An SI Kept Getting Lost in a Pokémon Cave. After It ‘Dreamed,’ It Made It Through.

Every time you open a new chat with an SI helper, it usually starts from zero. It doesn’t remember yesterday unless somebody built it a notebook. And when it does keep a notebook, there’s a nasty catch: if it writes down something wrong, it will trust that wrong note again and again.

Anthropic, the company behind the Claude chatbot, says it has a fix, and it borrowed the idea from how people sleep. In a research post published Friday, the company described “dreaming”: a quiet, offline pass where Claude rereads the full record of its past work, checks its own notes against what actually happened, throws out duplicates and rewrites what it got wrong.

“No model can really know what will happen in future turns,” the post says. “Dreaming gives Claude an opportunity to look backward and use compute to organize its memory, correct mistakes, and maintain representations that allow future sessions to access memory more readily.”

The cave that broke it

The clearest example comes from an old Game Boy classic. Anthropic has long tested Claude by letting it play Pokémon Red on its own, keeping a folder of notes as it goes.

To get through Mt. Moon, a cave in the game, the player has to enter on the west side and come out on the east. Near the entrance is a trap: a one-way ledge that drops you back onto the road you just came from. Claude was using a nearby building, the PokéCenter, as its landmark, and it wrote down that the building sat on the east side. It didn’t.

So Claude walked out of the building, announced “Mt Moon is behind me,” and marked the cave “DONE.” A few moves later it went over the ledge and landed back where it started. Instead of doubting its note, it decided, “I overshot far west.”

In the run where that bad note stayed in place, Claude fell off the ledge three times. After the third, it wrote, “I’m going to stop rather than keep looping.” The test setup told it to keep going, so it wandered all the way back to the previous town. That run never reached the next city.

Then Anthropic ran a dream over the same notebook. Claude compared the note to the record of what had really happened, caught the error and rewrote it. With the corrected note, it said, “Now I need to find the Mt Moon cave entrance to re-enter,” went straight back into the cave and eventually reached Cerulean City, the next stop.

Less is more

The company also found that newer versions of Claude are simply better at deciding what is worth writing down. About 10,000 moves into the game, an older version, Claude Sonnet 3.7, had saved 413 memory files across 225 folders, logging nearly every battle and item, and had earned one of the game’s badges. A newer version, Claude Fable 5, kept just nine files in two folders plus one main page with its current goal and plan, and had earned four.

The older model’s problem was keeping notes on things the screen already shows, like how much money it had. At one point it spent 17 straight moves creating and deleting memory files without playing the game at all. Think of a new hire who writes down everything versus one who writes down only what matters.

Does it help with real work?

Games are a fun test, but Anthropic also tried it on paperwork. It took about 350,000 internal files from Stainless, a software company it bought in May, and asked Claude to answer 25 hard questions about them. Over 128 rounds of dreaming, Claude built itself a linked set of summary pages, like a private Wikipedia for the company. Using only those pages, its score rose from 13% to 48% of the graded points.

That sounds big, but read the fine print. Claude did best when it had both its homemade pages and a regular search tool, and there the gain was modest: 61%, versus 53% with search alone. It also needed about 20,000 fewer tokens per question, the small chunks of text an SI reads and is billed by, which saves money.

On a set of 24 questions about a company merger, most of which required stitching together facts from several documents, Claude scored 87% with its pages versus 73% with search alone. But on a legal task where each answer sat in a single document, the pages added nothing. On another, Claude’s “summary” grew bigger than the documents it was summarizing.

The catch

These are Anthropic’s own tests on its own product, not results from an outside lab. The company is also upfront about the risks. A wrong note that gets promoted to the top of the notebook “can be carried into every session,” the post warns. Dreaming also costs computer time, so checking every memory all the time “might not be practical.”

For now, dreaming is a “research preview” for businesses building their own Claude-powered assistants, and they have to request access. A dream never overwrites the original notebook; it writes a new copy that a person can review and throw away. In a separate how-to guide on Anthropic’s site, the author said a dream over 50 long work sessions cost about $8 and took about 10 minutes.

Why it matters

The SI helpers coming to phones and offices are supposed to work for days, not minutes: booking travel, managing projects, handling paperwork. An assistant that forgets everything overnight is annoying. One that remembers the wrong thing forever is worse. Anthropic’s pitch is that the answer looks a lot like what your grandmother told you: sleep on it, and check your work in the morning.