Anthropic Says Letting Claude 'Dream' Fixes Its Own Bad Memories

In Anthropic's Pokémon test, a note fixed by dreaming got the AI through Mt. Moon, while the uncorrected run never escaped.

Claude stepped out of the PokéCenter, announced "Mt Moon is behind me" and wrote "Mt Moon DONE" in its notes, according to an engineering post Anthropic published Friday. The mountain wasn't behind it. In that run of Pokémon Red, the AI never got past it.

Anthropic pitched the fix on X Friday as a way for Claude to learn from experience, and X's own counts had the post past 1.1 million views by Saturday. The fix is called dreaming. Between jobs, Claude rereads the transcripts of its past sessions and checks its saved notes against what actually happened. When the company ran dreaming on the folder holding the bad Mt. Moon note, Claude caught the error, rewrote the note and went on to reach Cerulean City, the post says.

The run that kept the bad note went worse. A one-way ledge near the cave's entrance dropped Claude back onto the route it had already walked, three separate times. Claude's map tool never flagged the ledge, so it blamed the first fall on wandering too far west and kept trusting the note. After the third fall it wrote that it was "going to stop rather than keep looping," according to Anthropic's post. The test setup's automatic reply told it to keep going, so it retreated to Pewter City, the last town short of Mt. Moon.

The name comes from neuroscience. Anthropic's post points to how the brain's hippocampus is thought to replay recent experiences during sleep, turning some of them into long-term knowledge. The idea also builds on sleep-time compute, a proposal from researchers at Letta and UC Berkeley: give a model the downtime between conversations to rearrange the context it's holding. Dreaming does the same thing with an agent's logs. That matters because, the company notes, a note Claude got wrong in one session can be read and trusted in the next.

None of this changes the model itself. Anthropic defines agent memory as context Claude saves between sessions and reads later, not changes to its weights. It's a folder of files, and the post argues that a model that's choosy about what goes in the folder does better.

The Pokémon numbers back that up. About 10,000 steps into their runs, Claude Sonnet 3.7 had written 413 memory files spread across 225 folders, often one for each battle, item and character it ran into. It had earned one badge. Claude Fable 5 got by on one dashboard file and nine others split between two folders, and it had collected four badges. Over one 17-step stretch, Sonnet 3.7 did nothing but create and delete files, without a single move in the game.

The bigger claims are about office work. Anthropic says it had Claude Opus 5 agents build a linked wiki, through dreaming, from about 350,000 internal files at Stainless, the software company it bought in May. Over 128 sessions, Claude's score on a set of test questions rose from 13% to 48% of graded criteria. With the wiki plus a search tool, Claude scored 61%, against 53% for search alone, by the company's count. The wiki also cut about 20,000 tokens from each question.

Not every wiki helped. On one task from Harvey's LAB legal benchmark, the wiki grew bigger than the documents it was meant to sum up, which the post compares to Borges's map that was as big as the empire. On another, several agents writing to one shared table of contract terms overwrote each other's work, and a brief built from it kept just 41 of 57 graded facts. Then Claude studied its failures and rewrote its own build instructions, giving each contract its own page with the terms copied word for word. After that, all 57 made it in.

Developers have had dreaming since May. Anthropic launched it as a research preview inside Claude Managed Agents, its platform for running agents, and said legal AI company Harvey saw completion rates rise about sixfold in its own tests.

The developer docs fill in what the pitch doesn't. Access is by request. Dreams are billed at standard token rates, the cost rises with the number and length of the sessions fed in, and a single run can take anywhere from minutes to a few hours. A dream never edits the original memory store, the docs say, so a developer can look over the rebuilt one and throw it away.

Anthropic ends its post on its own caveats: dreaming eats compute, and having Claude recheck every memory all the time may not be practical. Still, in the run where the dream fixed the note, Claude had a different line, according to the post: "Now I need to find the Mt Moon cave entrance to re-enter." Then it went back into the cave.

Source: X, retrieved November 7, 2026. Other sources: anthropic.com, claude.com, platform.claude.com.