Anthropic Explains How Claude Agents Remember Like a Sleeping Brain

Anthropic has published a study on how its Claude agent remembers things across chats. It suggests AI can learn from the human brain, which sorts memories during sleep. The agent can review past records, fix mistakes, and build knowledge in its free time.

The study splits agent memory into four layers: chat logs, memory notes, knowledge like a wiki, and injected context. It adds a "dreaming" step, where the model reviews logs, fixes errors, and sharpens rules between chats. Memory, the team says, is really a guess about the future.

In tests, Claude Opus 5.5 and Sonnet 5.5 with memory scored 0.838 and 0.833, well above their no-memory baselines. In Pokemon Red, a Claude with just 10 memory files won 4 badges, while one that wrote 413 files got only 1. Keeping memories small and sharp works better.

In a test with about 350,000 company documents, the pass rate on questions rose from 13% to 48% as dreaming rounds grew. Anthropic says open questions remain, like when to save and how to avoid errors, but AI may soon turn scattered chats into lasting knowledge.