entries[j][0] older = entries[j][0] if newer is entries[i][0] else entries[i][0] to_delete_ids.add(older.id) to_merge_pairs.append((newer, older)) # 4. 执行合并删除 if to_delete_ids: for row_id in to_delete_ids: try: db.query(AgentVectorMemory).filter( AgentVectorMemory.id == row_id ).delete() deleted += 1 except Exception: db.rollback() db.commit() logger.info("Auto Dream：合并删除 %d 条重复记忆", deleted) merged = len(to_merge_pairs) # 5. 生成整合摘要（Dream Summary） dream_text = await _generate_dream_summary(rows[:50]) if dream_text: from app.services.embedding_service import embedding_service as es try: emb = await es.generate_embedding(dream_text) embedding_json = es.serialize_embedding(emb) if emb else "" dream_record = AgentVectorMemory( scope_kind="system", scope_id="auto_dream", session_key=f"dream_{today_str}", content_text=dream_text[:2000], embedding=embedding_json or None, metadata_={ "type": "dream_summary", "memory_type": "project", "source": "auto_dream", "dream_date": today_str, "merged_count": merged, "deleted_count": deleted, }, ) db.add(dream_record) db.commit() dreams = 1 logger.info("Auto Dream：已生成整合摘要 (%d 字)", len(dream_text)) except Exception as e: logger.warning("Auto Dream 摘要写入失败: %s", e) except Exception as e: logger.error("Auto Dream 整合失败: %s", e) if db: db.rollback() finally: if db: db.close() _last_dream_date = today_str elapsed = time.time() - start logger.info("Auto Dream 完成: merged=%d deleted=%d dreams=%d elapsed=%.1fs", merged, deleted, dreams, elapsed) return {"merged": merged, "deleted": deleted, "dreams": dreams, "elapsed_s": elapsed} async def _generate_dream_summary(rows: list) -> str: """ 用 LLM 从最近记忆生成整合摘要。 返回空字符串表示失败（不阻塞主流程）。 """ if not rows or len(rows) < 3: return "" # 提取记忆内容 items = [] for r in rows[-30:]: # 最近 30 条 content = r.content_text[:300] if r.content_text else "" meta = r.metadata_ or {} mem_type = meta.get("memory_type", "unknown") items.append(f"[{mem_type}] {content}") if not items: return "" prompt = ( "你是一个记忆整合助手。请分析以下 24 小时内的 Agent 对话记忆，\n" "生成一份简洁的每日摘要（200字以内），包含：\n" "1. 用户讨论了哪些主要话题\n" "2. 用户表达了哪些偏好或需求\n" "3. 有哪些值得保留的关键信息\n\n" "记忆条目：\n" ) + "\n".join(f"- {item}" for item in items) try: from openai import AsyncOpenAI from app.core.config import settings api_key = settings.DEEPSEEK_API_KEY or "" base_url = settings.DEEPSEEK_BASE_URL or "https://api.deepseek.com" if not api_key: return "" client = AsyncOpenAI(api_key=api_key, base_url=base_url) resp = await client.chat.completions.create( model="deepseek-v4-flash", messages=[{"role": "user", "content": prompt}], temperature=0.3, max_tokens=600, timeout=30, ) return resp.choices[0].message.content or "" except Exception as e: logger.warning("Auto Dream 摘要生成失败: %s", e) return ""