The components of agent memory

Anthropic said it has developed a “dreaming” technique that lets its Claude AI agents reorganize and consolidate memory during idle periods between user interactions, drawing on a sleep-time compute approach proposed by researchers at Letta and UC Berkeley. The company said agents can tag salient events in session logs as they occur or write them to a memory store of notes, treating memory as a prediction about what future sessions will need. Anthropic defines agent memory as knowledge retained across sessions, distinct from weight changes, in-context learning and compaction. The work matters because persistent memory is a key bottleneck for deploying agents on long-running tasks, and the evaluations Anthropic outlined are an early attempt to measure what each memory component contributes. Watch for whether dreaming improves agent performance on multi-session benchmarks and how rivals such as OpenAI and Google position their own memory architectures.