Anthropic details agent memory components and dreaming for consolidation

Anthropic's engineering team published how it thinks about agent memory, treating idle-time "dreaming" like sleep-time consolidation: check notes against session logs, fix mistakes, and build knowledge wikis. Using Claude Managed Agents on Continual Learning Bench (CL-Bench) 1.0, it shows how successive Claude models use a persistent memory folder to close a learn-and-correct loop; Pokémon runs contrast dumping everything into memory versus keeping a small, goal-focused set of notes. On an internal corpus of about 350,000 files (~420M tokens), scaling dreaming from 2 to 128 sessions raised graded criteria on a 25-question set from about 13% to 48%. Cookbooks for dreaming and knowledge wikis are linked.

Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.