Anthropic:智能体记忆的关键在于选择遗忘而非存储量
Anthropic says the key to agent memory is knowing what to forget, not how much to store
Anthropic 用 Pokémon Red 做了个很直观的实验:Fable 5 只留几份策略文件,进度是 Sonnet 3.7 存 400 多个文件时的四倍。做智能体长期记忆的开发者可以直接借鉴这套“选择性遗忘加 Dreaming 修正”的思路,比盲目堆上下文省钱得多。
Anthropic 研究了 Claude 如何在多轮会话之间保留知识,结论是“选择性笔记”在效果上远超“穷尽式存档”。在 Pokémon Red 测试中,较新的 Claude Fable 5 只保留少量策略文件(对比旧方案的 400 多个文件),推进进度达到 Sonnet 3.7 的四倍。核心机制是让智能体自己决定哪些信息值得写入记忆,而不是把所有上下文原样堆进存档。
Anthropic 还引入了一个名为“Dreaming”的后台流程,用来回溯性修正记忆中的错误内容,避免早期误判在后续会话中被反复引用。这套方案表明智能体长期记忆的瓶颈在于信息筛选质量,而非存储容量。对开发者而言,构建跨会话智能体时可以优先设计压缩、摘要和淘汰规则,而不是扩大上下文窗口或向量库规模。
Anthropic says the key to agent memory is knowing what to forget, not how much to store
Anthropic studied how Claude retains knowledge across sessions. Selective notes beat exhaustive archives by a wide margin. In a Pokémon Red test, the newer Claude Fable 5 got four times as far as the older Sonnet 3.7 by keeping just a handful of strategy files instead of more than 400. A "Dreaming" process retroactively corrects faulty memories. The article Anthropic says the key to agent memory is knowing what to forget, not how much to store appeared first on The Decoder .
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Claude, Anthropic, agents, memory management.
Source attribution: curated from Decoder (@Jonathan Kemper). Cross-source corroboration: 11 independent sources have covered this development (including shao__meng, IT之家, elvis), confirming industry-wide signal strength.