Anthropic Details Claude’s Dreaming Memory
Anthropic’s new Engineering post explains how Claude’s “dreaming” process reviews transcripts and memory stores between sessions to organize knowledge, correct stale notes, fill gaps, and build useful representations such as wikis. The approach aims to help agents learn from experience without bloating every context window.
The key insight is that agent memory needs an offline maintenance loop, not just a larger database. Dreaming makes memory a system that can validate, compress, and reshape itself as tasks evolve.
- –Memory is organized into a hierarchy spanning transcripts, notes, wikis, and selectively injected context.
- –Dreaming corrected a wrong Pokémon map note by checking it against the original transcript, changing the agent’s eventual behavior.
- –Search plus an evolving wiki improved document-question performance from 54% to 61%, while merger questions reached 87% versus 73% with search alone.
- –More memory is not automatically better: one wiki grew larger than its source documents and added no clear benefit for simple retrieval tasks.
- –Developers should optimize memory around task goals, provenance, and maintenance cadence while budgeting for the extra compute required to keep memories accurate.
DISCOVERED
1h ago
2026-11-06
PUBLISHED
1h ago
2026-11-06
RELEVANCE
AUTHOR
AnthropicAI