Anthropic shares research on the components of agent memory

Summary

Anthropic has published research on how it designs agent memory: knowledge Claude keeps between sessions. The core ideas are memory notes written by judging what is salient in the moment, and dreaming, where the agent re-reads session transcripts offline to fix mistaken notes. In the company's own experiments on a long-horizon Pokemon task, an older model wrote 413 memory files and earned one badge while a recent model wrote nine files and earned four; dreaming also corrected a wrongly noted location and got the run back on track.

Why it matters for our work

When agents handle long-horizon work, failures often trace back to memory design. A map of what to record and when to self-verify underpins the reliability of human-AI collaboration.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.