Claude Managed Agents: Dreaming — Agents That Learn From Past Sessions

What Is Dreaming?

Dreaming is a new research-preview capability in Claude Managed Agents that introduces scheduled background processing for agent self-improvement. Rather than relying on what an individual agent can observe within a single session, Dreaming reviews the accumulated history of past sessions and memory stores to extract patterns, refine stored knowledge, and curate higher-quality memories for future use.

How It Works

Dreaming runs as a scheduled background process — not during active task execution, but between sessions. During each Dreaming cycle, the system reviews the agent's past interactions and memory contents, looking for patterns that span multiple sessions: recurring mistakes, converged workflow strategies, and team preferences that emerge only when viewing the full history rather than any single run.

The output of a Dreaming cycle is updated memory content — restructured to maintain high signal as it grows and evolves over time. Developers have control over how involved this process is: Dreaming can be configured to apply memory updates automatically, or to surface proposed changes for human review before they are committed.

Dreaming works across both single-agent and multiagent systems. In multiagent contexts, it is especially powerful because it can pull shared learnings across the outputs of multiple agents, synthesizing cross-agent patterns that no individual subagent could observe on its own.

Why It Matters

Most agent systems are stateless between sessions — each run begins from the same baseline, regardless of what succeeded or failed in the past. Dreaming changes this by enabling genuine continuous improvement. Agents that run Dreaming accumulate institutional knowledge over time, becoming progressively more effective at their specific tasks without requiring manual intervention from developers.

Harvey, a legal AI platform, demonstrates the impact in production. Their agents handle drafting and document creation, and Dreaming enables those agents to remember filetype workarounds and tool-specific patterns across sessions. The result: completion rates went up approximately 6x in Harvey's internal tests, a direct consequence of agents learning from accumulated experience rather than starting fresh each time.

Status and Availability

Dreaming is currently in research preview. Developers interested in early access can request it at claude.com/form/claude-managed-agents. Documentation is available at platform.claude.com/docs/en/managed-agents/overview.