[{"data":1,"prerenderedAt":396},["ShallowReactive",2],{"/cn/open_source/home/memos_intro":3},{"id":4,"title":5,"avatar":6,"banner":6,"body":7,"category":6,"desc":389,"description":374,"extension":390,"links":6,"meta":391,"navigation":6,"path":392,"seo":393,"stem":394,"__hash__":395},"docs/cn/open_source/home/memos_intro.md","什么是 MemOS？",null,{"type":8,"value":9,"toc":373},"minimark",[10,18,26,29,43,46,58,63,66,75,85,88,93,99,102,105,127,134,137,140,160,170,173,180,242,245,256,260,277,281,307,310,327,332,335],[11,12,13],"p",{},[14,15],"img",{"alt":16,"src":17},"MemOS Architecture","https://statics.memtensor.com.cn/memos/memos-architecture.png",[11,19,20,21,25],{},"随着 LLMs 的发展，它们需要处理复杂任务——如多轮对话、长期规划、决策制定和个性化用户体验——",[22,23,24],"strong",{},"赋予它们结构化、管理和演进记忆的能力","对于实现真正的长期智能和适应性变得至关重要。",[11,27,28],{},"然而，大多数主流 LLMs 仍然严重依赖静态参数化记忆（模型权重）。这使得更新知识、跟踪记忆使用或积累演进的用户偏好变得困难。结果是什么？刷新知识成本高、行为脆弱以及个性化有限。",[11,30,31,34,35,38,39,42],{},[22,32,33],{},"MemOS"," 通过将记忆重新定义为具有统一结构、生命周期管理和调度逻辑的",[22,36,37],{},"核心模块化系统资源","来解决这些挑战。它提供了一个基于 Python 的层，位于您的 LLM 和外部知识源之间，实现",[22,40,41],{},"持久化、结构化和高效的记忆操作","。",[11,44,45],{},"使用 MemOS，您的 LLM 可以随时间保留知识，更稳健地管理上下文，并使用可解释和可审计的记忆进行推理——解锁更智能、可靠和自适应的 AI 行为。",[47,48,49],"note",{},[11,50,51,54,57],{},[22,52,53],{},"提示",[55,56],"br",{},"  MemOS 帮助弥合静态参数化权重和动态、用户特定记忆之间的差距。\n将其视为您代理的\"大脑\"，具有明文和激活记忆的即插即用模块。",[59,60,62],"h2",{"id":61},"为什么我们需要memos","为什么我们需要MemOS？",[11,64,65],{},"LLMs 强大，但严重依赖参数化记忆（权重），这些权重难以检查、更新或共享。\n典型的向量搜索 (RAG) 有助于检索外部事实，但缺乏统一治理、生命周期控制或跨代理共享。",[11,67,68,70,71,74],{},[22,69,33],{}," 改变了这一点。\n将其视为记忆的操作系统：\n就像操作系统调度 CPU、RAM 和文件一样，MemOS ",[22,72,73],{},"调度、转换和治理","多种记忆类型——从参数化权重到临时缓存再到明文、可追溯的知识。",[47,76,77],{},[11,78,79,82,84],{},[22,80,81],{},"深入理解",[55,83],{},"  MemOS 通过将参数化、激活和明文记忆融合到生命周期中，帮助您的 LLM 演进。",[59,86,87],{"id":87},"核心构建模块",[89,90,92],"h3",{"id":91},"memcubes","MemCubes",[11,94,95,98],{},[22,96,97],{},"灵活的容器","，容纳一种或多种记忆类型。\n每个用户、会话或代理都可以有自己的 MemCube——可交换、可重用和可追溯。",[89,100,101],{"id":101},"记忆生命周期",[11,103,104],{},"每个记忆单元可以流经以下状态：",[106,107,108],"ul",{},[109,110,111,114,115,114,118,114,121,114,124],"li",{},[22,112,113],{},"生成"," → ",[22,116,117],{},"激活",[22,119,120],{},"合并",[22,122,123],{},"归档",[22,125,126],{},"冻结",[11,128,129,130,133],{},"每个步骤都通过",[22,131,132],{},"来源跟踪","和审计日志进行版本控制。旧记忆可以\"时间机器\"回到之前的版本进行恢复或反事实模拟。",[89,135,136],{"id":136},"操作与治理",[11,138,139],{},"模块包括：",[106,141,142,148,154],{},[109,143,144,147],{},[22,145,146],{},"MemScheduler"," — 动态转换记忆类型以实现最佳复用。",[109,149,150,153],{},[22,151,152],{},"MemLifecycle"," — 管理状态转换、合并和归档。",[109,155,156,159],{},[22,157,158],{},"MemGovernance"," — 处理访问控制、编辑、合规性和审计跟踪。",[47,161,162],{},[11,163,164,167,169],{},[22,165,166],{},"合规提醒",[55,168],{},"    每个记忆单元都携带完整的来源元数据，因此您可以审计谁创建、修改或查询了它。",[59,171,172],{"id":172},"多视角记忆",[11,174,175,176,179],{},"MemOS 在生命周期中融合",[22,177,178],{},"三种记忆形式","：",[181,182,183,199],"table",{},[184,185,186],"thead",{},[187,188,189,193,196],"tr",{},[190,191,192],"th",{},"类型",[190,194,195],{},"描述",[190,197,198],{},"用例",[200,201,202,216,229],"tbody",{},[187,203,204,210,213],{},[205,206,207],"td",{},[22,208,209],{},"参数记忆",[205,211,212],{},"知识提炼到模型权重中",[205,214,215],{},"常青技能、稳定领域事实",[187,217,218,223,226],{},[205,219,220],{},[22,221,222],{},"激活记忆",[205,224,225],{},"用于推理复用的 KV caches 和隐藏状态",[205,227,228],{},"快速多轮聊天、低延迟生成",[187,230,231,236,239],{},[205,232,233],{},[22,234,235],{},"明文记忆",[205,237,238],{},"文本、文档、图、向量块、用户可见事实",[205,240,241],{},"语义搜索、演进、可解释记忆",[11,243,244],{},"随着时间的推移：",[106,246,247,250,253],{},[109,248,249],{},"频繁使用的明文记忆可以提炼为参数化权重。",[109,251,252],{},"稳定的上下文被提升为 KV cache 以快速注入。",[109,254,255],{},"使用频率低或过时的知识可以被降级。",[59,257,259],{"id":258},"memos-有什么不同","MemOS 有什么不同？",[106,261,262,265,268,271,274],{},[109,263,264],{},"混合检索 — 符号和语义混合检索、向量和图混合检索。",[109,266,267],{},"多代理和多用户图 — 私有和共享。",[109,269,270],{},"来源和审计跟踪 — 每个记忆单元都被治理和可解释。",[109,272,273],{},"自动 KV cache 提升以重用稳定上下文。",[109,275,276],{},"记忆的生命周期调度 — 减少陈旧事实或臃肿权重的调用。",[59,278,280],{"id":279},"适合谁","适合谁？",[106,282,283,290,297,304],{},[109,284,285,286,289],{},"需要",[22,287,288],{},"多轮、演进记忆","的对话代理",[109,291,292,293,296],{},"处理",[22,294,295],{},"合规性、领域更新和个性化","的企业级 Copilot",[109,298,299,300,303],{},"在",[22,301,302],{},"共享知识图","上协作的多代理系统",[109,305,306],{},"想要模块化、可查记忆而不是黑盒提示的 AI 构建者",[59,308,309],{"id":309},"关键要点",[11,311,312,314,315,318,319,322,323,326],{},[22,313,33],{}," 将您的 LLM 从\"只是预测 tokens\"\n升级为可以",[22,316,317],{},"记忆","、",[22,320,321],{},"推理","和",[22,324,325],{},"适应","的智能演进系统——\n就像您代理思维的操作系统。",[11,328,329],{},[22,330,331],{},"使用 MemOS，您的 AI 不仅仅是存储事实——它在成长。",[59,333,334],{"id":334},"主要特性",[106,336,337,343,349,355,361,367],{},[109,338,339,342],{},[22,340,341],{},"模块化记忆架构",": 支持明文、激活 (KV cache) 和参数化 (adapters/LoRA) 记忆。",[109,344,345,348],{},[22,346,347],{},"MemCube",": 所有记忆类型的统一容器，具有简单的加载/保存和 API 访问。",[109,350,351,354],{},[22,352,353],{},"MOS",": 面向 LLMs 的记忆增强系统，具有即插即用的记忆模块。",[109,356,357,360],{},[22,358,359],{},"基于图的后端",": 原生支持 Neo4j 和其他图数据库，用于结构化、可解释的记忆。",[109,362,363,366],{},[22,364,365],{},"易于集成",": 可与 HuggingFace、Ollama 和自定义 LLMs 配合使用。",[109,368,369,372],{},[22,370,371],{},"可扩展",": 添加您自己的记忆模块或后端。",{"title":374,"searchDepth":375,"depth":375,"links":376},"",2,[377,378,384,385,386,387,388],{"id":61,"depth":375,"text":62},{"id":87,"depth":375,"text":87,"children":379},[380,382,383],{"id":91,"depth":381,"text":92},3,{"id":101,"depth":381,"text":101},{"id":136,"depth":381,"text":136},{"id":172,"depth":375,"text":172},{"id":258,"depth":375,"text":259},{"id":279,"depth":375,"text":280},{"id":309,"depth":375,"text":309},{"id":334,"depth":375,"text":334},"**MemOS** 是为大语言模型 (LLMs) 和智能体打造的**记忆操作系统**。它将记忆视为**可管理、调度和解释的一级资源**，而不是隐藏在模型权重内部的不透明层。","md",{},"/cn/open_source/home/memos_intro",{"title":5,"description":374},"cn/open_source/home/memos_intro","cU67cHACxyhIBKF3oCR_yD0KAu7JQwfO-FSZEwLRvo8",1758590740983]