Primary sourceIT之家
View sourceEvent ·Multi-source
Anthropic 发布智能体记忆研究:空闲时"做梦"整理记忆可提升表现
Anthropic 于 11 月 6 日发布《The components of agent memory》,把智能体记忆分为会话日志、记忆笔记、知识表征和注入上下文四层,并引入空闲时回顾纠错的"做梦"机制。在 CL-Bench 数据库问答任务上,Claude Opus 5.5 和 Sonnet 5.5 使用记忆后平均得分分别为 0.838 和 0.833,无记忆基线只有 0.655 和 0.644。《宝可梦红》测试中,Claude Fable 5 仅维护 10 个记忆文件拿到 4 枚徽章,Sonnet 3.7 写下 413 个文件只拿 1 枚,说明"少而精"的记忆更有效。在约 35 万份企业文档测试里,"做梦"次数从 2 次扩展到 128 次后,25 个问题的达标率从 13% 升至 48%,Wiki 结合搜索工具后综合得分从 54% 升至 61%。
Current read
Anthropic 拆了 Claude 智能体的记忆架构,让模型在空闲时"做梦"纠错总结,CL-Bench 得分从 0.65 拉到 0.83,做智能体的可以抄作业。
11 sources62° AI heatLast updated 2026/11/6 23:25:14
Evidence
11 sourcesRelated source 1Anthropic: Research
View sourceRelated source 2AI Will
View sourceRelated source 3宝玉
View sourceRelated source 4elvis
View sourceRelated source 5SuperTechFans
View sourceRelated source 6shao__meng
View sourceRelated source 7kimmonismus
View sourceRelated source 8Genk
View sourceRelated source 9rohanpaul_ai
View sourceRelated source 10VnExpress Số hóa
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.