Memory and dreaming for self learning agents

Anthropic’s new "Memory" and "Dreaming" features for Claude Managed Agents enable autonomous systems to learn and improve across tasks, sessions, and multi-agent environments. Memory functions as a persistent, file-system-based storage layer that allows agents to carry forward strategies and avoid past mistakes. Complementing this, Dreaming acts as an out-of-band, decoupled feedback loop that periodically analyzes session transcripts to identify patterns and inefficiencies. By automatically curating and optimizing memory, Dreaming enables continuous self-learning, which has demonstrated significant gains in production scenarios, such as a sixfold increase in legal benchmark completion rates. This architecture shifts agent development from manual infrastructure management to a model where agents leverage shared, evolving knowledge to achieve higher-level, long-horizon objectives, effectively raising the floor for overall system performance.