Anthropic shares the Dreaming mechanism: letting Claude organize memories outside of tasks

Anthropic's latest research: more Agent memory is not always better; how Anthropic turns records into usable experience Shared by the Anthropic team @RLanceMartin, multiple experiments show that what an Agent needs is not just to save history, but also to look back at evidence, correct notes, organize experience, and find them when needed. Background organization can improve some tasks, but it is not a replacement for retrieval, and it may not always offset maintenance costs. https://t.co/H1FFfA8rkh # Memory is not equal to chat logs Memory here refers to context saved across sessions and readable later; it does not involve updating model weights, nor is it merely compressing the current conversation. The article discusses several interlocking steps: retaining raw records for traceability; writing down goals, plans, and verified experience while executing tasks; reviewing records in the background to correct errors, remove duplicates, add links, and build a knowledge base when necessary; and at read time, preloading only a small amount of important content while retrieving the rest on demand. The author calls background organization dreaming. It does not make the model "sleep"; rather, it separates organizing experience from the task at hand, giving the model dedicated time to process existing information. # What good memory should preserve In the CL-Bench database task, each question starts a new session, and only the memory directory is retained. After answering, the model receives the correct answer and then organizes experience. Stronger models begin to save verified rules and their basis, not just records of wrong answers; restricting it to "only record failures" instead weakened the gains. This shows that valuable memory is not just "what was done wrong last time," but also what rules hold, why they are credible, and when they will apply in the future. State that can be reread at any time does not need to be repeatedly copied into notes. # Experiments show: organization and search are complementary In an internal corpus experiment with about 350,000 files, 420 million tokens, and 25 evaluation questions, search alone achieved a rubric coverage rate of 53%; a knowledge base combined with search rose from 54% to 61% as the number of organization rounds increased, using about 20,000 fewer tokens per question. The 61% here is not "answered 61% of questions correctly," and the token savings also do not deduct the background cost of building the knowledge base. Gains were more obvious in cross-document integration tasks, while some tasks focused mainly on single-file retrieval showed no improvement, and the knowledge base was even larger than the original text. In another contract case, multiple agents co-writing one table caused overwrites. After changing to a separate page for each contract, with one agent responsible for each, and quoting clauses verbatim, the final report's key fact coverage rose from 41/57 to 57/57. The improvement came from organization and writing rules, not just from more thinking time.

View original

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。