Primary source宝玉
View sourceEvent ·Multi-source
Anthropic 拆解 Claude 跨会话记忆设计：靠"做梦"回头纠错整理笔记
Anthropic 工程团队 11 月 6 日发文，介绍给 Claude 设计跨会话记忆的方法。他们在会话结束后加入名为"做梦"的后台步骤，让 Claude 回翻原始对话、改错笔记并整理成 wiki。数据库测试里 Opus 5 加记忆后得分从 0.564 涨到 0.828；宝可梦测试中 Fable 5 只写 10 个记忆文件就拿到 4 枚徽章，而 Sonnet 3.7 写了 413 个文件只拿 1 枚。文章还给出开发者经验，比如自动加载层只取 MEMORY.md 前 200 行，多 agent 并行时别写同一文件。
Current read
Anthropic 公开了 Claude 记忆系统的设计细节，连宝可梦通关的测试数据都放了出来，做 agent 记忆的值得照着抄几条经验。
11 sources83° AI heatLast updated 2026/11/8 03:21:33
Evidence
11 sourcesRelated source 1IT之家
View sourceRelated source 2AIGCLINK
View sourceRelated source 3SuperTechFans
View sourceRelated source 4Claude Code: GitHub Releases
View sourceRelated source 5claudedevs
View sourceRelated source 6TestingCatalog
View sourceRelated source 7shao__meng
View sourceRelated source 8AI Will
View sourceRelated source 9Decoder
View sourceRelated source 10Detik Inet
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.