Anthropic details agent memory: dreaming organizes Claude's memory, wiki+search hits 61%
RLanceMartin · x · 2026-11-12
Anthropic's engineering team published "The components of agent memory," and the author previews a talk at AI Engineer covering it.
Key points
- Inspired by the brain: agents capture experiences in logs, while "consolidation" happens offline — reflecting on logs to fix mistakes and create notes/representations, dubbed dreaming (echoing Letta/UC Berkeley sleep-time compute).
- Claude uses dreaming to correct mistakes, build notes and new memory representations, and decide what gets preloaded into future sessions.
- Evals shown: with Opus 5 building wikis, graded criteria met rose from 54% to 61% for wiki+search as dreaming tokens scaled 2M→128M (search-only: 53%; wiki alone: 13%→48%); the wiki grew from 103 pages/766 links after 2 sessions to 222 pages/12,893 links after 128.
- Also covered: scaling a 1,000-agent team of Fable 5.2 to formalize Fermat's Last Theorem.
Related event: Anthropic Details Agent Memory: Letting Claude "Dream" to Consolidate(2 posts)→
More from coding & agent
- Harvard's new CS2680 walks students from Claude Code down the LLM serving stack — udmrzn · 2026-11-12
- Local vs Cloud AI Coding Agents: Benchmarks, Hardware Costs, Performance — luisdans · 2026-11-12
- The simplest heuristic for agent UX: expect it to behave like a capable coworker — andreisavu · 2026-11-12
- Agent Substrate 1.0 GA: microVM-sandboxed agent runtime with zero ambient credentials — rakyll · 2026-11-12
- Dev builds sketch-guessing app with new Gemini Decisions API, guesses terrible drawings — fofrAI · 2026-11-12
- Together Compiler: agent-driven continual compilation for AI inference — togethercompute · 2026-11-12