Anthropic Adds ‘Dreaming’ to Let Claude Agents Learn Across Sessions
SAN FRANCISCO — Anthropic is rolling out a new capability called “Dreaming” for its Claude Managed Agents platform, designed to let AI agents asynchronously review their past sessions and learn from mistakes, the company announced this week.
The feature works by reviewing completed agent sessions after the fact, consolidating memory entries, removing duplicates and outdated information, and distilling new insights that can be applied to future tasks, according to a report by The Decoder. The process runs asynchronously, meaning it operates in the background rather than during active agent sessions.
The name evokes parallels to biological sleep, during which the human brain consolidates memories and processes information.
Dreaming represents Anthropic’s latest push into cross-session learning for agentic AI systems. Most AI agents today operate within isolated sessions, losing context and accumulated knowledge once a task concludes. By enabling agents to carry forward refined insights from prior interactions, Anthropic said it aims to make its managed agents more capable over time without requiring users to manually curate agent memory.
The announcement comes alongside two additional features now entering public beta: Outcomes and Multiagent Orchestration.
Outcomes appears designed to give developers structured ways to evaluate and measure agent performance, while Multiagent Orchestration addresses the demand for systems in which multiple AI agents coordinate on complex tasks.
Anthropic, headquartered in San Francisco, has positioned Claude Managed Agents as an enterprise offering alongside agent platforms from OpenAI, Google DeepMind, and other AI providers.
