Anthropic describes Claude's dreaming process for organizing memory
Article timeline
The articles in this story. Times are ET.
meng shaoXAnthropic finds agent memory works best when organized, not just storedAIAnthropic's research says agent memory needs more than saved history, including evidence lookup, note correction, and experience organization. In an internal corpus test, a knowledge base combined with search raised criteria coverage from 54% to 61% as organization rounds increased. The author notes that offline organization helps some tasks but does not replace retrieval, and its maintenance cost may not always be offset.
elvisXElvis Saravia says agent memory should stay basic, mixing folders with episodic memoryAIElvis Saravia says no one-size-fits-all agent memory system exists, and that he uses a mix of organized folders and files with an episodic memory system alongside frontier models. He argues memory must be upgraded and maintained over time, and that agent memory systems should stay extremely basic to avoid confusing the agent. Anthropic separately published an Engineering Blog post on how its "dreaming" process helps Claude organize what it has learned and correct its mistakes.
AnthropicOfficialAnthropic explains how dreaming helps Claude organize and recall memoryAIAnthropic has published an Engineering Blog post describing how "dreaming" helps Claude organize what it has learned, correct its mistakes, and find needed information faster in later use. The post frames memory as what lets Claude learn from experience.
- Anthropic ResearchOfficialPickAnthropic explains agent memory components and dreaming for Claude
AIAnthropic describes how Claude keeps memory notes between sessions and uses a process it calls dreaming to review transcripts, correct wrong notes, and build wikis. Its evaluations compare memory-writing rules, the Pokémon Red and CL-Bench tasks, and a wiki built from 350,000 files, where scores on 25 questions rose from 13% to 48% as dreaming scaled from two sessions to 128.