Anthropic adds “dreaming” memory feature to Claude Managed Agents

In short: Anthropic is testing a new “dreaming” feature that helps Claude Managed Agents review past work and save useful details for future tasks.

What happened

Anthropic announced the feature at its Code with Claude developers’ conference in San Francisco. The company says “dreaming” means Claude looks back at recent sessions and picks out specific things worth saving in “memory” so it can use them later.

This matters because AI chat systems have a limited “context window,” which is like a short-term scratchpad where they keep the most recent parts of a conversation. On long projects, important details can fall out of that scratchpad. Many systems try to fix this by summarizing and trimming a single conversation, but Anthropic says dreaming can review past sessions and memory across multiple agents.

Dreaming is in a research preview and only works with Claude Managed Agents on the Claude Platform. Managed Agents are Anthropic’s more packaged option for developers who want several AI workers to collaborate on a task over minutes or hours, instead of building directly on a lower-level API (an API is a set of rules for software to talk to other software).

Anthropic says users can let the memory updates happen automatically, or review and approve the changes themselves. The company also said two other preview features, outcomes and multi-agent orchestration (coordinating several agents like a project manager), are now available to more developers.

Why it matters

If you use AI tools for longer, multi-step work, the biggest frustration is often that the system forgets earlier decisions and repeats mistakes. Dreaming is Anthropic’s attempt to make these agents better at keeping the right notes, like a team that writes down lessons learned after each meeting. Anthropic also said it will double rate limits for Pro and Max subscribers, including users of tools like Claude Code, which should reduce slowdowns when demand is high.

Source: Arstechnica