At its Code with Claude developer conference, Anthropic unveiled a research preview feature called 'dreaming' for Claude Managed Agents. This process involves reviewing recent sessions and selectively storing critical information in memory to improve future task performance. The feature addresses context window limitations in large language models by curating relevant memories, similar to human episodic recall.

Managed Agents—pre-built agent harnesses for complex multi-agent workflows—now gain this capability, though dreaming remains limited to Pro and Max subscribers with doubled 5-hour usage limits.