Anthropic shows how dreaming helps Claude fix its own memory
A review step between sessions lets Claude check its notes against what really happened and keep only what matters.

Anthropic published new research on November 6 about how Claude remembers things from one session to the next. Anthropic calls this agent memory. It means knowledge Claude gains from experience and keeps between sessions as notes it can read later.
The research focuses on a step called "dreaming". When people sleep, the brain replays the day and keeps what mattered. In a similar way, Claude can look back over its logs between sessions. It can fix wrong notes and write new ones.
One test used the video game Pokémon Red. An older model, Sonnet 3.7, saved almost everything. It wrote 413 memory files across 225 folders and earned one badge. Fable 5 wrote nine files in two folders, plus one main file, and earned four badges. Less memory, chosen well, went further.
Dreaming also fixed a real mistake. Claude wrote down the wrong side of a mountain for a landmark. Then it kept falling back onto the route it came from. A dreaming pass checked the note against the record of the session and rewrote it. With the fixed note, Claude reached the next city. Without it, the run never got there.
Dreaming can also build a wiki, a set of linked pages about what Claude knows. In a test with about 350,000 files from one company, scores on 25 questions grew from 13% to 48% as dreaming grew from two sessions to 128.
Open it and check the claim yourself. That is the point.