Claude gets a "dreaming" feature that learns from earlier tasks

What is new

Anthropic, the company behind the AI assistant Claude, has introduced a feature called "dreaming". It was announced on May 6, 2026. The name comes from the brain: when we sleep, we sort the day's experiences and store the most important things. Claude does something similar.

The feature is a scheduled task that runs between Claude's work sessions. It reviews what an AI agent has done and the memory the agent has stored. It then finds patterns and cleans up. Agent here means a Claude that carries out tasks on its own over time, such as processing documents or code.

It is important to know that the feature is part of Claude Managed Agents, which is a tool for developers and businesses. So it is not a button in the regular Claude chat. The feature is in "research preview", so only selected developers who have applied for access can try it.

What is clever

Normally, an AI forgets most things between tasks, or else its memory piles up with duplicates and outdated notes. Dreaming sorts through it. It can detect mistakes the agent repeats, workflows it ends up using, and preferences that recur across an entire team of agents. Then the next task can start smarter.

The developer decides how much control there should be. Memory updates can happen automatically, or they can be reviewed and approved before they take effect. That is an important safeguard, because you can then see what the AI has "learned".

At launch, Anthropic also mentions two other new features in the same tool, both in public beta. One is called Outcomes: you describe what a good task looks like, and a separate grader checks the result. Anthropic says it gives up to 10 percentage points better success than ordinary prompting. The other is multiple agents working together: a lead agent distributes the work to specialists who work simultaneously.

Cheaper or better?

The price has not been disclosed. Neither Anthropic nor the media we have seen mention what dreaming costs, and nothing has been said about a monthly price for ordinary users or a price per million tokens (pieces of text). Nothing has been said about availability in Denmark or the EU either.

As for whether it is better, we have only one concrete figure, and it comes from those involved. Legal firm Harvey says its agents' completion rate rose about six times in tests after it started using dreaming. That is a single customer example and not an independent measurement, so the figure should be taken with caution. We have not found a direct comparison with competitors' equivalent features at OpenAI or Google.

What it is good at

The feature is made for long-running work tasks, where agents work over weeks or collaborate with other agents. There, it is useful that yesterday's experience helps today, without anyone having to manually adjust the instructions every time.

On the other hand, it is not something ordinary Claude users can turn on right now. If your company has developers building their own AI agents, it is worth keeping an eye on. If you are an ordinary user, for now you just need to know that AI helpers will eventually get better at remembering what works.