[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"portal-settings:stajic:zh":3,"public-menus:all":38,"post:what-should-an-ai-agent-remember-forget-recompute-or-retrieve-again:zh":205,"related:post:what-should-an-ai-agent-remember-forget-recompute-or-retrieve-again:zh:1":1632},{"statusCode":4,"data":5,"message":37},200,{"tenantId":6,"lang":7,"defaultLang":8,"siteUrl":9,"contactEmail":10,"brandName":11,"logoUrl":12,"siteName":11,"siteDescription":13,"ogImage":10,"robotsIndex":14,"socialLinks":10,"reservedSlugs":10,"seoPolicy":15},"stajic","zh","de","https:\u002F\u002Fstajic.de",null,"Stajic Platform","\u002FLogo_Planet.svg","Stajic Portal",true,{"branding":16,"relatedContent":17,"crossDomainLinks":18},{"logoUrl":12},{"enabled":14},[19,22,25,28,31,34],{"url":20,"label":21,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Ffigure.rocks","figure.rocks",{"url":23,"label":24,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Floving.rocks","loving.rocks",{"url":26,"label":27,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.com","bazify.com",{"url":29,"label":30,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.de","bazify.de",{"url":32,"label":33,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.at","bazify.at",{"url":35,"label":36,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.ba","bazify.ba","Portal settings resolved",[39,45],{"id":40,"name":41,"location":42,"isActive":14,"isDefault":43,"items":44},1,"main-navigation","header",false,[],{"id":46,"name":47,"location":48,"isActive":14,"isDefault":14,"items":49},4,"main-menu","sidebar",[50,66,79,93,103,118,133],{"id":51,"title":52,"url":60,"target":61,"icon":62,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":64,"portfolioId":10,"children":65},"item-18",{"de":53,"en":54,"es":55,"fr":56,"it":54,"ru":57,"sr":58,"zh":59},"Startseite","Home","Inicio","Accueil","Главная","Почетна","首页","\u002Ffull-stack-web-developer-munich-performance-seo-and-maintainable-builds","_self","i-lucide-home","page",111,[],{"id":67,"title":68,"url":75,"target":61,"icon":76,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":77,"portfolioId":10,"children":78},"item-22",{"de":69,"en":69,"es":70,"fr":69,"it":71,"ru":72,"sr":73,"zh":74},"Vision","Visión","Visione","Видение","Визија","想象","\u002Fueber-uns-webdesign-muenchen-webaplikation","i-lucide-eye",113,[],{"id":80,"title":81,"url":89,"target":61,"icon":90,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":91,"portfolioId":10,"children":92},"item-19",{"de":82,"en":83,"es":84,"fr":83,"it":85,"ru":86,"sr":87,"zh":88},"Leistungen","Services","Servicios","Servizi","Услуги","Услуге","服务","\u002Fservices-dienstleistungen-muenchen","i-lucide-wrench",116,[],{"id":94,"title":95,"url":99,"target":61,"icon":100,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":101,"portfolioId":10,"children":102},"item-23",{"de":96,"en":96,"es":96,"fr":96,"it":96,"ru":97,"sr":97,"zh":98},"Blog","Блог","博客","\u002Fblog","i-lucide-book-open",112,[],{"id":104,"title":105,"url":114,"target":61,"icon":115,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":116,"portfolioId":10,"children":117},"item-32",{"de":106,"en":107,"es":108,"fr":109,"it":110,"ru":111,"sr":112,"zh":113},"Neue Technologien","New Technologies","Nuevas tecnologías","Nouvelles technologies","Nuove tecnologie","Новые технологии","Нове технологије","新技术！","\u002Fneue-webtechnologien","i-lucide-sparkles",122,[],{"id":119,"title":120,"url":129,"target":61,"icon":130,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":131,"portfolioId":10,"children":132},"item-20",{"de":121,"en":122,"es":123,"fr":124,"it":125,"ru":126,"sr":127,"zh":128},"Kontakt","Contact us!","Contacto","Contact","Contatto","Контакт","Контактирајте нас","联系我们！","\u002Fcontact","i-lucide-mail",115,[],{"id":134,"title":135,"url":144,"target":61,"icon":145,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":147},"item-21",{"de":136,"en":137,"es":138,"fr":139,"it":140,"ru":141,"sr":142,"zh":143},"Unsere Arbeit","Our Work","Nuestro trabajo","Nos réalisations","I nostri lavori","Наши работы","Наши радови","文件夹","\u002Fportfolio","i-lucide-briefcase",114,[148,161,175,181,193],{"id":149,"title":150,"url":144,"target":61,"icon":159,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":160},"item-24",{"de":151,"en":152,"es":153,"fr":154,"it":155,"ru":156,"sr":157,"zh":158},"Alle Projekte","All Projects","Todos los proyectos","Tous les projets","Tutti i progetti","Все проекты","Сви пројекти","所有项目","i-lucide-grid-3x3",[],{"id":162,"title":163,"url":171,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":174},"item-29",{"de":164,"en":165,"es":166,"fr":167,"it":168,"ru":169,"sr":170,"zh":143},"Local Roots, Global Reach","Local Roots - Global Reach","Empresa local ","Entreprise locale","Azienda locale","Местная компания","Локално предузеће глобално тржиште","\u002Fportfolio\u002Flocal-roots-global-reach-communication-media-systems-for-modern-business","i-lucide-folder","custom",[],{"id":176,"title":177,"url":179,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":180},"item-28",{"de":178,"en":178,"es":178,"fr":178,"it":178,"ru":178,"sr":178,"zh":178},"Solr Suggester","\u002Fportfolio\u002Fsolr-fuzzy-suggester-und-solr-infix-suggester-abfrage-ueber-ajax-und-filterung",[],{"id":182,"title":183,"url":191,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":192},"item-27",{"de":184,"en":185,"es":186,"fr":187,"it":188,"ru":189,"sr":190,"zh":185},"Firmenwebseite SEO","Company Website SEO","Sitio web corporativo SEO","Site web d’entreprise SEO","Sito web aziendale SEO","Корпоративный сайт SEO","Пословна веб-страница SEO","\u002Fportfolio\u002Fseo-sem-branding-mobile-webseite-muenchen",[],{"id":194,"title":195,"url":203,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":204},"item-31",{"de":196,"en":197,"es":198,"fr":199,"it":200,"ru":201,"sr":202,"zh":197},"Digitalisierungsportal","Digitalization Portal","Portal de digitalización","Portail de numérisation","Portale di digitalizzazione","Портал цифровизации","Портал за дигитализацију","\u002Fportfolio\u002Fdigitalisierungsportal-archiv-museum-bibliothek-ead-lido-mets-mods",[],{"statusCode":4,"data":206,"message":1631},{"id":207,"title":208,"slug":209,"content":210,"contentJson":211,"excerpt":813,"featuredImage":814,"featuredImageAlt":815,"featuredImageCaption":10,"featuredImageTitle":10,"featuredImageCopyright":10,"featuredImageAuthor":10,"featuredImageSourceUrl":10,"featuredImageLicense":10,"featuredImageIsAiGenerated":43,"status":816,"publishedAt":817,"createdAt":818,"updatedAt":819,"seoLocalePaths":820,"categories":829,"author":842,"translations":847},"470","AI代理应该记住、遗忘、重新计算还是再次检索什么？","what-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u003Cnav class=\"editorjs-toc\" data-editorjs-toc=\"true\" aria-label=\"目录\">\u003Cstrong class=\"editorjs-toc__title\">目录\u003C\u002Fstrong>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-0\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-5\" class=\"editorjs-toc__link\">真正的记忆问题不是存储——而是生命周期控制\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-9\" class=\"editorjs-toc__link\">对任何代理信息可采取的四种可能行动\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-11\" class=\"editorjs-toc__link\">记忆准入测试\u003C\u002Fa>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-1\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-15\" class=\"editorjs-toc__link\">1. 记住：能改善未来决策的持久知识\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-19\" class=\"editorjs-toc__link\">2. 重新读取或检索：具有外部真实来源的易变事实\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-23\" class=\"editorjs-toc__link\">3. 重新计算：计算比信任成本更低的派生信息\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-26\" class=\"editorjs-toc__link\">4. 遗忘、过期或取代：删除是一种能力\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-30\" class=\"editorjs-toc__link\">决策方法\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-32\" class=\"editorjs-toc__link\">示例：同一个智能体应使用不同的生命周期操作\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-34\" class=\"editorjs-toc__link\">记忆应存储条件，而不仅仅是结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-37\" class=\"editorjs-toc__link\">记忆写入应比记忆读取更昂贵\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-40\" class=\"editorjs-toc__link\">记忆质量至少包含五个维度\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-43\" class=\"editorjs-toc__link\">默认情况下不应放入持久记忆的内容\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-45\" class=\"editorjs-toc__link\">记忆是任务特定的——不存在通用的最优存储\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-48\" class=\"editorjs-toc__link\">什么会改变这个答案？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-52\" class=\"editorjs-toc__link\">局限性\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-55\" class=\"editorjs-toc__link\">结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-59\" class=\"editorjs-toc__link\">常见问题\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-61\" class=\"editorjs-toc__link\">术语表\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-63\" class=\"editorjs-toc__link\">主要来源与延伸阅读\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fnav>\n\u003Cp>长时间运行的AI代理会积累远超其应永久记住的信息量。对话、工具输出、中间计算、用户偏好、项目决策、搜索结果、系统状态、错误以及成功流程，在当下都可能看起来有用。将它们全部视为持久记忆会引发第二个问题：代理之后必须判断哪些已存储的信息仍然可信、最新、相关且可安全复用。\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--info my-6 rounded-xl border p-5 border-blue-300 bg-blue-50 dark:border-blue-900 dark:bg-blue-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">直接回答\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">AI代理应当<strong>记住那些持久、可复用、保留来源且重新发现成本高昂的信息<\u002Fstrong>；<strong>从权威来源重新读取或检索易变事实<\u002Fstrong>；<strong>在时效性重要时重新计算廉价的派生值<\u002Fstrong>；以及<strong>遗忘、过期或取代那些未来复用带来的风险大于价值的信息<\u002Fstrong>。正确的行动较少取决于信息是否“重要”，而更多取决于其易变性、权威性、派生成本、复用价值、敏感性以及修订行为。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">关于决策模型\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">下文中的“记住\u002F重读\u002F重算\u002F遗忘”模型以及记忆准入测试是本文提出的实用架构工具。它们并非正式的行业标准。它们旨在使代理记忆决策变得明确、可测试且可审计。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-5\">真正的记忆问题不是存储——而是生命周期控制\u003C\u002Fh2>\n\u003Cp>现代代理系统几乎可以存储任何内容：完整记录、摘要、嵌入、文件、数据库记录、工具轨迹、结构化事实、技能以及外部工件。因此，存储容量并非难点。难点在于决定什么值得保留、应保留多久，以及当现实发生变化时必须发生什么。\u003C\u002Fp>\n\u003Cp>OpenAI的会话记忆指南明确警告，将过多历史向前携带会造成分心、低效、上下文污染以及错误累积。Anthropic同样将上下文视为必须精心管理而非累积的有限资源。微软研究院也朝着同一方向前进：PlugMem将原始交互历史转换为可复用的结构化知识，而不是将完整历史视为同等有价值的记忆。\u003C\u002Fp>\n\u003Cp>其架构后果很简单：记忆需要准入策略、维护策略和退役策略。仅靠检索器无法提供这些语义。\u003C\u002Fp>\n\u003Ch2 id=\"section-9\">对任何代理信息可采取的四种可能行动\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">行动\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">使用时机\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">典型示例\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">主要风险\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">记住\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">该信息在未来任务中仍然有用，且可靠重建成本高昂或不可能\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">稳定的用户偏好、已接受的项目决策、可复用技能、已验证的长期约束\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">持久化错误、过时或过于宽泛的内容\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">重读\u002F检索\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">该信息具有可能变化的权威来源\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">权限、库存、策略版本、订单状态、产品价格、当前API文档\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">使用旧副本而非当前权威来源\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">重算\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">该信息是派生的，且重新计算成本足够低\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">基于当前源数据的总计、分数、排名、摘要，确定性转换\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">持久化过时的派生输出\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">遗忘\u002F过期\u002F取代\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">未来复用价值很小，或会带来隐私、过时、冲突或污染风险\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">临时工具输出、失败假设、被取代的决策、临时令牌、过时环境状态\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">丢失后来证明必要的信息\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-11\">记忆准入测试\u003C\u002Fh2>\n\u003Cp>在信息成为持久代理记忆之前，用六个属性对其进行测试。这些属性比模糊的重要性评分更有用，因为它们能预测信息随时间变化的行为。\u003C\u002Fp>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">决定信息是否属于记忆的六个属性\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">属性\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">问题\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">决策压力\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">易变性\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">权威性\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">复用价值\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">重建成本\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">敏感性\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">修订行为\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Caside class=\"editorjs-callout editorjs-callout--tip my-6 rounded-xl border p-5 border-violet-300 bg-violet-50 dark:border-violet-900 dark:bg-violet-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">一条实用规则\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">如果一个事实<strong>易变 + 在别处有权威来源 + 获取成本低<\u002Fstrong>，就不要将复制的值提升为长期记忆。应存储指针、标识符或检索路径。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch3 id=\"section-15\">1. 记住：能改善未来决策的持久知识\u003C\u002Fh3>\n\u003Cp>良好的持久记忆能减少重复工作，而不会把昨天的状态变成今天的真相。典型候选包括明确的用户偏好、持久的项目约束、决策及其理由、可复用流程、反复出现的失败模式，以及预计不会频繁变化的已验证事实。\u003C\u002Fp>\n\u003Cp>最强的记忆未必是原始记录。PlugMem在2026年的工作主张将交互历史转换为紧凑事实和可复用技能。微软的BREW同样将过去轨迹提炼为可检索的程序性知识，描述该做什么、何时适用以及需要注意什么。两者都指向一个有用的设计原则：存储可复用知识，而不仅仅是历史文本。\u003C\u002Fp>\n\u003Cp>被记住的条目还应保留来源。未来的代理应能区分“用户明确要求这一点”、“系统观察到这一点”、“某个来源陈述了这一点”以及“某个模型推断了这一点”。没有这种区分，记忆会逐渐将证据、解释和推测混为一个无差别的池子。\u003C\u002Fp>\n\u003Ch3 id=\"section-19\">2. 重新读取或检索：具有外部真实来源的易变事实\u003C\u002Fh3>\n\u003Cp>有些信息之所以有价值，恰恰是因为它会变化。当前权限、订单状态、库存、账户状态、服务健康状况、软件文档、价格、日程、法规和 API 行为，在用于重要决策之前，通常应从拥有这些信息的系统中重新读取。\u003C\u002Fp>\n\u003Cp>智能体可以记住某个来源存在、如何访问它，或哪些字段重要。但它不应假设过去检索到的值仍然具有权威性。这将关于从何处以及如何获取真相的记忆，与真相的缓存副本区分开来。\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">陈旧记忆陷阱\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">一个事实可能被完美地记住，但仍然是错误的。记忆质量不仅在于回忆的准确性；它还包括知道何时必须让位于全新的权威读取。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch3 id=\"section-23\">3. 重新计算：计算比信任成本更低的派生信息\u003C\u002Fh3>\n\u003Cp>派生信息应与源事实区别对待。如果一个值可以从当前输入确定性地重新计算，那么持久化结果可能会造成不必要的陈旧。总计、百分比、排名、资格标志、生成的摘要以及其他派生输出，在使用时通常应重新计算。\u003C\u002Fp>\n\u003Cp>关键的权衡是成本。如果重新计算成本高昂，系统可以将结果与确切的输入版本、时间戳、推导方法和失效条件一起缓存。如果重新计算成本低廉，新鲜度通常胜出。\u003C\u002Fp>\n\u003Ch3 id=\"section-26\">4. 遗忘、过期或取代：删除是一种能力\u003C\u002Fh3>\n\u003Cp>遗忘不一定是一种缺陷。它是一种控制机制。临时的工具输出、一次性的搜索结果、失败的假设、临时环境状态、中间推理产物、过时的用户偏好、过期的凭证以及被取代的决策，如果无限期地保持活跃，都可能成为负担。\u003C\u002Fp>\n\u003Cp>近期的记忆研究越来越认识到，无限制的积累会降低性能。微软 2026 年受人类启发的记忆架构明确包含了基于干扰的遗忘和巩固，而 PlugMem 报告称，原始历史记录可能会用低价值上下文压垮智能体。工程上的教训并不需要复制生物记忆：保留应该是有选择性的。\u003C\u002Fp>\n\u003Cp>在许多系统中，取代比立即删除更安全。旧决策仍然可审计，但检索默认使用新决策。这对于项目、政策、合规以及任何变更历史本身就是证据的工作流程都很重要。\u003C\u002Fp>\n\u003Ch2 id=\"section-30\">决策方法\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">决定信息项的生命周期\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. 对信息进行分类\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">它是权威状态、用户偏好、外部证据、派生输出、程序、观察还是模型推断？\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 确定真实来源\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">确定是否有其他系统或来源比记忆本身更具权威性。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 估计易变性\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">询问该项在下一次有意义的重用之前发生变化的可能性有多大。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 估计重用和重建成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">将未来价值与获取或重新创建信息的成本和可靠性进行比较。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 检查敏感性和范围\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">定义谁可以访问该信息、它可以在哪里持久化，以及持久化是否合理。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">6. 定义失效\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">指定过期、取代、冲突解决，或强制进行全新权威读取的条件。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">7\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">7. 选择操作\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">记住、重新读取\u002F检索、重新计算，或遗忘\u002F过期\u002F取代。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">8\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">8. 保留来源\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">存储足够的元数据，以区分源事实、用户陈述、观察、推导和模型推断。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-32\">示例：同一个智能体应使用不同的生命周期操作\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">信息\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">推荐操作\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">原因\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“用户偏好简洁的技术性回答。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">记住\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">稳定的偏好，具有高重用价值\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“部署当前已暂停。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">重新读取\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">当前操作状态可能发生变化\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“预计总成本为 48,620 欧元。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">根据当前输入重新计算\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">派生值应跟随源变化\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">昨天一个 20,000 token 的原始工具响应\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">遗忘或外部归档\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">直接重用价值低；上下文成本高\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">针对反复出现的构建失败的已确认解决方法\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">作为可重用程序记住\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">未来重用价值高，重新发现成本高\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型对服务器故障原因的猜测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">不要提升为持久事实\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">推断不是经过验证的证据\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">一个旧的项目决策后来被新决策取代\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">取代，保留审计历史\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">最新决策应胜出，同时不抹除来源\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">当前产品价格\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">再次检索\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">高易变性和外部权威\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">法律或政策解释\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">仅在有来源\u002F版本元数据时记住先前的分析；行动前重新检查权威性\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">适用性可能随时间和司法管辖区而变化\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-34\">记忆应存储条件，而不仅仅是结论\u003C\u002Fh2>\n\u003Cp>当持久记忆只存储结论而丢失了结论有效时的条件时，它就会变得危险。“使用每租户数据库”比“当监管隔离和租户特定生命周期要求超过运营开销时，使用每租户数据库”更弱。第二种形式保留了决策边界。\u003C\u002Fp>\n\u003Cp>这对于智能体学习到的程序更为重要。一个成功的工作流程不仅应捕获步骤，还应捕获前提条件、环境、工具版本、可观察的成功标准以及已知的失败模式。否则，在错误环境中检索到的记忆可能会自信地重现一个过时的解决方案。\u003C\u002Fp>\n\u003Ch2 id=\"section-37\">记忆写入应比记忆读取更昂贵\u003C\u002Fh2>\n\u003Cp>读取弱记忆可能损害一个答案。写入弱记忆可能损害许多未来的答案。这种不对称性表明写入路径应比读取路径更严格：对候选记忆进行分类、检查来源、检测矛盾、应用敏感性规则、定义作用域，并决定是否需要人工确认或外部验证。\u003C\u002Fp>\n\u003Cp>当智能体从自身生成的输出中写入记忆时，这一点尤为重要。生成的摘要可能包含压缩错误。工具故障可能被误解。一个看似合理的假设可能被存储为事实。如果这些输出在没有证据状态的情况下成为未来上下文，智能体可能创建自我强化的错误循环。\u003C\u002Fp>\n\u003Ch2 id=\"section-40\">记忆质量至少包含五个维度\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">维度\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">问题\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">保留质量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">系统是否保留了应当留存的信息？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">检索质量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">系统能否在需要时恢复正确的记忆？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">新鲜度质量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">系统是否知道存储的信息何时不再是最新的？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">来源质量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">系统能否区分来源、用户陈述、观察、推导和推断？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">退役质量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">当信息不应再影响决策时，系统能否使其过期、取代、限制或删除？\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>基准测试正开始区分这些关注点。微软的 MemGym 明确评估长时程智能体场景中的记忆，并报告记忆隔离分数，旨在减少推理、检索和工具使用能力的混杂影响。这一方向很重要，因为仅凭最终任务分数无法判断记忆本身是有帮助、有害还是无关。\u003C\u002Fp>\n\u003Ch2 id=\"section-43\">默认情况下不应放入持久记忆的内容\u003C\u002Fh2>\n\u003Cul>\u003Cli>原始思维链或隐藏推理产物。\u003C\u002Fli>\u003Cli>临时认证令牌、机密或凭据。\u003C\u002Fli>\u003Cli>未经核实的模型生成假设。\u003C\u002Fli>\u003Cli>存在实时权威系统的易变状态。\u003C\u002Fli>\u003Cli>缺少源输入且可低成本重新计算的派生值。\u003C\u002Fli>\u003Cli>仅因存储可用而保存的大型工具输出。\u003C\u002Fli>\u003Cli>已由更好的真相来源管理的信息的重复副本。\u003C\u002Fli>\u003Cli>没有明确持久化目的、访问范围和生命周期的敏感个人数据。\u003C\u002Fli>\u003Cli>没有明确版本或退役语义的已被取代的结论。\u003C\u002Fli>\u003Cli>仅对当前运行有用且没有可复用诊断价值的错误消息或故障状态。\u003C\u002Fli>\u003C\u002Ful>\n\u003Ch2 id=\"section-45\">记忆是任务特定的——不存在通用的最优存储\u003C\u002Fh2>\n\u003Cp>编码智能体受益于可复用流程、仓库约定、成功修复模式和项目决策。个人助理可能需要偏好、承诺和关系上下文。商务智能体更需要当前产品和交易状态，而非价格或库存的历史副本。研究智能体受益于来源出处、未解决的假设和明确的证据状态。\u003C\u002Fp>\n\u003Cp>微软研究院的 M-star 工作直接指出了这一点：为某一目的优化的记忆系统可能难以迁移到另一目的，而任务特定的记忆机制可能优于固定的通用设计。因此，记忆模式应遵循智能体必须做出的决策，而不是强加给每个智能体的通用模板。\u003C\u002Fp>\n\u003Ch2 id=\"section-48\">什么会改变这个答案？\u003C\u002Fh2>\n\u003Cp>当检索缓慢或昂贵、权威系统间歇性不可用、重新计算成本高昂、审计规则要求历史快照，或智能体必须离线运行时，这种平衡会发生变化。在这些情况下，可能需要缓存或持久化更多信息——但需附带版本、来源、时间戳和失效元数据。\u003C\u002Fp>\n\u003Cp>对于主要价值在于个性化的智能体，这种平衡也会发生变化。稳定的偏好可能值得记住，即使技术上可以再次询问。相反，在高风险领域，将观察或解释转化为持久记忆的门槛应高得多。\u003C\u002Fp>\n\u003Cp>未来的托管记忆平台可能会自动化整合、检索、遗忘和上下文构建。这可以减少实现工作，但并不能消除治理问题：哪些信息被允许影响未来决策，在什么条件下，以及系统何时必须回到当前的真相来源？\u003C\u002Fp>\n\u003Ch2 id=\"section-52\">局限性\u003C\u002Fh2>\n\u003Cp>在当前框架和研究中，“智能体记忆”没有统一定义。一些系统用该术语指代对话历史，另一些则指外部持久存储、结构化知识、学习到的流程、检查点或模型适配。本文中的决策模型侧重于操作生命周期语义，而非强制统一术语。\u003C\u002Fp>\n\u003Cp>四个生命周期操作也可能重叠。一个系统可以在同一工作流中记住稳定的摘要、保留指向来源的指针、重新读取易变字段并重新计算派生结果。该模型的目的不是为每个事实强制使用一种存储原语，而是使持久化的原因变得明确。\u003C\u002Fp>\n\u003Ch2 id=\"section-55\">结论\u003C\u002Fh2>\n\u003Cp>一个有用的智能体不是靠记住最多信息取胜，而是靠保留正确的信息、在现实可能变化时回到权威来源、重新计算那些再次推导更安全的内容，并淘汰不应再影响未来决策的信息。\u003C\u002Fp>\n\u003Cp>因此，对于每一条候选记忆，实际的问题不是“我们能存储这个吗？”，而是：如果这条信息保留下来，未来的决策会更可靠吗？如果答案取决于时效性、权威性、成本、敏感性或修订情况，就把这些条件编码进记忆生命周期，而不是只依赖检索。\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">从研究协议到通用AI推理框架\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">一种与领域无关的推理方法，用于将证据与假设分离、检验相互竞争的假设，并使用明确的验证规则。\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">阅读推理框架 →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-59\">常见问题\u003C\u002Fh2>\n\u003Csection class=\"editorjs-faq my-6 rounded-xl border border-gray-200 p-5 dark:border-gray-700\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">AI智能体记忆生命周期\u003C\u002Fh3>\u003Cdiv id=\"faq1\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI智能体应该长期记住哪些信息？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">优先选择那些持久、可复用、保留来源，并且重建成本高或不可靠的信息，例如稳定的用户偏好、已接受的项目决策、可复用的流程，以及经过验证的长期约束。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq2\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI智能体应该重新检索而不是记住哪些信息？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">具有权威外部来源的易变信息，通常应在用于重要决策前重新检索。例如权限、库存、当前价格、账户状态、策略版本、服务状态和当前文档。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq3\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI智能体应该在什么时候重新计算信息？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">当计算成本低而过期结果代价高时，应重新计算派生值。当重新计算成本高，并且缓存包含来源版本和失效条件时，持久化派生值才更有意义。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq4\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI智能体应该遗忘信息吗？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">应该。对于短暂、过时、敏感、低价值或具有误导性的信息，遗忘、过期和取代是有用的控制手段。无限制保留会产生噪声，并让过时或错误的信息持续影响未来决策。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq5\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">存储整个对话历史是一种好的记忆策略吗？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">仅靠它本身并不是。原始历史可以保留证据，但长期运行的智能体通常需要整理、结构化、摘要、可复用的事实或流程、检索以及生命周期规则，以避免低价值历史主导未来上下文。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-61\">术语表\u003C\u002Fh2>\n\u003Csection class=\"editorjs-glossary my-6 rounded-xl border border-gray-200 dark:border-gray-700 p-5\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">关键记忆生命周期术语\u003C\u002Fh3>\u003Cdl>\u003Cdiv id=\"memory-admission\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">记忆准入\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">决定信息是否被允许成为持久智能体记忆的决策过程。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"supersession\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">取代\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">将较旧的记忆或决策标记为已被较新信息替换，同时在需要时保留历史记录。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"invalidation\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">失效\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">使已存储或缓存的值在不刷新、不重新计算或不审查的情况下不安全复用的规则或事件。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"provenance\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">来源\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">描述信息来自何处、何时被观察到、由谁或什么断言，以及如何被转换的元数据。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"volatility\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">易变性\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">信息在存储时间与复用时间之间发生变化的可能性。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"reconstruction-cost\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">重建成本\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">为恢复或重新生成信息而不是存储信息所需的时间、金钱、计算、工具使用或不确定性。\u003C\u002Fdd>\u003C\u002Fdiv>\u003C\u002Fdl>\u003C\u002Fsection>\n\u003Ch2 id=\"section-63\">主要来源与延伸阅读\u003C\u002Fh2>\n\u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fcookbook\u002Fexamples\u002Fagents_sdk\u002Fsession_memory\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">OpenAI — 上下文工程：使用会话进行短期记忆管理\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于裁剪、摘要、长期运行上下文，以及过时细节和上下文污染等风险的指导。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-context-engineering-for-ai-agents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Anthropic — 面向AI智能体的有效上下文工程\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于整理、压缩、结构化笔记记录，以及在长期范围内维护有用智能体上下文的工程指导。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-harnesses-for-long-running-agents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Anthropic — 面向长期运行智能体的有效框架\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于在长期运行的智能体任务中跨上下文窗口保留进度和产物的实践工作。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fblog\u002Ffrom-raw-interaction-to-reusable-knowledge-rethinking-memory-for-ai-agents\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Microsoft Research — PlugMem\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于将原始智能体交互转化为结构化、可复用的事实和技能，而不是累积未区分历史的研究。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fmstar-every-task-deserves-its-own-memory-harness\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Microsoft Research — M★：每个任务都值得拥有自己的记忆框架\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">研究表明，任务特定的记忆机制可以优于固定的通用记忆设计。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fmemgym-a-long-horizon-memory-environment-for-llm-agents\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Microsoft Research — MemGym\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">一个用于在长期智能体环境中隔离和评估记忆性能的基准。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fhuman-inspired-memory-architecture-for-llm-agents\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Microsoft Research — 面向LLM智能体的类人记忆架构\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">探索持久智能体记忆中的巩固、基于干扰的遗忘、再巩固和检索的研究。\u003C\u002Fp>\u003C\u002Fa>",{"time":212,"blocks":213,"version":812},1790367146044,[214,222,228,236,243,248,253,258,263,268,299,304,309,351,358,363,368,373,378,383,388,393,400,405,410,415,420,425,430,435,440,472,477,521,526,531,536,541,546,551,556,578,583,588,606,611,616,621,626,631,636,641,646,651,656,661,666,671,680,685,711,716,743,748,758,767,776,785,794,803],{"id":215,"data":216,"type":220,"tunes":221},"XDf71jsthn",{"title":217,"maxLevel":218,"minLevel":219},"目录",3,2,"tableOfContents",{},{"id":223,"data":224,"type":226,"tunes":227},"intro",{"text":225},"长时间运行的AI代理会积累远超其应永久记住的信息量。对话、工具输出、中间计算、用户偏好、项目决策、搜索结果、系统状态、错误以及成功流程，在当下都可能看起来有用。将它们全部视为持久记忆会引发第二个问题：代理之后必须判断哪些已存储的信息仍然可信、最新、相关且可安全复用。","paragraph",{},{"id":229,"data":230,"type":234,"tunes":235},"direct",{"body":231,"title":232,"variant":233},"AI代理应当\u003Cstrong>记住那些持久、可复用、保留来源且重新发现成本高昂的信息\u003C\u002Fstrong>；\u003Cstrong>从权威来源重新读取或检索易变事实\u003C\u002Fstrong>；\u003Cstrong>在时效性重要时重新计算廉价的派生值\u003C\u002Fstrong>；以及\u003Cstrong>遗忘、过期或取代那些未来复用带来的风险大于价值的信息\u003C\u002Fstrong>。正确的行动较少取决于信息是否“重要”，而更多取决于其易变性、权威性、派生成本、复用价值、敏感性以及修订行为。","直接回答","info","callout",{},{"id":237,"data":238,"type":234,"tunes":242},"model-note",{"body":239,"title":240,"variant":241},"下文中的“记住\u002F重读\u002F重算\u002F遗忘”模型以及记忆准入测试是本文提出的实用架构工具。它们并非正式的行业标准。它们旨在使代理记忆决策变得明确、可测试且可审计。","关于决策模型","note",{},{"id":244,"data":245,"type":42,"tunes":247},"h-lifecycle",{"text":246,"level":219},"真正的记忆问题不是存储——而是生命周期控制",{},{"id":249,"data":250,"type":226,"tunes":252},"p-life-1",{"text":251},"现代代理系统几乎可以存储任何内容：完整记录、摘要、嵌入、文件、数据库记录、工具轨迹、结构化事实、技能以及外部工件。因此，存储容量并非难点。难点在于决定什么值得保留、应保留多久，以及当现实发生变化时必须发生什么。",{},{"id":254,"data":255,"type":226,"tunes":257},"p-life-2",{"text":256},"OpenAI的会话记忆指南明确警告，将过多历史向前携带会造成分心、低效、上下文污染以及错误累积。Anthropic同样将上下文视为必须精心管理而非累积的有限资源。微软研究院也朝着同一方向前进：PlugMem将原始交互历史转换为可复用的结构化知识，而不是将完整历史视为同等有价值的记忆。",{},{"id":259,"data":260,"type":226,"tunes":262},"p-life-3",{"text":261},"其架构后果很简单：记忆需要准入策略、维护策略和退役策略。仅靠检索器无法提供这些语义。",{},{"id":264,"data":265,"type":42,"tunes":267},"h-actions",{"text":266,"level":219},"对任何代理信息可采取的四种可能行动",{},{"id":269,"data":270,"type":297,"tunes":298},"table-actions",{"content":271,"stretched":43,"withHeadings":14},[272,277,282,287,292],[273,274,275,276],"行动","使用时机","典型示例","主要风险",[278,279,280,281],"记住","该信息在未来任务中仍然有用，且可靠重建成本高昂或不可能","稳定的用户偏好、已接受的项目决策、可复用技能、已验证的长期约束","持久化错误、过时或过于宽泛的内容",[283,284,285,286],"重读\u002F检索","该信息具有可能变化的权威来源","权限、库存、策略版本、订单状态、产品价格、当前API文档","使用旧副本而非当前权威来源",[288,289,290,291],"重算","该信息是派生的，且重新计算成本足够低","基于当前源数据的总计、分数、排名、摘要，确定性转换","持久化过时的派生输出",[293,294,295,296],"遗忘\u002F过期\u002F取代","未来复用价值很小，或会带来隐私、过时、冲突或污染风险","临时工具输出、失败假设、被取代的决策、临时令牌、过时环境状态","丢失后来证明必要的信息","table",{},{"id":300,"data":301,"type":42,"tunes":303},"h-admission",{"text":302,"level":219},"记忆准入测试",{},{"id":305,"data":306,"type":226,"tunes":308},"p-admission-intro",{"text":307},"在信息成为持久代理记忆之前，用六个属性对其进行测试。这些属性比模糊的重要性评分更有用，因为它们能预测信息随时间变化的行为。",{},{"id":310,"data":311,"type":349,"tunes":350},"admission-comparison",{"rows":312,"title":338,"layout":297,"columns":339},[313,318,322,326,330,334],{"id":314,"label":315,"values":316},"volatility","易变性",[317,317,317],"",{"id":319,"label":320,"values":321},"authority","权威性",[317,317,317],{"id":323,"label":324,"values":325},"reuse","复用价值",[317,317,317],{"id":327,"label":328,"values":329},"reconstruction","重建成本",[317,317,317],{"id":331,"label":332,"values":333},"sensitivity","敏感性",[317,317,317],{"id":335,"label":336,"values":337},"revision","修订行为",[317,317,317],"决定信息是否属于记忆的六个属性",[340,343,346],{"id":341,"label":342},"property","属性",{"id":344,"label":345},"question","问题",{"id":347,"label":348},"effect","决策压力","comparison",{},{"id":352,"data":353,"type":234,"tunes":357},"rule-volatile",{"body":354,"title":355,"variant":356},"如果一个事实\u003Cstrong>易变 + 在别处有权威来源 + 获取成本低\u003C\u002Fstrong>，就不要将复制的值提升为长期记忆。应存储指针、标识符或检索路径。","一条实用规则","tip",{},{"id":359,"data":360,"type":42,"tunes":362},"h-remember",{"text":361,"level":218},"1. 记住：能改善未来决策的持久知识",{},{"id":364,"data":365,"type":226,"tunes":367},"p-remember-1",{"text":366},"良好的持久记忆能减少重复工作，而不会把昨天的状态变成今天的真相。典型候选包括明确的用户偏好、持久的项目约束、决策及其理由、可复用流程、反复出现的失败模式，以及预计不会频繁变化的已验证事实。",{},{"id":369,"data":370,"type":226,"tunes":372},"p-remember-2",{"text":371},"最强的记忆未必是原始记录。PlugMem在2026年的工作主张将交互历史转换为紧凑事实和可复用技能。微软的BREW同样将过去轨迹提炼为可检索的程序性知识，描述该做什么、何时适用以及需要注意什么。两者都指向一个有用的设计原则：存储可复用知识，而不仅仅是历史文本。",{},{"id":374,"data":375,"type":226,"tunes":377},"p-remember-3",{"text":376},"被记住的条目还应保留来源。未来的代理应能区分“用户明确要求这一点”、“系统观察到这一点”、“某个来源陈述了这一点”以及“某个模型推断了这一点”。没有这种区分，记忆会逐渐将证据、解释和推测混为一个无差别的池子。",{},{"id":379,"data":380,"type":42,"tunes":382},"h-reread",{"text":381,"level":218},"2. 重新读取或检索：具有外部真实来源的易变事实",{},{"id":384,"data":385,"type":226,"tunes":387},"p-reread-1",{"text":386},"有些信息之所以有价值，恰恰是因为它会变化。当前权限、订单状态、库存、账户状态、服务健康状况、软件文档、价格、日程、法规和 API 行为，在用于重要决策之前，通常应从拥有这些信息的系统中重新读取。",{},{"id":389,"data":390,"type":226,"tunes":392},"p-reread-2",{"text":391},"智能体可以记住某个来源存在、如何访问它，或哪些字段重要。但它不应假设过去检索到的值仍然具有权威性。这将关于从何处以及如何获取真相的记忆，与真相的缓存副本区分开来。",{},{"id":394,"data":395,"type":234,"tunes":399},"stale-trap",{"body":396,"title":397,"variant":398},"一个事实可能被完美地记住，但仍然是错误的。记忆质量不仅在于回忆的准确性；它还包括知道何时必须让位于全新的权威读取。","陈旧记忆陷阱","warning",{},{"id":401,"data":402,"type":42,"tunes":404},"h-recompute",{"text":403,"level":218},"3. 重新计算：计算比信任成本更低的派生信息",{},{"id":406,"data":407,"type":226,"tunes":409},"p-recompute-1",{"text":408},"派生信息应与源事实区别对待。如果一个值可以从当前输入确定性地重新计算，那么持久化结果可能会造成不必要的陈旧。总计、百分比、排名、资格标志、生成的摘要以及其他派生输出，在使用时通常应重新计算。",{},{"id":411,"data":412,"type":226,"tunes":414},"p-recompute-2",{"text":413},"关键的权衡是成本。如果重新计算成本高昂，系统可以将结果与确切的输入版本、时间戳、推导方法和失效条件一起缓存。如果重新计算成本低廉，新鲜度通常胜出。",{},{"id":416,"data":417,"type":42,"tunes":419},"h-forget",{"text":418,"level":218},"4. 遗忘、过期或取代：删除是一种能力",{},{"id":421,"data":422,"type":226,"tunes":424},"p-forget-1",{"text":423},"遗忘不一定是一种缺陷。它是一种控制机制。临时的工具输出、一次性的搜索结果、失败的假设、临时环境状态、中间推理产物、过时的用户偏好、过期的凭证以及被取代的决策，如果无限期地保持活跃，都可能成为负担。",{},{"id":426,"data":427,"type":226,"tunes":429},"p-forget-2",{"text":428},"近期的记忆研究越来越认识到，无限制的积累会降低性能。微软 2026 年受人类启发的记忆架构明确包含了基于干扰的遗忘和巩固，而 PlugMem 报告称，原始历史记录可能会用低价值上下文压垮智能体。工程上的教训并不需要复制生物记忆：保留应该是有选择性的。",{},{"id":431,"data":432,"type":226,"tunes":434},"p-forget-3",{"text":433},"在许多系统中，取代比立即删除更安全。旧决策仍然可审计，但检索默认使用新决策。这对于项目、政策、合规以及任何变更历史本身就是证据的工作流程都很重要。",{},{"id":436,"data":437,"type":42,"tunes":439},"h-method",{"text":438,"level":219},"决策方法",{},{"id":441,"data":442,"type":470,"tunes":471},"decision-flow",{"steps":443,"title":468,"orientation":469},[444,447,450,453,456,459,462,465],{"label":445,"description":446},"1. 对信息进行分类","它是权威状态、用户偏好、外部证据、派生输出、程序、观察还是模型推断？",{"label":448,"description":449},"2. 确定真实来源","确定是否有其他系统或来源比记忆本身更具权威性。",{"label":451,"description":452},"3. 估计易变性","询问该项在下一次有意义的重用之前发生变化的可能性有多大。",{"label":454,"description":455},"4. 估计重用和重建成本","将未来价值与获取或重新创建信息的成本和可靠性进行比较。",{"label":457,"description":458},"5. 检查敏感性和范围","定义谁可以访问该信息、它可以在哪里持久化，以及持久化是否合理。",{"label":460,"description":461},"6. 定义失效","指定过期、取代、冲突解决，或强制进行全新权威读取的条件。",{"label":463,"description":464},"7. 选择操作","记住、重新读取\u002F检索、重新计算，或遗忘\u002F过期\u002F取代。",{"label":466,"description":467},"8. 保留来源","存储足够的元数据，以区分源事实、用户陈述、观察、推导和模型推断。","决定信息项的生命周期","auto","processFlow",{},{"id":473,"data":474,"type":42,"tunes":476},"h-examples",{"text":475,"level":219},"示例：同一个智能体应使用不同的生命周期操作",{},{"id":478,"data":479,"type":297,"tunes":520},"examples-table",{"content":480,"stretched":43,"withHeadings":14},[481,485,488,492,496,500,504,508,512,516],[482,483,484],"信息","推荐操作","原因",[486,278,487],"“用户偏好简洁的技术性回答。”","稳定的偏好，具有高重用价值",[489,490,491],"“部署当前已暂停。”","重新读取","当前操作状态可能发生变化",[493,494,495],"“预计总成本为 48,620 欧元。”","根据当前输入重新计算","派生值应跟随源变化",[497,498,499],"昨天一个 20,000 token 的原始工具响应","遗忘或外部归档","直接重用价值低；上下文成本高",[501,502,503],"针对反复出现的构建失败的已确认解决方法","作为可重用程序记住","未来重用价值高，重新发现成本高",[505,506,507],"模型对服务器故障原因的猜测","不要提升为持久事实","推断不是经过验证的证据",[509,510,511],"一个旧的项目决策后来被新决策取代","取代，保留审计历史","最新决策应胜出，同时不抹除来源",[513,514,515],"当前产品价格","再次检索","高易变性和外部权威",[517,518,519],"法律或政策解释","仅在有来源\u002F版本元数据时记住先前的分析；行动前重新检查权威性","适用性可能随时间和司法管辖区而变化",{},{"id":522,"data":523,"type":42,"tunes":525},"h-conditions",{"text":524,"level":219},"记忆应存储条件，而不仅仅是结论",{},{"id":527,"data":528,"type":226,"tunes":530},"p-cond-1",{"text":529},"当持久记忆只存储结论而丢失了结论有效时的条件时，它就会变得危险。“使用每租户数据库”比“当监管隔离和租户特定生命周期要求超过运营开销时，使用每租户数据库”更弱。第二种形式保留了决策边界。",{},{"id":532,"data":533,"type":226,"tunes":535},"p-cond-2",{"text":534},"这对于智能体学习到的程序更为重要。一个成功的工作流程不仅应捕获步骤，还应捕获前提条件、环境、工具版本、可观察的成功标准以及已知的失败模式。否则，在错误环境中检索到的记忆可能会自信地重现一个过时的解决方案。",{},{"id":537,"data":538,"type":42,"tunes":540},"h-write-cost",{"text":539,"level":219},"记忆写入应比记忆读取更昂贵",{},{"id":542,"data":543,"type":226,"tunes":545},"p-write-1",{"text":544},"读取弱记忆可能损害一个答案。写入弱记忆可能损害许多未来的答案。这种不对称性表明写入路径应比读取路径更严格：对候选记忆进行分类、检查来源、检测矛盾、应用敏感性规则、定义作用域，并决定是否需要人工确认或外部验证。",{},{"id":547,"data":548,"type":226,"tunes":550},"p-write-2",{"text":549},"当智能体从自身生成的输出中写入记忆时，这一点尤为重要。生成的摘要可能包含压缩错误。工具故障可能被误解。一个看似合理的假设可能被存储为事实。如果这些输出在没有证据状态的情况下成为未来上下文，智能体可能创建自我强化的错误循环。",{},{"id":552,"data":553,"type":42,"tunes":555},"h-quality",{"text":554,"level":219},"记忆质量至少包含五个维度",{},{"id":557,"data":558,"type":297,"tunes":577},"quality-table",{"content":559,"stretched":43,"withHeadings":14},[560,562,565,568,571,574],[561,345],"维度",[563,564],"保留质量","系统是否保留了应当留存的信息？",[566,567],"检索质量","系统能否在需要时恢复正确的记忆？",[569,570],"新鲜度质量","系统是否知道存储的信息何时不再是最新的？",[572,573],"来源质量","系统能否区分来源、用户陈述、观察、推导和推断？",[575,576],"退役质量","当信息不应再影响决策时，系统能否使其过期、取代、限制或删除？",{},{"id":579,"data":580,"type":226,"tunes":582},"p-quality-1",{"text":581},"基准测试正开始区分这些关注点。微软的 MemGym 明确评估长时程智能体场景中的记忆，并报告记忆隔离分数，旨在减少推理、检索和工具使用能力的混杂影响。这一方向很重要，因为仅凭最终任务分数无法判断记忆本身是有帮助、有害还是无关。",{},{"id":584,"data":585,"type":42,"tunes":587},"h-not-store",{"text":586,"level":219},"默认情况下不应放入持久记忆的内容",{},{"id":589,"data":590,"type":604,"tunes":605},"not-store-list",{"meta":591,"items":592,"style":603},{},[593,594,595,596,597,598,599,600,601,602],"原始思维链或隐藏推理产物。","临时认证令牌、机密或凭据。","未经核实的模型生成假设。","存在实时权威系统的易变状态。","缺少源输入且可低成本重新计算的派生值。","仅因存储可用而保存的大型工具输出。","已由更好的真相来源管理的信息的重复副本。","没有明确持久化目的、访问范围和生命周期的敏感个人数据。","没有明确版本或退役语义的已被取代的结论。","仅对当前运行有用且没有可复用诊断价值的错误消息或故障状态。","unordered","list",{},{"id":607,"data":608,"type":42,"tunes":610},"h-task-specific",{"text":609,"level":219},"记忆是任务特定的——不存在通用的最优存储",{},{"id":612,"data":613,"type":226,"tunes":615},"p-task-1",{"text":614},"编码智能体受益于可复用流程、仓库约定、成功修复模式和项目决策。个人助理可能需要偏好、承诺和关系上下文。商务智能体更需要当前产品和交易状态，而非价格或库存的历史副本。研究智能体受益于来源出处、未解决的假设和明确的证据状态。",{},{"id":617,"data":618,"type":226,"tunes":620},"p-task-2",{"text":619},"微软研究院的 M-star 工作直接指出了这一点：为某一目的优化的记忆系统可能难以迁移到另一目的，而任务特定的记忆机制可能优于固定的通用设计。因此，记忆模式应遵循智能体必须做出的决策，而不是强加给每个智能体的通用模板。",{},{"id":622,"data":623,"type":42,"tunes":625},"h-change",{"text":624,"level":219},"什么会改变这个答案？",{},{"id":627,"data":628,"type":226,"tunes":630},"p-change-1",{"text":629},"当检索缓慢或昂贵、权威系统间歇性不可用、重新计算成本高昂、审计规则要求历史快照，或智能体必须离线运行时，这种平衡会发生变化。在这些情况下，可能需要缓存或持久化更多信息——但需附带版本、来源、时间戳和失效元数据。",{},{"id":632,"data":633,"type":226,"tunes":635},"p-change-2",{"text":634},"对于主要价值在于个性化的智能体，这种平衡也会发生变化。稳定的偏好可能值得记住，即使技术上可以再次询问。相反，在高风险领域，将观察或解释转化为持久记忆的门槛应高得多。",{},{"id":637,"data":638,"type":226,"tunes":640},"p-change-3",{"text":639},"未来的托管记忆平台可能会自动化整合、检索、遗忘和上下文构建。这可以减少实现工作，但并不能消除治理问题：哪些信息被允许影响未来决策，在什么条件下，以及系统何时必须回到当前的真相来源？",{},{"id":642,"data":643,"type":42,"tunes":645},"h-limitations",{"text":644,"level":219},"局限性",{},{"id":647,"data":648,"type":226,"tunes":650},"p-limit-1",{"text":649},"在当前框架和研究中，“智能体记忆”没有统一定义。一些系统用该术语指代对话历史，另一些则指外部持久存储、结构化知识、学习到的流程、检查点或模型适配。本文中的决策模型侧重于操作生命周期语义，而非强制统一术语。",{},{"id":652,"data":653,"type":226,"tunes":655},"p-limit-2",{"text":654},"四个生命周期操作也可能重叠。一个系统可以在同一工作流中记住稳定的摘要、保留指向来源的指针、重新读取易变字段并重新计算派生结果。该模型的目的不是为每个事实强制使用一种存储原语，而是使持久化的原因变得明确。",{},{"id":657,"data":658,"type":42,"tunes":660},"h-conclusion",{"text":659,"level":219},"结论",{},{"id":662,"data":663,"type":226,"tunes":665},"p-conclusion-1",{"text":664},"一个有用的智能体不是靠记住最多信息取胜，而是靠保留正确的信息、在现实可能变化时回到权威来源、重新计算那些再次推导更安全的内容，并淘汰不应再影响未来决策的信息。",{},{"id":667,"data":668,"type":226,"tunes":670},"p-conclusion-2",{"text":669},"因此，对于每一条候选记忆，实际的问题不是“我们能存储这个吗？”，而是：如果这条信息保留下来，未来的决策会更可靠吗？如果答案取决于时效性、权威性、成本、敏感性或修订情况，就把这些条件编码进记忆生命周期，而不是只依赖检索。",{},{"id":672,"data":673,"type":678,"tunes":679},"internal-reasoning",{"url":674,"title":675,"excerpt":676,"ctaLabel":677},"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework","从研究协议到通用AI推理框架","一种与领域无关的推理方法，用于将证据与假设分离、检验相互竞争的假设，并使用明确的验证规则。","阅读推理框架","referralArticle",{},{"id":681,"data":682,"type":42,"tunes":684},"h-faq",{"text":683,"level":219},"常见问题",{},{"id":686,"data":687,"type":686,"tunes":710},"faq",{"items":688,"title":709},[689,693,697,701,705],{"id":690,"answer":691,"question":692},"faq1","优先选择那些持久、可复用、保留来源，并且重建成本高或不可靠的信息，例如稳定的用户偏好、已接受的项目决策、可复用的流程，以及经过验证的长期约束。","AI智能体应该长期记住哪些信息？",{"id":694,"answer":695,"question":696},"faq2","具有权威外部来源的易变信息，通常应在用于重要决策前重新检索。例如权限、库存、当前价格、账户状态、策略版本、服务状态和当前文档。","AI智能体应该重新检索而不是记住哪些信息？",{"id":698,"answer":699,"question":700},"faq3","当计算成本低而过期结果代价高时，应重新计算派生值。当重新计算成本高，并且缓存包含来源版本和失效条件时，持久化派生值才更有意义。","AI智能体应该在什么时候重新计算信息？",{"id":702,"answer":703,"question":704},"faq4","应该。对于短暂、过时、敏感、低价值或具有误导性的信息，遗忘、过期和取代是有用的控制手段。无限制保留会产生噪声，并让过时或错误的信息持续影响未来决策。","AI智能体应该遗忘信息吗？",{"id":706,"answer":707,"question":708},"faq5","仅靠它本身并不是。原始历史可以保留证据，但长期运行的智能体通常需要整理、结构化、摘要、可复用的事实或流程、检索以及生命周期规则，以避免低价值历史主导未来上下文。","存储整个对话历史是一种好的记忆策略吗？","AI智能体记忆生命周期",{},{"id":712,"data":713,"type":42,"tunes":715},"h-glossary",{"text":714,"level":219},"术语表",{},{"id":717,"data":718,"type":717,"tunes":742},"glossary",{"title":719,"entries":720},"关键记忆生命周期术语",[721,725,729,733,737,739],{"term":722,"anchor":723,"definition":724},"记忆准入","memory-admission","决定信息是否被允许成为持久智能体记忆的决策过程。",{"term":726,"anchor":727,"definition":728},"取代","supersession","将较旧的记忆或决策标记为已被较新信息替换，同时在需要时保留历史记录。",{"term":730,"anchor":731,"definition":732},"失效","invalidation","使已存储或缓存的值在不刷新、不重新计算或不审查的情况下不安全复用的规则或事件。",{"term":734,"anchor":735,"definition":736},"来源","provenance","描述信息来自何处、何时被观察到、由谁或什么断言，以及如何被转换的元数据。",{"term":315,"anchor":314,"definition":738},"信息在存储时间与复用时间之间发生变化的可能性。",{"term":328,"anchor":740,"definition":741},"reconstruction-cost","为恢复或重新生成信息而不是存储信息所需的时间、金钱、计算、工具使用或不确定性。",{},{"id":744,"data":745,"type":42,"tunes":747},"h-sources",{"text":746,"level":219},"主要来源与延伸阅读",{},{"id":749,"data":750,"type":756,"tunes":757},"src-openai-session",{"link":751,"meta":752},"https:\u002F\u002Fdevelopers.openai.com\u002Fcookbook\u002Fexamples\u002Fagents_sdk\u002Fsession_memory",{"image":753,"title":754,"description":755},{"url":317},"OpenAI — 上下文工程：使用会话进行短期记忆管理","关于裁剪、摘要、长期运行上下文，以及过时细节和上下文污染等风险的指导。","linkTool",{},{"id":759,"data":760,"type":756,"tunes":766},"src-anthropic-context",{"link":761,"meta":762},"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-context-engineering-for-ai-agents",{"image":763,"title":764,"description":765},{"url":317},"Anthropic — 面向AI智能体的有效上下文工程","关于整理、压缩、结构化笔记记录，以及在长期范围内维护有用智能体上下文的工程指导。",{},{"id":768,"data":769,"type":756,"tunes":775},"src-anthropic-harness",{"link":770,"meta":771},"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-harnesses-for-long-running-agents",{"image":772,"title":773,"description":774},{"url":317},"Anthropic — 面向长期运行智能体的有效框架","关于在长期运行的智能体任务中跨上下文窗口保留进度和产物的实践工作。",{},{"id":777,"data":778,"type":756,"tunes":784},"src-ms-plugmem",{"link":779,"meta":780},"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fblog\u002Ffrom-raw-interaction-to-reusable-knowledge-rethinking-memory-for-ai-agents\u002F",{"image":781,"title":782,"description":783},{"url":317},"Microsoft Research — PlugMem","关于将原始智能体交互转化为结构化、可复用的事实和技能，而不是累积未区分历史的研究。",{},{"id":786,"data":787,"type":756,"tunes":793},"src-ms-mstar",{"link":788,"meta":789},"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fmstar-every-task-deserves-its-own-memory-harness\u002F",{"image":790,"title":791,"description":792},{"url":317},"Microsoft Research — M★：每个任务都值得拥有自己的记忆框架","研究表明，任务特定的记忆机制可以优于固定的通用记忆设计。",{},{"id":795,"data":796,"type":756,"tunes":802},"src-ms-memgym",{"link":797,"meta":798},"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fmemgym-a-long-horizon-memory-environment-for-llm-agents\u002F",{"image":799,"title":800,"description":801},{"url":317},"Microsoft Research — MemGym","一个用于在长期智能体环境中隔离和评估记忆性能的基准。",{},{"id":804,"data":805,"type":756,"tunes":811},"src-ms-human-memory",{"link":806,"meta":807},"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fhuman-inspired-memory-architecture-for-llm-agents\u002F",{"image":808,"title":809,"description":810},{"url":317},"Microsoft Research — 面向LLM智能体的类人记忆架构","探索持久智能体记忆中的巩固、基于干扰的遗忘、再巩固和检索的研究。",{},"2.31","长时间运行的代理不应记住所有内容。本文提供了一个实用的生命周期模型，用于决定哪些内容应属于持久记忆、哪些内容应重新检索、哪些内容重新计算更安全，以及哪些内容应过期或被取代。","\u002Fuploads\u002F2026\u002F09\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again-1790351131087-iehz28.webp","what-should-an-ai-agent-remember-forget-recompute-or-retrieve-again-1790351131087-iehz28","PUBLISHED","2026-09-25T09:43:00.000Z","2026-09-25T15:43:41.228Z","2026-09-25T20:20:04.123Z",{"en":821,"de":822,"sr":823,"es":824,"fr":825,"it":826,"ru":827,"zh":828},"\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u002Fde\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u002Fsr\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u002Fes\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u002Ffr\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u002Fit\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u002Fru\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","\u002Fzh\u002Fblog\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again",[830,834,838],{"id":831,"name":832,"slug":833},57,"数据边界","data-boundaries",{"id":835,"name":836,"slug":837},84,"策略与数据边界","policy-and-data",{"id":839,"name":840,"slug":841},54,"威胁模型","threat-model",{"id":843,"login":844,"email":845,"displayName":846},"20","rooth8233","aleksandar@stajic.de","Aleksandar Stajić",[848,1339],{"lang":849,"title":850,"content":851,"contentJson":852,"excerpt":1338},"en","What Should an AI Agent Remember, Forget, Recompute or Retrieve Again?","{\"time\":1790351469618,\"blocks\":[{\"id\":\"XDf71jsthn\",\"type\":\"tableOfContents\",\"data\":{\"title\":\"Contents\",\"minLevel\":2,\"maxLevel\":3},\"tunes\":{}},{\"id\":\"intro\",\"type\":\"paragraph\",\"data\":{\"text\":\"Long-running AI agents accumulate far more information than they should permanently remember. Conversations, tool outputs, intermediate calculations, user preferences, project decisions, search results, system state, mistakes, and successful procedures can all look useful in the moment. Treating all of them as durable memory creates a second problem: the agent must later decide which stored information is still trustworthy, current, relevant, and safe to reuse.\"},\"tunes\":{}},{\"id\":\"direct\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Direct answer\",\"body\":\"An AI agent should \u003Cstrong>remember information that is durable, reusable, provenance-preserving, and expensive to rediscover\u003C\u002Fstrong>; \u003Cstrong>re-read or retrieve volatile facts from their authoritative source\u003C\u002Fstrong>; \u003Cstrong>recompute cheap derived values when freshness matters\u003C\u002Fstrong>; and \u003Cstrong>forget, expire, or supersede information whose future reuse creates more risk than value\u003C\u002Fstrong>. The correct action depends less on whether information is “important” and more on its volatility, authority, derivation cost, reuse value, sensitivity, and revision behaviour.\"},\"tunes\":{}},{\"id\":\"model-note\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"About the decision model\",\"body\":\"The Remember \u002F Re-read \u002F Recompute \u002F Forget model and the Memory Admission Test below are practical architecture tools proposed in this article. They are not formal industry standards. They are designed to make agent-memory decisions explicit, testable, and auditable.\"},\"tunes\":{}},{\"id\":\"h-lifecycle\",\"type\":\"header\",\"data\":{\"text\":\"The real memory problem is not storage — it is lifecycle control\",\"level\":2},\"tunes\":{}},{\"id\":\"p-life-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Modern agent systems can store almost anything: full transcripts, summaries, embeddings, files, database records, tool traces, structured facts, skills, and external artifacts. Storage capacity is therefore not the hard part. The hard part is deciding what deserves to survive, how long it should survive, and what must happen when reality changes.\"},\"tunes\":{}},{\"id\":\"p-life-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"OpenAI's session-memory guidance explicitly warns that carrying too much history forward can create distraction, inefficiency, context poisoning, and compounding errors. Anthropic similarly treats context as a finite resource that must be curated rather than accumulated. Microsoft Research has moved in the same direction: PlugMem converts raw interaction history into reusable structured knowledge instead of treating the complete history as equally valuable memory.\"},\"tunes\":{}},{\"id\":\"p-life-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The architectural consequence is simple: memory needs an admission policy, a maintenance policy, and a retirement policy. A retriever alone does not provide those semantics.\"},\"tunes\":{}},{\"id\":\"h-actions\",\"type\":\"header\",\"data\":{\"text\":\"Four possible actions for any piece of agent information\",\"level\":2},\"tunes\":{}},{\"id\":\"table-actions\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Action\",\"Use when\",\"Typical examples\",\"Primary risk\"],[\"Remember\",\"The information remains useful across future tasks and is costly or impossible to reconstruct reliably\",\"Stable user preference, accepted project decision, reusable skill, verified long-term constraint\",\"Persisting something false, stale, or too broad\"],[\"Re-read \u002F Retrieve\",\"The information has an authoritative source that may change\",\"Permissions, inventory, policy version, order state, product price, current API documentation\",\"Using an old copy instead of current authority\"],[\"Recompute\",\"The information is derived and inexpensive enough to calculate again\",\"Totals, scores, rankings, summaries from current source data, deterministic transformations\",\"Persisting stale derived output\"],[\"Forget \u002F Expire \u002F Supersede\",\"Future reuse has little value or creates privacy, staleness, conflict, or contamination risk\",\"Transient tool output, failed hypothesis, superseded decision, temporary token, obsolete environment state\",\"Losing information that later proves necessary\"]]},\"tunes\":{}},{\"id\":\"h-admission\",\"type\":\"header\",\"data\":{\"text\":\"The Memory Admission Test\",\"level\":2},\"tunes\":{}},{\"id\":\"p-admission-intro\",\"type\":\"paragraph\",\"data\":{\"text\":\"Before information becomes durable agent memory, test it against six properties. These properties are more useful than a vague importance score because they predict how the information behaves over time.\"},\"tunes\":{}},{\"id\":\"admission-comparison\",\"type\":\"comparison\",\"data\":{\"title\":\"Six properties that decide whether information belongs in memory\",\"layout\":\"table\",\"columns\":[{\"id\":\"property\",\"label\":\"Property\"},{\"id\":\"question\",\"label\":\"Question\"},{\"id\":\"effect\",\"label\":\"Decision pressure\"}],\"rows\":[{\"id\":\"volatility\",\"label\":\"Volatility\",\"values\":[\"\",\"\",\"\"]},{\"id\":\"authority\",\"label\":\"Authority\",\"values\":[\"\",\"\",\"\"]},{\"id\":\"reuse\",\"label\":\"Reuse value\",\"values\":[\"\",\"\",\"\"]},{\"id\":\"reconstruction\",\"label\":\"Reconstruction cost\",\"values\":[\"\",\"\",\"\"]},{\"id\":\"sensitivity\",\"label\":\"Sensitivity\",\"values\":[\"\",\"\",\"\"]},{\"id\":\"revision\",\"label\":\"Revision behaviour\",\"values\":[\"\",\"\",\"\"]}]},\"tunes\":{}},{\"id\":\"rule-volatile\",\"type\":\"callout\",\"data\":{\"variant\":\"tip\",\"title\":\"A practical rule\",\"body\":\"If a fact is \u003Cstrong>volatile + authoritative elsewhere + cheap to fetch\u003C\u002Fstrong>, do not promote a copied value into long-term memory. Store the pointer, identifier, or retrieval path instead.\"},\"tunes\":{}},{\"id\":\"h-remember\",\"type\":\"header\",\"data\":{\"text\":\"1. Remember: durable knowledge that improves future decisions\",\"level\":3},\"tunes\":{}},{\"id\":\"p-remember-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Good durable memory reduces repeated work without turning yesterday's state into today's truth. Typical candidates include explicit user preferences, durable project constraints, decisions and their rationale, reusable procedures, recurring failure patterns, and verified facts that are not expected to change frequently.\"},\"tunes\":{}},{\"id\":\"p-remember-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The strongest memories are not necessarily raw transcripts. PlugMem's 2026 work argues for converting interaction history into compact facts and reusable skills. Microsoft's BREW similarly distills past trajectories into retrievable procedural knowledge describing what to do, when it applies, and what to watch out for. Both point toward a useful design principle: store reusable knowledge, not merely historical text.\"},\"tunes\":{}},{\"id\":\"p-remember-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"A remembered item should also retain provenance. A future agent should be able to distinguish “the user explicitly requested this,” “the system observed this,” “a source stated this,” and “a model inferred this.” Without that distinction, memory gradually converts evidence, interpretation, and speculation into one undifferentiated pool.\"},\"tunes\":{}},{\"id\":\"h-reread\",\"type\":\"header\",\"data\":{\"text\":\"2. Re-read or retrieve: volatile facts with an external source of truth\",\"level\":3},\"tunes\":{}},{\"id\":\"p-reread-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Some information is valuable precisely because it changes. Current permissions, order state, inventory, account status, service health, software documentation, prices, schedules, regulations, and API behaviour should normally be re-read from the system that owns them before consequential use.\"},\"tunes\":{}},{\"id\":\"p-reread-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The agent may remember that a source exists, how to access it, or what fields matter. It should not assume that an old retrieved value remains authoritative. This separates memory of where and how to obtain truth from a cached copy of truth.\"},\"tunes\":{}},{\"id\":\"stale-trap\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"The stale-memory trap\",\"body\":\"A fact can be perfectly remembered and still be wrong. Memory quality is not only recall accuracy; it also includes knowing when recall must yield to a fresh authoritative read.\"},\"tunes\":{}},{\"id\":\"h-recompute\",\"type\":\"header\",\"data\":{\"text\":\"3. Recompute: derived information that is cheaper to calculate than to trust\",\"level\":3},\"tunes\":{}},{\"id\":\"p-recompute-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Derived information deserves different treatment from source facts. If a value can be deterministically recalculated from current inputs, persisting the result may create unnecessary staleness. Totals, percentages, rankings, eligibility flags, generated summaries, and other derived outputs should often be recomputed when used.\"},\"tunes\":{}},{\"id\":\"p-recompute-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The key trade-off is cost. If recomputation is expensive, the system may cache the result together with the exact input version, timestamp, derivation method, and invalidation conditions. If recomputation is cheap, freshness usually wins.\"},\"tunes\":{}},{\"id\":\"h-forget\",\"type\":\"header\",\"data\":{\"text\":\"4. Forget, expire, or supersede: deletion is a capability\",\"level\":3},\"tunes\":{}},{\"id\":\"p-forget-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Forgetting is not necessarily a defect. It is a control mechanism. Transient tool outputs, one-off search results, failed hypotheses, temporary environment state, intermediate reasoning artifacts, obsolete user preferences, expired credentials, and superseded decisions can all become liabilities if they remain active indefinitely.\"},\"tunes\":{}},{\"id\":\"p-forget-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Recent memory research increasingly recognizes that unbounded accumulation can degrade performance. Microsoft's 2026 human-inspired memory architecture explicitly includes interference-based forgetting and consolidation, while PlugMem reports that raw histories can overwhelm agents with low-value context. The engineering lesson does not require copying biological memory: retention should be selective.\"},\"tunes\":{}},{\"id\":\"p-forget-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"In many systems, supersession is safer than immediate deletion. The old decision remains auditable, but retrieval defaults to the new decision. This matters for projects, policies, compliance, and any workflow where the history of change is itself evidence.\"},\"tunes\":{}},{\"id\":\"h-method\",\"type\":\"header\",\"data\":{\"text\":\"The decision method\",\"level\":2},\"tunes\":{}},{\"id\":\"decision-flow\",\"type\":\"processFlow\",\"data\":{\"title\":\"Decide the lifecycle of an information item\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Classify the information\",\"description\":\"Is it authoritative state, user preference, external evidence, derived output, procedure, observation, or model inference?\"},{\"label\":\"2. Identify the source of truth\",\"description\":\"Determine whether another system or source remains more authoritative than the memory itself.\"},{\"label\":\"3. Estimate volatility\",\"description\":\"Ask how likely the item is to change before the next meaningful reuse.\"},{\"label\":\"4. Estimate reuse and reconstruction cost\",\"description\":\"Compare future value with the cost and reliability of fetching or recreating the information.\"},{\"label\":\"5. Check sensitivity and scope\",\"description\":\"Define who may access the information, where it may persist, and whether persistence is justified.\"},{\"label\":\"6. Define invalidation\",\"description\":\"Specify expiry, supersession, conflict resolution, or a condition that forces a fresh authoritative read.\"},{\"label\":\"7. Choose the action\",\"description\":\"Remember, re-read\u002Fretrieve, recompute, or forget\u002Fexpire\u002Fsupersede.\"},{\"label\":\"8. Preserve provenance\",\"description\":\"Store enough metadata to distinguish source fact, user statement, observation, derivation, and model inference.\"}]},\"tunes\":{}},{\"id\":\"h-examples\",\"type\":\"header\",\"data\":{\"text\":\"Examples: the same agent should use different lifecycle actions\",\"level\":2},\"tunes\":{}},{\"id\":\"examples-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Information\",\"Recommended action\",\"Why\"],[\"“The user prefers concise technical answers.”\",\"Remember\",\"Stable preference with high reuse value\"],[\"“The deployment is currently paused.”\",\"Re-read\",\"Current operational state can change\"],[\"“The total projected cost is €48,620.”\",\"Recompute from current inputs\",\"Derived value should follow source changes\"],[\"A 20,000-token raw tool response from yesterday\",\"Forget or archive externally\",\"Low direct reuse; high context cost\"],[\"A confirmed workaround for a recurring build failure\",\"Remember as reusable procedure\",\"High future reuse and expensive rediscovery\"],[\"A model guess about why a server failed\",\"Do not promote to durable fact\",\"Inference is not verified evidence\"],[\"An old project decision later replaced by a new one\",\"Supersede, retain audit history\",\"The latest decision should win without erasing provenance\"],[\"A current product price\",\"Retrieve again\",\"High volatility and external authority\"],[\"A legal or policy interpretation\",\"Remember the prior analysis only with source\u002Fversion metadata; re-check authority before action\",\"Applicability can change with time and jurisdiction\"]]},\"tunes\":{}},{\"id\":\"h-conditions\",\"type\":\"header\",\"data\":{\"text\":\"Memory should store conditions, not only conclusions\",\"level\":2},\"tunes\":{}},{\"id\":\"p-cond-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A durable memory becomes dangerous when it stores only the conclusion and loses the conditions under which the conclusion was valid. “Use database-per-tenant” is weaker than “Use database-per-tenant when regulatory isolation and tenant-specific lifecycle requirements outweigh operational overhead.” The second form preserves the decision boundary.\"},\"tunes\":{}},{\"id\":\"p-cond-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"This matters even more for agent-learned procedures. A successful workflow should capture not only the steps but also the preconditions, environment, tool version, observable success criteria, and known failure modes. Otherwise a memory retrieved in the wrong environment can confidently reproduce an obsolete solution.\"},\"tunes\":{}},{\"id\":\"h-write-cost\",\"type\":\"header\",\"data\":{\"text\":\"A memory write should be more expensive than a memory read\",\"level\":2},\"tunes\":{}},{\"id\":\"p-write-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Reading a weak memory can damage one answer. Writing a weak memory can damage many future answers. The asymmetry suggests a stricter write path than read path: classify the candidate, check provenance, detect contradictions, apply sensitivity rules, define scope, and decide whether human confirmation or external validation is required.\"},\"tunes\":{}},{\"id\":\"p-write-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is especially important when an agent writes memories from its own generated output. A generated summary can contain compression errors. A tool failure can be misinterpreted. A plausible hypothesis can be stored as a fact. If those outputs become future context without evidence status, the agent can create a self-reinforcing error loop.\"},\"tunes\":{}},{\"id\":\"h-quality\",\"type\":\"header\",\"data\":{\"text\":\"Memory quality has at least five dimensions\",\"level\":2},\"tunes\":{}},{\"id\":\"quality-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Dimension\",\"Question\"],[\"Retention quality\",\"Did the system preserve the information that should survive?\"],[\"Retrieval quality\",\"Can the system recover the right memory when it matters?\"],[\"Freshness quality\",\"Does the system know when stored information is no longer current?\"],[\"Provenance quality\",\"Can the system distinguish source, user statement, observation, derivation, and inference?\"],[\"Retirement quality\",\"Can the system expire, supersede, restrict, or remove information when it should no longer influence decisions?\"]]},\"tunes\":{}},{\"id\":\"p-quality-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Benchmarks are starting to separate these concerns. Microsoft's MemGym explicitly evaluates memory in long-horizon agentic settings and reports memory-isolated scores intended to reduce confounding from reasoning, retrieval, and tool-use ability. That direction is important because a final task score alone cannot tell you whether memory itself helped, harmed, or was irrelevant.\"},\"tunes\":{}},{\"id\":\"h-not-store\",\"type\":\"header\",\"data\":{\"text\":\"What not to put into durable memory by default\",\"level\":2},\"tunes\":{}},{\"id\":\"not-store-list\",\"type\":\"list\",\"data\":{\"style\":\"unordered\",\"meta\":{},\"items\":[\"Raw chain-of-thought or hidden reasoning artifacts.\",\"Temporary authentication tokens, secrets, or credentials.\",\"Model-generated hypotheses that have not been verified.\",\"Volatile state that has a live authoritative system.\",\"Cheaply recomputable derived values without their source inputs.\",\"Large tool outputs merely because storage is available.\",\"Duplicate copies of information already governed by a better source of truth.\",\"Sensitive personal data without a clear persistence purpose, access scope, and lifecycle.\",\"Superseded conclusions without explicit version or retirement semantics.\",\"Error messages or failure states that are only useful for the current run and have no reusable diagnostic value.\"]},\"tunes\":{}},{\"id\":\"h-task-specific\",\"type\":\"header\",\"data\":{\"text\":\"Memory is task-specific — there is no universal optimal store\",\"level\":2},\"tunes\":{}},{\"id\":\"p-task-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A coding agent benefits from reusable procedures, repository conventions, successful repair patterns, and project decisions. A personal assistant may need preferences, commitments, and relationship context. A commerce agent needs current product and transaction state far more than historical copies of price or inventory. A research agent benefits from source provenance, unresolved hypotheses, and explicit evidence status.\"},\"tunes\":{}},{\"id\":\"p-task-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft Research's M-star work makes this point directly: memory systems optimized for one purpose may transfer poorly to another, and task-specific memory mechanisms can outperform a fixed general-purpose design. The memory schema should therefore follow the decisions the agent must make, not a universal template imposed on every agent.\"},\"tunes\":{}},{\"id\":\"h-change\",\"type\":\"header\",\"data\":{\"text\":\"What would change this answer?\",\"level\":2},\"tunes\":{}},{\"id\":\"p-change-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The balance changes when retrieval is slow or expensive, authoritative systems are intermittently unavailable, recomputation is costly, audit rules require historical snapshots, or the agent must operate offline. In those cases, more information may need to be cached or persisted — but with version, provenance, timestamp, and invalidation metadata.\"},\"tunes\":{}},{\"id\":\"p-change-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The balance also changes for agents whose primary value is personalization. A stable preference may be worth remembering even if it could technically be asked again. Conversely, in high-risk domains, the threshold for converting an observation or interpretation into durable memory should be much higher.\"},\"tunes\":{}},{\"id\":\"p-change-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Future managed-memory platforms may automate consolidation, retrieval, forgetting, and context construction. That can reduce implementation work, but it does not remove the governance question: which information is allowed to influence future decisions, under what conditions, and when must the system return to the current source of truth?\"},\"tunes\":{}},{\"id\":\"h-limitations\",\"type\":\"header\",\"data\":{\"text\":\"Limitations\",\"level\":2},\"tunes\":{}},{\"id\":\"p-limit-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"There is no single definition of “agent memory” across current frameworks and research. Some systems use the term for conversation history, others for external persistent stores, structured knowledge, learned procedures, checkpoints, or model adaptation. The decision model in this article focuses on operational lifecycle semantics rather than enforcing one vocabulary.\"},\"tunes\":{}},{\"id\":\"p-limit-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The four lifecycle actions can also overlap. A system may remember a stable summary, retain a pointer to the source, re-read volatile fields, and recompute a derived result in one workflow. The purpose of the model is not to force one storage primitive per fact, but to make the reason for persistence explicit.\"},\"tunes\":{}},{\"id\":\"h-conclusion\",\"type\":\"header\",\"data\":{\"text\":\"Conclusion\",\"level\":2},\"tunes\":{}},{\"id\":\"p-conclusion-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A useful agent does not win by remembering the most. It wins by preserving the right information, returning to authoritative sources when reality can change, recalculating what is safer to derive again, and retiring information that should no longer influence future decisions.\"},\"tunes\":{}},{\"id\":\"p-conclusion-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The practical question for every candidate memory is therefore not “Can we store this?” but: Will future decisions be more reliable if this survives? If the answer depends on freshness, authority, cost, sensitivity, or revision, encode those conditions into the memory lifecycle instead of trusting retrieval alone.\"},\"tunes\":{}},{\"id\":\"internal-reasoning\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Fstajic.de\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework\",\"title\":\"From Research Protocol to a General AI Reasoning Framework\",\"excerpt\":\"A domain-independent reasoning method for separating evidence from assumptions, testing competing hypotheses, and using explicit validation rules.\",\"ctaLabel\":\"Read the reasoning framework\"},\"tunes\":{}},{\"id\":\"h-faq\",\"type\":\"header\",\"data\":{\"text\":\"FAQ\",\"level\":2},\"tunes\":{}},{\"id\":\"faq\",\"type\":\"faq\",\"data\":{\"title\":\"AI agent memory lifecycle\",\"items\":[{\"id\":\"faq1\",\"question\":\"What information should an AI agent remember long term?\",\"answer\":\"Prefer information that is durable, reusable, provenance-preserving, and expensive or unreliable to reconstruct, such as stable user preferences, accepted project decisions, reusable procedures, and verified long-term constraints.\"},{\"id\":\"faq2\",\"question\":\"What should an AI agent retrieve again instead of remembering?\",\"answer\":\"Volatile information with an authoritative external source should normally be retrieved again before consequential use. Examples include permissions, inventory, current prices, account state, policy versions, service status, and current documentation.\"},{\"id\":\"faq3\",\"question\":\"When should an AI agent recompute information?\",\"answer\":\"Recompute derived values when the calculation is cheap and stale results would be costly. Persisting a derived value makes more sense when recomputation is expensive and the cache includes the source version and invalidation conditions.\"},{\"id\":\"faq4\",\"question\":\"Should AI agents forget information?\",\"answer\":\"Yes. Forgetting, expiry, and supersession are useful controls for transient, obsolete, sensitive, low-value, or misleading information. Unbounded retention can create noise and allow stale or incorrect information to keep influencing future decisions.\"},{\"id\":\"faq5\",\"question\":\"Is storing the entire conversation history a good memory strategy?\",\"answer\":\"Not by itself. Raw history can preserve evidence, but long-running agents usually need curation, structure, summaries, reusable facts or procedures, retrieval, and lifecycle rules so that low-value history does not dominate future context.\"}]},\"tunes\":{}},{\"id\":\"h-glossary\",\"type\":\"header\",\"data\":{\"text\":\"Glossary\",\"level\":2},\"tunes\":{}},{\"id\":\"glossary\",\"type\":\"glossary\",\"data\":{\"title\":\"Key memory lifecycle terms\",\"entries\":[{\"term\":\"Memory admission\",\"definition\":\"The decision process that determines whether information is allowed to become persistent agent memory.\",\"anchor\":\"memory-admission\"},{\"term\":\"Supersession\",\"definition\":\"Marking an older memory or decision as replaced by newer information while preserving the historical record where needed.\",\"anchor\":\"supersession\"},{\"term\":\"Invalidation\",\"definition\":\"A rule or event that makes a stored or cached value unsafe to reuse without refresh, recomputation, or review.\",\"anchor\":\"invalidation\"},{\"term\":\"Provenance\",\"definition\":\"Metadata describing where information came from, when it was observed, who or what asserted it, and how it was transformed.\",\"anchor\":\"provenance\"},{\"term\":\"Volatility\",\"definition\":\"The likelihood that information will change between the time it is stored and the time it is reused.\",\"anchor\":\"volatility\"},{\"term\":\"Reconstruction cost\",\"definition\":\"The time, money, computation, tool use, or uncertainty required to recover or regenerate information instead of storing it.\",\"anchor\":\"reconstruction-cost\"}]},\"tunes\":{}},{\"id\":\"h-sources\",\"type\":\"header\",\"data\":{\"text\":\"Primary sources and further reading\",\"level\":2},\"tunes\":{}},{\"id\":\"src-openai-session\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevelopers.openai.com\u002Fcookbook\u002Fexamples\u002Fagents_sdk\u002Fsession_memory\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"OpenAI — Context Engineering: Short-Term Memory Management with Sessions\",\"description\":\"Guidance on trimming, summarization, long-running context, and risks such as stale details and context poisoning.\"}},\"tunes\":{}},{\"id\":\"src-anthropic-context\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-context-engineering-for-ai-agents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Anthropic — Effective Context Engineering for AI Agents\",\"description\":\"Engineering guidance on curation, compaction, structured note-taking, and maintaining useful agent context over long horizons.\"}},\"tunes\":{}},{\"id\":\"src-anthropic-harness\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-harnesses-for-long-running-agents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Anthropic — Effective Harnesses for Long-Running Agents\",\"description\":\"Practical work on preserving progress and artifacts across context windows in long-running agent tasks.\"}},\"tunes\":{}},{\"id\":\"src-ms-plugmem\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fblog\u002Ffrom-raw-interaction-to-reusable-knowledge-rethinking-memory-for-ai-agents\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft Research — PlugMem\",\"description\":\"Research on transforming raw agent interactions into structured reusable facts and skills rather than accumulating undifferentiated history.\"}},\"tunes\":{}},{\"id\":\"src-ms-mstar\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fmstar-every-task-deserves-its-own-memory-harness\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft Research — M★: Every Task Deserves Its Own Memory Harness\",\"description\":\"Research showing that task-specific memory mechanisms can outperform fixed general-purpose memory designs.\"}},\"tunes\":{}},{\"id\":\"src-ms-memgym\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fmemgym-a-long-horizon-memory-environment-for-llm-agents\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft Research — MemGym\",\"description\":\"A benchmark for isolating and evaluating memory performance in long-horizon agent environments.\"}},\"tunes\":{}},{\"id\":\"src-ms-human-memory\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fhuman-inspired-memory-architecture-for-llm-agents\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft Research — Human-Inspired Memory Architecture for LLM Agents\",\"description\":\"Research exploring consolidation, interference-based forgetting, reconsolidation, and retrieval in persistent agent memory.\"}},\"tunes\":{}}],\"version\":\"2.31.6\"}",{"time":853,"blocks":854,"version":1337},1790351469618,[855,859,863,868,873,877,881,885,889,893,922,926,930,960,965,969,973,977,981,985,989,993,998,1002,1006,1010,1014,1018,1022,1026,1030,1059,1063,1106,1110,1114,1118,1122,1126,1130,1134,1155,1159,1163,1178,1182,1186,1190,1194,1198,1202,1206,1210,1214,1218,1222,1226,1230,1237,1241,1261,1265,1286,1290,1297,1304,1311,1317,1324,1330],{"id":215,"data":856,"type":220,"tunes":858},{"title":857,"maxLevel":218,"minLevel":219},"Contents",{},{"id":223,"data":860,"type":226,"tunes":862},{"text":861},"Long-running AI agents accumulate far more information than they should permanently remember. Conversations, tool outputs, intermediate calculations, user preferences, project decisions, search results, system state, mistakes, and successful procedures can all look useful in the moment. Treating all of them as durable memory creates a second problem: the agent must later decide which stored information is still trustworthy, current, relevant, and safe to reuse.",{},{"id":229,"data":864,"type":234,"tunes":867},{"body":865,"title":866,"variant":233},"An AI agent should \u003Cstrong>remember information that is durable, reusable, provenance-preserving, and expensive to rediscover\u003C\u002Fstrong>; \u003Cstrong>re-read or retrieve volatile facts from their authoritative source\u003C\u002Fstrong>; \u003Cstrong>recompute cheap derived values when freshness matters\u003C\u002Fstrong>; and \u003Cstrong>forget, expire, or supersede information whose future reuse creates more risk than value\u003C\u002Fstrong>. The correct action depends less on whether information is “important” and more on its volatility, authority, derivation cost, reuse value, sensitivity, and revision behaviour.","Direct answer",{},{"id":237,"data":869,"type":234,"tunes":872},{"body":870,"title":871,"variant":241},"The Remember \u002F Re-read \u002F Recompute \u002F Forget model and the Memory Admission Test below are practical architecture tools proposed in this article. They are not formal industry standards. They are designed to make agent-memory decisions explicit, testable, and auditable.","About the decision model",{},{"id":244,"data":874,"type":42,"tunes":876},{"text":875,"level":219},"The real memory problem is not storage — it is lifecycle control",{},{"id":249,"data":878,"type":226,"tunes":880},{"text":879},"Modern agent systems can store almost anything: full transcripts, summaries, embeddings, files, database records, tool traces, structured facts, skills, and external artifacts. Storage capacity is therefore not the hard part. The hard part is deciding what deserves to survive, how long it should survive, and what must happen when reality changes.",{},{"id":254,"data":882,"type":226,"tunes":884},{"text":883},"OpenAI's session-memory guidance explicitly warns that carrying too much history forward can create distraction, inefficiency, context poisoning, and compounding errors. Anthropic similarly treats context as a finite resource that must be curated rather than accumulated. Microsoft Research has moved in the same direction: PlugMem converts raw interaction history into reusable structured knowledge instead of treating the complete history as equally valuable memory.",{},{"id":259,"data":886,"type":226,"tunes":888},{"text":887},"The architectural consequence is simple: memory needs an admission policy, a maintenance policy, and a retirement policy. A retriever alone does not provide those semantics.",{},{"id":264,"data":890,"type":42,"tunes":892},{"text":891,"level":219},"Four possible actions for any piece of agent information",{},{"id":269,"data":894,"type":297,"tunes":921},{"content":895,"stretched":43,"withHeadings":14},[896,901,906,911,916],[897,898,899,900],"Action","Use when","Typical examples","Primary risk",[902,903,904,905],"Remember","The information remains useful across future tasks and is costly or impossible to reconstruct reliably","Stable user preference, accepted project decision, reusable skill, verified long-term constraint","Persisting something false, stale, or too broad",[907,908,909,910],"Re-read \u002F Retrieve","The information has an authoritative source that may change","Permissions, inventory, policy version, order state, product price, current API documentation","Using an old copy instead of current authority",[912,913,914,915],"Recompute","The information is derived and inexpensive enough to calculate again","Totals, scores, rankings, summaries from current source data, deterministic transformations","Persisting stale derived output",[917,918,919,920],"Forget \u002F Expire \u002F Supersede","Future reuse has little value or creates privacy, staleness, conflict, or contamination risk","Transient tool output, failed hypothesis, superseded decision, temporary token, obsolete environment state","Losing information that later proves necessary",{},{"id":300,"data":923,"type":42,"tunes":925},{"text":924,"level":219},"The Memory Admission Test",{},{"id":305,"data":927,"type":226,"tunes":929},{"text":928},"Before information becomes durable agent memory, test it against six properties. These properties are more useful than a vague importance score because they predict how the information behaves over time.",{},{"id":310,"data":931,"type":349,"tunes":959},{"rows":932,"title":951,"layout":297,"columns":952},[933,936,939,942,945,948],{"id":314,"label":934,"values":935},"Volatility",[317,317,317],{"id":319,"label":937,"values":938},"Authority",[317,317,317],{"id":323,"label":940,"values":941},"Reuse value",[317,317,317],{"id":327,"label":943,"values":944},"Reconstruction cost",[317,317,317],{"id":331,"label":946,"values":947},"Sensitivity",[317,317,317],{"id":335,"label":949,"values":950},"Revision behaviour",[317,317,317],"Six properties that decide whether information belongs in memory",[953,955,957],{"id":341,"label":954},"Property",{"id":344,"label":956},"Question",{"id":347,"label":958},"Decision pressure",{},{"id":352,"data":961,"type":234,"tunes":964},{"body":962,"title":963,"variant":356},"If a fact is \u003Cstrong>volatile + authoritative elsewhere + cheap to fetch\u003C\u002Fstrong>, do not promote a copied value into long-term memory. Store the pointer, identifier, or retrieval path instead.","A practical rule",{},{"id":359,"data":966,"type":42,"tunes":968},{"text":967,"level":218},"1. Remember: durable knowledge that improves future decisions",{},{"id":364,"data":970,"type":226,"tunes":972},{"text":971},"Good durable memory reduces repeated work without turning yesterday's state into today's truth. Typical candidates include explicit user preferences, durable project constraints, decisions and their rationale, reusable procedures, recurring failure patterns, and verified facts that are not expected to change frequently.",{},{"id":369,"data":974,"type":226,"tunes":976},{"text":975},"The strongest memories are not necessarily raw transcripts. PlugMem's 2026 work argues for converting interaction history into compact facts and reusable skills. Microsoft's BREW similarly distills past trajectories into retrievable procedural knowledge describing what to do, when it applies, and what to watch out for. Both point toward a useful design principle: store reusable knowledge, not merely historical text.",{},{"id":374,"data":978,"type":226,"tunes":980},{"text":979},"A remembered item should also retain provenance. A future agent should be able to distinguish “the user explicitly requested this,” “the system observed this,” “a source stated this,” and “a model inferred this.” Without that distinction, memory gradually converts evidence, interpretation, and speculation into one undifferentiated pool.",{},{"id":379,"data":982,"type":42,"tunes":984},{"text":983,"level":218},"2. Re-read or retrieve: volatile facts with an external source of truth",{},{"id":384,"data":986,"type":226,"tunes":988},{"text":987},"Some information is valuable precisely because it changes. Current permissions, order state, inventory, account status, service health, software documentation, prices, schedules, regulations, and API behaviour should normally be re-read from the system that owns them before consequential use.",{},{"id":389,"data":990,"type":226,"tunes":992},{"text":991},"The agent may remember that a source exists, how to access it, or what fields matter. It should not assume that an old retrieved value remains authoritative. This separates memory of where and how to obtain truth from a cached copy of truth.",{},{"id":394,"data":994,"type":234,"tunes":997},{"body":995,"title":996,"variant":398},"A fact can be perfectly remembered and still be wrong. Memory quality is not only recall accuracy; it also includes knowing when recall must yield to a fresh authoritative read.","The stale-memory trap",{},{"id":401,"data":999,"type":42,"tunes":1001},{"text":1000,"level":218},"3. Recompute: derived information that is cheaper to calculate than to trust",{},{"id":406,"data":1003,"type":226,"tunes":1005},{"text":1004},"Derived information deserves different treatment from source facts. If a value can be deterministically recalculated from current inputs, persisting the result may create unnecessary staleness. Totals, percentages, rankings, eligibility flags, generated summaries, and other derived outputs should often be recomputed when used.",{},{"id":411,"data":1007,"type":226,"tunes":1009},{"text":1008},"The key trade-off is cost. If recomputation is expensive, the system may cache the result together with the exact input version, timestamp, derivation method, and invalidation conditions. If recomputation is cheap, freshness usually wins.",{},{"id":416,"data":1011,"type":42,"tunes":1013},{"text":1012,"level":218},"4. Forget, expire, or supersede: deletion is a capability",{},{"id":421,"data":1015,"type":226,"tunes":1017},{"text":1016},"Forgetting is not necessarily a defect. It is a control mechanism. Transient tool outputs, one-off search results, failed hypotheses, temporary environment state, intermediate reasoning artifacts, obsolete user preferences, expired credentials, and superseded decisions can all become liabilities if they remain active indefinitely.",{},{"id":426,"data":1019,"type":226,"tunes":1021},{"text":1020},"Recent memory research increasingly recognizes that unbounded accumulation can degrade performance. Microsoft's 2026 human-inspired memory architecture explicitly includes interference-based forgetting and consolidation, while PlugMem reports that raw histories can overwhelm agents with low-value context. The engineering lesson does not require copying biological memory: retention should be selective.",{},{"id":431,"data":1023,"type":226,"tunes":1025},{"text":1024},"In many systems, supersession is safer than immediate deletion. The old decision remains auditable, but retrieval defaults to the new decision. This matters for projects, policies, compliance, and any workflow where the history of change is itself evidence.",{},{"id":436,"data":1027,"type":42,"tunes":1029},{"text":1028,"level":219},"The decision method",{},{"id":441,"data":1031,"type":470,"tunes":1058},{"steps":1032,"title":1057,"orientation":469},[1033,1036,1039,1042,1045,1048,1051,1054],{"label":1034,"description":1035},"1. Classify the information","Is it authoritative state, user preference, external evidence, derived output, procedure, observation, or model inference?",{"label":1037,"description":1038},"2. Identify the source of truth","Determine whether another system or source remains more authoritative than the memory itself.",{"label":1040,"description":1041},"3. Estimate volatility","Ask how likely the item is to change before the next meaningful reuse.",{"label":1043,"description":1044},"4. Estimate reuse and reconstruction cost","Compare future value with the cost and reliability of fetching or recreating the information.",{"label":1046,"description":1047},"5. Check sensitivity and scope","Define who may access the information, where it may persist, and whether persistence is justified.",{"label":1049,"description":1050},"6. Define invalidation","Specify expiry, supersession, conflict resolution, or a condition that forces a fresh authoritative read.",{"label":1052,"description":1053},"7. Choose the action","Remember, re-read\u002Fretrieve, recompute, or forget\u002Fexpire\u002Fsupersede.",{"label":1055,"description":1056},"8. Preserve provenance","Store enough metadata to distinguish source fact, user statement, observation, derivation, and model inference.","Decide the lifecycle of an information item",{},{"id":473,"data":1060,"type":42,"tunes":1062},{"text":1061,"level":219},"Examples: the same agent should use different lifecycle actions",{},{"id":478,"data":1064,"type":297,"tunes":1105},{"content":1065,"stretched":43,"withHeadings":14},[1066,1070,1073,1077,1081,1085,1089,1093,1097,1101],[1067,1068,1069],"Information","Recommended action","Why",[1071,902,1072],"“The user prefers concise technical answers.”","Stable preference with high reuse value",[1074,1075,1076],"“The deployment is currently paused.”","Re-read","Current operational state can change",[1078,1079,1080],"“The total projected cost is €48,620.”","Recompute from current inputs","Derived value should follow source changes",[1082,1083,1084],"A 20,000-token raw tool response from yesterday","Forget or archive externally","Low direct reuse; high context cost",[1086,1087,1088],"A confirmed workaround for a recurring build failure","Remember as reusable procedure","High future reuse and expensive rediscovery",[1090,1091,1092],"A model guess about why a server failed","Do not promote to durable fact","Inference is not verified evidence",[1094,1095,1096],"An old project decision later replaced by a new one","Supersede, retain audit history","The latest decision should win without erasing provenance",[1098,1099,1100],"A current product price","Retrieve again","High volatility and external authority",[1102,1103,1104],"A legal or policy interpretation","Remember the prior analysis only with source\u002Fversion metadata; re-check authority before action","Applicability can change with time and jurisdiction",{},{"id":522,"data":1107,"type":42,"tunes":1109},{"text":1108,"level":219},"Memory should store conditions, not only conclusions",{},{"id":527,"data":1111,"type":226,"tunes":1113},{"text":1112},"A durable memory becomes dangerous when it stores only the conclusion and loses the conditions under which the conclusion was valid. “Use database-per-tenant” is weaker than “Use database-per-tenant when regulatory isolation and tenant-specific lifecycle requirements outweigh operational overhead.” The second form preserves the decision boundary.",{},{"id":532,"data":1115,"type":226,"tunes":1117},{"text":1116},"This matters even more for agent-learned procedures. A successful workflow should capture not only the steps but also the preconditions, environment, tool version, observable success criteria, and known failure modes. Otherwise a memory retrieved in the wrong environment can confidently reproduce an obsolete solution.",{},{"id":537,"data":1119,"type":42,"tunes":1121},{"text":1120,"level":219},"A memory write should be more expensive than a memory read",{},{"id":542,"data":1123,"type":226,"tunes":1125},{"text":1124},"Reading a weak memory can damage one answer. Writing a weak memory can damage many future answers. The asymmetry suggests a stricter write path than read path: classify the candidate, check provenance, detect contradictions, apply sensitivity rules, define scope, and decide whether human confirmation or external validation is required.",{},{"id":547,"data":1127,"type":226,"tunes":1129},{"text":1128},"This is especially important when an agent writes memories from its own generated output. A generated summary can contain compression errors. A tool failure can be misinterpreted. A plausible hypothesis can be stored as a fact. If those outputs become future context without evidence status, the agent can create a self-reinforcing error loop.",{},{"id":552,"data":1131,"type":42,"tunes":1133},{"text":1132,"level":219},"Memory quality has at least five dimensions",{},{"id":557,"data":1135,"type":297,"tunes":1154},{"content":1136,"stretched":43,"withHeadings":14},[1137,1139,1142,1145,1148,1151],[1138,956],"Dimension",[1140,1141],"Retention quality","Did the system preserve the information that should survive?",[1143,1144],"Retrieval quality","Can the system recover the right memory when it matters?",[1146,1147],"Freshness quality","Does the system know when stored information is no longer current?",[1149,1150],"Provenance quality","Can the system distinguish source, user statement, observation, derivation, and inference?",[1152,1153],"Retirement quality","Can the system expire, supersede, restrict, or remove information when it should no longer influence decisions?",{},{"id":579,"data":1156,"type":226,"tunes":1158},{"text":1157},"Benchmarks are starting to separate these concerns. Microsoft's MemGym explicitly evaluates memory in long-horizon agentic settings and reports memory-isolated scores intended to reduce confounding from reasoning, retrieval, and tool-use ability. That direction is important because a final task score alone cannot tell you whether memory itself helped, harmed, or was irrelevant.",{},{"id":584,"data":1160,"type":42,"tunes":1162},{"text":1161,"level":219},"What not to put into durable memory by default",{},{"id":589,"data":1164,"type":604,"tunes":1177},{"meta":1165,"items":1166,"style":603},{},[1167,1168,1169,1170,1171,1172,1173,1174,1175,1176],"Raw chain-of-thought or hidden reasoning artifacts.","Temporary authentication tokens, secrets, or credentials.","Model-generated hypotheses that have not been verified.","Volatile state that has a live authoritative system.","Cheaply recomputable derived values without their source inputs.","Large tool outputs merely because storage is available.","Duplicate copies of information already governed by a better source of truth.","Sensitive personal data without a clear persistence purpose, access scope, and lifecycle.","Superseded conclusions without explicit version or retirement semantics.","Error messages or failure states that are only useful for the current run and have no reusable diagnostic value.",{},{"id":607,"data":1179,"type":42,"tunes":1181},{"text":1180,"level":219},"Memory is task-specific — there is no universal optimal store",{},{"id":612,"data":1183,"type":226,"tunes":1185},{"text":1184},"A coding agent benefits from reusable procedures, repository conventions, successful repair patterns, and project decisions. A personal assistant may need preferences, commitments, and relationship context. A commerce agent needs current product and transaction state far more than historical copies of price or inventory. A research agent benefits from source provenance, unresolved hypotheses, and explicit evidence status.",{},{"id":617,"data":1187,"type":226,"tunes":1189},{"text":1188},"Microsoft Research's M-star work makes this point directly: memory systems optimized for one purpose may transfer poorly to another, and task-specific memory mechanisms can outperform a fixed general-purpose design. The memory schema should therefore follow the decisions the agent must make, not a universal template imposed on every agent.",{},{"id":622,"data":1191,"type":42,"tunes":1193},{"text":1192,"level":219},"What would change this answer?",{},{"id":627,"data":1195,"type":226,"tunes":1197},{"text":1196},"The balance changes when retrieval is slow or expensive, authoritative systems are intermittently unavailable, recomputation is costly, audit rules require historical snapshots, or the agent must operate offline. In those cases, more information may need to be cached or persisted — but with version, provenance, timestamp, and invalidation metadata.",{},{"id":632,"data":1199,"type":226,"tunes":1201},{"text":1200},"The balance also changes for agents whose primary value is personalization. A stable preference may be worth remembering even if it could technically be asked again. Conversely, in high-risk domains, the threshold for converting an observation or interpretation into durable memory should be much higher.",{},{"id":637,"data":1203,"type":226,"tunes":1205},{"text":1204},"Future managed-memory platforms may automate consolidation, retrieval, forgetting, and context construction. That can reduce implementation work, but it does not remove the governance question: which information is allowed to influence future decisions, under what conditions, and when must the system return to the current source of truth?",{},{"id":642,"data":1207,"type":42,"tunes":1209},{"text":1208,"level":219},"Limitations",{},{"id":647,"data":1211,"type":226,"tunes":1213},{"text":1212},"There is no single definition of “agent memory” across current frameworks and research. Some systems use the term for conversation history, others for external persistent stores, structured knowledge, learned procedures, checkpoints, or model adaptation. The decision model in this article focuses on operational lifecycle semantics rather than enforcing one vocabulary.",{},{"id":652,"data":1215,"type":226,"tunes":1217},{"text":1216},"The four lifecycle actions can also overlap. A system may remember a stable summary, retain a pointer to the source, re-read volatile fields, and recompute a derived result in one workflow. The purpose of the model is not to force one storage primitive per fact, but to make the reason for persistence explicit.",{},{"id":657,"data":1219,"type":42,"tunes":1221},{"text":1220,"level":219},"Conclusion",{},{"id":662,"data":1223,"type":226,"tunes":1225},{"text":1224},"A useful agent does not win by remembering the most. It wins by preserving the right information, returning to authoritative sources when reality can change, recalculating what is safer to derive again, and retiring information that should no longer influence future decisions.",{},{"id":667,"data":1227,"type":226,"tunes":1229},{"text":1228},"The practical question for every candidate memory is therefore not “Can we store this?” but: Will future decisions be more reliable if this survives? If the answer depends on freshness, authority, cost, sensitivity, or revision, encode those conditions into the memory lifecycle instead of trusting retrieval alone.",{},{"id":672,"data":1231,"type":678,"tunes":1236},{"url":1232,"title":1233,"excerpt":1234,"ctaLabel":1235},"https:\u002F\u002Fstajic.de\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework","From Research Protocol to a General AI Reasoning Framework","A domain-independent reasoning method for separating evidence from assumptions, testing competing hypotheses, and using explicit validation rules.","Read the reasoning framework",{},{"id":681,"data":1238,"type":42,"tunes":1240},{"text":1239,"level":219},"FAQ",{},{"id":686,"data":1242,"type":686,"tunes":1260},{"items":1243,"title":1259},[1244,1247,1250,1253,1256],{"id":690,"answer":1245,"question":1246},"Prefer information that is durable, reusable, provenance-preserving, and expensive or unreliable to reconstruct, such as stable user preferences, accepted project decisions, reusable procedures, and verified long-term constraints.","What information should an AI agent remember long term?",{"id":694,"answer":1248,"question":1249},"Volatile information with an authoritative external source should normally be retrieved again before consequential use. Examples include permissions, inventory, current prices, account state, policy versions, service status, and current documentation.","What should an AI agent retrieve again instead of remembering?",{"id":698,"answer":1251,"question":1252},"Recompute derived values when the calculation is cheap and stale results would be costly. Persisting a derived value makes more sense when recomputation is expensive and the cache includes the source version and invalidation conditions.","When should an AI agent recompute information?",{"id":702,"answer":1254,"question":1255},"Yes. Forgetting, expiry, and supersession are useful controls for transient, obsolete, sensitive, low-value, or misleading information. Unbounded retention can create noise and allow stale or incorrect information to keep influencing future decisions.","Should AI agents forget information?",{"id":706,"answer":1257,"question":1258},"Not by itself. Raw history can preserve evidence, but long-running agents usually need curation, structure, summaries, reusable facts or procedures, retrieval, and lifecycle rules so that low-value history does not dominate future context.","Is storing the entire conversation history a good memory strategy?","AI agent memory lifecycle",{},{"id":712,"data":1262,"type":42,"tunes":1264},{"text":1263,"level":219},"Glossary",{},{"id":717,"data":1266,"type":717,"tunes":1285},{"title":1267,"entries":1268},"Key memory lifecycle terms",[1269,1272,1275,1278,1281,1283],{"term":1270,"anchor":723,"definition":1271},"Memory admission","The decision process that determines whether information is allowed to become persistent agent memory.",{"term":1273,"anchor":727,"definition":1274},"Supersession","Marking an older memory or decision as replaced by newer information while preserving the historical record where needed.",{"term":1276,"anchor":731,"definition":1277},"Invalidation","A rule or event that makes a stored or cached value unsafe to reuse without refresh, recomputation, or review.",{"term":1279,"anchor":735,"definition":1280},"Provenance","Metadata describing where information came from, when it was observed, who or what asserted it, and how it was transformed.",{"term":934,"anchor":314,"definition":1282},"The likelihood that information will change between the time it is stored and the time it is reused.",{"term":943,"anchor":740,"definition":1284},"The time, money, computation, tool use, or uncertainty required to recover or regenerate information instead of storing it.",{},{"id":744,"data":1287,"type":42,"tunes":1289},{"text":1288,"level":219},"Primary sources and further reading",{},{"id":749,"data":1291,"type":756,"tunes":1296},{"link":751,"meta":1292},{"image":1293,"title":1294,"description":1295},{"url":317},"OpenAI — Context Engineering: Short-Term Memory Management with Sessions","Guidance on trimming, summarization, long-running context, and risks such as stale details and context poisoning.",{},{"id":759,"data":1298,"type":756,"tunes":1303},{"link":761,"meta":1299},{"image":1300,"title":1301,"description":1302},{"url":317},"Anthropic — Effective Context Engineering for AI Agents","Engineering guidance on curation, compaction, structured note-taking, and maintaining useful agent context over long horizons.",{},{"id":768,"data":1305,"type":756,"tunes":1310},{"link":770,"meta":1306},{"image":1307,"title":1308,"description":1309},{"url":317},"Anthropic — Effective Harnesses for Long-Running Agents","Practical work on preserving progress and artifacts across context windows in long-running agent tasks.",{},{"id":777,"data":1312,"type":756,"tunes":1316},{"link":779,"meta":1313},{"image":1314,"title":782,"description":1315},{"url":317},"Research on transforming raw agent interactions into structured reusable facts and skills rather than accumulating undifferentiated history.",{},{"id":786,"data":1318,"type":756,"tunes":1323},{"link":788,"meta":1319},{"image":1320,"title":1321,"description":1322},{"url":317},"Microsoft Research — M★: Every Task Deserves Its Own Memory Harness","Research showing that task-specific memory mechanisms can outperform fixed general-purpose memory designs.",{},{"id":795,"data":1325,"type":756,"tunes":1329},{"link":797,"meta":1326},{"image":1327,"title":800,"description":1328},{"url":317},"A benchmark for isolating and evaluating memory performance in long-horizon agent environments.",{},{"id":804,"data":1331,"type":756,"tunes":1336},{"link":806,"meta":1332},{"image":1333,"title":1334,"description":1335},{"url":317},"Microsoft Research — Human-Inspired Memory Architecture for LLM Agents","Research exploring consolidation, interference-based forgetting, reconsolidation, and retrieval in persistent agent memory.",{},"2.31.6","Long-running agents should not remember everything. This article provides a practical lifecycle model for deciding what belongs in durable memory, what should be retrieved again, what is safer to recompute, and what should expire or be superseded.",{"lang":7,"title":208,"content":210,"contentJson":1340,"excerpt":813},{"time":212,"blocks":1341,"version":812},[1342,1345,1348,1351,1354,1357,1360,1363,1366,1369,1378,1381,1384,1404,1407,1410,1413,1416,1419,1422,1425,1428,1431,1434,1437,1440,1443,1446,1449,1452,1455,1467,1470,1484,1487,1490,1493,1496,1499,1502,1505,1515,1518,1521,1526,1529,1532,1535,1538,1541,1544,1547,1550,1553,1556,1559,1562,1565,1568,1571,1580,1583,1593,1596,1601,1606,1611,1616,1621,1626],{"id":215,"data":1343,"type":220,"tunes":1344},{"title":217,"maxLevel":218,"minLevel":219},{},{"id":223,"data":1346,"type":226,"tunes":1347},{"text":225},{},{"id":229,"data":1349,"type":234,"tunes":1350},{"body":231,"title":232,"variant":233},{},{"id":237,"data":1352,"type":234,"tunes":1353},{"body":239,"title":240,"variant":241},{},{"id":244,"data":1355,"type":42,"tunes":1356},{"text":246,"level":219},{},{"id":249,"data":1358,"type":226,"tunes":1359},{"text":251},{},{"id":254,"data":1361,"type":226,"tunes":1362},{"text":256},{},{"id":259,"data":1364,"type":226,"tunes":1365},{"text":261},{},{"id":264,"data":1367,"type":42,"tunes":1368},{"text":266,"level":219},{},{"id":269,"data":1370,"type":297,"tunes":1377},{"content":1371,"stretched":43,"withHeadings":14},[1372,1373,1374,1375,1376],[273,274,275,276],[278,279,280,281],[283,284,285,286],[288,289,290,291],[293,294,295,296],{},{"id":300,"data":1379,"type":42,"tunes":1380},{"text":302,"level":219},{},{"id":305,"data":1382,"type":226,"tunes":1383},{"text":307},{},{"id":310,"data":1385,"type":349,"tunes":1403},{"rows":1386,"title":338,"layout":297,"columns":1399},[1387,1389,1391,1393,1395,1397],{"id":314,"label":315,"values":1388},[317,317,317],{"id":319,"label":320,"values":1390},[317,317,317],{"id":323,"label":324,"values":1392},[317,317,317],{"id":327,"label":328,"values":1394},[317,317,317],{"id":331,"label":332,"values":1396},[317,317,317],{"id":335,"label":336,"values":1398},[317,317,317],[1400,1401,1402],{"id":341,"label":342},{"id":344,"label":345},{"id":347,"label":348},{},{"id":352,"data":1405,"type":234,"tunes":1406},{"body":354,"title":355,"variant":356},{},{"id":359,"data":1408,"type":42,"tunes":1409},{"text":361,"level":218},{},{"id":364,"data":1411,"type":226,"tunes":1412},{"text":366},{},{"id":369,"data":1414,"type":226,"tunes":1415},{"text":371},{},{"id":374,"data":1417,"type":226,"tunes":1418},{"text":376},{},{"id":379,"data":1420,"type":42,"tunes":1421},{"text":381,"level":218},{},{"id":384,"data":1423,"type":226,"tunes":1424},{"text":386},{},{"id":389,"data":1426,"type":226,"tunes":1427},{"text":391},{},{"id":394,"data":1429,"type":234,"tunes":1430},{"body":396,"title":397,"variant":398},{},{"id":401,"data":1432,"type":42,"tunes":1433},{"text":403,"level":218},{},{"id":406,"data":1435,"type":226,"tunes":1436},{"text":408},{},{"id":411,"data":1438,"type":226,"tunes":1439},{"text":413},{},{"id":416,"data":1441,"type":42,"tunes":1442},{"text":418,"level":218},{},{"id":421,"data":1444,"type":226,"tunes":1445},{"text":423},{},{"id":426,"data":1447,"type":226,"tunes":1448},{"text":428},{},{"id":431,"data":1450,"type":226,"tunes":1451},{"text":433},{},{"id":436,"data":1453,"type":42,"tunes":1454},{"text":438,"level":219},{},{"id":441,"data":1456,"type":470,"tunes":1466},{"steps":1457,"title":468,"orientation":469},[1458,1459,1460,1461,1462,1463,1464,1465],{"label":445,"description":446},{"label":448,"description":449},{"label":451,"description":452},{"label":454,"description":455},{"label":457,"description":458},{"label":460,"description":461},{"label":463,"description":464},{"label":466,"description":467},{},{"id":473,"data":1468,"type":42,"tunes":1469},{"text":475,"level":219},{},{"id":478,"data":1471,"type":297,"tunes":1483},{"content":1472,"stretched":43,"withHeadings":14},[1473,1474,1475,1476,1477,1478,1479,1480,1481,1482],[482,483,484],[486,278,487],[489,490,491],[493,494,495],[497,498,499],[501,502,503],[505,506,507],[509,510,511],[513,514,515],[517,518,519],{},{"id":522,"data":1485,"type":42,"tunes":1486},{"text":524,"level":219},{},{"id":527,"data":1488,"type":226,"tunes":1489},{"text":529},{},{"id":532,"data":1491,"type":226,"tunes":1492},{"text":534},{},{"id":537,"data":1494,"type":42,"tunes":1495},{"text":539,"level":219},{},{"id":542,"data":1497,"type":226,"tunes":1498},{"text":544},{},{"id":547,"data":1500,"type":226,"tunes":1501},{"text":549},{},{"id":552,"data":1503,"type":42,"tunes":1504},{"text":554,"level":219},{},{"id":557,"data":1506,"type":297,"tunes":1514},{"content":1507,"stretched":43,"withHeadings":14},[1508,1509,1510,1511,1512,1513],[561,345],[563,564],[566,567],[569,570],[572,573],[575,576],{},{"id":579,"data":1516,"type":226,"tunes":1517},{"text":581},{},{"id":584,"data":1519,"type":42,"tunes":1520},{"text":586,"level":219},{},{"id":589,"data":1522,"type":604,"tunes":1525},{"meta":1523,"items":1524,"style":603},{},[593,594,595,596,597,598,599,600,601,602],{},{"id":607,"data":1527,"type":42,"tunes":1528},{"text":609,"level":219},{},{"id":612,"data":1530,"type":226,"tunes":1531},{"text":614},{},{"id":617,"data":1533,"type":226,"tunes":1534},{"text":619},{},{"id":622,"data":1536,"type":42,"tunes":1537},{"text":624,"level":219},{},{"id":627,"data":1539,"type":226,"tunes":1540},{"text":629},{},{"id":632,"data":1542,"type":226,"tunes":1543},{"text":634},{},{"id":637,"data":1545,"type":226,"tunes":1546},{"text":639},{},{"id":642,"data":1548,"type":42,"tunes":1549},{"text":644,"level":219},{},{"id":647,"data":1551,"type":226,"tunes":1552},{"text":649},{},{"id":652,"data":1554,"type":226,"tunes":1555},{"text":654},{},{"id":657,"data":1557,"type":42,"tunes":1558},{"text":659,"level":219},{},{"id":662,"data":1560,"type":226,"tunes":1561},{"text":664},{},{"id":667,"data":1563,"type":226,"tunes":1564},{"text":669},{},{"id":672,"data":1566,"type":678,"tunes":1567},{"url":674,"title":675,"excerpt":676,"ctaLabel":677},{},{"id":681,"data":1569,"type":42,"tunes":1570},{"text":683,"level":219},{},{"id":686,"data":1572,"type":686,"tunes":1579},{"items":1573,"title":709},[1574,1575,1576,1577,1578],{"id":690,"answer":691,"question":692},{"id":694,"answer":695,"question":696},{"id":698,"answer":699,"question":700},{"id":702,"answer":703,"question":704},{"id":706,"answer":707,"question":708},{},{"id":712,"data":1581,"type":42,"tunes":1582},{"text":714,"level":219},{},{"id":717,"data":1584,"type":717,"tunes":1592},{"title":719,"entries":1585},[1586,1587,1588,1589,1590,1591],{"term":722,"anchor":723,"definition":724},{"term":726,"anchor":727,"definition":728},{"term":730,"anchor":731,"definition":732},{"term":734,"anchor":735,"definition":736},{"term":315,"anchor":314,"definition":738},{"term":328,"anchor":740,"definition":741},{},{"id":744,"data":1594,"type":42,"tunes":1595},{"text":746,"level":219},{},{"id":749,"data":1597,"type":756,"tunes":1600},{"link":751,"meta":1598},{"image":1599,"title":754,"description":755},{"url":317},{},{"id":759,"data":1602,"type":756,"tunes":1605},{"link":761,"meta":1603},{"image":1604,"title":764,"description":765},{"url":317},{},{"id":768,"data":1607,"type":756,"tunes":1610},{"link":770,"meta":1608},{"image":1609,"title":773,"description":774},{"url":317},{},{"id":777,"data":1612,"type":756,"tunes":1615},{"link":779,"meta":1613},{"image":1614,"title":782,"description":783},{"url":317},{},{"id":786,"data":1617,"type":756,"tunes":1620},{"link":788,"meta":1618},{"image":1619,"title":791,"description":792},{"url":317},{},{"id":795,"data":1622,"type":756,"tunes":1625},{"link":797,"meta":1623},{"image":1624,"title":800,"description":801},{"url":317},{},{"id":804,"data":1627,"type":756,"tunes":1630},{"link":806,"meta":1628},{"image":1629,"title":809,"description":810},{"url":317},{},"Post erfolgreich abgerufen",{"items":1633,"source":1676,"manualIds":1677,"manualMatchedIds":1678},[1634,1641,1648,1655,1662,1669],{"id":1635,"slug":1636,"title":1637,"excerpt":1638,"featuredImage":1639,"publishedAt":1640},"466","the-gpu-is-not-the-product-future-proof-private-ai-architecture","GPU 不是产品：面向未来的私有 AI 架构","私有 AI 基础设施不应围绕单一 GPU 或单一模型来设计。更具韧性的做法是将快速推理 GPU、内存充裕的 AI 系统、物理 AI 节点以及可选的前沿云模型，统一置于一个具备能力感知的路由层之后。","\u002Fuploads\u002F2026\u002F09\u002Fthe-gpu-is-not-the-product-future-proof-private-ai-architecture-1790140878812-8hsl39.webp","2026-09-23T01:19:00.000Z",{"id":1642,"slug":1643,"title":1644,"excerpt":1645,"featuredImage":1646,"publishedAt":1647},"476","mcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained","MCP vs A2A vs UCP vs AP2 vs A2UI：智能体协议栈详解","MCP、A2A、UCP、AP2 和 A2UI 常被描述为相互竞争的智能体标准。它们大多解决的是不同的互操作性问题。本指南将每个协议映射到其实际标准化的边界，并展示它们如何在同一个生产系统中协同工作。","\u002Fuploads\u002F2026\u002F09\u002Fmcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained-1790352625869-2ezle0.webp","2026-09-25T12:09:00.000Z",{"id":1649,"slug":1650,"title":1651,"excerpt":1652,"featuredImage":1653,"publishedAt":1654},"468","ai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context","AI代理记忆不是RAG：如何区分记忆、检索、状态和上下文","代理记忆、RAG、状态和上下文经常被当作可以互换的概念来使用。它们并不是。这个实用的架构模型将这四个层次区分开来，展示了每一层各自应处的位置，并解释了当系统将它们合并为一层时会出现什么问题。","\u002Fuploads\u002F2026\u002F09\u002Fai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context-1790350560308-np0xy6.webp","2026-09-25T11:34:00.000Z",{"id":1656,"slug":1657,"title":1658,"excerpt":1659,"featuredImage":1660,"publishedAt":1661},"471","how-to-know-whether-an-ai-agent-actually-used-the-right-evidence","如何判断一个AI智能体是否真正使用了正确的证据","AI代理可以引用来源，却仍然使用错误的证据。本文介绍一种实用方法，用于核查主张支持、来源权威性、适用性、出处，以及证据是否实际影响了答案。","\u002Fuploads\u002F2026\u002F09\u002Fhow-to-know-whether-an-ai-agent-actually-used-the-right-evidence-1790351317188-o5z9ve.webp","2026-09-25T11:47:00.000Z",{"id":1663,"slug":1664,"title":1665,"excerpt":1666,"featuredImage":1667,"publishedAt":1668},"467","the-answer-validity-boundary-the-missing-layer-between-relevance-and-reliable-ai-answers","答案有效性边界：相关性到可靠AI答案之间缺失的层级","一个来源可能相关、权威，但对于所提出的问题仍然是错误的。缺失的层次是适用性：答案成立的条件，以及迫使其被重新考虑的变化。本文介绍了“答案有效性边界”这一面向人类、AI搜索和RAG系统的来源设计模式。","\u002Fuploads\u002F2026\u002F09\u002Fthe-answer-validity-boundary-the-missing-layer-between-relevance-and-reliable-ai-answers-1790272901306-1g5jly.webp","2026-09-24T11:59:00.000Z",{"id":1670,"slug":1671,"title":1672,"excerpt":1673,"featuredImage":1674,"publishedAt":1675},"478","what-is-rag-the-simplest-explanation-of-how-it-works","什么是RAG？对其工作原理的最简单解释","RAG听起来很复杂，但想法很简单：在AI回答之前，它先从知识源查找有用的信息，并将该信息提供给语言模型。本指南使用一个简单的思维模型来解释RAG、LLM、状态、记忆和工具。","\u002Fuploads\u002F2026\u002F09\u002Fwhat-is-rag-the-simplest-explanation-of-how-it-works-1790377492124-khjagt.webp","2026-09-25T19:03:00.000Z","fallback",[],[]]