Anthropic explains Claude agent memory: 'dreaming' uses idle compute to organize memory
tomssilver · x · 2026-11-09
Tom Silver shares Eric Horvitz's 2001 paper Principles and Applications of Continual Computation, an early take on sleep-time compute, alongside Anthropic's new engineering post on agent memory. Anthropic borrows the hippocampus/neocortex consolidation metaphor: agents can use idle time between interactions to 'dream'—replay logs, fix mistakes, write memory notes, and decide what to preload into future sessions. The post walks through Claude's memory components and the evaluations they used to study memory organization, building on the sleep-time compute idea from Letta and UC Berkeley.
More from coding & agent
- Swarm agent training needs thousands of agents, ambitious problems and cost math — samiramanabi · 2026-11-09
- product-engineering: MIT-licensed agent skills for shipping products with Claude agents — PawelHuryn · 2026-11-09
- MIT-licensed skill embeds Claude agents that can propose but not execute Stripe changes — PawelHuryn · 2026-11-09
- Google's frontier AI product head Tulsee Doshi on why PMs are suddenly valuable again — lennysan · 2026-11-09
- GitHub Copilot app adds Projects: give it a goal and it coordinates cross-repo work — 0xkarasy · 2026-11-09
- Termilly 1.6 launches: manage terminals and agent sessions from anywhere — alex_under___ · 2026-11-09