Anthropic explains how Claude agent memory works, including offline dreaming

Anthropic details in a company research post how Claude can retain knowledge between sessions using memory files. Its approach includes a process called dreaming, in which the agent reviews session logs during idle periods to correct mistakes and reorganize notes.

Anthropic defines agent memory as knowledge gained from user interactions and assigned tasks that is saved as context for later sessions. This excludes changes to model weights, learning within one conversation and summaries used to replace older context.

In a Continual Learning Bench test, newer Claude models created validated rules from earlier errors and used them to answer later questions. Anthropic found that fixed instructions about what to record could reduce the benefit of memory for stronger models, while older models improved when prompted to read their notes and test their theories.

A Pokémon Red evaluation also showed the value of selective memory. Sonnet 3.7 created 413 files across 225 folders and earned one badge, while Fable 5 wrote nine files in two folders plus a single top-level file, and earned four badges. Fable 5 focused its notes on goals, navigation plans and tested strategies rather than information already visible in the game.

The post's key design rule, as CellCog highlighted: "it helps to tell Claude what memory is for, but not how to use it."