Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明Anthropic study: Claude AI agents can get smarter by 'dreaming'


AI free token/AI免费积分
感谢IT之家网友 惊讶仙人 的线索投递!

IT之家 11 月 7 日消息,Anthropic 昨日(11 月 6 日)发布研究文章《The components of agent memory》,平台拆解 Claude 智能体如何跨会话记忆、筛选、纠错和整理知识。

该研究指出 AI 智能体可借鉴人脑“睡眠整理记忆”的机制,在空闲时间反思历史记录、纠错并构建知识结构,从而显著提高长周期任务表现。

该研究将智能体记忆分为会话日志、记忆笔记、知识表征(如 Wiki)和注入上下文四层,并引入“做梦”(dreaming)机制,让模型在两次交互之间的空闲时间回顾日志、修正错误、提炼规则和更新知识库。

Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明

研究团队认为,AI 可像人脑在睡眠中巩固记忆一样,将即时经验逐步转化为可复用的长期知识。Claude 会在会话中标记关键事件或写入记忆笔记,但并非所有信息都值得保存,它需要判断哪些内容在将来真正有用。

Anthropic 指出,记忆本质上是一种对将来的预测,例如记录一条已验证的规则,意味着模型预期将来会再次依赖它。

在持续学习基准 CL-Bench 的数据库问答任务中,具备记忆能力的 Claude 模型表现显著优于无记忆基线。Claude Opus 5.5 和 Sonnet 5.5 在使用记忆后,平均得分分别达到 0.838 和 0.833,而二者无记忆基线分别为 0.655 和 0.644。更强的模型可以自主形成“失败 — 反思 — 验证 — 总结规则”的学习闭环。

Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明

在《宝可梦红》游戏测试中,Anthropic 还发现“少而精”的记忆更有效。Claude Fable 5 仅维护 10 个记忆文件,却取得 4 枚徽章;而 Sonnet 3.7 写下 413 个文件,仅取得 1 枚徽章。研究认为,模型应优先保存策略、计划和经测试的规则,而非不断变化的显示屏状态。

Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明

“做梦”机制还展现出纠错能力。当 Claude 错误记录了宝可梦中心的位置并因而迷路时,平台借助比对原始会话记录修正记忆,让其最终成功抵达下一座城市;未修正记忆的运行则反复折返,未能完成目标。

Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明

Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明

在约 35 万份企业文档的测试中,Claude 借助多轮“做梦”逐步构建 Wiki。随着“做梦”规模从 2 次会话扩展到 128 次,25 个难题的达标率从 13% 提高至 48%;将 Wiki 与搜索工具结合后,综合得分从 54% 升至 61%。

Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明

Anthropic 研究称 Claude AI 智能体可借「做梦」变得更聪明

Anthropic 表示,智能体记忆仍面临何时写入、如何保持更新、如何避免错误累积等挑战。将来 AI 或将更善于在完成任务后“回顾经验”,把零散交互转化为可迁移、可验证的长期知识。

IT之家附上参考地址

相关阅读:

本文来源:https://www.ithome.com/1/023/613.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Thanks to IT Home user 惊讶仙人 for the tip!

IT Home, November 7 — Anthropic yesterday (November 6) published a research article titled "The components of agent memory," breaking down how Claude agents remember, filter, correct errors, and organize knowledge across sessions.

The study notes that AI agents can draw on a human-brain-like mechanism of “sleep-based memory consolidation,” reflecting on past records, correcting mistakes, and building knowledge structures during idle time, thereby significantly improving performance on long-horizon tasks.

The study divides agent memory into four layers: session logs, memory notes, knowledge representations (such as a Wiki), and injected context, and introduces a “dreaming” mechanism that lets the model review logs, fix errors, distill rules, and update the knowledge base during idle time between interactions.

Anthropic study: Claude AI agents can get smarter by 'dreaming'

The research team believes that, like the human brain consolidating memories during sleep, AI can gradually turn immediate experience into reusable long-term knowledge. Claude will mark key events during a session or write memory notes, but not all information is worth keeping—it needs to judge what will truly be useful in the future.

Anthropic notes that memory is essentially a form of prediction about the future; for example, recording a verified rule means the model expects to rely on it again later.

In the database question-answering task of the continual learning benchmark CL-Bench, the memory-enabled Claude model significantly outperformed the no-memory baseline. With memory, Claude Opus 5.5 and Sonnet 5.5 achieved average scores of 0.838 and 0.833 respectively, while their no-memory baselines were 0.655 and 0.644. Stronger models can autonomously form a learning loop of “failure—reflection—verification—rule summarization.”

Anthropic study: Claude AI agents can get smarter by 'dreaming'

In Pokémon Red gameplay tests, Anthropic also found that “less but better” memory is more effective. Claude Fable 5 maintained only 10 memory files but obtained 4 badges, whereas Sonnet 3.7 wrote 413 files and obtained only 1 badge. The study suggests that models should prioritize saving strategies, plans, and tested rules rather than constantly changing on-screen states.

Anthropic study: Claude AI agents can get smarter by 'dreaming'

The “dreaming” mechanism also demonstrated error-correction ability. When Claude incorrectly recorded the location of a Pokémon Center and got lost as a result, the platform corrected its memory by comparing against the original session record, enabling it to finally reach the next city; runs without memory correction repeatedly backtracked and failed to complete the objective.

Anthropic study: Claude AI agents can get smarter by 'dreaming'

Anthropic study: Claude AI agents can get smarter by 'dreaming'

In tests across about 350,000 enterprise documents, Claude used multiple rounds of “dreaming” to gradually build a Wiki. As the scale of “dreaming” expanded from 2 sessions to 128 sessions, the success rate on 25 challenging tasks rose from 13% to 48%; after combining the Wiki with search tools, the overall score increased from 54% to 61%.

Anthropic study: Claude AI agents can get smarter by 'dreaming'

Anthropic study: Claude AI agents can get smarter by 'dreaming'

Anthropic says agent memory still faces challenges such as when to write, how to stay up to date, and how to avoid error accumulation. In the future, AI may become better at “reviewing experience” after completing tasks, turning scattered interactions into transferable, verifiable long-term knowledge.

IT Home provides the reference link below.

Related reading:

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry