Anthropic's 'dreaming' memory lifts Opus 5.5 to 0.833 on CL-Bench

Anthropic's Nov 6 post is notable for naming a three-part structure rather than a single memory feature. The first two layers happen during a session: Claude writes salient memory notes, and stronger models like Opus 5.5 judge salience themselves. The third layer, 'dreaming', happens away from the user, consolidating knowledge during idle time — which the post calls the first public description of Anthropic's sleep-time compute concept.

The post also pairs the architecture with measurements. With memory enabled, Opus 5.5 rose from 0.644 to 0.833 on CL-Bench. In a Pokemon eval, Claude Fable 5 earned four badges with 10 focused files, while Sonnet 3.7 earned one badge from 413 files — a contrast that, per the body, accompanied the memory-enabled setup.

For readers outside engineering, the practical thread is that memory is being split into writing, judging, and consolidating phases, with the consolidation happening off the user's clock. The body does not say when or whether these capabilities will be generally available, so the timing remains to be seen.