Anthropic Research Unveils Three-Part Agent Memory Architecture, Lifts Opus 5.5 CL-Bench From 0.644 to 0.833

Anthropic's Nov 6 research post breaks agent memory into three layers: Claude writes salient memory notes during sessions, stronger models like Opus 5.5 judge salience themselves, then 'dreaming' consolidates knowledge during idle time (the first public description of Anthropic's sleep-time compute concept). With memory enabled, Opus 5.5 jumped from 0.644 to 0.833 on CL-Bench, and in a Pokemon eval Claude Fable 5 earned four badges with 10 focused files versus Sonnet 3.7's one badge from 413 files.