Claude's "Dreaming" Explained: What It Is, Where It Works, and How Developers Should Use It
Dreaming is a scheduled memory process in Claude Managed Agents that reviews past sessions and curates what agents remember. It is not background reasoning, and it is not an API switch.
Corrected on September 28, 2026. An earlier version of this article described dreaming as agents reasoning continuously in the background, and advised enabling extended thinking to get it. Both were wrong. This version describes the feature as Anthropic documents it.
In May 2026 Anthropic added a feature called dreaming to Claude Managed Agents, its hosted platform for running agents. In Anthropic's words, dreaming is "a scheduled process in Claude Managed Agents that reviews agent sessions and memory stores, extracts patterns, and curates memories so agents improve over time." It launched as a research preview that developers request access to.
What it actually does
Managed Agents can keep a memory store: notes an agent writes down as it works. Over time that memory gets messy, with duplicates, outdated facts and lessons scattered across sessions. Dreaming runs between sessions, on a schedule, and tidies it: it reads recent sessions and the memory store, merges and updates what is there, drops what is stale, and pulls out patterns, including learnings shared across agents. Anthropic's summary: memory captures what an agent learns as it works; dreaming refines that memory between sessions.
You choose how much control to keep: dreaming can update memory automatically, or you can review its changes before they land.
What it is not
- It is not the model thinking in the background while nobody is using it. It is a memory-maintenance job.
- It is not a setting on the ordinary Messages API. An agent you built yourself on the API does not get it by changing a parameter, and turning on extended thinking has nothing to do with it.
Alongside dreaming, Anthropic announced outcomes (a separate grader checks an agent's work against a rubric you write), multiagent orchestration (a lead agent delegating to specialists) and webhooks for finished tasks.
How developers should use it
- Decide what an agent should remember. Customer preferences and resolved edge cases, yes. Personal data it does not need, no. Memory is data you are responsible for.
- Start in review mode for client work, especially in regulated industries, so a person approves memory changes until you trust them.
- Audit memory periodically. Read what the agent now "knows" and correct anything wrong; a wrong memory repeats its mistake in every session.
- Measure with outcomes. Define a rubric for the agent's work and check whether results improve after dreaming runs.
- Not on Managed Agents? The idea still applies: a scheduled job that reviews your agent's logs and updates the instructions or knowledge it works from, with a person approving the changes.
Practical takeaway: If you run client agents that repeat mistakes or forget what they learned, a curated memory is the fix, whether through dreaming on Managed Agents or a review job of your own. Keep a person approving what gets remembered.
Source: Anthropic, "New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration" (May 2026).
- 1Request dreaming access for Claude Managed Agents to automatically curate and deduplicate your agent's memory stores over time.
- 2Use dreaming to extract patterns from agent sessions so your agents can improve decision-making without manual memory management.
- 3Enable dreaming in production agents handling repetitive tasks to reduce memory clutter and enhance long-term performance.
Ready to implement AI in your business?
Our team builds the AI systems you just read about. Start with a free 30-minute discovery meeting.