How Claude Repairs Its Own Memory Between Sessions

Anthropic published a paper on how its AI models handle long-term memory. The company calls the process "dreaming." When the AI model has nothing else to do, it looks back over past interactions, meaning session logs, and edits its old notes. To show this in action, Anthropic let Claude play Pokemon. The older version, Claude Sonnet 3.7, built up 413 separate memory files. It spread them across 225 folders. That version earned only one badge in the game, which is the game's progress marker. The newer version, Claude Fable 5, created just 9 files in 2 folders. There was also one shared file called "global.md." This version earned 4 badges.

This is worth keeping in mind when we design our own AI systems. If an AI model piles up too many details in long-term memory, it can drown in them. That is what happened to Claude Sonnet 3.7. In a separate experiment, Anthropic found something useful. Telling the AI model to check its own notes before answering nearly doubled its score on a test. The researchers also had it recheck those notes for errors.