Anthropic study finds Claude's memory gains depend heavily on what gets saved

Anthropic's engineering team published research on November 6 examining how persistent memory affects Claude's performance on agent tasks. In a database question benchmark where each question starts a fresh session, giving Claude Opus 4.7 a memory directory raised its score from 0.383 to 0.700. However, instructing the model to record only its failures erased much of that improvement for some models. The study also found that Sonnet 4.6 often ignored its memory files, and that explicitly prompting it to read them nearly doubled its score. The authors recommend keeping the always-loaded top layer of memory small, since injected notes consume tokens and can propagate errors.