{"type":"doc","content":[{"type":"heading","attrs":{"id":"a3ce2d32-29d4-4396-b2d6-60d2b9937c4f","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"新闻导语"}]},{"type":"paragraph","attrs":{"id":"4b8d65fc-982a-4a20-900b-e65083dd5a87","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"2026年8月，AI Agent正从“无状态应答”迈向“有记忆协作”，但“记忆失控”成为长程任务可靠性的致命短板。DeepMind研究显示，78%的Agent在多轮交互后出现上下文污染或关键信息遗忘，导致决策漂移与幻觉激增。行业共识转向：Agent记忆不能靠窗口堆砌，而需分层治理——工作记忆精准可控、长期记忆可检索可验证、过期记忆主动遗忘。可管理、可追溯、可衰减的记忆架构，已成为智能体胜任复杂任务的认知基石。"}]},{"type":"heading","attrs":{"id":"de38b377-6059-412b-9120-096b3a3bbd06","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"一、痛点剖析：为什么你的Agent总是“记不住、想不起、忘不掉”？"}]},{"type":"heading","attrs":{"id":"e44ace68-6bba-4e8e-aabf-7b4f44889aeb","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"1. “上下文爆炸”：塞得越多，错得越离谱"}]},{"type":"paragraph","attrs":{"id":"1064ae3e-ede8-46e1-a5b5-4cf27730b97b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"现象"},{"type":"text","text":" ：为让Agent“记住一切”，将过去50轮对话全塞入Prompt，结果模型注意力分散，对最新指令响应迟钝；历史消息中夹杂过时的工具返回和中间推理，Agent误将其当作当前事实；Token消耗飙升，单次调用成本超预算3倍，延迟从2秒涨至15秒。"}]},{"type":"paragraph","attrs":{"id":"8a95ff1f-ae30-4487-8bd8-0b5ec6d4191f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"根因"},{"type":"text","text":" ：缺乏工作记忆的动态压缩与优先级管理。所有历史信息平等对待，无重要性分级；缺少摘要与提炼机制，原始消息无限累积；未区分“事实”“假设”“待验证”等语义类型，噪声与信号混杂。"}]},{"type":"heading","attrs":{"id":"5ea1fb49-f2f7-49a4-a247-f2699866e033","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"2. “检索失准”：找到了，但不是你要的"}]},{"type":"paragraph","attrs":{"id":"61c0efa9-e957-421f-9018-514f66e60388","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"现象"},{"type":"text","text":" ：用户问“上周的项目进度”，Agent检索到三个月前的同名项目文档；向量相似度高的内容被优先召回，但与当前问题语义无关；检索结果未经时效性与权威性校验，Agent引用了已废弃的流程规范。"}]},{"type":"paragraph","attrs":{"id":"f183189b-8acb-4d6c-aca9-d3168f284ffd","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"根因"},{"type":"text","text":" ：检索策略单一且缺乏上下文感知。仅依赖向量相似度，忽略时间、来源、置信度等元数据；查询未结合当前对话意图重写，原始query歧义大；召回结果无重排序与验证环节，Top-K直接喂给模型。"}]},{"type":"heading","attrs":{"id":"f2a7c83a-c3c1-47c4-b4d6-de2b85b000b5","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"3. “遗忘缺失”：该忘的忘不了，该更新的没更新"}]},{"type":"paragraph","attrs":{"id":"a2d8ad30-b44e-4cf8-be45-27e6b0bd0834","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"现象"},{"type":"text","text":" ：用户明确说“之前说的地址作废”，但Agent后续仍使用旧地址生成合同；政策变更后，Agent继续引用旧版条款回答合规问题；临时调试信息残留在长期记忆中，污染正式业务推理。"}]},{"type":"paragraph","attrs":{"id":"63d68723-c3a4-48b6-a791-a48ee4287980","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"根因"},{"type":"text","text":" ：缺乏主动遗忘与记忆更新机制。记忆只有“写入”没有“删除/覆盖”；无版本管理与有效期标记；用户显式纠正未被转化为记忆修正操作。"}]},{"type":"heading","attrs":{"id":"9530043c-0d86-45d6-bc25-189f5f16100a","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"二、技术解密：2026 Agent记忆三层治理架构"}]},{"type":"codeBlock","attrs":{"id":"553f8983-ebc3-473f-9029-75c6a2772ea8","language":"javascript","theme":"atom-one-dark","runtimes":0,"isHoverDragHandle":false,"key":"","languageByAi":"javascript"},"content":[{"type":"text","text":"┌─────────────────────────────────────────────────────────────────────┐n│ 2026 Agent Memory Governance Architecture │n├─────────────────────────────────────────────────────────────────────┤n│[Agent Runtime / Reasoning Loop] │n│↓│n│[Layer 1: 工作记忆层] ← Dynamic Compression / Priority Queue│n│ ├─ 滑动窗口 重要性评分 自动摘要│n│ ├─ 语义类型标记（fact/hypothesis/pending/verified）│n│ └─ Token预算控制与实时裁剪│n│↓│n│[Layer 2: 长期记忆层] ← Hybrid Retrieval / Metadata Filtering │n│ ├─ 向量 关键词 知识图谱混合检索 │n│ ├─ 查询重写与意图对齐 │n│ └─ 时效性/权威性/置信度加权重排序│n│↓│n│[Layer 3: 遗忘与更新层] ← TTL / Explicit Override / Versioning│n│ ├─ 自动过期（TTL）与手动失效 │n│ ├─ 用户纠正 → 记忆修正闭环│n│ └─ 版本链与变更审计│n└─────────────────────────────────────────────────────────────────────┘n"}]},{"type":"heading","attrs":{"id":"6271f5bf-bae2-49b2-b52e-58810e3d0e6c","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"三、硬核实战1：工作记忆动态压缩与优先级引擎"}]},{"type":"paragraph","attrs":{"id":"f591b34f-2e64-41d2-bb82-cbb10dbb9cf3","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"让Agent在有限上下文中“记住该记的，放下该放的”。"}]},{"type":"heading","attrs":{"id":"0b5c5f8a-37eb-4978-9b7f-fb8029954f9c","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"3.1 环境准备"}]},{"type":"codeBlock","attrs":{"id":"86d2ea51-3f23-47a9-a1cc-675996c38a62","language":"javascript","theme":"atom-one-dark","runtimes":0,"isHoverDragHandle":false,"key":"","languageByAi":"javascript"},"content":[{"type":"text","text":"pip install pydantic tiktoken transformers redisn# 部署: Redis (工作记忆缓存) LLM (摘要服务) Token计数器n"}]},{"type":"heading","attrs":{"id":"1baa8741-2d23-4e39-9984-d4563bc7cc13","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"3.2 核心代码实现"}]},{"type":"paragraph","attrs":{"id":"c9ef0fdc-007c-49ed-86ac-01669cc57f7f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"创建 "},{"type":"text","marks":[{"type":"code"}],"text":"working_memory_engine.py"},{"type":"text","text":" ："}]},{"type":"codeBlock","attrs":{"id":"996506b1-97d7-4a8b-9897-30b206d9098f","language":"javascript","theme":"atom-one-dark","runtimes":0,"isHoverDragHandle":false,"key":"","languageByAi":"javascript"},"content":[{"type":"text","text":""""nworking_memory_engine.py - Agent工作记忆动态压缩与优先级引擎n技术栈: Pydantic / tiktoken / Transformers / Redisn"""nfrom typing import Dict, List, Any, Optional, Tuplenfrom pydantic import BaseModel, Fieldnfrom enum import Enumnimport asyncionimport timenimport jsonnimport hashlibnimport tiktokennfrom dataclasses import dataclass, fieldnnnclass SemanticType(str, Enum):nFACT = "fact"# 已验证事实nHYPOTHESIS = "hypothesis"# 待验证假设nPENDING = "pending"# 等待外部确认nVERIFIED = "verified"# 经工具/用户确认nINSTRUCTION = "instruction"# 用户指令nTOOL_OUTPUT = "tool_output"# 工具返回nSUMMARY = "summary"# 历史摘要nnn@dataclassnclass MemoryUnit:n"""工作记忆单元"""nunit_id: str = field(default_factory=lambda: f"wm-{uuid.uuid4().hex[:8]}")ncontent: str = ""nsemantic_type: SemanticType = SemanticType.FACTnimportance: float = 0.5 # 0-1, 越高越重要ncreated_at: float = field(default_factory=time.time)nlast_accessed: float = field(default_factory=time.time)naccess_count: int = 0nsource_turn: int = 0# 来自第几轮对话ntoken_count: int = 0nmetadata: Dict[str, Any] = field(default_factory=dict)nnnclass WorkingMemoryManager:n"""工作记忆管理器"""nnMAX_TOKEN_BUDGET = 8000 # 工作记忆Token上限nRECENCY_WEIGHT = 0.3nIMPORTANCE_WEIGHT = 0.4nACCESS_FREQ_WEIGHT = 0.2nTYPE_WEIGHT = 0.1 # 语义类型基础权重nnTYPE_BASE_SCORES = {nSemanticType.INSTRUCTION: 0.9,nSemanticType.VERIFIED: 0.8,nSemanticType.FACT: 0.6,nSemanticType.PENDING: 0.5,nSemanticType.TOOL_OUTPUT: 0.4,nSemanticType.HYPOTHESIS: 0.3,nSemanticType.SUMMARY: 0.7,n}nndef __init__(self, cache_client, llm_summarizer, tokenizer):nself.cache = cache_client# Redisnself.summarizer = llm_summarizernself.tokenizer = tokenizernnasync def add_unit(self, session_id: str, unit: MemoryUnit) -> None:n"""添加记忆单元"""nunit.token_count = len(self.tokenizer.encode(unit.content))nunit.importance = self._compute_importance(unit)nnawait self.cache.zadd(nf"wm:{session_id}",n{unit.unit_id: unit.importance}n)nawait self.cache.hset(nf"wm:unit:{unit.unit_id}",nmapping={"data": json.dumps(unit.__dict__, default=str)}n)nn# 检查是否超预算，触发压缩nawait self._enforce_budget(session_id)nnasync def get_context(self, session_id: str) -> List[MemoryUnit]:n"""获取当前工作记忆上下文（按优先级排序）"""nunit_ids = await self.cache.zrevrange(f"wm:{session_id}", 0, -1)nnunits = []nfor uid in unit_ids:ndata = await self.cache.hgetall(f"wm:unit:{uid}")nif data: 31268.t.kuaisou.comnunit = MemoryUnit(**json.loads(data["data"]))nunit.last_accessed = time.time()nunit.access_count = 1nunits.append(unit)nn# 更新访问统计nfor u in units:nawait self.cache.hset(nf"wm:unit:{u.unit_id}",nmapping={"data": json.dumps(u.__dict__, default=str)}n)nnreturn unitsnnasync def _enforce_budget(self, session_id: str) -> None:n"""强制Token预算，必要时压缩旧记忆"""ncurrent_tokens = await self._total_tokens(session_id)nnif current_tokens <= self.MAX_TOKEN_BUDGET:nreturnnn# 获取低优先级单元nlow_priority = await self.cache.zrangebyscore(nf"wm:{session_id}", "-inf", 0.4, start=0, num=10n)nnif len(low_priority) >= 3:n# 将多个低优先级单元压缩为一个摘要nunits_to_compress = []nfor uid in low_priority[:5]:ndata = await self.cache.hgetall(f"wm:unit:{uid}")nif data:nunits_to_compress.append(MemoryUnit(**json.loads(data["data"])))nnsummary_content = await self.summarizer.summarize(n[u.content for u in units_to_compress],nfocus="key facts and decisions"n)nn# 移除原单元nfor u in units_to_compress:nawait self.cache.zrem(f"wm:{session_id}", u.unit_id)nawait self.cache.delete(f"wm:unit:{u.unit_id}")nn# 插入摘要单元nsummary_unit = MemoryUnit(ncontent=summary_content,nsemantic_type=SemanticType.SUMMARY,nimportance=0.7,nsource_turn=min(u.source_turn for u in units_to_compress),nmetadata={"compressed_from": [u.unit_id for u in units_to_compress]}n)nawait self.add_unit(session_id, summary_unit)nelse:n# 无法压缩，移除最低优先级单元nlowest = await self.cache.zrange(f"wm:{session_id}", 0, 0)nif lowest:nawait self.cache.zrem(f"wm:{session_id}", lowest[0])nawait self.cache.delete(f"wm:unit:{lowest[0]}")nndef _compute_importance(self, unit: MemoryUnit) -> float:n"""计算记忆单元综合重要性"""nrecency = max(0, 1 - (time.time() - unit.created_at) / 3600)ntype_score = self.TYPE_BASE_SCORES.get(unit.semantic_type, 0.5)nnscore = (nself.RECENCY_WEIGHT * recency nself.IMPORTANCE_WEIGHT * unit.importance nself.ACCESS_FREQ_WEIGHT * min(unit.access_count / 5, 1.0) nself.TYPE_WEIGHT * type_scoren)nreturn round(min(score, 1.0), 3)nnasync def _total_tokens(self, session_id: str) -> int:nunit_ids = await self.cache.zrange(f"wm:{session_id}", 0, -1)ntotal = 0nfor uid in unit_ids:ndata = await self.cache.hget(f"wm:unit:{uid}", "data")nif data:nunit = json.loads(data)ntotal = unit.get("token_count", 0)nreturn totalnnnimport uuidn"}]},{"type":"heading","attrs":{"id":"3c6593b5-2ac4-4dfa-9df6-a2ef1c309cd1","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"3.3 专业性点评"}]},{"type":"paragraph","attrs":{"id":"202ef55f-988c-4c6d-b838-17b993d9fc9a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"此方案将工作记忆从“FIFO队列”升级为“优先级驱动 动态压缩 预算约束”的认知资源管理系统。核心设计亮点：1）"},{"type":"text","marks":[{"type":"bold"}],"text":"重要性必须是多维计算的"},{"type":"text","text":" ——不能只看时间远近，指令和已验证事实天然比假设更重要，高频访问的内容应被保留；2）"},{"type":"text","marks":[{"type":"bold"}],"text":"压缩必须是语义保真的"},{"type":"text","text":" ——不是简单截断，而是用LLM将多条低优先级消息提炼为一条摘要，保留关键决策和事实；3）"},{"type":"text","marks":[{"type":"bold"}],"text":"Token预算必须硬性执行"},{"type":"text","text":" ——超出即触发压缩或淘汰，不能靠“希望模型自己能忽略无关内容”；4）"},{"type":"text","marks":[{"type":"bold"}],"text":"语义类型是治理的抓手"},{"type":"text","text":" ——不同类型有不同的基础权重和生命周期，这是实现精细化记忆管理的前提。"}]},{"type":"heading","attrs":{"id":"1d3e1ec7-79b4-4b22-ae5e-169047f16bf2","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"四、硬核实战2：混合检索增强与记忆遗忘引擎"}]},{"type":"paragraph","attrs":{"id":"d0acb9ce-91f6-478b-b69b-c5ca053698dc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"让Agent“找得准、信得过、忘得掉”。"}]},{"type":"heading","attrs":{"id":"90393eb3-7efd-48ba-adfc-02252acca2db","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"4.1 核心代码实现"}]},{"type":"paragraph","attrs":{"id":"325a8f2f-261f-4fd3-9e52-e68f2436730d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"创建 "},{"type":"text","marks":[{"type":"code"}],"text":"long_term_memory_engine.py"},{"type":"text","text":" ："}]},{"type":"codeBlock","attrs":{"id":"b6987a92-9d51-4f1d-9ecc-a476c563df88","language":"javascript","theme":"atom-one-dark","runtimes":0,"isHoverDragHandle":false,"key":"","languageByAi":"javascript"},"content":[{"type":"text","text":""""nlong_term_memory_engine.py - Agent长期记忆检索与遗忘引擎n技术栈: Elasticsearch / Qdrant / Neo4j / Redisn"""nfrom typing import Dict, List, Any, Optional, Tuplenfrom pydantic import BaseModel, Fieldnfrom enum import Enumnimport asyncionimport timenimport jsonnimport hashlibnfrom dataclasses import dataclass, fieldnfrom datetime import datetime, timedeltannnclass RetrievalStrategy(str, Enum):nVECTOR_ONLY = "vector"nKEYWORD_ONLY = "keyword"nHYBRID = "hybrid"nGRAPH_TRAVERSAL = "graph"nnnclass MemoryValidity(str, Enum):nACTIVE = "active"nDEPRECATED = "deprecated"nEXPIRED = "expired"nSUPERSEDED = "superseded"nnn@dataclassnclass LongTermMemoryEntry: 31272.t.kuaisou.comn"""长期记忆条目"""nentry_id: str = field(default_factory=lambda: f"ltm-{uuid.uuid4().hex[:10]}")ncontent: str = ""nembedding: List[float] = field(default_factory=list)nkeywords: List[str] = field(default_factory=list)nn# 元数据nsource: str = ""# 来源系统/文档ncreated_at: float = field(default_factory=time.time)nvalid_until: Optional[float] = None# TTL过期时间nversion: int = 1nvalidity: MemoryValidity = MemoryValidity.ACTIVEnn# 质量标注nconfidence: float = 0.8 # 置信度nauthority_score: float = 0.7# 来源权威性nverified_by: Optional[str] = None# 验证者/工具nn# 关联nsupersedes: Optional[str] = None# 替代的旧条目IDnrelated_entries: List[str] = field(default_factory=list)ntags: List[str] = field(default_factory=list)nnn@dataclassnclass RetrievalResult:n"""检索结果"""nentry: LongTermMemoryEntrynrelevance_score: float = 0.0nretrieval_method: str = ""nrerank_score: float = 0.0nwarnings: List[str] = field(default_factory=list)nnnclass HybridRetriever:n"""混合检索引擎"""nndef __init__(self, vector_store, keyword_store, graph_store, llm_client):nself.vector = vector_store# Qdrant / Pineconenself.keyword = keyword_store# Elasticsearchnself.graph = graph_store# Neo4jnself.llm = llm_clientnnasync def retrieve(nself, nquery: str,nsession_context: Optional[List[str]] = None,nfilters: Optional[Dict[str, Any]] = None,ntop_k: int = 10,nstrategy: RetrievalStrategy = RetrievalStrategy.HYBRIDn) -> List[RetrievalResult]: 31271.t.kuaisou.comn"""混合检索"""n# Step 1: 查询重写（结合会话上下文）nrewritten_query = await self._rewrite_query(query, session_context)nn# Step 2: 多路召回ncandidates = []nif strategy in [RetrievalStrategy.VECTOR_ONLY, RetrievalStrategy.HYBRID]:nvec_results = await self.vector.search(nrewritten_query, top_k=top_k * 2, filters=filtersn)ncandidates.extend([nRetrievalResult(entry=r, relevance_score=r.score, retrieval_method="vector")nfor r in vec_resultsn])nnif strategy in [RetrievalStrategy.KEYWORD_ONLY, RetrievalStrategy.HYBRID]:nkw_results = await self.keyword.search(nrewritten_query, top_k=top_k * 2, filters=filtersn)ncandidates.extend([nRetrievalResult(entry=r, relevance_score=r.score, retrieval_method="keyword")nfor r in kw_resultsn])nn# Step 3: 去重nseen_ids = set(31270.t.kuaisou.com)nunique_candidates = []nfor c in candidates:nif c.entry.entry_id not in seen_ids:nseen_ids.add(c.entry.entry_id)nunique_candidates.append(c)nn# Step 4: 重排序（综合考虑相关性 时效 权威 置信度）nreranked = await self._rerank(unique_candidates, rewritten_query)nn# Step 5: 过滤无效条目并附加警告nfinal = []nfor r in reranked[:top_k]:nwarnings = self._check_validity(r.entry)nr.warnings = warningsnif r.entry.validity == MemoryValidity.ACTIVE or warnings:nfinal.append(r)nnreturn finalnnasync def _rewrite_query(nself, query: str, context: Optional[List[str]]n) -> str:n"""结合上下文重写查询"""nif not context:nreturn querynnprompt = f"""Rewrite this query for better retrieval, incorporating conversation context.nOriginal Query: {query}nRecent Context: {' | '.join(context[-3:])}nnRewritten Query (more specific, disambiguated):"""nnreturn await self.llm.chat(prompt)nnasync def _rerank(nself, candidates: List[RetrievalResult], query: strn) -> List[RetrievalResult]:n"""多维度重排序"""nfor c in candidates:ne = c.entryn# 时效性衰减（越新越高）nage_hours = (time.time() - e.created_at) / 3600nrecency = max(0, 1 - age_hours / (30 * 24))# 30天衰减nn# 综合分数nc.rerank_score = (n0.4 * c.relevance_score n0.2 * recency n0.2 * e.authority_score n0.2 * e.confidencen)nncandidates.sort(key=lambda x: x.rerank_score, reverse=True)nreturn candidatesnndef _check_validity(self, entry: LongTermMemoryEntry) -> List[str]:n"""检查条目有效性"""nwarnings = []nif entry.validity == MemoryValidity.DEPRECATED:nwarnings.append("This entry is deprecated")nif entry.validity == MemoryValidity.SUPERSEDED:nwarnings.append(f"Superseded by {entry.supersedes}")nif entry.valid_until and time.time() > entry.valid_until:nwarnings.append("Entry has expired")nif entry.confidence < 0.5:nwarnings.append("Low confidence entry")nreturn warningsnnnclass MemoryForgettingEngine:n"""记忆遗忘与更新引擎"""nndef __init__(self, memory_store, audit_logger):nself.store = memory_storenself.audit = audit_loggernnasync def explicit_override(nself, old_entry_id: str, new_content: str, nreason: str, overridden_by: strn) -> str:n"""显式覆盖：用户纠正或政策更新"""nold_entry = await self.store.get(old_entry_id)nif not old_entry:nraise ValueError(f"Entry {old_entry_id} not found")nn# 标记旧条目为被替代nold_entry.validity = MemoryValidity.SUPERSEDEDnold_entry.supersedes = None# 它被别人替代nawait self.store.update(old_entry)nn# 创建新条目nnew_entry = LongTermMemoryEntry(ncontent=new_content,nsource=old_entry.source,nversion=old_entry.version 1,nconfidence=0.95,# 显式纠正，高置信度nsupersedes=old_entry_id,ntags=old_entry.tags ["corrected"],nverified_by=overridden_byn)nnew_id = await self.store.insert(new_entry)nn# 审计nawait self.audit.log("memory_override", {n"old_entry_id": old_entry_id,n"new_entry_id": new_id,n"reason": reason,n"overridden_by": overridden_by,n"timestamp": time.time()n})nnreturn new_idnnasync def apply_ttl_policy(self) -> int:n"""执行TTL过期策略"""nnow = time.time(31269.t.kuaisou.com)nexpired = await self.store.find_expired(now)nncount = 0nfor entry in expired:nif entry.validity != MemoryValidity.EXPIRED:nentry.validity = MemoryValidity.EXPIREDnawait self.store.update(entry)ncount = 1nnif count > 0:nawait self.audit.log("ttl_cleanup", {n"expired_count": count,n"timestamp": nown})nnreturn countnnasync def soft_delete(self, entry_id: str, reason: str) -> bool:n"""软删除（保留审计痕迹）"""nentry = await self.store.get(entry_id)nif not entry:nreturn Falsennentry.validity = MemoryValidity.DEPRECATEDnawait self.store.update(entry)nnawait self.audit.log("soft_delete", {n"entry_id": entry_id,n"reason": reason,n"timestamp": time.time()n})nreturn Truennnimport uuidn"}]},{"type":"heading","attrs":{"id":"80de0dce-eaa0-4f9c-b969-5141932ba717","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"4.2 专业性点评"}]},{"type":"paragraph","attrs":{"id":"26ce2342-68ea-4316-abf4-fee18457e63c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"此方案将长期记忆从“向量搜索黑盒”升级为“混合检索 多维重排 主动遗忘”的可信知识管理系统。核心设计要点：1）"},{"type":"text","marks":[{"type":"bold"}],"text":"检索必须结合上下文重写查询"},{"type":"text","text":" ——用户说“那个项目”时，必须结合最近对话消歧，否则召回必然失准；2）"},{"type":"text","marks":[{"type":"bold"}],"text":"重排序必须超越纯语义相似度"},{"type":"text","text":" ——时效性、来源权威性、置信度都是决定“该不该信”的关键因子，不能只看embedding距离；3）"},{"type":"text","marks":[{"type":"bold"}],"text":"无效条目不能静默丢弃"},{"type":"text","text":" ——要附警告返回，让Agent和用户可以判断是否采纳，而非替用户做决定；4）"},{"type":"text","marks":[{"type":"bold"}],"text":"遗忘必须是可审计的操作"},{"type":"text","text":" ——无论是TTL过期还是人工覆盖，都要留痕，这是合规与溯源的基础。"}]},{"type":"heading","attrs":{"id":"e3541983-653f-4cb8-9408-f31233741420","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"五、生产环境避坑指南：Agent记忆治理五大铁律"}]},{"type":"heading","attrs":{"id":"a21d26eb-41c7-433a-9d5e-1314553b0342","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"1. 工作记忆必须有Token硬预算，不能靠模型自觉"}]},{"type":"paragraph","attrs":{"id":"4e2aab1c-f3f8-423e-8091-133b550051e0","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"坑"},{"type":"text","text":" ：认为“模型上下文窗口够大就不用管”，结果关键指令被淹没在历史噪声中；Token费用失控，单次调用成本是预期的5倍；延迟随对话轮数线性增长，用户体验恶化。"}]},{"type":"paragraph","attrs":{"id":"84fd82be-0914-4e59-beb9-52e1697a2fb4","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"对策"},{"type":"text","text":" ：设定明确的Token预算（如8K），超出即触发压缩或淘汰；压缩策略可配置（摘要优先 vs 淘汰优先）；实时监控Token使用率，超80%预警。"}]},{"type":"heading","attrs":{"id":"58908a7b-7caa-4d4c-904e-daf6255b3d81","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"2. 语义类型标注必须在写入时完成，不能事后补"}]},{"type":"paragraph","attrs":{"id":"a25f72b0-1317-4dde-b86b-93ff8991be5e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"坑"},{"type":"text","text":" ：所有记忆都标为“fact”，无法区分假设与已验证结论；后续想做差异化处理时发现没有类型字段，只能全量重新标注；类型定义模糊，不同开发者标注标准不一。"}]},{"type":"paragraph","attrs":{"id":"8750b8e6-8989-46d1-9b2d-b139b4323251","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"对策"},{"type":"text","text":" ：定义清晰的语义类型枚举及判定规则；写入时由Agent或预处理模块自动标注，人工抽检校准；类型定义纳入团队规范文档，定期Review。"}]},{"type":"heading","attrs":{"id":"a01e601f-a18d-42fc-a765-6cc987745c26","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"3. 检索结果必须带元数据警告，不能裸喂模型"}]},{"type":"paragraph","attrs":{"id":"94d7de2e-df92-4848-9888-f819024594cb","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"坑"},{"type":"text","text":" ：Agent引用了已过期的政策条款，因为检索结果没标明“已废止”；低置信度的猜测被当作事实输出，引发用户投诉；无法追溯某条信息的来源和验证状态。"}]},{"type":"paragraph","attrs":{"id":"17651037-6d92-4bed-aa3c-3829478ba25b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"对策"},{"type":"text","text":" ：每条检索结果必须携带validity、confidence、source、version等元数据；Agent Prompt中明确要求“对带警告的内容需谨慎引用并注明不确定性”；前端展示时对警告内容做视觉区分。"}]},{"type":"heading","attrs":{"id":"08b2ae9a-3185-4fb3-93a2-88260db21aeb","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"4. 用户纠正必须闭环到记忆修正，不能只改当次回复"}]},{"type":"paragraph","attrs":{"id":"c2f3301e-07d4-4bcd-b247-794894014800","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"坑"},{"type":"text","text":" ：用户说“地址错了”，Agent改了本次回复但下次还用旧地址；同样的错误反复出现，用户失去耐心；纠正记录散落在对话日志中，无法系统化利用。"}]},{"type":"paragraph","attrs":{"id":"69751d8d-0f73-446b-867e-018d2b6c34dc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"对策"},{"type":"text","text":" ：建立“纠正→记忆覆盖→验证”闭环流程；用户纠正自动触发explicit_override，生成新版本记忆；后续相同查询优先返回新版本，并标记“已根据用户反馈更新”。"}]},{"type":"heading","attrs":{"id":"84956a2b-4330-4972-9d9e-763b5cfe8f30","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"5. 记忆变更必须全量审计，不能无痕操作"}]},{"type":"paragraph","attrs":{"id":"9ee40a29-09c3-46aa-a394-79e1ca645e0c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"坑"},{"type":"text","text":" ：记忆被误删或误改，无法恢复；合规审查时无法证明“Agent使用的信息在当时是有效的”；多人协作时不知道谁改了什么记忆。"}]},{"type":"paragraph","attrs":{"id":"33238b04-932a-4e33-aa2e-158ec80feab5","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"对策"},{"type":"text","text":" ：所有写入、覆盖、删除、过期操作写入不可篡改审计日志；日志包含操作人、原因、前后快照；支持按时间线回溯任意时刻的记忆状态。"}]},{"type":"heading","attrs":{"id":"c4b20dea-cb6a-4a2d-ba73-44f246515f6a","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"六、结语：记忆治理是智能体从应答机器走向可信伙伴的认知契约"}]},{"type":"paragraph","attrs":{"id":"ed62763e-c353-485d-87f7-929585cb59ef","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"当AI Agent从“一问一答”进化为“持续协作”，记忆就不再是“可选的增强功能”，而是“可靠性的根基”。2026年的竞争分水岭，不在于谁的Agent记得更多，而在于谁的Agent“记得对、信得准、忘得妥”——能让用户确信“它不会拿过时信息误导我”，能让合规官看到“它的知识始终处于受控状态”，能让团队相信“它的认知是可维护的工程产物”。"}]},{"type":"paragraph","attrs":{"id":"5af2be02-a94c-4d05-89ac-d2340f5a4c03","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"工作记忆赋予了Agent以专注力，长期记忆赋予了Agent以知识深度，遗忘机制赋予了Agent以认知弹性。这三者共同构成了Agent记忆工程的“认知三角”。那些仍认为“把对话历史全塞进去就行”的团队，终将在第100轮对话时被自己的记忆反噬。"}]},{"type":"paragraph","attrs":{"id":"c8dfd57e-0162-4342-a299-5c8be3cfb29a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"真正的AI工程化，不是让Agent拥有无限记忆，而是建立一套让记忆“可管理、可信赖、可演化”的治理体系，让每一段记忆都服务于当下决策，让每一次遗忘都为新知腾出空间，在智能体承担越来越重认知责任的时代，以治理换取信任，以节制赢得可靠。"}]},{"type":"heading","attrs":{"id":"9bb3a64d-6fd2-460b-93f8-f3754e9e9e78","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"参考资料"}]},{"type":"orderedList","attrs":{"id":"58f5a4ec-b2ce-4194-8267-864461b233f4","start":1,"isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"244d6d07-9721-4001-b6c7-cdeda6754204"},"content":[{"type":"paragraph","attrs":{"id":"c757b75a-7eb9-4066-84d5-6cf10341b58b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"DeepMind, "},{"type":"text","marks":[{"type":"italic"}],"text":"Memory Management in Long-Horizon AI Agents: Failure Modes & Solutions"},{"type":"text","text":" , 2026."}]}]},{"type":"listItem","attrs":{"id":"7c8331df-d637-45c6-ad63-cfc5d234f4da"},"content":[{"type":"paragraph","attrs":{"id":"b91fc6e2-8839-45ee-a0dc-5d0bb3436893","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"Anthropic, "},{"type":"text","marks":[{"type":"italic"}],"text":"Context Window Optimization: Compression, Prioritization & Budget Control"},{"type":"text","text":" , 2026."}]}]},{"type":"listItem","attrs":{"id":"92cf03d2-4c2c-46a0-a931-7656b4c36903"},"content":[{"type":"paragraph","attrs":{"id":"e47d5009-d101-4334-956e-e3b5fff11693","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"LangChain, "},{"type":"text","marks":[{"type":"italic"}],"text":"Hybrid Retrieval for Production RAG: Beyond Vector Similarity"},{"type":"text","text":" , 2026."}]}]},{"type":"listItem","attrs":{"id":"51b4939b-164c-4c63-a6a5-d0a61a205a25"},"content":[{"type":"paragraph","attrs":{"id":"74c7932b-c6b1-49d7-b673-c92023021414","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"Microsoft Research, "},{"type":"text","marks":[{"type":"italic"}],"text":"Active Forgetting & Memory Updating in Conversational Agents"},{"type":"text","text":" , ACL 2026."}]}]},{"type":"listItem","attrs":{"id":"9b2a3a19-7de2-44b7-803c-58d67bd4f2d2"},"content":[{"type":"paragraph","attrs":{"id":"111724eb-316d-4e68-816c-1523340af654","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"ISO/IEC, "},{"type":"text","marks":[{"type":"italic"}],"text":"AI System Knowledge Management & Traceability Standard"},{"type":"text","text":" , 42500:2026."}]}]},{"type":"listItem","attrs":{"id":"41bd7e64-6470-4822-8c39-dda0fe0e3bad"},"content":[{"type":"paragraph","attrs":{"id":"e7a125f5-1454-4644-a686-036409e52433","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"OpenAI, "},{"type":"text","marks":[{"type":"italic"}],"text":"Best Practices for Agent Memory: Lessons from GPT-Agent Deployments"},{"type":"text","text":" , 2026."}]}]}]},{"type":"paragraph","attrs":{"id":"4532d5ad-ab14-4f29-b708-969df78dad19","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]}]}","createTime":1786456364,"ext":{"closeTextLink":0,"comment_ban":0,"description":"","focusRead":0},"favNum":0,"html":"","isOriginal":0,"likeNum":0,