Anthropic tests 'dreaming' so Claude can clean up its own memory notes

What 'dreaming' means here

As interestingengineering.com reported, Anthropic, the company behind the Claude AI assistant, has published research on a process it calls "dreaming." The name is a loose borrowing. It has little to do with human dreams, though it is inspired by the way people's brains are thought to sort and consolidate memories during sleep.

In practice, Claude uses idle time to look back over earlier interactions. It hunts for incorrect information, tightens up the notes it has already written, and builds organized knowledge that later sessions can use.

The problem it targets is simple. Claude can save useful details in memory files while it works, but a new chat does not automatically carry over what the last one learned. Notes are the bridge, and the quality of those notes decides whether the bridge holds.

Less memory, more badges

The first challenge is deciding what to write down. Save everything and you get clutter. Save too little and a future session lacks the context it needs. To test this, Anthropic used the video game Pokémon Red, where Claude must explore the game world and earn badges. Researchers compared two models after about 10,000 gameplay steps.

Sonnet 3.7 recorded nearly everything. Its memory held 413 files spread over 225 folders, close to 978,000 characters in total. It earned one badge.

The other model, which the source calls Claude Fable 5, kept just 10 files in three locations, about 43,400 characters. It earned four badges. Its notes focused on navigation strategies and lessons from play, plus one central file listing current goals and planned next moves.

The smaller set also spared the model from re-reading facts it could already get from the game's own interface, while still keeping discoveries that would otherwise vanish between sessions.

Fixing a note that was wrong

Memory has a nasty failure mode: a wrong note gets treated as fact. Anthropic showed this with another Pokémon run. Claude had written down the wrong location for a Pokémon Center (the game's healing stop) near Mt. Moon. After leaving the building, it misjudged where it was, kept walking over a one-way ledge and ended up back on an earlier route.

An offline review changed that. Claude compared its saved notes with the original transcript of the session, spotted the bad location and corrected the file. With the fix, it went back into Mt. Moon and eventually reached Cerulean City.

The run that kept the faulty memory hit the same obstacle again and again and never finished the journey. It is a small, game-sized example, but it shows why checking notes against the original record is useful.

Turning a document pile into a wiki

Anthropic also asked whether dreaming could turn large collections of documents into a connected knowledge base, something like a wiki where pages link to related pages. One test used roughly 350,000 internal documents from the software company Stainless, about 420 million tokens. (Tokens are the small chunks of text AI models read.)

Over 128 sessions, Claude gradually built and refined a wiki from that material. On 25 evaluation questions, its score rose from 13 percent to 48 percent.

A separate test used merger documents. For questions that needed facts from several documents, Claude answered with 87 percent accuracy using its wiki. Searching the documents directly reached 73 percent.

Dreaming can also reshuffle what is where: details Claude needs often can be moved into its immediately available context, while less relevant material stays in searchable files.

Limits and what comes next

Anthropic does not present this as solved. Maintaining memory uses computing resources, and too much information can swamp future sessions. The researchers say agents, meaning AI systems that carry out multi-step tasks, need to learn when to keep information and when to go back and review it.

Their results suggest that better memory management could help Claude handle more complicated work that stretches across several sessions. Note that this is a research write-up of experiments, mostly in a game and in document tests. The source gives no price, release date or product plan.