Claude "dreams": How Anthropic teaches its AI agents to tidy up their memories

What it's about

Can an AI dream? Anthropic, the company behind the Claude language model, at least uses the word. On November 6, 2026, it published a research paper titled "The components of agent memory". The paper looks at how AI agents retain, verify and organize knowledge across multiple work sessions. AI agents are programs that handle multi-step tasks on their own instead of just answering one question at a time.

The centerpiece is a process Anthropic calls "dreaming". When Claude has nothing to do, it goes back over earlier sessions, spots wrong entries in its notes, fixes them and turns the results into organized knowledge for future conversations. According to Interesting Engineering, this has little to do with human dreams. Anthropic does take its cue from a process in the brain, though: the hippocampus records experiences in real time, and during sleep they get consolidated into long-term knowledge in the cerebral cortex.

The feature itself isn't brand new. As Crypto Briefing reports, Anthropic first showed dreaming on May 6, 2026, at its "Code with Claude" developer conference in San Francisco. Back then it was pitched as a feature for Claude Managed Agents, meaning agents run by Anthropic. The new study now adds experiments and numbers.

The problem: what's worth remembering?

While it works, Claude can save useful information to memory files. These are plain text notes that a later session can read again. The hard part is deciding what goes in. Write everything down and you drown in clutter. Write too little and future sessions are missing important context. Anthropic calls this salience: how important a piece of information actually is.

The team tested this with the Game Boy classic Pokémon Red. Claude had to explore the game world and collect badges. After roughly 10,000 game steps, the researchers compared two models:

Interesting Engineering puts Fable 5 at 10 files, while Blockchain.news says nine. Either way, the picture is clear: the older model wrote down almost everything and barely made progress. Fable 5 mostly recorded navigation strategies and lessons from the game, and kept one central file with its current goals and planned next steps. That meant it didn't have to process information twice when it was already on the game screen. Useful discoveries still carried over from one session to the next.

In another test, Continual Learning Bench (CL-Bench) 1.0, which measures how well a model keeps learning over time, Claude had to answer database queries and take notes for later. According to Blockchain.news, Claude Opus 5 and Opus 5.5 did better than Sonnet 4.6 and Opus 4.7 at "closing the learning loop". In practice, that means they recorded verified rules and noted possible pitfalls.

Dreaming fixes false memories

Notes can also be wrong. That gets dangerous when later sessions treat them as established fact. Another Pokémon experiment showed exactly this: Claude had written down the wrong location for a Pokémon Center near Mt. Moon. So after leaving the building, it didn't know where it was. Again and again it jumped off a ledge that can only be crossed in one direction and ended up back on an earlier route.

The fix was a review after the fact. Claude compared its notes against the original session transcript, found the wrong location and corrected the file. Then it headed back into Mt. Moon and eventually reached Cerulean City. The version with the faulty note kept failing at the same obstacle and never made it.

From scratchpad to company wiki

Dreaming is also meant to turn large document collections into linked knowledge bases structured like a wiki: individual pages that reference each other. For one experiment, Anthropic used around 350,000 internal documents from software company Stainless, about 420 million tokens in total. Tokens are the word and syllable fragments language models use to process text. Over 128 sessions, Claude built and refined a wiki from them step by step. On 25 evaluation questions, accuracy rose from 13 to 48 percent. According to Blockchain.news, that's the jump from two to 128 dreaming sessions.

A second test involved documents from a corporate acquisition. On questions whose answers required pulling information from several documents, Claude got 87 percent right using its wiki. Plain document search managed 73 percent. Dreaming can also reshuffle information: frequently needed material moves into the directly available context, meaning what the model has in front of it for every request. Less frequently needed material stays in searchable files.

How it works in practice

Under the hood is a Dreams API, a programming interface for developers. It takes an existing memory store plus 1 to 100 session transcripts and runs asynchronously, in the background, without interrupting a running agent. According to Crypto Briefing, it merges duplicate entries, is supposed to remove contradictions and finds patterns, such as mistakes the agent keeps making or workflows that reliably succeed. The output is a cleaned-up, reorganized memory store. The original data stays untouched.

One key point: dreaming doesn't change Claude's model weights. So the model doesn't learn in the strict sense. Whatever it "learns" ends up as a plain text note or a structured "playbook", a kind of set of instructions. Crypto Briefing sums it up nicely: the agent doesn't become a different model, it gets a better-organized notebook. You can choose whether changes are applied automatically or only after you approve them.

At the launch of the research preview, an early testing phase, dreaming supported the models claude-opus-4-8, claude-opus-4-7 and claude-sonnet-4-6, according to Crypto Briefing. Companies have to apply for access. Other features announced at the same conference have since moved into a broader public beta, but dreaming hasn't. Anthropic's most impressive number so far comes from an early test with legal AI company Harvey: on realistic legal benchmarks, the task completion rate reportedly went up sixfold after dreaming.

Our take

Memory is widely seen as one of the biggest weaknesses of today's AI agents. Every session basically starts from scratch, and long tasks spanning days or weeks often fail because knowledge gets lost or passed on incorrectly. Anthropic's message: just saving things isn't enough. Useful memory also needs evidence-based verification, correction, organization and targeted retrieval. That's a sensible, well-supported view. Another plus is that the notes are plain text, so people can read what the agent thinks it knows.

There are caveats, though. Almost all the numbers come from Anthropic itself and from hand-picked test scenarios. The Pokémon example also compares two very different model generations, so the gap isn't down to memory alone. The Harvey figure comes from a single partner in a single industry, and Crypto Briefing calls it "a signal, not a guarantee". Anthropic itself admits that maintaining memory costs compute and that too much information can overwhelm later sessions. And a system that summarizes and condenses can also cement mistakes, drawing the wrong lesson and filing it away neatly. Keeping the originals and allowing manual approval of changes is meant to prevent exactly that. Whether it's enough in day-to-day use remains to be seen. OpenAI and Google are also working on persistent memory for their assistants. But with this study, Anthropic is laying out in unusual detail how it plans to tackle the problem.