Event ·Multi-source

Anthropic 分享 Dreaming:让 Claude 在任务外离线整理 Agent 记忆

Anthropic 研究员 Lance Martin 分享了名为 dreaming 的后台记忆整理方法:不更新模型权重,而是在任务外回顾记录、纠错、去重并组织知识。在约 35 万份文件、4.2 亿 token 的内部语料实验中,单靠搜索的评分标准覆盖率为 53%,wiki 配合搜索提升到 61%,每题少用约 2 万 token。在 CL-Bench 数据库任务中,较强模型开始保存经过验证的规则和依据,只记失败的设定反而削弱收益。合同案例里,改为每份合同单独一页并逐字引用条款后,最终报告的关键事实覆盖从 41/57 提升到 57/57。

Current read

Anthropic 把 Agent 的记忆整理比作做梦:后台纠错、归纳规则、建 wiki,搜索覆盖率从 53% 涨到 61%,做 Agent 的都该看看细节。

11 sources64° AI heatLast updated 2026/11/7 08:03:56

Evidence

11 sources
Related source 4Gorden Sun
View source
Related source 5kimmonismus
View source
Related source 7SuperTechFans
View source
Related source 8Latent.Space
View source
Related source 9VnExpress Số hóa
View source

Conflict 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.