Anthropic’s Four-Layer Agent Memory and Dream Workflow
Two Chinese technical commentaries explain Anthropic’s four-layer agent-memory design: transcripts, memory notes, memory representations such as linked wikis, and injected context. They describe Dream as an offline maintenance process that checks notes against source records, fills gaps, removes duplicates and conflicts, and reorganizes memory. Reported evaluations include improved task scores with memory and a Stainless corpus criterion-satisfaction rate rising from 13% to 48%, but the articles did not independently reproduce these vendor claims.
The reports provide a concrete framework for building persistent agent memory while distinguishing reusable knowledge from raw chat history, model-weight changes, and context compaction; they also highlight evidence verification, retrieval cost, and stale-memory risks.
Agent 记忆不是聊天记录,而是为未来准备的上下文
——Anthropic 的四层记忆架构与 Dream 思想
Agent Memory 系列·上篇:为什么需要长期记忆,为什么需要分层,以及 Dream 究竟在整理什么。 本文聚焦 Anthropic 的四层记忆架构与 Dream 思想,并参考 Dream workflow 材料解释核心概念;具体实现将在下一篇展开。实验数字均来自原文,本文没有独立复测,不应把它们视为不受条件限制的性能承诺。
内容摘要
Agent Memory(智能体记忆)不是把所有聊天记录重新塞给模型,而是为未来行动准备可信、可找到、恰好够用的上下文。
- 四层架构:会话记录保存证据,记忆笔记提炼经验,记忆表征组织知识,注入上下文自动加载小型入口。
- Dream 的作用:在离线时核验错误、补全缺口、整理重复和冲突,并重组不同层级的记忆。
- 评价标准:看下游任务效果、检索成本、证据支持和完整性,而不是记忆文件的数量。
#Anthropic#Claude#Claude Managed Agents#Claude Opus 5#Claude Sonnet 3.7#Claude Fable 5#Claude Opus 4.7#Claude Sonnet 4.6#Claude Sonnet 5