Anthropic details agent memory: 'dreaming' lets Claude organize its experience

daniel_mac8 · x · 2026-11-19

danielmac8 highlights Anthropic's engineering post on agent memory, arguing externalized memory systems are the practical path to continual learning—recent models natively use wiki-style memory better with less instruction, so algorithmic breakthroughs on catastrophic forgetting may not be needed. The post explains a hippocampus/neocortex-style division of labor (logs as immediate capture), a 'dreaming' process based on sleep-time compute (Letta/UC Berkeley) where idle-time reflection fixes mistakes, creates notes, and shapes what's preloaded into future sessions, plus the evaluations Anthropic uses to measure each memory component's contribution.

Original post →

More from coding & agent

coding & agent channel →