[{"data":1,"prerenderedAt":461},["ShallowReactive",2],{"docs-page-cn-/cn/usecase/blog/keep_bad_assumptions_out_of_agent_memory":3,"surround-cn-/cn/usecase/blog/keep_bad_assumptions_out_of_agent_memory":449},{"id":4,"title":5,"avatar":6,"banner":6,"body":7,"category":6,"desc":6,"description":427,"extension":443,"lead":6,"links":6,"meta":444,"navigation":6,"path":445,"seo":446,"stem":447,"__hash__":448},"docs/cn/usecase/blog/keep_bad_assumptions_out_of_agent_memory.md","如何让错误假设远离 Agent 记忆",null,{"type":8,"value":9,"toc":426},"minimark",[10,21,24,27,30,44,47,50,53,58,66,69,72,75,78,80,84,87,90,93,145,148,151,153,157,165,167,171,186,189,192,194,197,206,209,211,214,229,238,240,243,252,265,267,270,284,287,289,292,295,329,335,337,340,343,405,408,411,413,417],[11,12,13],"blockquote",{},[14,15,16,20],"p",{},[17,18,19],"strong",{},"导读："," 写入时就有误的记忆，很难靠检索质量补救。生产 Agent 需要在信息成为长期记忆前，区分用户陈述、环境事实、模型推断、操作流程和高影响状态。记忆张量 MemTensor 的 MemOS 提供系统层架构，让这些生命周期控制成为明确的操作。",[22,23],"hr",{},[14,25,26],{},"Agent 可能把一次任务中的错误假设带入下一次任务。检查它学到了什么、保留来源，并在条件变化后重新审视记忆，有助于阻止错误传播。MemOS 提供从处理新信息到纠正已有记忆的工具，支持这一过程。",[14,28,29],{},"Agent 答错时，人们通常先检查召回内容。但问题可能更早就已出现：记忆建立在不完整观察上，或保存了一条从未验证过的结论。",[14,31,32,33,43],{},"在论文 ",[34,35,39],"a",{"href":36,"rel":37},"https://arxiv.org/html/2609.11060v1",[38],"nofollow",[40,41,42],"em",{},"Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents"," 中，微软研究人员让记忆整理组件在任务结束后以只读方式访问环境。它可以核查存疑陈述，在后续任务使用前修改或跳过相应记录。",[14,45,46],{},"在论文包含模式变更的 40 道 CLBench 问题实验中，研究者使用 GPT-5.4 和 GitHub Copilot SDK 运行框架。记忆加环境探查系统的平均通过率为 73%，仅使用记忆为 70%，无记忆基线为 39%。加入探查后，每题平均查询次数从 5.6 降至 4.7，任务 Agent 成本从 1.99 美元降至 1.68 美元；这些成本未包含单独的蒸馏与整理阶段。",[14,48,49],{},"这些结果对应研究者的具体设置，也引出一个实际问题：Agent 把学到的内容传给下一次任务前，应该先核查什么？",[51,52],"br",{},[54,55,57],"h2",{"id":56},"一次有用观察怎样变成误导性规则","一次有用观察，怎样变成误导性规则",[14,59,60,61,65],{},"假设数据库 Agent 在 ",[62,63,64],"code",{},"customers_current"," 表中找到所需记录，完成任务后保存了一条说明：“查询活跃账户时使用 customers_current。”",[14,67,68],{},"这次查询可能只适用于某个地区或报告周期。保存的说明省略了条件，另一个 Agent 就可能将它用于更广泛的问题。检查表定义与相关业务规则，可以帮助明确建议的适用范围。",[14,70,71],{},"即使核查仔细，说明也可能过时。一个月后，这张表可能被更新频率更低的兼容视图替代。后续 Agent 需要识别变化并更新记忆。",[14,73,74],{},"类似问题还包括：接口已修复却继续推荐临时绕行方案；把临时审批流程保存为永久流程；把失败命令记成成功方案。政策可能在摘要时丢失生效日期，针对一个客户的规则也可能被推广到所有客户。",[14,76,77],{},"这些记忆都缺少后续任务所需的证据、条件或更新。检索能找到说明，但 Agent 还需要足够的信息判断它是否适用。",[51,79],{},[54,81,83],{"id":82},"让结论可复用前先核查内容","让结论可复用前，先核查内容",[14,85,86],{},"常见记忆流程从对话或任务历史出发，抽取有用信息并保存，供以后检索。在结论成为可复用建议之前，可以增加验证步骤。",[14,88,89],{},"原始对话和工具结果仍可作为证据保留。需要决定的是：哪些结论供未来任务使用，以及附带什么限制。保留这一区别，也便于团队在出错后检查摘要如何生成。",[14,91,92],{},"核查方式应与信息类型匹配。",[94,95,96,109],"table",{},[97,98,99],"thead",{},[100,101,102,106],"tr",{},[103,104,105],"th",{},"信息",[103,107,108],{},"核查内容",[110,111,112,121,129,137],"tbody",{},[100,113,114,118],{},[115,116,117],"td",{},"用户陈述或偏好",[115,119,120],{},"保留谁在何时、何种背景下说过。“我喜欢简洁的周报摘要”可以直接记录，并在用户修改时更新。",[100,122,123,126],{},[115,124,125],{},"环境事实",[115,127,128],{},"检查相应系统，如数据库模式、代码仓库、API 规格或政策文档，并记录观察的版本或时间。",[100,130,131,134],{},[115,132,133],{},"Agent 推断",[115,135,136],{},"保留证据并标记为推断。“这位客户可能对价格敏感”应与客户明确表达的内容区分。",[100,138,139,142],{},[115,140,141],{},"流程或技能",[115,143,144],{},"保留前提条件与成功证据，采用适合该流程的测试或验收条件。",[14,146,147],{},"敏感程度与影响范围适用于以上所有类别。报告长短偏好所需的审核，可以少于一条可能改变访问权限或授权支付的流程。",[14,149,150],{},"只读检查往往足以消除疑问。编码 Agent 可以查看符号定义，数据库 Agent 可以检查模式或查询有限样本。有副作用的流程，则需要适当测试环境或其他成功证据。",[51,152],{},[54,154,156],{"id":155},"memos-如何支持这一流程","MemOS 如何支持这一流程",[14,158,159,164],{},[34,160,163],{"href":161,"rel":162},"https://github.com/MemTensor/MemOS",[38],"MemOS","提供添加、查找、纠正和删除记忆的操作。开发者可以将这些操作与自己的业务系统核查结合起来。应用选择权威来源，并实现前文所述的环境检查。",[51,166],{},[168,169,170],"h3",{"id":170},"让记忆保留来源",[14,172,173,174,179,180,185],{},"开源服务中的 ",[34,175,178],{"href":176,"rel":177},"https://memos-docs.openmem.net/cn/open_source/modules/mem_reader/",[38],"MemReader","将对话、文档与图片处理为带来源信息的记忆条目。开发者可以使用添加接口的",[34,181,184],{"href":182,"rel":183},"https://memos-docs.openmem.net/cn/api-reference/add-memories/",[38],"元数据","附带来源位置、应用提供的验证结果等信息。",[14,187,188],{},"对数据库示例，可以记录数据库模式版本、核查时间，以及这张表适用的业务背景。后续回答异常时，团队就有排查起点。",[14,190,191],{},"验证应覆盖抽取后的结论。仅检查输入文档，可能遗漏从文档生成短记忆时引入的错误。",[51,193],{},[168,195,196],{"id":196},"区分偏好与推断",[14,198,199,200,205],{},"MemOS Cloud 的",[34,201,204],{"href":202,"rel":203},"https://memos-docs.openmem.net/cn/api_docs/core/extract_memory/",[38],"记忆抽取接口","区分明确偏好与根据行为推断的偏好，并返回推断依据。",[14,207,208],{},"这样，应用可以分别处理“我喜欢经济型酒店”和根据一次预订作出的猜测。开发者可判断推断是否适合推荐，或是否需要用户确认后再用于影响更大的操作。",[51,210],{},[168,212,213],{"id":213},"信息变化时更新记忆",[14,215,216,217,222,223,228],{},"MemOS Cloud 在",[34,218,221],{"href":219,"rel":220},"https://memos-docs.openmem.net/cn/changelog/",[38],"记忆维护","中支持冲突与重复处理。",[34,224,227],{"href":225,"rel":226},"https://memos-docs.openmem.net/cn/memos_cloud/introduction/time_awareness/",[38],"时间感知机制","保留变化状态，帮助检索区分当前事实与早期记录。",[14,230,231,232,237],{},"这些能力处理已经进入记忆系统的变化。数据库示例中，应用仍需发现模式变更或提供新证据，再通过",[34,233,236],{"href":234,"rel":235},"https://memos-docs.openmem.net/cn/memos_cloud/mem_operations/add_feedback/",[38],"反馈","纠正建议，或使用相应记忆操作更新、删除受影响条目。",[51,239],{},[168,241,242],{"id":242},"确定记忆可以在哪里使用",[14,244,245,246,251],{},"开源服务通过 ",[34,247,250],{"href":248,"rel":249},"https://memos-docs.openmem.net/cn/open_source/modules/mos/overview/",[38],"MemCube","组织记忆。请求可以指定读取或写入的空间，让应用隔离项目记忆，或向多个 Agent 开放选定信息。应用负责将这些选择与授权规则关联。",[14,253,254,255,260,261,264],{},"会话上下文承担不同职责。在开源",[34,256,259],{"href":257,"rel":258},"https://memos-docs.openmem.net/cn/api-reference/search-memories/",[38],"检索接口","中，",[62,262,263],{},"session_id"," 帮助优先返回相关记忆，不构成会话间的强制隔离边界。",[51,266],{},[168,268,269],{"id":269},"检查后台处理结果",[14,271,272,273,277,278,283],{},"记忆写入和纠错可以在后台运行。开源服务提供",[34,274,276],{"href":248,"rel":275},[38],"调度与状态操作","来跟踪处理。MemOS Cloud 的",[34,279,282],{"href":280,"rel":281},"https://memos-docs.openmem.net/cn/api_docs/message/get_status/",[38],"任务状态接口","报告处理状态，并在完成后返回受影响的记忆 ID 与变更摘要。",[14,285,286],{},"应用由此可以从提交跟踪到结果，再查看受影响条目，核对后续检索返回的内容。",[51,288],{},[54,290,291],{"id":291},"把核查接入实际流程",[14,293,294],{},"先选择一种准确性对任务有重要影响的记忆，例如数据库使用建议或经常复用的客服流程。",[296,297,298,305,311,317,323],"ol",{},[299,300,301,304],"li",{},[17,302,303],{},"抽取候选内容。"," 识别值得保存的结论或流程，保留来源。",[299,306,307,310],{},[17,308,309],{},"核查证据。"," 检查权威系统，验证候选内容依赖的条件。证据不完整时，缩小结论范围或转交审核。",[299,312,313,316],{},[17,314,315],{},"同时保存适用条件。"," 记录适用范围、核查时间，以及再次核查的触发条件，让已通过审核的候选内容参与复用。",[299,318,319,322],{},[17,320,321],{},"检查结果。"," 写入或纠正后，检查操作结果，以及后续任务能取得的记录。",[299,324,325,328],{},[17,326,327],{},"来源变化后重新审视。"," 数据库迁移、政策修订或 API 变化，应触发相关记忆的复核。",[14,330,331,332,334],{},"对 ",[62,333,64],{},"，有用的记录应包含适用地区与报告周期、建议依据，以及已核查的模式版本。涉及这张表的迁移，为复核提供了具体触发条件。",[51,336],{},[54,338,339],{"id":339},"追踪记忆如何进入下一次任务",[14,341,342],{},"团队也需要观察流程是否有效。以下问题可指导应用与记忆服务之间的监测。",[94,344,345,355],{},[97,346,347],{},[100,348,349,352],{},[103,350,351],{},"阶段",[103,353,354],{},"检查内容",[110,356,357,365,373,381,389,397],{},[100,358,359,362],{},[115,360,361],{},"候选审核",[115,363,364],{},"候选来自哪些来源，核查了什么证据，为什么接受、修订或拒绝",[100,366,367,370],{},[115,368,369],{},"写入",[115,371,372],{},"添加或修改了哪些记录，存在哪里，处理是否完成",[100,374,375,378],{},[115,376,377],{},"检索",[115,379,380],{},"返回了哪些记忆，哪些实际进入模型上下文",[100,382,383,386],{},[115,384,385],{},"使用",[115,387,388],{},"任务完成情况、重复错误、用户纠正，以及与无记忆基线的对比",[100,390,391,394],{},[115,392,393],{},"维护",[115,395,396],{},"哪些条目受来源变化影响，哪些核查已逾期，纠正是否进入后续检索",[100,398,399,402],{},[115,400,401],{},"删除",[115,403,404],{},"记忆存储与应用管理的副本、缓存是否完成删除",[14,406,407],{},"关联这些记录有助于排查失败。召回的记忆可能从未进入模型，进入提示词的记忆也未必能解释后续动作。追踪记录提供证据，确认因果仍需要测试。",[14,409,410],{},"首次评估可以先保存一条核查过的记忆，在后续任务中使用，再修改底层信息。确认纠正在后续检索中出现，Agent 也能用新信息完成任务。这检验了持续工作中最有价值的部分：把有用知识带向未来，并持续允许纠正。",[51,412],{},[54,414,416],{"id":415},"关于-memtensor-与-memos","关于 MemTensor 与 MemOS",[14,418,419,420,425],{},"MemTensor 为 AI 应用与 Agent 开发记忆基础设施，工作覆盖 ",[34,421,424],{"href":422,"rel":423},"https://memos-docs.openmem.net/cn/usecase/blog/agent_memory_is_not_rag/",[38],"Memory³ 研究、MemOS、Agent 记忆基础设施和记忆原生模型研究","。MemOS 通过 API、开源组件和集成方案，将这些工作用于跨任务的信息存储、检索与维护。",{"title":427,"searchDepth":428,"depth":428,"links":429},"",2,[430,431,432,440,441,442],{"id":56,"depth":428,"text":57},{"id":82,"depth":428,"text":83},{"id":155,"depth":428,"text":156,"children":433},[434,436,437,438,439],{"id":170,"depth":435,"text":170},3,{"id":196,"depth":435,"text":196},{"id":213,"depth":435,"text":213},{"id":242,"depth":435,"text":242},{"id":269,"depth":435,"text":269},{"id":291,"depth":428,"text":291},{"id":339,"depth":428,"text":339},{"id":415,"depth":428,"text":416},"md",{},"/cn/usecase/blog/keep_bad_assumptions_out_of_agent_memory",{"title":5,"description":427},"cn/usecase/blog/keep_bad_assumptions_out_of_agent_memory","JyDP-EQjVkYXx3yuBJlL4dRirTHZoA65DYr4cyzVSos",[450,456],{"title":451,"path":452,"stem":453,"framework":6,"module":6,"class":454,"target":-1,"active":455,"defaultOpen":455,"children":-1,"description":-1},"Claude 让记忆可编辑，生产 Agent 仍需要可治理的记忆层","/cn/usecase/blog/claude_editable_memory_governed_layer","usecase/blog/claude_editable_memory_governed_layer",[],false,{"title":457,"path":458,"stem":459,"framework":6,"module":6,"class":460,"target":-1,"active":455,"defaultOpen":455,"children":-1,"description":-1},"持久化 Agent 记忆，应当让下一次任务更轻松","/cn/usecase/blog/persistent_agent_memory_next_task","usecase/blog/persistent_agent_memory_next_task",[],1793163788619]