AI · 9 November 2026 · 4 min read
Claude’s memory got a tidy-up pass. The badge count is one run of a game
Anthropic published on 6 November a note on how Claude keeps notes between sessions, and on an offline pass it calls dreaming, which checks those notes against the logs. In one comparison, a model that wrote nine files got further in a game than a model that wrote 413. One run is not a result.
Anthropic published on 6 November, written by Lance Martin, a note on what it means by agent memory: knowledge Claude picks up from a task and keeps between sessions, as context it can read later. The note sets aside changes to the model’s weights, learning inside a single conversation, and the summaries that merely free space in the window. The interesting part is a pass it calls dreaming. Between sessions, Claude can test its notes against the original transcripts, merge duplicates, and promote a note so the next session starts with it already loaded. A pass that tidies a notebook is not a new model.
The game is the part with a number a reader can hold. Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5, each about 10,000 steps into Pokémon. Sonnet 3.7 wrote 413 memory files across 225 folders, and earned one badge. Fable 5 wrote nine files in two folders, plus a top file called global.md, and earned four badges. One run against one run is an illustration. It is not a trial with a count of attempts. The note’s other measured pass used Opus 5 on a corpus from the software firm Stainless, about 350,000 files and about 420 million tokens. As dreaming was scaled from two sessions to 128, the score on 25 questions rose from 13 per cent to 48 per cent of the graded criteria. That is Anthropic’s test, on Anthropic’s questions.