Spaces:
Running
Running
File size: 23,704 Bytes
17fba62 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 | #!/usr/bin/env python3
"""
虫群v7 — 记忆模型(Memory Model)
参数化个人记忆系统:将交互记录编码为可检索的结构
设计思路:
- 传统方案:对话存数据库,检索靠关键词匹配
- 虫群方案:记忆以向量索引+时序结构存储,检索更精准
- 未来方向:将记忆编码为模型参数(模型即数据库)
- 当前实现:轻量级向量索引 + 时序衰减 + 关键词增强
"""
import hashlib
import json
import logging
import math
import os
import re
import time
from datetime import datetime
from typing import Dict, List, Optional, Tuple
logger = logging.getLogger(__name__)
# ============================================================
# 记忆条目
# ============================================================
class MemoryEntry:
"""单条记忆"""
def __init__(self, user_message: str, ai_response: str,
intent: str = "", route: str = "",
timestamp: str = ""):
self.entry_id = hashlib.md5(
f"{user_message}{ai_response}{time.time()}".encode()
).hexdigest()[:12]
self.user_message = user_message
self.ai_response = ai_response
self.intent = intent
self.route = route
self.timestamp = timestamp or datetime.now().isoformat()
self.created_at = time.time()
# 关键词提取(简易版,替代分词)
self.keywords = self._extract_keywords(user_message)
# 访问计数(用于衰减/强化)
self.access_count = 0
self.last_access = self.created_at
def _extract_keywords(self, text: str) -> List[str]:
"""简易中文关键词提取:去停用词 + 长词优先"""
# 去标点
cleaned = re.sub(r'[,。!?、;:""''()\s]', ' ', text)
# 按空格和常见分隔符切分
words = re.split(r'[\s,.\-!?;:]+', cleaned)
# 过滤短词
keywords = [w for w in words if len(w) >= 2]
return keywords[:20] # 最多20个
def to_dict(self) -> Dict:
return {
"entry_id": self.entry_id,
"user_message": self.user_message[:100],
"ai_response": self.ai_response[:100],
"intent": self.intent,
"keywords": self.keywords[:5],
"timestamp": self.timestamp,
"access_count": self.access_count,
}
# ============================================================
# 记忆模型核心
# ============================================================
class MemoryModel:
"""
记忆模型 — 参数化个人记忆
特性:
1. retrieve_context(): 获取上下文摘要(给元模型用)
2. retrieve(): 精确检索记忆条目
3. store_interaction(): 存储新交互
4. 时间衰减: 越早的记忆权重越低
5. 关键词增强: 匹配关键词越多权重越高
"""
# 记忆存储路径
DATA_DIR = "/home/admin/swarm/data/memory"
MAX_ENTRIES = 5000 # 最大记忆条数
DECAY_HALF_LIFE = 86400 * 7 # 衰减半衰期: 7天
def __init__(self, user_id: str = "default"):
self.user_id = user_id
self._entries: List[MemoryEntry] = []
self._keyword_index: Dict[str, List[str]] = {} # keyword -> [entry_ids]
self._intent_index: Dict[str, List[str]] = {} # intent -> [entry_ids]
self._loaded = False
# 统计
self._store_count = 0
self._retrieve_count = 0
self._hit_count = 0
# 延迟加载
self._load()
# ============================================================
# 核心接口
# ============================================================
def retrieve_context(self, message: str) -> str:
"""
获取与当前消息相关的记忆上下文文本
返回格式: "之前你问过XXX,我回答了YYY"
"""
self._retrieve_count += 1
entries = self.retrieve(message, top_k=3)
if not entries:
return ""
self._hit_count += 1
parts = []
for e in entries:
msg = e.get("user_message", "") if isinstance(e, dict) else e.user_message
resp = e.get("ai_response", "") if isinstance(e, dict) else e.ai_response
parts.append(f"之前你问过「{msg[:50]}」,回答是「{resp[:50]}」")
return ";".join(parts)
def retrieve(self, query: str, top_k: int = 5) -> List[Dict]:
"""
检索相关记忆条目
算法: 关键词匹配 + 时间衰减 + 访问频次
"""
self._retrieve_count += 1
if not self._entries:
return []
# 提取查询关键词
query_words = set()
cleaned = re.sub(r'[,。!?、;:""''()\s]', ' ', query)
for w in re.split(r'[\s,.\-!?;:]+', cleaned):
if len(w) >= 2:
query_words.add(w)
# 计算每条记忆的得分
scored: List[Tuple[float, MemoryEntry]] = []
now = time.time()
for entry in self._entries:
score = 0.0
# 关键词匹配分 (0~0.5)
overlap = query_words & set(entry.keywords)
if overlap:
match_ratio = len(overlap) / max(len(query_words), 1)
score += match_ratio * 0.5
# 时间衰减分 (0~0.3): 越新越高
age_seconds = now - entry.created_at
decay = math.exp(-0.693 * age_seconds / self.DECAY_HALF_LIFE)
score += decay * 0.3
# 访问频次分 (0~0.2): 常访问的记忆更重要
freq_score = min(entry.access_count / 10.0, 1.0)
score += freq_score * 0.2
if score > 0.05: # 最低阈值
scored.append((score, entry))
# 排序取top_k
scored.sort(key=lambda x: x[0], reverse=True)
results = []
for score, entry in scored[:top_k]:
entry.access_count += 1
entry.last_access = now
d = entry.to_dict()
d["_score"] = round(score, 3)
results.append(d)
if results:
self._hit_count += 1
return results
def store_interaction(self, message: str, result: str,
analysis=None) -> str:
"""
存储一条交互记忆
Args:
message: 用户消息
result: AI回复
analysis: TaskAnalysis对象(可选)
Returns:
entry_id
"""
self._store_count += 1
intent = analysis.intent if analysis else ""
route = analysis.route if analysis else ""
entry = MemoryEntry(
user_message=message,
ai_response=result,
intent=intent,
route=route,
)
self._entries.append(entry)
# 更新关键词索引
for kw in entry.keywords:
if kw not in self._keyword_index:
self._keyword_index[kw] = []
self._keyword_index[kw].append(entry.entry_id)
# 更新意图索引
if intent:
if intent not in self._intent_index:
self._intent_index[intent] = []
self._intent_index[intent].append(entry.entry_id)
# 超过上限时淘汰最旧且最少访问的
if len(self._entries) > self.MAX_ENTRIES:
self._evict()
# 定期持久化(每10次存储)
if self._store_count % 10 == 0:
self._save()
return entry.entry_id
# ============================================================
# 淘汰策略
# ============================================================
def _evict(self, count: int = 100):
"""淘汰最旧且最少访问的记忆"""
now = time.time()
def evict_score(entry: MemoryEntry) -> float:
"""得分越低越先淘汰"""
age_days = (now - entry.created_at) / 86400
freq = entry.access_count
# 旧 + 低频 = 低分
return freq / (1 + age_days)
self._entries.sort(key=evict_score, reverse=True)
removed = self._entries[self.MAX_ENTRIES - count:]
self._entries = self._entries[:self.MAX_ENTRIES - count]
# 重建索引
self._rebuild_index()
logger.debug(f"记忆淘汰: 移除{len(removed)}条")
def _rebuild_index(self):
"""重建关键词索引"""
self._keyword_index.clear()
self._intent_index.clear()
for entry in self._entries:
for kw in entry.keywords:
if kw not in self._keyword_index:
self._keyword_index[kw] = []
self._keyword_index[kw].append(entry.entry_id)
if entry.intent:
if entry.intent not in self._intent_index:
self._intent_index[entry.intent] = []
self._intent_index[entry.intent].append(entry.entry_id)
# ============================================================
# 持久化
# ============================================================
def _load(self):
"""从磁盘加载记忆"""
if self._loaded:
return
filepath = os.path.join(self.DATA_DIR, f"{self.user_id}.json")
if not os.path.exists(filepath):
self._loaded = True
return
try:
with open(filepath, "r", encoding="utf-8") as f:
data = json.load(f)
for item in data.get("entries", []):
entry = MemoryEntry(
user_message=item.get("user_message", ""),
ai_response=item.get("ai_response", ""),
intent=item.get("intent", ""),
route=item.get("route", ""),
timestamp=item.get("timestamp", ""),
)
entry.access_count = item.get("access_count", 0)
entry.created_at = item.get("created_at", time.time())
self._entries.append(entry)
self._rebuild_index()
logger.info(f"记忆加载: {len(self._entries)}条 (用户: {self.user_id})")
except Exception as e:
logger.warning(f"记忆加载失败: {e}")
self._loaded = True
def _save(self):
"""持久化记忆到磁盘"""
os.makedirs(self.DATA_DIR, exist_ok=True)
filepath = os.path.join(self.DATA_DIR, f"{self.user_id}.json")
try:
data = {
"user_id": self.user_id,
"version": 1,
"entries": [
{
"entry_id": e.entry_id,
"user_message": e.user_message,
"ai_response": e.ai_response,
"intent": e.intent,
"route": e.route,
"timestamp": e.timestamp,
"created_at": e.created_at,
"access_count": e.access_count,
"keywords": e.keywords,
}
for e in self._entries[-self.MAX_ENTRIES:]
]
}
with open(filepath, "w", encoding="utf-8") as f:
json.dump(data, f, ensure_ascii=False, indent=2)
logger.debug(f"记忆保存: {len(self._entries)}条")
except Exception as e:
logger.warning(f"记忆保存失败: {e}")
# ============================================================
# 统计
# ============================================================
def get_stats(self) -> Dict:
"""获取记忆统计"""
return {
"user_id": self.user_id,
"total_entries": len(self._entries),
"keyword_index_size": len(self._keyword_index),
"intent_index_size": len(self._intent_index),
"store_count": self._store_count,
"retrieve_count": self._retrieve_count,
"hit_count": self._hit_count,
"hit_rate": round(
self._hit_count / max(self._retrieve_count, 1), 3
),
}
# ============================================================
# v7.1增强: 实时训练 + 记忆矩阵 + 超长对话
# ============================================================
def encode_realtime(self, message: str, result: str, analysis=None):
"""
实时训练: 每次交互都更新记忆参数
实现策略:
- 维护一个微型训练缓冲区(最近N条交互)
- 当缓冲区满时触发微调(模拟)
- 关键: 高频记忆权重增大,低频记忆权重衰减
"""
# 先正常存储
entry_id = self.store_interaction(message, result, analysis)
# 更新记忆强度参数
self._update_memory_params(message, result)
# 检查是否需要触发微调
if self._store_count % 20 == 0:
self._micro_finetune()
return entry_id
def _update_memory_params(self, message: str, result: str):
"""
更新记忆参数(模拟参数化存储)
思路: 不是真的训练模型,而是维护一个"参数字典"
key=概念, value=强度权重
频繁出现的概念权重增大 → 类似TF-IDF的反向操作
"""
if not hasattr(self, '_param_dict'):
self._param_dict: Dict[str, float] = {}
# 提取概念(用关键词替代)
concepts = self._extract_concepts(message + " " + result)
for concept in concepts:
if concept in self._param_dict:
# 已有概念: 增强权重(但不超过上限)
self._param_dict[concept] = min(
self._param_dict[concept] * 1.1, 2.0
)
else:
# 新概念: 初始权重
self._param_dict[concept] = 1.0
# 所有概念轻微衰减(防止无限增长)
for k in list(self._param_dict.keys()):
self._param_dict[k] *= 0.999
if self._param_dict[k] < 0.1:
del self._param_dict[k]
def _extract_concepts(self, text: str) -> List[str]:
"""提取概念(复用关键词提取)"""
cleaned = re.sub(r'[,。!?、;:""''()\s]', ' ', text)
words = re.split(r'[\s,.\-!?;:]+', cleaned)
return [w for w in words if len(w) >= 2][:10]
def _micro_finetune(self):
"""
微调(模拟): 基于近期交互微调记忆权重
真实实现: 用缓冲区数据做1-2步梯度下降
当前实现: 强化近期高频记忆、衰减低频记忆
"""
if not hasattr(self, '_param_dict') or not self._param_dict:
return
# 按权重排序,保留top 500概念
sorted_params = sorted(
self._param_dict.items(), key=lambda x: x[1], reverse=True
)
self._param_dict = dict(sorted_params[:500])
logger.debug(f"记忆微调: 保留{len(self._param_dict)}个概念参数")
def fork_for_task(self, task_name: str) -> 'MemoryModel':
"""
为特定任务复制一个专属记忆模型
用法: 物理课备课时fork一个physics记忆模型
课上只用物理相关记忆,课后合并回主记忆
Returns:
新的MemoryModel实例(副本)
"""
# 创建副本
forked = MemoryModel.__new__(MemoryModel)
forked.user_id = f"{self.user_id}__{task_name}"
forked._entries = [] # 不复制全部,只建空壳
forked._keyword_index = {}
forked._intent_index = {}
forked._loaded = True
forked._store_count = 0
forked._retrieve_count = 0
forked._hit_count = 0
# 复制相关记忆(按意图过滤)
task_keywords = self._extract_concepts(task_name)
for entry in self._entries:
overlap = set(entry.keywords) & set(task_keywords)
if overlap:
forked._entries.append(entry)
for kw in entry.keywords:
if kw not in forked._keyword_index:
forked._keyword_index[kw] = []
forked._keyword_index[kw].append(entry.entry_id)
# 复制参数字典
if hasattr(self, '_param_dict'):
forked._param_dict = {
k: v for k, v in self._param_dict.items()
if k in task_keywords or v > 1.0
}
else:
forked._param_dict = {}
logger.info(f"记忆分叉: {task_name}, 复制{len(forked._entries)}条相关记忆")
return forked
def merge_from(self, other: 'MemoryModel'):
"""
合并另一个记忆模型(从分叉回归)
用法: 课上积累的物理记忆,课后合并回主记忆
"""
merged = 0
for entry in other._entries:
# 检查是否已存在
existing_ids = {e.entry_id for e in self._entries}
if entry.entry_id not in existing_ids:
self._entries.append(entry)
for kw in entry.keywords:
if kw not in self._keyword_index:
self._keyword_index[kw] = []
self._keyword_index[kw].append(entry.entry_id)
merged += 1
# 合并参数
if hasattr(other, '_param_dict') and hasattr(self, '_param_dict'):
for k, v in other._param_dict.items():
if k in self._param_dict:
self._param_dict[k] = max(self._param_dict[k], v)
else:
self._param_dict[k] = v
self._save()
logger.info(f"记忆合并: +{merged}条新记忆")
def get_unlimited_context(self, query: str, max_tokens: int = 4000) -> str:
"""
超长对话支持: 通过记忆压缩实现无限上下文
原理:
- 传统: 把所有历史拼上去(受token限制)
- 虫群: 只检索最相关的记忆+最近N轮(滑动窗口)
- 压缩: 旧记忆自动摘要,保持信息密度
"""
parts = []
char_count = 0
# 1. 最近N轮对话(滑动窗口,最近10条)
recent = self._entries[-10:]
for e in reversed(recent):
line = f"用户: {e.user_message[:100]}\n助手: {e.ai_response[:100]}"
if char_count + len(line) > max_tokens:
break
parts.insert(0, line)
char_count += len(line)
# 2. 相关历史记忆(按相关性)
relevant = self.retrieve(query, top_k=5)
seen_ids = {e.entry_id for e in recent}
for r in relevant:
rid = r.get("entry_id", "")
if rid in seen_ids:
continue
line = f"[历史] 用户: {r['user_message'][:80]}\n[历史] 助手: {r['ai_response'][:80]}"
if char_count + len(line) > max_tokens:
break
parts.insert(0, line)
char_count += len(line)
seen_ids.add(rid)
# 3. 概念参数摘要
if hasattr(self, '_param_dict') and self._param_dict:
top_concepts = sorted(
self._param_dict.items(), key=lambda x: x[1], reverse=True
)[:10]
concept_str = "核心概念: " + ", ".join(
f"{k}({v:.1f})" for k, v in top_concepts
)
parts.insert(0, concept_str)
return "\n---\n".join(parts)
def get_param_stats(self) -> Dict:
"""获取记忆参数统计"""
params = getattr(self, '_param_dict', {})
return {
"concept_count": len(params),
"top_concepts": sorted(
params.items(), key=lambda x: x[1], reverse=True
)[:5] if params else [],
"avg_weight": sum(params.values()) / max(len(params), 1) if params else 0,
}
# ============================================================
# 记忆矩阵: 多任务记忆的统一管理
# ============================================================
class MemoryMatrix:
"""
记忆矩阵 — 管理用户的所有任务记忆
结构:
主记忆(General) — 日常交互
├── 物理记忆(Physics) — 备课专用
├── 编程记忆(Coding) — 开发专用
└── ...按需创建
用法:
matrix = MemoryMatrix("user_001")
matrix.activate("physics") # 激活物理记忆
matrix.store("...", "...") # 存入当前激活记忆
matrix.deactivate() # 回到主记忆
"""
def __init__(self, user_id: str = "default"):
self.user_id = user_id
self._main = MemoryModel(user_id)
self._forks: Dict[str, MemoryModel] = {}
self._active: Optional[str] = None # 当前激活的任务名
def activate(self, task_name: str):
"""激活一个任务记忆(不存在则创建)"""
if task_name not in self._forks:
# 从主记忆分叉
self._forks[task_name] = self._main.fork_for_task(task_name)
logger.info(f"创建任务记忆: {task_name}")
self._active = task_name
def deactivate(self):
"""停用任务记忆,合并回主记忆"""
if self._active and self._active in self._forks:
self._main.merge_from(self._forks[self._active])
# 保留分叉不删除(下次可复用)
self._active = None
@property
def current(self) -> MemoryModel:
"""获取当前活跃的记忆模型"""
if self._active and self._active in self._forks:
return self._forks[self._active]
return self._main
def store(self, message: str, result: str, analysis=None):
"""存入当前活跃记忆"""
self.current.store_interaction(message, result, analysis)
def retrieve(self, query: str, top_k: int = 5) -> List[Dict]:
"""从当前活跃记忆检索"""
return self.current.retrieve(query, top_k)
def get_context(self, query: str) -> str:
"""获取上下文(优先当前任务+补充主记忆)"""
task_ctx = self.current.retrieve_context(query)
if self._active:
main_ctx = self._main.retrieve_context(query)
if task_ctx and main_ctx:
return f"[任务:{self._active}] {task_ctx}\n[通用] {main_ctx}"
return task_ctx
def list_tasks(self) -> List[str]:
"""列出所有任务记忆"""
return list(self._forks.keys())
def get_status(self) -> Dict:
"""获取矩阵状态"""
tasks = {}
for name, mem in self._forks.items():
tasks[name] = mem.get_stats()
return {
"user_id": self.user_id,
"active_task": self._active,
"main_stats": self._main.get_stats(),
"task_count": len(self._forks),
"tasks": tasks,
}
|