Giving an AI agent memory sounds simple: let it write notes and read them next time. Anthropic's new research post shows why it is not, and why the answer has changed as models have improved. The short version is that stronger Claude models save far less, choose better what to save, and do worse when told exactly what to write down.
The clearest example comes from Pokémon Red. About 10,000 steps into the game, Claude Sonnet 3.7 had written 413 memory files across 225 folders, including separate files for individual battles, items and characters, and had earned one badge. Claude Fable 5 wrote nine files in two folders plus a single top-level dashboard, and earned four. Much of what the older model saved, such as money or map position, can be read off the screen anyway, and at one point it spent 17 steps reorganising files without making a move. Opus 5.5 kept a small, constantly rewritten set of notes and finished the game. On a database question task from Continual Learning Bench 1.0, Opus 5 was the first model to close the loop, writing rules along with how it knew they were true; Opus 5.5 and Sonnet 5.5 did the same.
The guidance for builders is specific. Telling Claude what memory is for, without dictating how to use it, worked best for strong models. Imposing a rule such as "record only failures" cost Opus 5 and Opus 4.7 much of their gain from memory. Older models still need steering: prompting Sonnet 4.6 to read its memory and test its theories nearly doubled its score, while Sonnet 5 read its memory unprompted nine times out of ten.
The second idea is dreaming, reviewing past transcripts offline to fix bad notes. In one Pokémon run, Claude wrongly recorded which side of Mt. Moon a landmark was on and kept falling back down the same ledge; after dreaming checked the note against the transcript, it went back into the cave and reached the next city. On a corpus of about 350,000 internal files, building a linked wiki over 128 dreaming sessions lifted scores on 25 questions from 13% to 48%, and wiki plus search beat search alone, 61% to 53%, while saving about 20,000 tokens per question.
For developers on the Claude Developer Platform, much of this runs through Managed Agents and is directly usable. The caveats are that every result comes from Anthropic's own tasks and games, some comparisons rest on single runs, and the post itself shows a wiki growing larger than the documents it summarised, a reminder that more memory is not always better memory.