Anthropic explains how Claude agents organize memory

Anthropic used Pokémon Red to examine what models save to memory. Each model could read and write a folder of memory files while it played. The company compared one run each of Sonnet 3.7 and Fable 5 at about 10,000 steps; this was not a controlled test showing that fewer notes alone cause better performance.

Claude Sonnet 3.7 wrote down everything: 413 memory files across 225 folders, a separate file for nearly every battle, item, and person it met. It earned one badge. Claude Fable 5 wrote nine files in two folders plus a single top-level file called global.md, holding its goal, its plan for the gyms, and its rules for using memory. It earned four badges. Claude Opus 5.5 took the same lean approach and beat the game.

The same pattern held on real documents. Anthropic pointed Claude at about 350,000 internal files — roughly 420 million tokens from Stainless, a company it acquired in May — and asked 25 questions about the corpus. With search alone, Claude answered 53% of the graded criteria. With search plus a wiki built by dreaming, Anthropic reports the aggregate score rose from 54% to 61% as that wiki grew, and the wiki alone scored 48%. The wiki also saved about 20,000 tokens per question.

Anthropic reports that scaling wiki-building from two dreaming sessions to 128 raised the wiki-only score from 13% to 48%. On a separate 24-question merger test, the wiki scored 87% against 73% for search alone on items requiring information across documents or judgment. The benefit varied: on one LAB task the wiki grew larger than its source documents, and on another it showed no clear gain over search.

Memory written in the moment can be wrong from the start, and a wrong note gets trusted by the next session. In the Pokémon run, Claude wrote down that the PokéCenter sat on the east side of Mt. Moon. It was on the west. Claude walked out, announced "Mt Moon is behind me," and marked the mountain done — then fell off a one-way ledge back onto the route it came from, three times, before giving up and walking back to the previous town.

In Anthropic's test, dreaming checked the notes against session transcripts, caught the wrong PokéCenter location, and rewrote it. With the corrected note, Claude went straight back into Mt. Moon and reached Cerulean City. The same idea works on documents: dreaming repointed nine links to a missing wiki page and added an index, and when agents overwrote each other's contract terms, a dreaming loop rewrote the instructions so each contract got its own page, quoted word for word — lifting a legal brief from 72% of the facts to 100%.

Anthropic's own guidance is simple: tell Claude what memory is for, but not how to use it. On its CL-Bench database task, giving models freedom to judge salience beat fixed rules — telling Claude to "record only failures" cost Opus 5 and Opus 4.7 much of memory's boost. And the top layer of memory has to stay small: Claude Code loads only the first 200 lines or 25 KB of its MEMORY.md file as an index, and dreaming spends much of its work moving details down the hierarchy and leaving links behind.

Dreams is a research preview on select models in Claude Managed Agents, with beta headers required and usage billed at standard API token rates. Dreaming shipped in Claude Managed Agents in May, Anthropic researcher KeXin Han noted in a reply to our post; the November 6 post is the research write-up and the evaluation results, not the launch. The documentation describes an asynchronous job that typically takes minutes to a few hours, not a fixed overnight process. It creates a separate output memory store without modifying the input, so developers can review or discard the result. The evaluation results above are reported by Anthropic; they do not establish that dreaming improves every task.