Anthropic is giving Claude a more deliberate way to improve by letting it revisit what it has already done. The new Dreaming feature is designed to help the AI system sift through past sessions, reorganize memory, and surface patterns that can make future work cleaner and more consistent.
The update matters because Anthropic is not simply trying to expand storage. The company wants agents built on its platform to recognize repeated mistakes, refine behavior over time, and turn accumulated experience into something more useful for the next task.
How Dreaming works
Dreaming is built for Claude Managed Agents, the platform developers use to build and deploy AI agents through an API for production-scale tasks. Anthropic positions that platform as a way to speed up launches by handling much of the backend deployment work.
Within that setup, Dreaming runs on a scheduled, asynchronous basis. Claude reviews older sessions and its memory store, then removes duplicated, outdated, or conflicting information that has built up during normal use.
According to Anthropic, agent memory can become cluttered as more information is written over time. Repeated entries, contradictions, and irrelevant details can pile up until the memory store becomes harder to use effectively.
What Claude actually produces
One Dreaming run can examine transcripts from as many as 100 previous sessions together with an existing memory store. From there, Claude generates a new memory version by extracting patterns and identifying insights that might not be obvious if each interaction were reviewed separately.
The original memory store is not overwritten. Instead, Claude creates a separate output memory store that developers can review, keep, or discard.
That approach gives developers more control over how learning is applied. The “dream” does not automatically replace earlier data, so the new memory can be checked before it is used in later workflows.
Anthropic also says the process can be guided with custom instructions. For example, Claude can be told to focus on programming preferences while setting aside temporary debugging conversations.
Why Anthropic is adding it
The company says the feature is meant to help agents identify mistakes that keep happening. It is also intended to spot the workflows that consistently become the final choice and to understand shared preferences within a team.
That makes Dreaming especially relevant for long-running projects. It also has a role in multi-agent systems, where several AI agents work together on connected tasks.
In that setting, the goal is not just to preserve history. Anthropic wants the accumulated record to become a more structured source of decision-making for future work.
Outcomes expands the same idea
Alongside Dreaming, Anthropic is also expanding another feature called Outcomes. This feature lets an AI agent judge its own work against a set goal and try again if the result falls short.
Anthropic says a separate scoring system evaluates the output independently. That is meant to keep the review process from being influenced by the way the agent reasoned while completing the task.
Together, Dreaming and Outcomes show a broader push toward internal reflection in Claude. One feature focuses on memory and pattern recognition, while the other focuses on whether the work itself meets the intended target.
Still limited for now
Dreaming currently supports only Claude Opus 4.7 and Claude Sonnet 4.6. Anthropic is offering it as a research preview for developers using special beta headers on Managed Agents.
Access is not open to everyone yet. Developers need to request access before they can try it.
Anthropic also warns that the process is not always smooth. It can fail if the memory store is too large, if sessions are unavailable, or if the process runs out of time.
Processing is not instant either. The company says it can take tens of minutes depending on how much data is being analyzed, and pricing follows standard API token rates based on the number of sessions and the size of the data involved.
With Dreaming and Outcomes, Anthropic is pushing Claude toward a more self-correcting model of work. The emphasis is no longer only on faster responses or larger context, but on the ability to revisit experience, detect patterns, and improve step by step.
