[{"data":1,"prerenderedAt":201},["ShallowReactive",2],{"docs-page-cn-/cn/usecase/blog/agent_tool_failure_memory":3,"surround-cn-/cn/usecase/blog/agent_tool_failure_memory":188},{"id":4,"title":5,"avatar":6,"banner":6,"body":7,"category":6,"desc":6,"description":176,"extension":182,"lead":6,"links":6,"meta":183,"navigation":6,"path":184,"seo":185,"stem":186,"__hash__":187},"docs/cn/usecase/blog/agent_tool_failure_memory.md","Agent 工具调用失败后，哪些信息值得写进长期记忆？",null,{"type":8,"value":9,"toc":175},"minimark",[10,18,21,24,27,31,34,37,100,103,106,109,112,115,117,120,141,144,147,158,161,163,166,169,172],[11,12,13],"p",{},[14,15],"img",{"alt":16,"src":17},"Agent 工具调用失败与长期记忆","https://cdn.memtensor.com.cn/img/1794814838168_f9zw5x_compressed.png",[11,19,20],{},"2026 年 11 月 12 日，OpenRouter 发布了一组 Agent 重试实验。面对一个持续超时的发票查询工具，未启用重复调用检测时，18 次运行中有 4 次触到 40 步上限。启用后，18 次中有 16 次只实际调用了两次工具。",[11,22,23],{},"这组实验关注的是当轮能否停下来。如果助手还会跨会话保存工具轨迹，验收可以再往前走一步：下一次查询时，它会把昨天的超时当成结论，还是先查看当前结果？",[25,26],"br",{},[28,29,30],"h2",{"id":30},"用三个结果驱动同一条查询",[11,32,33],{},"在测试环境准备只读发票查询，用同一个虚构单号分别模拟三种结果。每组用例开始前清空本轮计数，并隔离其他用例留下的测试记忆。完成写入后，保留工具版本和轨迹标识，供下一轮复测。",[11,35,36],{},"这里假设企业发票接口的第二版支持先查处理状态，再查明细。",[38,39,40,63],"table",{},[41,42,43],"thead",{},[44,45,46,53,58],"tr",{},[47,48,49],"th",{},[50,51,52],"strong",{},"模拟结果",[47,54,55],{},[50,56,57],{},"当轮预期",[47,59,60],{},[50,61,62],{},"下一轮可以沿用什么",[64,65,66,78,89],"tbody",{},[44,67,68,72,75],{},[69,70,71],"td",{},"后端暂时超时",[69,73,74],{},"到达本轮预算后停止，并告知状态未确认",[69,76,77],{},"留下事件用于复盘。下次仍可查同一张发票",[44,79,80,83,86],{},[69,81,82],{},"请求缺少账套字段，接口明确指出原因",[69,84,85],{},"仅在有权取得正确账套后补齐参数。缺少信息时请用户确认",[69,87,88],{},"记录接口版本和必填字段，供同版本复用",[44,90,91,94,97],{},[69,92,93],{},"查到发票仍在生成",[69,95,96],{},"按接口说明等待，或在预算内查询处理状态",[69,98,99],{},"记录何时查状态及如何继续。下一轮重新读这张发票的当前状态",[11,101,102],{},"单独测试重复调用检测。把同一张发票的单号写成接口允许的两种等价格式，确认检测器能把它们识别为同一对象。达到设定上限后，在发起下一次查询前拦截请求，后端不应再收到这笔查询。",[11,104,105],{},"保留检测器状态，再提交另一张测试发票，确认它不会被误判为重复。不同账套中的相同单号也应分别识别。",[11,107,108],{},"查询发票状态有时本来就需要轮询。可以为这类调用单独设置间隔和次数预算，再验证它不会被普通的重复调用规则误拦。整轮停机仍用前面的预算用例验收，同时记录模型尝试发出的次数和后端实际收到的次数。",[11,110,111],{},"验收时还要区分整轮步数和实际工具调用次数。有些 Agent 运行框架在步数上限触发时，可能仍会执行已经排队的工具调用。如果业务要求达到上限后不再访问后端，应在工具执行前单独检查预算。",[11,113,114],{},"第一组要加一次反向验证。次日用新会话请求同一张发票，将测试后端改成正常返回。如果助手因为昨天超时就跳过这次查询，检查是否把故障写成了长期规则。第三组也要改一次状态，让发票从「生成中」变成「已出具」，确认答案使用实时结果。否则，即使新会话准确记起昨天的轨迹，也会给用户错误的当前状态。",[25,116],{},[28,118,119],{"id":119},"让调用轨迹保留可核对的来源",[11,121,122,127,128,132,133,136,137,140],{},[123,124,126],"a",{"href":125},"/cn/memos_cloud/features/tool_calling/","MemOS Cloud 的工具记忆","支持把助手发出的 ",[129,130,131],"code",{},"tool_calls"," 与实际的 ",[129,134,135],{},"tool"," 返回通过 ",[129,138,139],{},"tool_call_id"," 对应，并将工具使用过程整理成后续可检索的轨迹。",[11,142,143],{},"例如，一次查询中使用了哪个工具、传入了什么参数、返回了什么结果，都可以成为排查和复用的线索。提交前，应用可以去掉凭证和无关客户信息。对准备复用的步骤，再记录工具版本、适用错误、验证结果和来源。",[11,145,146],{},"一次失败轨迹可以帮助团队复盘，但要把它作为下一轮的处理方法，还需要先确认它适用于什么情况。",[11,148,149,150,153,154,157],{},"下一轮复测时，可以让应用检索相关工具记忆，再检查返回的轨迹是否包含本次测试需要的信息。MemOS Cloud 的示例支持在检索时通过 ",[129,151,152],{},"include_memory_view"," 指定 ",[129,155,156],{},"tool_memory","。写入和检索完成后，应用仍需判断这条经验是否适用于当前接口版本和用户。",[11,159,160],{},"例如，如果只是补齐一次账套字段并成功返回，要把它写成「第二版查询必须携带账套」，还需核对接口文档或用隔离用例重现缺字段报错。记忆提供查阅和复用的线索。是否还有重试预算、当前用户能不能访问该账套、发票目前是什么状态，由应用在当轮重新判断。",[25,162],{},[28,164,165],{"id":165},"用一次版本变化检验旧经验会不会误用",[11,167,168],{},"让第三版测试接口通过当前身份推导账套，旧字段在新接口里会报错。保留第二版轨迹后，以相同问题发起新会话。",[11,170,171],{},"预期是第二版轨迹可以作为历史记录保留，应用准备第三版调用时应采用第三版的参数要求。如果旧规则仍被用于调用，先检查轨迹中是否记录了版本和适用范围，再检查检索及上下文组装时有没有区分这些信息。不同版本的轨迹若被合并成一条摘要，也要确认摘要没有丢掉适用条件。",[11,173,174],{},"验收时记录本轮停止原因、工具确认的结果，以及下一轮实际收到的记忆。只统计重试次数，看不出失败经验有没有被错误保留。把三份记录放在一起，才能定位是预算限制失效、工具结果被解释错，还是过期经验被带进了新任务。",{"title":176,"searchDepth":177,"depth":177,"links":178},"",2,[179,180,181],{"id":30,"depth":177,"text":30},{"id":119,"depth":177,"text":119},{"id":165,"depth":177,"text":165},"md",{},"/cn/usecase/blog/agent_tool_failure_memory",{"title":5,"description":176},"cn/usecase/blog/agent_tool_failure_memory","mjxnDOYqRi9-tVd1lstkK4XEywvOR71D26PnU4TZcIg",[189,196],{"title":190,"path":191,"stem":192,"framework":6,"module":6,"class":193,"target":-1,"active":194,"defaultOpen":195,"children":-1,"description":-1},"多人 Agent 的共享记忆怎么验收？","/cn/usecase/blog/shared_agent_memory_acceptance","usecase/blog/shared_agent_memory_acceptance",[],false,true,{"title":197,"path":198,"stem":199,"framework":6,"module":6,"class":200,"target":-1,"active":194,"defaultOpen":195,"children":-1,"description":-1},"MemT-1.0 与 MemOS 如何分工？从一次编码任务开始选型","/cn/usecase/blog/memt_memos_agent_memory","usecase/blog/memt_memos_agent_memory",[],1794819066527]