Claude can now 'dream' as Anthropic races to build ‘self improving’ AI agents

Anthropic has added a new feature for Claude which gives the AI chatbot human like ability. The new feature called ‘Dreams’ was introduced by the company during its Code with Claude developers’ conference.

But why do agents need to ‘dream’? Anthropic notes in a suport page that as AI agents work alongside users over time, they continuously write information to their memory stores. However, the company says these ‘writes’ are ‘local and incremental’ and the memory stores can accumulate duplicates and can become messy over multiple sessions.

“Dreaming is a scheduled process that reviews your agent sessions and memory stores, extracts patterns, and curates memories so your agents improve over time. You decide how much control you want: dreaming can update memory automatically, or you can review changes before they land.” Anthropic wrote in a blogpost

Anthropic also notes that the original memory story is never modified and Claude instead creates a separate output memory store that developers can review, keep or discard.

“Dreaming surfaces patterns that a single agent can’t see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across a team. It also restructures memory so it stays high-signal as it evolves. This is especially useful for long-running work and multiagent orchestration.” the company added in a blogpost

Anthropic notes that Dreams can ‘tens of minutes’ torun depending on the size of the data. Developers can also provide custom instructions during the dreaming process, such as asking Claude to focus only on coding preferences while ignoring temporary debugging conversations.

Dreams currently supports only Claude Opus 4.7 and Claude Sonnet 4.6 models and is available through special beta headers for developers running the company's Managed Agents platform.

Anthropic also states that dreams can fail for several reasons such as oversized memory stores, unavailable sessions or runtime timeouts.

The AI startup noted that Dreams will be billed at the standard API token rates and costs will increase based on the number and size of the sessions being analyzed.

Anthropic also shared the example of legal AI company Harvey that used dreaming to help their agents filetype workarounds and tool-specific patterns across long-form drafting sessions, that resulted in a 6 times increase in task completion rates.

Anthropic's push towards ‘self improving AI agents’:

Anthropic says Dreams is part of a broader push towards ‘self-improving’ AI agents. Alongside the feature, the company also introduced a new ‘Outcomes’ feature, which allows AI agents to evaluate their own work against a predefined metric and retry tasks if the result doesn’t meet expectations

“With outcomes, you write a rubric describing what success looks like and the agent works toward it. A separate grader evaluates the output against your criteria in its own context window, so it isn't influenced by the agent's reasoning. When something isn't right, the grader pinpoints what needs to change and the agent takes another pass.” Anthropic explained