• Sources: primary
  • Summary: Anthropic describes agent memory as notes written in session plus dreaming, an offline pass that checks notes against transcripts, builds linked wikis and promotes a small index into future context. On a 420M-token internal corpus, wiki-only scores rose from 13% to 48% as dreaming scaled from 2 to 128 sessions, and search plus wiki reached 61% versus 53% for search alone. Telling strong models to record only failures cut their gain from memory, while explicit steering helped older models.
  • Why it matters: Agent builders get evidence to keep injected memory small, let strong models judge what to save, and verify notes against logs.