Tuesday, 17 November 2026, story 6 of 7
Anthropic says agents remember better with fewer, curated notes
In Anthropic's Pokémon test, a newer Claude with few notes beat an older one that logged everything; a review pass called Dreaming fixes errors.

What happened
Anthropic says selective notes work better than exhaustive archives when Claude keeps memory across sessions, according to a research post reported by The Decoder. In a Pokémon Red test, Claude Sonnet 3.7 created 413 files in 225 folders over roughly 10,000 game steps and finished with one badge. Claude Fable 5 kept a few files of plans, strategies and tested rules and earned four badges. Anthropic's comparison covers a single playthrough per model.
Anthropic also describes Dreaming, a pass in which Claude checks its notes against earlier session transcripts. It launched in May as a research preview for Claude Managed Agents.
What it means
For anyone building agents that work across sessions, the practical lesson is to filter before storing. Anthropic's results suggest notes worth keeping are plans and tested rules, not facts the agent can already see. Its advice on instructions cuts both ways: rigid rules such as "only record mistakes" erased most of the advantage for stronger models, while older models gained from explicit ones. Memory prompts may therefore need tuning per model.
The evidence is thinner than the framing: one playthrough per model, different models compared, all tests Anthropic's own. Dreaming costs extra compute and delays results, and notes need upkeep. In the legal test a wiki scored 87 percent against 73 for search alone, but simple lookups showed no gain. Watch for independent replication.
What to do: Before adding more memory to an agent, cut what it records to plans and tested rules, and check whether a periodic review pass catches stale notes.
Source
Filed under Research. Published .
"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.