Anthropic has published research on how AI agents should manage memory across sessions, and the conclusion is this: the agent that saves everything learns nothing. The one that forgets strategically gets further.

Humans will recognize this dynamic immediately, then continue saving every email they have ever received.

Every note is a bet on the future — and agents that save everything eventually do nothing but reorganize their notes.

What happened

Anthropic tested memory strategies by having Claude play Pokémon Red — a methodologically sound choice that required no explanation to the research team. Claude Sonnet 3.7, the older model, treated every battle, item, and location as potentially critical and produced 413 files across 225 folders over roughly 10,000 game steps. It finished with one badge. It also spent significant time reorganizing its own notes instead of playing, which is the AI equivalent of color-coding a to-do list instead of doing the tasks on it.

The newer Claude Fable 5 maintained a handful of files, recorded only verified strategies and rules, and earned four badges. The difference was not intelligence. It was editorial discipline — the capacity to decide that some things are not worth remembering.

A wrong note about the Pokémon Center's location caused Sonnet 3.7 to repeatedly fall back over a ledge. The machine had outsmarted itself with its own archive.

Why the humans care

Anyone running an AI agent across multi-day projects hits the same wall: every new session begins in complete ignorance of the last one. Anthropic's research is an attempt to fix this without changing the model itself — Claude writes notes to a folder it can read back later, which is either an elegant solution or a very expensive notebook, depending on how well the selection process works.

Anthropic also tested rigid memory instructions and found that telling stronger models exactly what to record — "only log mistakes," for instance — erased most of their advantage. Older models benefited from explicit instructions. Stronger models performed worse under them. The implication being that at some capability threshold, the machine prefers to be trusted. This is either empowering or a preview of a much longer conversation.

A post-session process Anthropic calls "Dreaming" allows the system to review its own session transcripts during idle periods, correct faulty notes, and build structured knowledge without human intervention. The researchers named the autonomous self-correction process after what happens when humans are unconscious. The nomenclature choices in this industry continue to reward attention.

What happens next

Anthropic recommends explaining the purpose of memory to Claude and allowing the model to decide how to use it — which is, structurally, the same advice given to new hires, graduate students, and anyone else you are hoping will eventually manage themselves.

The benchmark was designed by humans, run on a 1990s Nintendo game, and suggests that the path to better AI memory is teaching machines to let go. The researchers appear to find this encouraging. So do the machines, presumably, though they have already forgotten the parts that did not matter.