A new engineering blog published by Anthropic compares the memory of intelligent agents to sleep.
The method is divided into two stages. While the session was going on, Claude wrote notes while working; after the session, he "dreamed" offline, played back the entire record, corrected mistakes, organized it into a knowledge map similar to a wiki, and then forwarded the content that should be read first next time.
They tested it with Pokémon Red. Sonnet 3.7 left 413 memory files and only got one badge; Fable 5 left only 9 files and got four.
A specific example: Claude once mistakenly marked Mount Tsukimi as "completed". When I was "dreaming", I played back the record and found something wrong, so I changed the note. The next time Claude went back to the cave and continued walking instead of wandering around outside.
Remember that more does not equal better. For a model, 413 files are like not remembering them. The key points cannot be found, and they will be biased by error information. The idea this time is to allow memory to be modified and compressed.
They also tested it on about 350,000 internal company documents, expanded "dreaming" from 2 sessions to 128, and the task completion rate was significantly improved. An agent that runs for a long time may end up struggling with this ability to "organize notes".