Anthropic study: Agent memory requires complementary background organization and retrieval
English text is being prepared.
The Anthropic team shared research on Agent memory, noting that memory is not just about saving history, but also about re-checking evidence, correcting notes, organizing experience, and retrieving on demand; background organization is called dreaming. In an internal corpus experiment with about 350,000 files, 420 million tokens, and 25 evaluation questions, the scoring-criteria coverage of search alone was 53%; the knowledge base combined with search improved from 54% to 61% as organization rounds increased, using about 20,000 fewer tokens per question, but the saved tokens did not deduct the cost of building the background knowledge base. In another contract case, after changing to one separate page per contract, with one agent responsible and quoting clauses verbatim, key-fact coverage improved from 41/57 to 57/57.
Why read this
AI score 60
The original text provides experimental data on the complementarity of organization and retrieval, and notes that the cost of building the background knowledge base was not deducted, allowing readers to judge the actual benefit boundaries of memory solutions.
Topics
View original postMemory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. https://t.co/uUEN4bNkr9
View original
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. https://t.co/uUEN4bNkr9
Body
Anthropic 最新研究：Agent 的记忆不是越多越好，Anthropic 如何把记录整理成可用经验
来自 Anthropic 团队 @RLanceMartin 分享，通过多组实验说明：Agent 需要的不只是保存历史，还要回查证据、纠正笔记、整理经验，并在需要时找到它们。后台整理能够改善部分任务，但不是检索的替代品，也未必总能抵消维护成本。
https://t.co/H1FFfA8rkh
# 记忆不等于聊天记录
这里的记忆是跨会话保存、以后可以读取的上下文，不涉及更新模型权重，也不只是压缩当前对话。
文章讨论了几个配合使用的环节：保留原始记录供追溯；执行任务时写下目标、计划和已验证的经验；后台回看记录，纠错、去重、补链接，必要时建立知识库；读取时只提前加载少量重要内容，其余按需检索。
作者把后台整理称为 dreaming。它并不是让模型“睡觉”，而是把整理经验与眼前任务分开，给模型专门的时间处理已有信息。
# 好记忆要保存什么
在 CL-Bench 数据库任务中，每道题开启新会话，只有记忆目录保留。模型作答后会收到正确答案，再整理经验。较强模型开始保存经过验证的规则和依据，而不只是记录错题；限定“只记失败”，反而削弱了收益。
这说明有价值的记忆不只是“上次做错了什么”，还包括什么规则成立、为什么可信、以后何时适用。能随时重新读取的状态，则不必反复抄进笔记。
# 实验显示：整理与搜索互补
在约 35 万份文件、4.2 亿 token、25 个评测问题的内部语料实验中，单靠搜索的评分标准覆盖率为 53%；知识库结合搜索，随着整理轮次增加，从 54% 提升到 61%，每题少用约 2 万 token。
这里的 61% 不是“答对了 61% 的问题”，节省的 token 也没有扣除后台建库成本。跨文档整合任务收益更明显，部分以单文件检索为主的任务则没有改善，知识库甚至比原文更大。
另一个合同案例中，多代理共写一张表导致覆盖。改成每份合同单独一页、一个代理负责，并逐字引用条款后，最终报告的关键事实覆盖从 41/57 提升到 57/57。改善来自组织和写入规则，不只是增加思考时间。