Anthropic research details how Claude AI agents can improve intelligence through simulated memory consolidation

Anthropic released a study, "The components of agent memory," detailing how Claude agents manage cross-session memory, knowledge refinement, and error correction. The research proposes AI agents can emulate human memory consolidation during rest by reviewing past logs to boost long-term task performance. The system utilizes four memory layers: session logs, memory notes, knowledge representations (like a Wiki), and injected context. A 'dreaming' mechanism allows the model to review logs between interactions to refine rules and update its knowledge base. Empirical results showed marked improvements; in CL-Bench Q&A tasks, Claude Opus 5.5 and Sonnet 5.5 scored 0.838 and 0.833, respectively, versus 0.655 and 0.644 for non-memory baselines. In a Pokémon Red test, Claude Fable 5 needed only 10 memory files for 4 badges, while Sonnet 3.7 required 413 files for just 1 badge. Furthermore, testing across 350,000 corporate documents showed that scaling the 'dreaming' process from 2 to 128 sessions increased the success rate for 25 questions from 13% to 48%, with Wiki integration boosting the score to 61%.