Anthropic introduced a new feature called dreaming for its Claude Managed Agents on May 6, 2026, during the Code with Claude conference in San Francisco. This capability allows AI agents to asynchronously review past session transcripts to reorganize their memory, merge duplicate notes, and identify recurring patterns or mistakes.
The dreaming process functions in the background without modifying the underlying model weights of the AI. Instead, the system generates plain-text notes and structured playbooks. Users maintain control over the process, with the option to require manual approval before any memory updates are applied to the agent's store.
During early testing with the legal AI firm Harvey, the application of the dreaming feature reportedly increased task completion rates on legal benchmarks by sixfold. The system is capable of processing between one and 100 session transcripts in a single run while keeping original input data intact.
The feature is currently available in a research preview for three models: claude-opus-4-8, claude-opus-4-7, and claude-sonnet-4-6. Access to the tool is restricted, and organizations must submit a request to use it, as it is not yet available for general deployment.
