Anthropic explains the components of agent memory — and why agents should 'dream' between sessions
brada · x · 2026-11-07
Anthropic's engineering team published a long-form post on how they think about agent memory for Claude, with brada summarizing the key insight as "agents should dream."
Core idea: borrow the brain's division of labor — the hippocampus captures experiences in the moment, while consolidation happens offline during sleep. For agents, logs/transcripts are raw records; idle time between interactions is used to "dream" (building on Letta/UC Berkeley's sleep-time compute), reflecting on logs to fix mistakes, produce notes, build new memory representations, and decide what gets preloaded into future sessions.
Key points:
- Agent memory = knowledge Claude gains from experience with users and tasks
- Dreaming can correct errors, create notes, and shape new memory structures
- The team shares several evals used to explore agent memory and quantify each component's contribution
This is Anthropic's most complete public walkthrough of its memory architecture methodology and evaluation approach.
More from coding & agent
- Game studio runs AI agent team on agile sprints: 284 messages and counting — draginol · 2026-11-07
- macOS CI at $0.16/min pays for a Mac Mini in 10 days — why CI prices should drop 50%+ — gajesh · 2026-11-07
- OpenAI ships API starter projects for agents, voice, and more, remixable in Codex — OpenAIDevs · 2026-11-07
- OpenAI Developers plugin adds Showcase tab; Pro/Business subscribers can claim API credits — OpenAIDevs · 2026-11-07
- Google resets everyone's Gemini and Argon quota on Antigravity ahead of the weekend — rseroter · 2026-11-07
- GTEx MCP Server gives AI assistants 25 tools to query genomics across 54 tissue types — modelcontextprotocol · 2026-11-07