Anthropic says effective AI memory depends on selective forgetting
Anthropic tested how Claude can preserve information in note files and use it during later sessions. The experiments suggest that recording strategies, tested rules and plans works better than saving every event. In a Pokémon Red trial, Claude Sonnet 3.7 created 413 files in 225 folders over roughly 10,000 steps and earned one badge, while Claude Fable 5 kept its memory to a few files and earned four. The comparison covers only one playthrough per model.
Anthropic also described Dreaming, a process that reviews old session transcripts during idle periods. Claude compares its notes with the transcripts, corrects inaccurate entries and organizes knowledge into more structured collections. In the Pokémon test, this process fixed a mistaken note about a landmark near a mountain, allowing Claude to continue from Mt. Moon to Cerulean City.
Dreaming was also used to build a linked wiki from about 350,000 Stainless documents. On questions requiring information from multiple documents, the wiki produced an 87 percent score versus 73 percent for search alone. Anthropic says memory still requires maintenance: note-taking can distract agents, Dreaming uses additional compute, and loading too many memories can spread errors. Dreaming was introduced as a research preview for Claude Managed Agents and initially supported Claude Opus 4.7 and Sonnet 4.6.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.