Anthropic Taught Its Agents to Dream Between Conversations
让智能体「睡一觉」：Anthropic 谈记忆怎么写
In the idle time after a session, Claude reviews its own logs, fixes errors and writes rules. In a Pokémon test, ten memory files beat 413.
Anthropic has published a study on how its agents remember things between conversations, and the central idea borrows from sleep. In the gaps between sessions, the model goes back over its own logs, corrects mistakes, extracts rules and updates what it knows — a process the researchers call “dreaming”.
The paper separates agent memory into four layers: a log of what happened in each session, short notes the model writes for itself, a structured knowledge base such as a wiki, and the slice of all this that gets injected into the current context.
The framing that makes the design legible is their claim that memory is a prediction about the future. Writing down a rule you have verified is a bet that you will need it again — which means the hard part is not storage but judgement about what deserves to be kept.
The benchmarks suggest the approach works. On a database question-answering task in a continual-learning suite, Claude Opus 5.5 scored 0.838 with memory against 0.655 without; Sonnet 5.5 went from 0.644 to 0.833. The stronger model was also the one able to close the loop on its own: fail, reflect, verify, write the rule.
A test in Pokémon Red produced the more interesting result. Claude Fable 5 kept ten memory files and earned four badges; Sonnet 3.7 wrote 413 files and earned one. Less, here, was more: the useful memory was strategy, plans and tested rules, not a running description of whatever was on screen.
Dreaming also turned out to be a correction mechanism. When one run misrecorded the location of a Pokémon Centre and kept getting lost, the process compared the memory against the original session log, fixed the entry, and the agent reached the next town. The uncorrected run kept doubling back.
On a corpus of about 350,000 enterprise documents, repeated dreaming built a wiki that improved as it went: scaling from two sessions to 128 raised the pass rate on 25 questions from 13% to 48%, and adding a search tool took the combined score from 54% to 61%. Anthropic is careful about what remains unsolved — when to write, how to keep entries current, and how to stop errors compounding.
中文翻译（全文）
Anthropic 发表了一项关于智能体如何在会话之间记住事情的研究，核心思路借自睡眠。在两次交互之间的空档里，模型回看自己的日志、修正错误、提炼规则并更新知识——研究者把这个过程称为「做梦」。
论文把智能体记忆分成四层：每次会话发生了什么日志、模型写给自己的简短笔记、诸如 Wiki 这样的结构化知识库，以及这些内容中被注入当前上下文的那一部分。
让这套设计变得可理解的关键，是他们的一个说法：记忆本质上是对未来的预测。写下一则已被验证的规则，就是在赌你之后还会用到它——这也意味着难的不是存储，而是判断什么值得留下。
基准测试显示这套做法有效。在一个持续学习套件的数据库问答任务上，带记忆的 Claude Opus 5.5 得分 0.838，不带记忆是 0.655；Sonnet 5.5 则从 0.644 升到 0.833。更强的模型也恰好是能自己跑完闭环的那个：失败、反思、验证、写下规则。
《宝可梦红》的测试给出了更有意思的结果。Claude Fable 5 只维护十个记忆文件，拿到四枚徽章；Sonnet 3.7 写了 413 个文件，只拿到一枚。在这里少即是多：有用的记忆是策略、计划和经过检验的规则，而不是对屏幕上当前画面的流水描述。
「做梦」同时被证明是一种纠错机制。有一次运行错记了宝可梦中心的位置、反复迷路，这个过程把记忆与原始会话日志比对、改正了条目，智能体随后抵达下一座城市；未做纠正的那次运行则一直在原地折返。
在约 35 万份企业文档的语料上，多轮「做梦」逐步搭起一个越做越好的 Wiki：从 2 次会话扩到 128 次，25 道问题的达标率从 13% 升到 48%；再叠加搜索工具后，综合得分从 54% 升到 61%。Anthropic 也谨慎说明了尚未解决的问题——何时写入、如何保持条目更新，以及怎样避免错误累积。
重点词汇
- consolidate/kənˈsɒlɪdeɪt/v.— to combine separate things into a stronger whole巩固；整合Sleep consolidates memory.
- context/ˈkɒntekst/n.— the information a model is given for the current task上下文A slice of memory is injected into context.
- benchmark/ˈbentʃmɑːk/n.— a standard test used for comparison基准测试The benchmarks suggest the approach works.
- verify/ˈverɪfaɪ/v.— to check that something is true验证It verifies before writing a rule.
- extract/ɪkˈstrækt/v.— to pull something out of a larger body提炼；提取The model extracts rules from logs.
- compound/kəmˈpaʊnd/v.— to make a problem worse by adding to it累积；加剧Errors can compound over time.
- log/lɒɡ/n.— a record of events日志It reviews its own logs.
- judgement/ˈdʒʌdʒmənt/n.— the ability to decide well判断力The hard part is judgement.
- session/ˈseʃn/n.— one continuous period of interaction会话Memory persists across sessions.
- loop/luːp/n.— a cycle that returns to its start闭环；循环It closes the learning loop itself.
阅读测验
共 5 道题目
What is “dreaming” in this research?
How does the paper frame memory?
In the Pokémon test, the model with only ten memory files performed better than the one with 413.
What did the benchmarks show? (Select all that apply)
Scaling dreaming from two sessions to 128 raised the pass rate from 13% to _____ per cent.
评论
0 条讨论
登录后发表评论
立即登录暂无评论
成为第一个分享想法的人吧！