Skip to content
The Decoder· Jonathan Kemper·Indexed by this site · Originally published· AI updates →· Its place in the scan record →

Anthropic says the key to agent memory is knowing what to forget, not how much to store

Anthropic says the key to agent memory is knowing what to forget, not how much to store

AI summary

Anthropic published a research article saying that the key to agent memory across sessions is selective recording rather than full archiving, and that letting Claude review its notes afterward can correct errors. In the Pokémon Red test, Claude Sonnet 3.7 saved everything, creating 413 files over about 10,000 steps of play and earning only 1 badge; Claude Fable 5 recorded only plans, strategies and verified rules, earning 4 badges, but the comparison covered only one playthrough per model. Its Dreaming mechanism lets Claude check its notes against old session records when idle, and only after correcting erroneous notes did it get through Mt. Moon to Cerulean City. In a wiki legal test built from about 350,000 internal documents from the acquired Stainless, Claude scored 87% on questions spanning multiple documents, versus 73% with search alone. The article also notes that rigid instructions offset most of the advantages of strong models, and that note-taking distracts attention and Dreaming consumes extra compute.

Read originalthe-decoder.com