Agent 记忆系统：架构、设计与LangChain 实践

一、背景：为什么 Agent 需要记忆系统？

1.1 问题的根源：LLM 的无状态本质

大语言模型（LLM）本质上是一个无状态函数：给定一段输入 Prompt，输出一段文本。它不具备跨会话的持久记忆，每次调用都从零开始。这在单轮问答场景下问题不大，但当我们把 LLM 包装成 Agent（智能体）时，问题就暴露了：

• 上下文窗口有限：即便是 128K token 的长上下文，也无法承载数月的交互历史、海量工具调用记录、外部知识库；
• 多轮对话失忆：用户在第 20 轮提到"上周说的那个方案"，模型完全不知道；
• 个性化缺失：Agent 无法记住用户偏好、习惯、历史决策；
• 跨任务知识断层：Agent 在处理任务 B 时，无法利用任务 A 中积累的领域知识；
• 推理链丢失：复杂任务需要多步规划，中间推理过程无法持久化。

1.2 Agent 记忆系统要解决的核心问题

| 问题

描述

持久化

信息跨 session、跨进程持续存在

| |

检索效率

从海量历史中快速找到相关记忆

| |

遗忘机制

避免记忆无限膨胀，合理淘汰过时信息

| |

记忆类型区分

事实性知识 vs 情节记忆 vs 工作记忆的分层管理

| |

一致性

避免记忆冲突和矛盾

二、技术架构与设计思路

2.1 记忆分类体系（仿人类认知）

Agent 记忆系统借鉴了认知科学对人类记忆的分类，并在工程实现上有明确的对应关系：

<span leaf="">记忆层级 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 对应实现 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 读写特征</span><br><span leaf="">─────────────────────────────────────────────────────────────────────</span><br><span leaf="">感知缓冲（Sensory Buffer） → 原始 Prompt / 传感器输入 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;只写，极短暂</span><br><span leaf="">工作记忆（Working Memory） → LLM Context Window / Scratchpad &nbsp;读写，单次生命周期</span><br><span leaf="">情节记忆（Episodic Memory）→ 向量数据库（时序 + 语义索引） &nbsp; &nbsp; 异步写，语义检索</span><br><span leaf="">语义记忆（Semantic Memory）→ 知识图谱 / 结构化 KV Store &nbsp; &nbsp; &nbsp; 读多写少，强一致</span><br><span leaf="">过程记忆（Procedural Mem）→ ToolSpec / Prompt 模板库 &nbsp; &nbsp; &nbsp; &nbsp; 低频写，高频读</span><br><span leaf="">外部记忆（External Mem） &nbsp;→ 文件系统 / 数据库 / 外部 API &nbsp; &nbsp; &nbsp;按需访问</span>

这一分层不是单纯的学术分类，它直接决定了读写路径、存储选型、一致性级别的取舍：

• 工作记忆是"热数据"，必须在 token 预算内；
• 情节记忆是"温数据"，牺牲精确换取规模；
• 语义/过程记忆是"冷数据"，强调结构化与可靠性。

2.2 整体架构设计

<span leaf="">┌─────────────────────────────────────────────────────────────────────┐</span><br><span leaf="">│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; Agent Runtime &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│</span><br><span leaf="">│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│</span><br><span leaf="">│ &nbsp;┌──────────────────────────────────────────────────────────────┐ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;Working Memory Layer &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; ┌─────────────────────┐ &nbsp; ┌────────────────────────────┐ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; │ &nbsp;In-context Window &nbsp;│ &nbsp; │ &nbsp;Scratchpad（思维链草稿） &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; │ &nbsp;(token budget管理) │ &nbsp; │ &nbsp;(ReAct / CoT 中间推理) &nbsp; &nbsp;│ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; └──────────┬──────────┘ &nbsp; └────────────┬───────────────┘ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;└──────────────┼─────────────────────────────┼─────────────────┘ &nbsp; │</span><br><span leaf="">│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp;超出阈值触发压缩 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ 推理结束写回 &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">│ &nbsp;┌──────────────▼─────────────────────────────▼─────────────────┐ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; Memory Manager &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;┌───────────────┐ &nbsp;┌────────────────┐ &nbsp;┌────────────────┐ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ Episodic Store│ &nbsp;│ Semantic Store │ &nbsp;│Procedural Store│ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ · 事件流写入 &nbsp; │ &nbsp;│ · 实体/关系 &nbsp; &nbsp;│ &nbsp;│ · Tool注册表 &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ · 向量化索引 &nbsp; │ &nbsp;│ · 知识三元组 &nbsp; │ &nbsp;│ · Prompt模板 &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ · TTL过期管理 &nbsp;│ &nbsp;│ · 版本化更新 &nbsp; │ &nbsp;│ · 经验规则库 &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;└───────┬───────┘ &nbsp;└───────┬────────┘ &nbsp;└───────┬────────┘ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;┌───────▼──────────────────▼────────────────────▼────────┐ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; Memory Router &amp; Scheduler &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ &nbsp; · 读：多路归并（语义 + 时序 + 重要性三维排序） &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ &nbsp; · 写：异步落盘 + 重要性过滤 + 冲突检测 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ &nbsp; · 整合（Consolidation）：定时触发 LLM 摘要压缩 &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;└───────────────────────────┬─────────────────────────────┘ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;└──────────────────────────────┼────────────────────────────────┘ &nbsp; │</span><br><span leaf="">│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">│ &nbsp;┌──────────────────────────────▼────────────────────────────────┐ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;Storage Backend &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;┌──────────────┐ &nbsp;┌─────────────────┐ &nbsp;┌─────────────────┐ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ &nbsp;向量数据库 &nbsp; │ &nbsp;│ &nbsp; 关系/KV数据库 &nbsp;│ &nbsp;│ &nbsp; &nbsp;知识图谱 &nbsp; &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ Chroma/Milvus│ &nbsp;│ Postgres/Redis &nbsp;│ &nbsp;│ Neo4j/NebulaGraph│ │ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ Pinecone/ &nbsp; &nbsp;│ &nbsp;│ SQLite/DynamoDB │ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;│ Weaviate &nbsp; &nbsp; │ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp;│ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;│ &nbsp;└──────────────┘ &nbsp;└─────────────────┘ &nbsp;└─────────────────┘ &nbsp;│ &nbsp; │</span><br><span leaf="">│ &nbsp;└────────────────────────────────────────────────────────────────┘ &nbsp; │</span><br><span leaf="">└─────────────────────────────────────────────────────────────────────┘</span>

2.3 Memory Manager 核心设计：写路径

写路径是记忆系统质量的决定因素，需解决三个关键问题：写什么、怎么写、何时写。

2.3.1 选择性写入与重要性过滤

不加区分地将所有对话写入向量库会带来两个问题：检索时噪音过多（低质记忆稀释高质结果）、存储膨胀。工程上的解决方案是写前过滤：

<span leaf="">原始交互 → [重要性评分器] → 评分 ≥ 阈值？</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; ├─ Yes → [结构化提取] → 向量化 → 落库</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; └─ No &nbsp;→ 丢弃 或 仅存摘要</span>

重要性评分可以用 LLM 打分（成本高但准确），也可以用轻量规则（关键词命中、实体密度、用户显式确认）：

<span><span leaf=""># 规则 + LLM 混合的重要性过滤器示例</span></span><br><span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">ImportanceFilter</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; KEYWORD_BOOST = [</span><span><span leaf="">"记住"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"重要"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"注意"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"下次"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"规范"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"错误"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"修复"</span></span><span leaf="">]</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">score</span></span><span leaf="">(</span><span><span leaf="">self, text:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, use_llm:&nbsp;</span><span><span leaf="">bool</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">False</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">float</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 规则分（快速）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; rule_score =&nbsp;</span><span><span leaf="">sum</span></span><span leaf="">(</span><span><span leaf="">0.1</span></span><span leaf="">&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;kw&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.KEYWORD_BOOST&nbsp;</span><span><span leaf="">if</span></span><span leaf="">&nbsp;kw&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;text)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; rule_score =&nbsp;</span><span><span leaf="">min</span></span><span leaf="">(rule_score,&nbsp;</span><span><span leaf="">0.5</span></span><span leaf="">) &nbsp;</span><span><span leaf=""># 上限 0.5</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 内容密度分：实体数 / token数（简单启发式）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; density_score =&nbsp;</span><span><span leaf="">min</span></span><span leaf="">(</span><span><span leaf="">len</span></span><span leaf="">(re.findall(</span><span><span leaf="">r'[A-Z][a-z]+|[\u4e00-\u9fa5]{2,}'</span></span><span leaf="">, text)) /&nbsp;</span><span><span leaf="">50</span></span><span leaf="">,&nbsp;</span><span><span leaf="">0.3</span></span><span leaf="">)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">if</span></span><span leaf="">&nbsp;</span><span><span leaf="">not</span></span><span leaf="">&nbsp;use_llm:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;rule_score + density_score</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># LLM 精确打分（高价值路径才触发）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; llm_score =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._llm_importance(text) &nbsp;</span><span><span leaf=""># 返回 0~1</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;</span><span><span leaf="">0.4</span></span><span leaf="">&nbsp;* (rule_score + density_score) +&nbsp;</span><span><span leaf="">0.6</span></span><span leaf="">&nbsp;* llm_score</span>

2.3.2 记忆的结构化提取（Memory Extraction）

原始对话文本直接向量化效果差——同义表达导致召回不稳定。生产系统通常在写入前做结构化提取，将非结构化对话转成标准记忆单元：

<span leaf="">原始文本："用户提到他们团队使用 GitFlow 分支策略，主要用 Java + Spring Boot，</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 讨厌用 Lombok，以前在项目 X 里用过 Kafka 踩了坑"</span><br><br><span leaf="">结构化提取结果：</span><br><span leaf="">[</span><br><span leaf="">&nbsp; { type: "preference", &nbsp;subject: "user", key: "branch_strategy", value: "GitFlow" },</span><br><span leaf="">&nbsp; { type: "tech_stack", &nbsp;subject: "user", items: ["Java", "Spring Boot"] },</span><br><span leaf="">&nbsp; { type: "dislike", &nbsp; &nbsp; subject: "user", key: "library", value: "Lombok" },</span><br><span leaf="">&nbsp; { type: "experience", &nbsp;subject: "user", domain: "Kafka", outcome: "negative" }</span><br><span leaf="">]</span>

每条结构化记忆单独向量化，检索精度远高于整段文本。

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_core.prompts&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;ChatPromptTemplate</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_core.output_parsers&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;JsonOutputParser</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;pydantic&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;BaseModel</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;typing&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;</span><span><span leaf="">List</span></span><span leaf="">,&nbsp;</span><span><span leaf="">Literal</span></span><br><br><span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">MemoryUnit</span></span><span leaf="">(</span><span><span leaf="">BaseModel</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">type</span></span><span leaf="">:&nbsp;</span><span><span leaf="">Literal</span></span><span leaf="">[</span><span><span leaf="">"preference"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"tech_stack"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"experience"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"fact"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"dislike"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"goal"</span></span><span leaf="">]</span><br><span leaf="">&nbsp; &nbsp; subject:&nbsp;</span><span><span leaf="">str</span></span><br><span leaf="">&nbsp; &nbsp; content:&nbsp;</span><span><span leaf="">str</span></span><br><span leaf="">&nbsp; &nbsp; confidence:&nbsp;</span><span><span leaf="">float</span></span><span leaf="">&nbsp;&nbsp;</span><span><span leaf=""># 0~1，提取置信度</span></span><br><br><span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">MemoryExtractor</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""将对话转换为结构化记忆单元"""</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, llm</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.chain = (</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; ChatPromptTemplate.from_template(</span><span><span leaf="">"""</span><br><span leaf="">从以下对话中提取值得长期记忆的信息单元，以 JSON 数组返回：</span><br><span leaf="">- 只提取明确表达的信息，不推断</span><br><span leaf="">- 每条记忆简洁完整，独立可理解</span><br><span leaf="">- 置信度：用户明确陈述=1.0，推断=0.5</span><br><br><span leaf="">对话内容：</span><br><span leaf="">{text}</span><br><br><span leaf="">输出格式：[{{"type": "...", "subject": "...", "content": "...", "confidence": 0.0}}]</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; """</span></span><span leaf="">)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; | llm</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; | JsonOutputParser()</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">extract</span></span><span leaf="">(</span><span><span leaf="">self, text:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">List</span></span><span leaf="">[MemoryUnit]:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; raw =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.chain.invoke({</span><span><span leaf="">"text"</span></span><span leaf="">: text})</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;[MemoryUnit(**item)&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;item&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;raw&nbsp;</span><span><span leaf="">if</span></span><span leaf="">&nbsp;item.get(</span><span><span leaf="">"confidence"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">0</span></span><span leaf="">) &gt;=&nbsp;</span><span><span leaf="">0.6</span></span><span leaf="">]</span>

2.3.3 记忆冲突检测与更新

记忆系统面临的一个棘手问题：用户说"我喜欢 Python"，三个月后说"我现在主要写 Go 了"——新记忆与旧记忆冲突，不处理会导致 Agent 输出自相矛盾。

解决方案是在写入时做冲突检测：

<span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">ConflictAwareMemoryWriter</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""写入时检测并解决记忆冲突"""</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, vectorstore, llm</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore = vectorstore</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.llm = llm</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">write</span></span><span leaf="">(</span><span><span leaf="">self, new_memory: MemoryUnit</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 1. 检索高相似度的已有记忆（可能冲突）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; candidates =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.similarity_search(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; new_memory.content, k=</span><span><span leaf="">5</span></span><span leaf="">, score_threshold=</span><span><span leaf="">0.85</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 2. 对高相似记忆做冲突判断</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;old_doc&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;candidates:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; conflict_type =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._detect_conflict(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; old_doc.page_content, new_memory.content</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">if</span></span><span leaf="">&nbsp;conflict_type ==&nbsp;</span><span><span leaf="">"UPDATE"</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 更新：标记旧记忆为过期，写入新记忆</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.delete([old_doc.metadata[</span><span><span leaf="">"id"</span></span><span leaf="">]])</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._add_with_provenance(new_memory, supersedes=old_doc.metadata[</span><span><span leaf="">"id"</span></span><span leaf="">])</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">elif</span></span><span leaf="">&nbsp;conflict_type ==&nbsp;</span><span><span leaf="">"COMPLEMENT"</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 互补：两者共存，添加关联标记</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._add_with_provenance(new_memory, related=old_doc.metadata[</span><span><span leaf="">"id"</span></span><span leaf="">])</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 3. 无冲突，直接写入</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._add_with_provenance(new_memory)</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">_detect_conflict</span></span><span leaf="">(</span><span><span leaf="">self, old_text:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, new_text:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">str</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; prompt =&nbsp;</span><span><span leaf="">f"""判断两条信息的关系，只返回: UPDATE / COMPLEMENT / UNRELATED</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; 旧信息：</span><span><span leaf="">{old_text}</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; 新信息：</span><span><span leaf="">{new_text}</span></span><span leaf="">"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.llm.invoke(prompt).content.strip()</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">_add_with_provenance</span></span><span leaf="">(</span><span><span leaf="">self, memory: MemoryUnit, **kwargs</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; metadata = {</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"id"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">(uuid.uuid4()),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"type"</span></span><span leaf="">: memory.</span><span><span leaf="">type</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"timestamp"</span></span><span leaf="">: time.time(),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; **kwargs</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; }</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.add_texts([memory.content], metadatas=[metadata])</span>

2.4 Memory Manager 核心设计：读路径

2.4.1 三维加权检索

检索不是简单的 Top-K 向量相似度，需要综合三个维度：

<span leaf="">最终得分 = α × Relevance(q, m) + β × Importance(m) + γ × Recency(m, t)</span>

其中：

• Relevance：向量余弦相似度，衡量语义相关程度
• Importance：写入时计算并存储的重要性分数
• Recency：指数衰减函数 decay^(elapsed_hours)，近期记忆得分更高

权重 α/β/γ 可根据场景动态调整：

• 精确事实查询（“我之前说的接口地址是？"）→ 高 α（相关性优先）
• 周期性摘要（“今天做了什么？"）→ 高 γ（时效性优先）
• 决策辅助（“这个方案之前评估过吗？"）→ 高 β（重要性优先）

<span><span leaf="">import</span></span><span leaf="">&nbsp;math</span><br><span><span leaf="">import</span></span><span leaf="">&nbsp;numpy&nbsp;</span><span><span leaf="">as</span></span><span leaf="">&nbsp;np</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;dataclasses&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;dataclass</span><br><br><span><span leaf="">@dataclass</span></span><br><span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">RetrievalConfig</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; alpha:&nbsp;</span><span><span leaf="">float</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">0.5</span></span><span leaf="">&nbsp; &nbsp;</span><span><span leaf=""># 相关性权重</span></span><br><span leaf="">&nbsp; &nbsp; beta:&nbsp;</span><span><span leaf="">float</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">0.3</span></span><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 重要性权重</span></span><br><span leaf="">&nbsp; &nbsp; gamma:&nbsp;</span><span><span leaf="">float</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">0.2</span></span><span leaf="">&nbsp; &nbsp;</span><span><span leaf=""># 时效性权重</span></span><br><span leaf="">&nbsp; &nbsp; decay:&nbsp;</span><span><span leaf="">float</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">0.99</span></span><span leaf="">&nbsp;&nbsp;</span><span><span leaf=""># 时效衰减因子（per hour）</span></span><br><span leaf="">&nbsp; &nbsp; top_k:&nbsp;</span><span><span leaf="">int</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">5</span></span><br><br><span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">WeightedMemoryRetriever</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, vectorstore, config: RetrievalConfig = RetrievalConfig(</span><span></span><span leaf="">)</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vs = vectorstore</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.cfg = config</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">retrieve</span></span><span leaf="">(</span><span><span leaf="">self, query:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, context:&nbsp;</span><span><span leaf="">dict</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">None</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">list</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 动态权重调整（根据查询意图）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; cfg =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._adjust_weights(query, context)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 召回候选集（比 top_k 多取，留出重排空间）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; candidates =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vs.similarity_search_with_score(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; query, k=cfg.top_k *&nbsp;</span><span><span leaf="">4</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; scored = []</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;doc, raw_similarity&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;candidates:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; meta = doc.metadata</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 相关性：FAISS 返回 L2 距离，需转换为相似度</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; relevance =&nbsp;</span><span><span leaf="">1</span></span><span leaf="">&nbsp;/ (</span><span><span leaf="">1</span></span><span leaf="">&nbsp;+ raw_similarity)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 重要性：写入时预计算并存储</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; importance = meta.get(</span><span><span leaf="">"importance"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">0.5</span></span><span leaf="">)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 时效性：指数衰减</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; hours = (time.time() - meta.get(</span><span><span leaf="">"timestamp"</span></span><span leaf="">, time.time())) /&nbsp;</span><span><span leaf="">3600</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; recency = math.</span><span><span leaf="">pow</span></span><span leaf="">(cfg.decay, hours)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; score = cfg.alpha * relevance + cfg.beta * importance + cfg.gamma * recency</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; scored.append((doc, score))</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; scored.sort(key=</span><span><span leaf="">lambda</span></span><span leaf="">&nbsp;x: x[</span><span><span leaf="">1</span></span><span leaf="">], reverse=</span><span><span leaf="">True</span></span><span leaf="">)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;[doc&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;doc, _&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;scored[:cfg.top_k]]</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">_adjust_weights</span></span><span leaf="">(</span><span><span leaf="">self, query:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, context:&nbsp;</span><span><span leaf="">dict</span></span></span><span leaf="">) -&gt; RetrievalConfig:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""根据查询意图动态调整权重"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; cfg = RetrievalConfig(**</span><span><span leaf="">vars</span></span><span leaf="">(</span><span><span leaf="">self</span></span><span leaf="">.cfg))</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">if</span></span><span leaf="">&nbsp;context&nbsp;</span><span><span leaf="">and</span></span><span leaf="">&nbsp;context.get(</span><span><span leaf="">"intent"</span></span><span leaf="">) ==&nbsp;</span><span><span leaf="">"recent_summary"</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; cfg.gamma, cfg.alpha =&nbsp;</span><span><span leaf="">0.5</span></span><span leaf="">,&nbsp;</span><span><span leaf="">0.3</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">elif</span></span><span leaf="">&nbsp;context&nbsp;</span><span><span leaf="">and</span></span><span leaf="">&nbsp;context.get(</span><span><span leaf="">"intent"</span></span><span leaf="">) ==&nbsp;</span><span><span leaf="">"fact_lookup"</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; cfg.alpha, cfg.gamma =&nbsp;</span><span><span leaf="">0.7</span></span><span leaf="">,&nbsp;</span><span><span leaf="">0.1</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;cfg</span>

2.4.2 分层检索融合（Hierarchical Retrieval Fusion）

不同类型的记忆应该走不同的检索路径，最终在 Working Memory 中融合：

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_core.runnables&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;RunnableParallel, RunnableLambda</span><br><br><span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">HierarchicalMemoryRetriever</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""多层记忆并行检索 + 融合"""</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, episodic_retriever, semantic_store, procedural_store</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.episodic = episodic_retriever &nbsp; &nbsp; &nbsp;</span><span><span leaf=""># 向量检索</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.semantic = semantic_store &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 知识图谱/结构化查询</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.procedural = procedural_store &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 工具/规范查询</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">retrieve_all</span></span><span leaf="">(</span><span><span leaf="">self, query:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, agent_state:&nbsp;</span><span><span leaf="">dict</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">dict</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 并行检索三类记忆（减少延迟）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; parallel = RunnableParallel({</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"episodic"</span></span><span leaf="">: &nbsp; &nbsp;RunnableLambda(</span><span><span leaf="">lambda</span></span><span leaf="">&nbsp;q:&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.episodic.retrieve(q)),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"semantic"</span></span><span leaf="">: &nbsp; &nbsp;RunnableLambda(</span><span><span leaf="">lambda</span></span><span leaf="">&nbsp;q:&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.semantic.query(q)),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"procedural"</span></span><span leaf="">: &nbsp;RunnableLambda(</span><span><span leaf="">lambda</span></span><span leaf="">&nbsp;q:&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.procedural.lookup(q)),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; })</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; raw = parallel.invoke(query)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 融合：根据 token 预算截断，重要性优先</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._merge_with_budget(raw, budget=</span><span><span leaf="">2000</span></span><span leaf="">) &nbsp;</span><span><span leaf=""># token预算</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">_merge_with_budget</span></span><span leaf="">(</span><span><span leaf="">self, memories:&nbsp;</span><span><span leaf="">dict</span></span><span leaf="">, budget:&nbsp;</span><span><span leaf="">int</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">str</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; sections = []</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; remaining = budget</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 优先级：语义记忆（事实可靠）&gt; 情节记忆（相关经验）&gt; 过程记忆（工具使用）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; order = [</span><span><span leaf="">"semantic"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"episodic"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"procedural"</span></span><span leaf="">]</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;key&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;order:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;item&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;memories.get(key, []):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; text =&nbsp;</span><span><span leaf="">f"[</span><span><span leaf="">{key}</span></span><span leaf="">]&nbsp;</span><span><span leaf="">{item}</span></span><span leaf="">"</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; tokens =&nbsp;</span><span><span leaf="">len</span></span><span leaf="">(text) //&nbsp;</span><span><span leaf="">4</span></span><span leaf="">&nbsp;&nbsp;</span><span><span leaf=""># 粗略 token 估算</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">if</span></span><span leaf="">&nbsp;remaining - tokens &gt;&nbsp;</span><span><span leaf="">0</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; sections.append(text)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; remaining -= tokens</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">else</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">break</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;</span><span><span leaf="">"\n\n"</span></span><span leaf="">.join(sections)</span>

2.5 记忆整合（Consolidation）：从情节到语义的升华

Consolidation 是记忆系统中最具技术价值的机制——将大量碎片化的情节记忆提炼成高层的语义记忆，类比人类睡眠中的记忆巩固过程。

这个过程分两步：

1. 摘要压缩（Summarization）：将 N 条相关情节记忆合并为一条压缩摘要，降低存储成本
2. 反思提炼（Reflection）：对压缩摘要做更高层的归纳，提取行为规律和洞见

<span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">MemoryConsolidator</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""</span><br><span leaf="">&nbsp; &nbsp; 记忆整合器：情节记忆 → 语义记忆</span><br><span leaf="">&nbsp; &nbsp; 建议通过 cron 定时触发（如每天凌晨），或在情节记忆超过阈值时触发</span><br><span leaf="">&nbsp; &nbsp; """</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; REFLECTION_PROMPT =&nbsp;</span><span><span leaf="">"""</span><br><span leaf="">以下是 Agent 近期的记忆片段，请从中提炼出：</span><br><span leaf="">1. 用户的核心偏好和习惯规律（不超过5条）</span><br><span leaf="">2. 值得注意的重复性问题或错误模式</span><br><span leaf="">3. 可以优化 Agent 未来行为的洞见</span><br><br><span leaf="">记忆片段：</span><br><span leaf="">{memories}</span><br><br><span leaf="">以 JSON 格式输出：{{"preferences": [], "patterns": [], "insights": []}}</span><br><span leaf="">&nbsp; &nbsp; """</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, episodic_store, semantic_store, llm, threshold:&nbsp;</span><span><span leaf="">int</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">50</span></span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.episodic = episodic_store</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.semantic = semantic_store</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.llm = llm</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.threshold = threshold &nbsp;</span><span><span leaf=""># 超过多少条情节记忆触发整合</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">should_consolidate</span></span><span leaf="">(</span><span><span leaf="">self</span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">bool</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.episodic.count() &gt;=&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.threshold</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">consolidate</span></span><span leaf="">(</span><span><span leaf="">self, entity_id:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 1. 取出待整合的情节记忆（按时间分组）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; episodes =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.episodic.fetch_recent(entity_id, limit=</span><span><span leaf="">self</span></span><span leaf="">.threshold)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 2. 摘要压缩（多条 → 1条）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; combined_text =&nbsp;</span><span><span leaf="">"\n"</span></span><span leaf="">.join([e.content&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;e&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;episodes])</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; summary =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.llm.invoke(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">f"请将以下记忆片段压缩为一段连贯的摘要，保留关键信息：\n</span><span><span leaf="">{combined_text}</span></span><span leaf="">"</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; ).content</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 3. 反思提炼（提取高层洞见）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; reflection_result =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.llm.invoke(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.REFLECTION_PROMPT.</span><span><span leaf="">format</span></span><span leaf="">(memories=combined_text)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; ).content</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; reflection = json.loads(reflection_result)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 4. 写入语义记忆</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.semantic.upsert(entity_id, {</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"summary"</span></span><span leaf="">: summary,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"preferences"</span></span><span leaf="">: reflection[</span><span><span leaf="">"preferences"</span></span><span leaf="">],</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"patterns"</span></span><span leaf="">: reflection[</span><span><span leaf="">"patterns"</span></span><span leaf="">],</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"insights"</span></span><span leaf="">: reflection[</span><span><span leaf="">"insights"</span></span><span leaf="">],</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"consolidated_at"</span></span><span leaf="">: time.time(),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"source_count"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">len</span></span><span leaf="">(episodes)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; })</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 5. 清理已整合的情节记忆（可选：标记为已整合而非删除，保留溯源）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.episodic.mark_consolidated([e.</span><span><span leaf="">id</span></span><span leaf="">&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;e&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;episodes])</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;reflection</span>

2.6 记忆的生命周期完整流图

<span leaf="">用户输入/工具调用</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; ▼</span><br><span leaf="">&nbsp; ┌─────────────┐</span><br><span leaf="">&nbsp; │ &nbsp;感知缓冲 &nbsp; │ ← 原始输入，仅在当前帧有效</span><br><span leaf="">&nbsp; └──────┬──────┘</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ 流入 Working Memory</span><br><span leaf="">&nbsp; ┌──────▼──────┐</span><br><span leaf="">&nbsp; │ &nbsp;工作记忆 &nbsp; │ ← token 消耗在此发生</span><br><span leaf="">&nbsp; │ &nbsp;(上下文窗口)│</span><br><span leaf="">&nbsp; └──────┬──────┘</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ 超出 token 预算 or 轮次结束</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp;▼</span><br><span leaf="">&nbsp; ┌───────────────────────────────── ─┐</span><br><span leaf="">&nbsp; │ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;Memory Manager &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp;[写路径] &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp;重要性评分 → 过滤 → 结构化提取 &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp; &nbsp; &nbsp;→ 冲突检测 → 向量化 → 落库 &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp;[整合触发]（定时 or 阈值） &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp;情节记忆 → 摘要压缩 → 反思提炼 &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp; &nbsp; &nbsp;→ 写入语义记忆 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│</span><br><span leaf="">&nbsp; │ &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp;[遗忘触发]（定时清理） &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp;TTL 过期 → 删除 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; │ &nbsp;低重要性 + 长时未访问 → 软删除 &nbsp; &nbsp; │</span><br><span leaf="">&nbsp; └────────────────┬──────────────────┘</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;│ 下次请求时读路径</span><br><span leaf="">&nbsp; ┌────────────────▼──────────────────────┐</span><br><span leaf="">&nbsp; │ &nbsp; 三维加权检索（Relevance+Importance+ &nbsp;│</span><br><span leaf="">&nbsp; │ &nbsp; Recency）→ 分层融合 → 注入 Prompt &nbsp; &nbsp;│</span><br><span leaf="">&nbsp; └───────────────────────────────────────┘</span>

2.7 存储选型决策矩阵

不同记忆类型对存储有不同要求，以下是选型决策矩阵：

| 记忆类型

查询模式

一致性要求

推荐存储

备注

工作记忆

顺序读写

强一致

In-memory（Python dict）

无需持久化

| |

情节记忆

语义相似度检索

最终一致

Chroma / Milvus / Weaviate

需支持 metadata 过滤

| |

语义记忆（KV型）

精确 Key 查找

强一致

Redis / DynamoDB

读高性能

| |

语义记忆（关系型）

实体关系图遍历

强一致

Neo4j / NebulaGraph

适合复杂关系推理

| |

过程记忆

分类检索

强一致

PostgreSQL + pgvector

混合查询场景

| |

事件溯源记忆

时序 + 回放

强一致（追加）

Kafka / EventStore

不可变，只追加

pgvector 方案（结构化 + 向量一体） 在中小规模场景下尤为实用，避免多系统维护成本：

<span><span leaf="">-- pgvector：在 Postgres 中直接存向量，同时保留结构化字段</span></span><br><span><span leaf="">CREATE</span></span><span leaf="">&nbsp;EXTENSION IF&nbsp;</span><span><span leaf="">NOT</span></span><span leaf="">&nbsp;</span><span><span leaf="">EXISTS</span></span><span leaf="">&nbsp;vector;</span><br><br><span><span leaf="">CREATE TABLE</span></span><span leaf="">&nbsp;agent_memories (</span><br><span leaf="">&nbsp; &nbsp; id &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;UUID&nbsp;</span><span><span leaf="">PRIMARY KEY</span></span><span leaf="">&nbsp;</span><span><span leaf="">DEFAULT</span></span><span leaf="">&nbsp;gen_random_uuid(),</span><br><span leaf="">&nbsp; &nbsp; entity_id &nbsp;&nbsp;</span><span><span leaf="">VARCHAR</span></span><span leaf="">(</span><span><span leaf="">64</span></span><span leaf="">)&nbsp;</span><span><span leaf="">NOT NULL</span></span><span leaf="">, &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">-- 用户/Agent ID</span></span><br><span leaf="">&nbsp; &nbsp; memory_type&nbsp;</span><span><span leaf="">VARCHAR</span></span><span leaf="">(</span><span><span leaf="">32</span></span><span leaf="">)&nbsp;</span><span><span leaf="">NOT NULL</span></span><span leaf="">, &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;</span><span><span leaf="">-- episodic/semantic/procedural</span></span><br><span leaf="">&nbsp; &nbsp; content &nbsp; &nbsp; TEXT&nbsp;</span><span><span leaf="">NOT NULL</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; embedding &nbsp; VECTOR(</span><span><span leaf="">1536</span></span><span leaf="">), &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;</span><span><span leaf="">-- OpenAI text-embedding-3-small</span></span><br><span leaf="">&nbsp; &nbsp; importance &nbsp;</span><span><span leaf="">FLOAT</span></span><span leaf="">&nbsp;</span><span><span leaf="">DEFAULT</span></span><span leaf="">&nbsp;</span><span><span leaf="">0.5</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">timestamp</span></span><span leaf="">&nbsp; &nbsp;TIMESTAMPTZ&nbsp;</span><span><span leaf="">DEFAULT</span></span><span leaf="">&nbsp;NOW(),</span><br><span leaf="">&nbsp; &nbsp; expires_at &nbsp;TIMESTAMPTZ, &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">-- TTL 过期时间</span></span><br><span leaf="">&nbsp; &nbsp; is_active &nbsp;&nbsp;</span><span><span leaf="">BOOLEAN</span></span><span leaf="">&nbsp;</span><span><span leaf="">DEFAULT</span></span><span leaf="">&nbsp;</span><span><span leaf="">TRUE</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; metadata &nbsp; &nbsp;JSONB&nbsp;</span><span><span leaf="">DEFAULT</span></span><span leaf="">&nbsp;</span><span><span leaf="">'{}'</span></span><br><span leaf="">);</span><br><br><span><span leaf="">-- 向量索引（HNSW，低延迟近似最近邻）</span></span><br><span><span leaf="">CREATE</span></span><span leaf="">&nbsp;INDEX&nbsp;</span><span><span leaf="">ON</span></span><span leaf="">&nbsp;agent_memories</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">USING</span></span><span leaf="">&nbsp;hnsw (embedding vector_cosine_ops)</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">WITH</span></span><span leaf="">&nbsp;(m&nbsp;</span><span><span leaf="">=</span></span><span leaf="">&nbsp;</span><span><span leaf="">16</span></span><span leaf="">, ef_construction&nbsp;</span><span><span leaf="">=</span></span><span leaf="">&nbsp;</span><span><span leaf="">64</span></span><span leaf="">);</span><br><br><span><span leaf="">-- 复合索引：快速过滤特定实体的活跃记忆</span></span><br><span><span leaf="">CREATE</span></span><span leaf="">&nbsp;INDEX&nbsp;</span><span><span leaf="">ON</span></span><span leaf="">&nbsp;agent_memories (entity_id, memory_type, is_active,&nbsp;</span><span><span leaf="">timestamp</span></span><span leaf="">&nbsp;</span><span><span leaf="">DESC</span></span><span leaf="">);</span><br><br><span><span leaf="">-- 三维加权检索（SQL 实现）</span></span><br><span><span leaf="">SELECT</span></span><br><span leaf="">&nbsp; &nbsp; content,</span><br><span leaf="">&nbsp; &nbsp; importance,</span><br><span leaf="">&nbsp; &nbsp; (</span><span><span leaf="">1</span></span><span leaf="">&nbsp;</span><span><span leaf="">-</span></span><span leaf="">&nbsp;(embedding&nbsp;</span><span><span leaf="">&lt;=&gt;</span></span><span leaf="">&nbsp;$query_embedding))&nbsp;</span><span><span leaf="">AS</span></span><span leaf="">&nbsp;relevance, &nbsp;&nbsp;</span><span><span leaf="">-- 余弦相似度</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">EXP</span></span><span leaf="">(</span><span><span leaf="">-0.01</span></span><span leaf="">&nbsp;</span><span><span leaf="">*</span></span><span leaf="">&nbsp;</span><span><span leaf="">EXTRACT</span></span><span leaf="">(EPOCH&nbsp;</span><span><span leaf="">FROM</span></span><span leaf="">&nbsp;(NOW()&nbsp;</span><span><span leaf="">-</span></span><span leaf="">&nbsp;</span><span><span leaf="">timestamp</span></span><span leaf="">))&nbsp;</span><span><span leaf="">/</span></span><span leaf="">&nbsp;</span><span><span leaf="">3600</span></span><span leaf="">)&nbsp;</span><span><span leaf="">AS</span></span><span leaf="">&nbsp;recency,</span><br><span leaf="">&nbsp; &nbsp; (</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">0.5</span></span><span leaf="">&nbsp;</span><span><span leaf="">*</span></span><span leaf="">&nbsp;(</span><span><span leaf="">1</span></span><span leaf="">&nbsp;</span><span><span leaf="">-</span></span><span leaf="">&nbsp;(embedding&nbsp;</span><span><span leaf="">&lt;=&gt;</span></span><span leaf="">&nbsp;$query_embedding))&nbsp;</span><span><span leaf="">+</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">0.3</span></span><span leaf="">&nbsp;</span><span><span leaf="">*</span></span><span leaf="">&nbsp;importance&nbsp;</span><span><span leaf="">+</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">0.2</span></span><span leaf="">&nbsp;</span><span><span leaf="">*</span></span><span leaf="">&nbsp;</span><span><span leaf="">EXP</span></span><span leaf="">(</span><span><span leaf="">-0.01</span></span><span leaf="">&nbsp;</span><span><span leaf="">*</span></span><span leaf="">&nbsp;</span><span><span leaf="">EXTRACT</span></span><span leaf="">(EPOCH&nbsp;</span><span><span leaf="">FROM</span></span><span leaf="">&nbsp;(NOW()&nbsp;</span><span><span leaf="">-</span></span><span leaf="">&nbsp;</span><span><span leaf="">timestamp</span></span><span leaf="">))&nbsp;</span><span><span leaf="">/</span></span><span leaf="">&nbsp;</span><span><span leaf="">3600</span></span><span leaf="">)</span><br><span leaf="">&nbsp; &nbsp; )&nbsp;</span><span><span leaf="">AS</span></span><span leaf="">&nbsp;final_score</span><br><span><span leaf="">FROM</span></span><span leaf="">&nbsp;agent_memories</span><br><span><span leaf="">WHERE</span></span><span leaf="">&nbsp;entity_id&nbsp;</span><span><span leaf="">=</span></span><span leaf="">&nbsp;$entity_id</span><br><span leaf="">&nbsp;&nbsp;</span><span><span leaf="">AND</span></span><span leaf="">&nbsp;is_active&nbsp;</span><span><span leaf="">=</span></span><span leaf="">&nbsp;</span><span><span leaf="">TRUE</span></span><br><span leaf="">&nbsp;&nbsp;</span><span><span leaf="">AND</span></span><span leaf="">&nbsp;(expires_at&nbsp;</span><span><span leaf="">IS</span></span><span leaf="">&nbsp;</span><span><span leaf="">NULL</span></span><span leaf="">&nbsp;</span><span><span leaf="">OR</span></span><span leaf="">&nbsp;expires_at&nbsp;</span><span><span leaf="">&gt;</span></span><span leaf="">&nbsp;NOW())</span><br><span><span leaf="">ORDER</span></span><span leaf="">&nbsp;</span><span><span leaf="">BY</span></span><span leaf="">&nbsp;final_score&nbsp;</span><span><span leaf="">DESC</span></span><br><span leaf="">LIMIT&nbsp;</span><span><span leaf="">5</span></span><span leaf="">;</span>

三、核心技术实现（LangChain 示例）

3.1 工作记忆：ConversationBufferMemory

最基础的记忆形式，将历史对话塞进 Prompt。

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain.memory&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;ConversationBufferMemory</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;langchain.chains&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;ConversationChain</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_openai&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;ChatOpenAI</span><br><br><span leaf="">llm = ChatOpenAI(model=</span><span><span leaf="">"gpt-4o"</span></span><span leaf="">)</span><br><br><span leaf="">memory = ConversationBufferMemory(</span><br><span leaf="">&nbsp; &nbsp; memory_key=</span><span><span leaf="">"chat_history"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; return_messages=</span><span><span leaf="">True</span></span><br><span leaf="">)</span><br><br><span leaf="">chain = ConversationChain(</span><br><span leaf="">&nbsp; &nbsp; llm=llm,</span><br><span leaf="">&nbsp; &nbsp; memory=memory,</span><br><span leaf="">&nbsp; &nbsp; verbose=</span><span><span leaf="">True</span></span><br><span leaf="">)</span><br><br><span leaf="">chain.invoke({</span><span><span leaf="">"input"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"我叫张三，是一名后端工程师"</span></span><span leaf="">})</span><br><span leaf="">chain.invoke({</span><span><span leaf="">"input"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"帮我写一个 Python 快速排序"</span></span><span leaf="">})</span><br><span><span leaf=""># 第二轮调用时，模型仍然知道用户是张三</span></span>

缺点： 上下文无限增长，超出 token 限制后崩溃。

3.2 滑动窗口记忆：ConversationBufferWindowMemory

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain.memory&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;ConversationBufferWindowMemory</span><br><br><span leaf="">memory = ConversationBufferWindowMemory(</span><br><span leaf="">&nbsp; &nbsp; k=</span><span><span leaf="">10</span></span><span leaf="">, &nbsp;</span><span><span leaf=""># 只保留最近 10 轮对话</span></span><br><span leaf="">&nbsp; &nbsp; memory_key=</span><span><span leaf="">"chat_history"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; return_messages=</span><span><span leaf="">True</span></span><br><span leaf="">)</span>

适用场景： 对话轮次多、上下文时效性高（近期对话更重要）的场景。

3.3 摘要记忆：ConversationSummaryMemory

超过一定长度后，用 LLM 对历史进行压缩摘要：

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain.memory&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;ConversationSummaryMemory</span><br><br><span leaf="">memory = ConversationSummaryMemory(</span><br><span leaf="">&nbsp; &nbsp; llm=llm,</span><br><span leaf="">&nbsp; &nbsp; memory_key=</span><span><span leaf="">"chat_history"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; return_messages=</span><span><span leaf="">True</span></span><br><span leaf="">)</span><br><br><span><span leaf=""># 当对话积累到一定量，memory 会自动调用 LLM 生成摘要：</span></span><br><span><span leaf=""># "用户自我介绍为张三，后端工程师，讨论了快速排序实现，</span></span><br><span><span leaf=""># &nbsp;并询问了 Redis 缓存最佳实践..."</span></span>

3.4 向量检索记忆：VectorStoreRetrieverMemory

这是 Agent 记忆系统最重要的组件之一——通过语义相似度检索历史记忆：

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain.memory&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;VectorStoreRetrieverMemory</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_community.vectorstores&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;Chroma</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_openai&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;OpenAIEmbeddings</span><br><br><span><span leaf=""># 初始化向量数据库</span></span><br><span leaf="">embedding = OpenAIEmbeddings()</span><br><span leaf="">vectorstore = Chroma(</span><br><span leaf="">&nbsp; &nbsp; collection_name=</span><span><span leaf="">"agent_memory"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; embedding_function=embedding,</span><br><span leaf="">&nbsp; &nbsp; persist_directory=</span><span><span leaf="">"./memory_db"</span></span><br><span leaf="">)</span><br><br><span leaf="">retriever = vectorstore.as_retriever(</span><br><span leaf="">&nbsp; &nbsp; search_kwargs={</span><span><span leaf="">"k"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">5</span></span><span leaf="">} &nbsp;</span><span><span leaf=""># 每次检索最相关的 5 条记忆</span></span><br><span leaf="">)</span><br><br><span leaf="">memory = VectorStoreRetrieverMemory(</span><br><span leaf="">&nbsp; &nbsp; retriever=retriever,</span><br><span leaf="">&nbsp; &nbsp; memory_key=</span><span><span leaf="">"relevant_history"</span></span><br><span leaf="">)</span><br><br><span><span leaf=""># 存入记忆</span></span><br><span leaf="">memory.save_context(</span><br><span leaf="">&nbsp; &nbsp; inputs={</span><span><span leaf="">"input"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"用户偏好：喜欢简洁的代码风格，讨厌过度注释"</span></span><span leaf="">},</span><br><span leaf="">&nbsp; &nbsp; outputs={</span><span><span leaf="">"output"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"已记录用户编码偏好"</span></span><span leaf="">}</span><br><span leaf="">)</span><br><br><span><span leaf=""># 检索：当用户问到代码风格时，自动召回相关记忆</span></span><br><span leaf="">relevant = memory.load_memory_variables(</span><br><span leaf="">&nbsp; &nbsp; inputs={</span><span><span leaf="">"input"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"帮我重构这段代码"</span></span><span leaf="">}</span><br><span leaf="">)</span><br><span><span leaf=""># 自动返回「用户偏好：喜欢简洁的代码风格」相关记忆</span></span>

3.5 组合记忆：CombinedMemory

生产环境中通常需要多种记忆类型组合使用：

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain.memory&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;CombinedMemory, ConversationSummaryMemory</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;langchain.memory&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;VectorStoreRetrieverMemory</span><br><br><span><span leaf=""># 短期：摘要记忆（近期对话）</span></span><br><span leaf="">summary_memory = ConversationSummaryMemory(</span><br><span leaf="">&nbsp; &nbsp; llm=llm,</span><br><span leaf="">&nbsp; &nbsp; input_key=</span><span><span leaf="">"input"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; memory_key=</span><span><span leaf="">"recent_context"</span></span><br><span leaf="">)</span><br><br><span><span leaf=""># 长期：向量记忆（语义检索历史）</span></span><br><span leaf="">vector_memory = VectorStoreRetrieverMemory(</span><br><span leaf="">&nbsp; &nbsp; retriever=retriever,</span><br><span leaf="">&nbsp; &nbsp; input_key=</span><span><span leaf="">"input"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; memory_key=</span><span><span leaf="">"long_term_context"</span></span><br><span leaf="">)</span><br><br><span leaf="">combined = CombinedMemory(memories=[summary_memory, vector_memory])</span>

3.6 自定义重要性评分记忆

Generative Agents 论文提出的重要性 + 时效性 + 相关性三维加权检索：

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_core.documents&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;Document</span><br><span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_community.vectorstores&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;FAISS</span><br><span><span leaf="">import</span></span><span leaf="">&nbsp;time</span><br><span><span leaf="">import</span></span><span leaf="">&nbsp;math</span><br><br><span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">ScoredMemory</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""带重要性评分和时效衰减的记忆系统"""</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, llm, embedding</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.llm = llm</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore = FAISS.from_texts([</span><span><span leaf="">"init"</span></span><span leaf="">], embedding)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.memories = [] &nbsp;</span><span><span leaf=""># 元数据存储</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">_importance_score</span></span><span leaf="">(</span><span><span leaf="">self, text:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">float</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""让 LLM 评估记忆重要性 (1-10)"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; prompt =&nbsp;</span><span><span leaf="">f"""评估以下信息的重要性，返回 1-10 的整数，只返回数字：</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; 信息：</span><span><span leaf="">{text}</span></span><span leaf="">"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; score =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.llm.invoke(prompt).content</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;</span><span><span leaf="">float</span></span><span leaf="">(score.strip()) /&nbsp;</span><span><span leaf="">10.0</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">_recency_score</span></span><span leaf="">(</span><span><span leaf="">self, timestamp:&nbsp;</span><span><span leaf="">float</span></span><span leaf="">, decay_factor:&nbsp;</span><span><span leaf="">float</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">0.99</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">float</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""时效性衰减分数"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; hours_passed = (time.time() - timestamp) /&nbsp;</span><span><span leaf="">3600</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;math.</span><span><span leaf="">pow</span></span><span leaf="">(decay_factor, hours_passed)</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">add_memory</span></span><span leaf="">(</span><span><span leaf="">self, text:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; importance =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._importance_score(text)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; metadata = {</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"timestamp"</span></span><span leaf="">: time.time(),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"importance"</span></span><span leaf="">: importance,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"text"</span></span><span leaf="">: text</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; }</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.memories.append(metadata)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.add_texts([text], metadatas=[metadata])</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">retrieve</span></span><span leaf="">(</span><span><span leaf="">self, query:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, top_k:&nbsp;</span><span><span leaf="">int</span></span><span leaf="">&nbsp;=&nbsp;</span><span><span leaf="">5</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">list</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""综合相关性 + 重要性 + 时效性排序"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; docs_scores =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.similarity_search_with_score(query, k=</span><span><span leaf="">20</span></span><span leaf="">)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; results = []</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;doc, relevance_score&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;docs_scores:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; meta = doc.metadata</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; recency =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">._recency_score(meta[</span><span><span leaf="">"timestamp"</span></span><span leaf="">])</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; importance = meta[</span><span><span leaf="">"importance"</span></span><span leaf="">]</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 综合得分（权重可调）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; final_score = (</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">0.5</span></span><span leaf="">&nbsp;* (</span><span><span leaf="">1</span></span><span leaf="">&nbsp;- relevance_score) + &nbsp;</span><span><span leaf=""># 相关性（转为正向）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">0.3</span></span><span leaf="">&nbsp;* importance + &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 重要性</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">0.2</span></span><span leaf="">&nbsp;* recency &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;</span><span><span leaf=""># 时效性</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; results.append((doc.page_content, final_score))</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; results.sort(key=</span><span><span leaf="">lambda</span></span><span leaf="">&nbsp;x: x[</span><span><span leaf="">1</span></span><span leaf="">], reverse=</span><span><span leaf="">True</span></span><span leaf="">)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;results[:top_k]</span>

四、行业通用解决方案

4.1 客服与对话场景

问题： 客服 Agent 需要记住同一用户在多次咨询中描述的问题、偏好，避免用户重复陈述。

解决方案架构：

<span leaf="">用户请求</span><br><span leaf="">&nbsp; &nbsp; │</span><br><span leaf="">&nbsp; &nbsp; ▼</span><br><span leaf="">用户身份识别（user_id）</span><br><span leaf="">&nbsp; &nbsp; │</span><br><span leaf="">&nbsp; &nbsp; ├─► 加载短期记忆（本次会话摘要）</span><br><span leaf="">&nbsp; &nbsp; ├─► 检索长期记忆（该用户历史问题向量检索）</span><br><span leaf="">&nbsp; &nbsp; └─► 检索用户画像（结构化：偏好、购买记录、等级）</span><br><span leaf="">&nbsp; &nbsp; │</span><br><span leaf="">&nbsp; &nbsp; ▼</span><br><span leaf="">LLM 生成回复（融合三类记忆）</span><br><span leaf="">&nbsp; &nbsp; │</span><br><span leaf="">&nbsp; &nbsp; ▼</span><br><span leaf="">记忆写回（重要信息提取 → 向量化存储）</span>

<span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">CustomerServiceMemory</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""客服场景记忆系统"""</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, user_id:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, vectorstore, db</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.user_id = user_id</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore = vectorstore</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.db = db &nbsp;</span><span><span leaf=""># 结构化用户画像数据库</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">load_context</span></span><span leaf="">(</span><span><span leaf="">self, current_query:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">dict</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 1. 结构化画像</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; profile =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.db.get_user_profile(</span><span><span leaf="">self</span></span><span leaf="">.user_id)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 2. 语义检索历史问题（相关问题优先）</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; collection =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.as_retriever(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; search_kwargs={</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"k"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">3</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"filter"</span></span><span leaf="">: {</span><span><span leaf="">"user_id"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.user_id}</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; }</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; history_docs = collection.invoke(current_query)</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;{</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"user_profile"</span></span><span leaf="">: profile,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"relevant_history"</span></span><span leaf="">: [d.page_content&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;d&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;history_docs]</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; }</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">save_interaction</span></span><span leaf="">(</span><span><span leaf="">self, query:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, response:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, summary:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf=""># 只存摘要，减少存储成本</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.add_texts(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; [summary],</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; metadatas=[{</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"user_id"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.user_id,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"timestamp"</span></span><span leaf="">: time.time(),</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"type"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"interaction_summary"</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; }]</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; )</span>

4.2 代码开发辅助场景

问题： 开发助手 Agent 需要记住项目结构、技术栈、团队编码规范、历史 Bug 经验。

解决方案： 分层记忆设计

<span><span leaf="">class</span></span><span leaf="">&nbsp;</span><span><span leaf="">DevAssistantMemory</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""开发助手记忆系统"""</span></span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp; MEMORY_TYPES = {</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"project_context"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"项目元数据（技术栈、架构）"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"code_pattern"</span></span><span leaf="">: &nbsp; &nbsp;</span><span><span leaf="">"编码模式与最佳实践"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"bug_experience"</span></span><span leaf="">: &nbsp;</span><span><span leaf="">"历史 Bug 与修复经验"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"team_convention"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"团队规范与约定"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; }</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">__init__</span></span><span leaf="">(</span><span><span leaf="">self, project_id:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, vectorstore</span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.project_id = project_id</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore = vectorstore</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">add_bug_experience</span></span><span leaf="">(</span><span><span leaf="">self, bug_desc:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, root_cause:&nbsp;</span><span><span leaf="">str</span></span><span leaf="">, fix:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">):</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""记录 Bug 修复经验，供未来类似问题检索"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; content =&nbsp;</span><span><span leaf="">f"""</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; Bug描述：</span><span><span leaf="">{bug_desc}</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; 根本原因：</span><span><span leaf="">{root_cause}</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; 修复方案：</span><span><span leaf="">{fix}</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; """</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.add_texts(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; [content],</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; metadatas=[{</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"project_id"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.project_id,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"type"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"bug_experience"</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"timestamp"</span></span><span leaf="">: time.time()</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; }]</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><br><span leaf="">&nbsp; &nbsp;&nbsp;</span><span><span leaf="">def</span></span><span leaf="">&nbsp;</span><span><span leaf="">retrieve_relevant_experience</span></span><span leaf="">(</span><span><span leaf="">self, current_issue:&nbsp;</span><span><span leaf="">str</span></span></span><span leaf="">) -&gt;&nbsp;</span><span><span leaf="">list</span></span><span leaf="">:</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"""遇到新问题时，检索历史经验"""</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; docs =&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.vectorstore.similarity_search(</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; current_issue,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; k=</span><span><span leaf="">3</span></span><span leaf="">,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">filter</span></span><span leaf="">={</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"project_id"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">self</span></span><span leaf="">.project_id,</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">"type"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"bug_experience"</span></span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; }</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp; )</span><br><span leaf="">&nbsp; &nbsp; &nbsp; &nbsp;&nbsp;</span><span><span leaf="">return</span></span><span leaf="">&nbsp;[d.page_content&nbsp;</span><span><span leaf="">for</span></span><span leaf="">&nbsp;d&nbsp;</span><span><span leaf="">in</span></span><span leaf="">&nbsp;docs]</span>

五、相关技术对比与关联

5.1 记忆系统 vs RAG（检索增强生成）

| 维度

记忆系统

RAG

数据来源

交互历史、Agent 自身经验

外部静态知识库

| |

更新频率

实时、持续更新

批量离线更新

| |

个性化

用户/Agent 专属

通常共享

| |

遗忘机制

需要（TTL、重要性淘汰）

通常不需要

| |

目标

经验积累与个性化

知识注入

结合使用： 记忆系统 + RAG 是互补的：RAG 提供领域知识底座，记忆系统提供个人化、历史化的动态上下文。

5.2 与知识图谱的关系

知识图谱可以作为语义记忆的存储后端，提供结构化的实体关系记忆：

<span><span leaf="">from</span></span><span leaf="">&nbsp;langchain_community.graphs&nbsp;</span><span><span leaf="">import</span></span><span leaf="">&nbsp;Neo4jGraph</span><br><br><span leaf="">graph = Neo4jGraph(url=</span><span><span leaf="">"bolt://localhost:7687"</span></span><span leaf="">, ...)</span><br><br><span><span leaf=""># 将记忆以图结构存储：用户 → 偏好 → 技术栈</span></span><br><span leaf="">graph.query(</span><span><span leaf="">"""</span><br><span leaf="">&nbsp; &nbsp; MERGE (u:User {id: $user_id})</span><br><span leaf="">&nbsp; &nbsp; MERGE (t:Technology {name: $tech})</span><br><span leaf="">&nbsp; &nbsp; MERGE (u)-[:PREFERS {strength: $strength}]-&gt;(t)</span><br><span leaf="">"""</span></span><span leaf="">, {</span><span><span leaf="">"user_id"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"u001"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"tech"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">"Python"</span></span><span leaf="">,&nbsp;</span><span><span leaf="">"strength"</span></span><span leaf="">:&nbsp;</span><span><span leaf="">0.9</span></span><span leaf="">})</span>

5.3 与 ReAct / Plan-and-Execute 的关系

记忆系统是 Agent 推理框架的基础设施：

• ReAct Agent：每步 Thought/Action/Observation 可以写入情节记忆，下次遇到类似任务直接复用推理链
• Plan-and-Execute：计划阶段检索历史相似任务的执行经验，优化规划质量

六、前沿进展与发展展望

6.1 当前主要研究方向

① Generative Agents（斯坦福，2023）
提出完整的 Agent 记忆架构：记忆流（Memory Stream）+ 反思（Reflection）+ 计划（Planning）。反思机制让 Agent 定期对记忆进行高层归纳，形成抽象洞见。

② MemGPT（2023）
借鉴操作系统虚拟内存思想，把 LLM 上下文窗口视为内存，外部存储视为磁盘，Agent 自主管理内存的换入换出。

③ A-MEM（2024）
动态记忆网络，让 Agent 在记忆之间建立关联链接，类似人类联想记忆。

6.2 技术挑战

| 挑战

描述

研究方向

记忆一致性

相互矛盾的记忆如何处理

记忆冲突检测与合并

| |

隐私安全

记忆可能包含敏感信息

差分隐私、记忆脱敏

| |

幻觉记忆

Agent 可能"记住"从未发生的事

记忆验证机制

| |

跨 Agent 记忆共享

多 Agent 协作时记忆同步

分布式记忆协议

| |

长期记忆压缩质量

LLM 摘要可能丢失关键细节

结构化提取 + 向量双存

6.3 发展趋势

近期：

• 记忆系统框架标准化（LangChain Memory v2、LlamaIndex Memory Modules）
• 向量数据库与关系数据库的融合（如 pgvector）
• 端侧轻量记忆系统（手机/IoT Agent）

中期：

• 记忆与微调的融合：重要记忆直接更新模型权重（LoRA 动态更新）
• 跨模态记忆：图像、音频、代码统一向量空间
• 联邦记忆：多端记忆隐私保护同步

长期展望：

• 真正连续学习的 Agent：记忆系统 + 在线学习，Agent 在使用中持续进化
• 记忆作为 Agent 的"人格"基础：稳定、可解释、可审计

七、总结

Agent 记忆系统是让 AI 从"工具"走向"伙伴"的关键基础设施。

| 维度

核心要点

架构

工作记忆 + 情节记忆 + 语义记忆 + 过程记忆分层设计

| |

存储

向量数据库（语义检索）+ 关系数据库（结构化）+ KG（关系网络）

| |

检索

相关性 × 重要性 × 时效性三维加权

| |

遗忘

TTL + LRU + 重要性阈值，避免记忆膨胀

| |

工程

LangChain Memory 模块提供开箱即用实现

| |

未来

迈向持续学习、跨模态、隐私保护的统一记忆平台

记忆系统的成熟，将使 Agent 真正具备"经验”——不再是每次归零的问答机器，而是能从每一次交互中成长的智能实体。

持续更新中，关注我获取更多干货

文章目录