Anthropic published agent memory research on November 6 laying out four layers from transcripts to injected context
Anthropic's study divides agent memory into transcripts, memory notes, knowledge representations and injected context, and introduces a dreaming mechanism that reviews and consolidates during idle time, lifting Opus 5.5 from 0.655 to 0.838 on the CL-Bench database task.
On November 6, Anthropic published the research article The components of agent memory, systematically breaking down how Claude agents remember, filter, correct and organize knowledge across sessions. The study argues that AI agents can borrow the mechanism of the brain organizing memory during sleep, reflecting on historical records during idle time, correcting errors and building knowledge structures, which improves performance on long-horizon tasks.
The research divides agent memory into four layers, session transcripts, memory notes, knowledge representations such as wikis, and injected context, and introduces a dreaming mechanism that lets the model review transcripts, fix errors, distill rules and update the knowledge base during idle time between interactions. Claude marks important events or writes memory notes during a session, but not all information is worth keeping, so it has to judge what will genuinely be useful in the future. Anthropic notes that memory is essentially a prediction about the future, since recording a verified rule means the model expects to rely on it again.
On the database question-answering task in the CL-Bench continual learning benchmark, Claude models with memory performed clearly better than stateless baselines. After using memory, Claude Opus 5.5 and Sonnet 5.5 averaged scores of 0.838 and 0.833 respectively, against stateless baselines of 0.655 and 0.644. The study also found that for stronger models, letting Claude decide for itself what is worth remembering usually works better than prescribing fixed rules in advance.