Anthropic has unveiled a new feature called “dreaming” for its Claude Managed Agents. This development was announced during the company’s Code with Claude conference. The goal is to help AI agents better remember past interactions and improve their performance over long tasks or complex projects.

What is Dreaming and How Does It Work?

Dreaming allows Claude Managed Agents to review recent events and identify important information to store in their memory. This process occurs on a schedule and helps the AI decide what details are worth remembering for future interactions. It’s like giving the AI a way to reflect on what it has done and learn from it.

This feature is still in research preview and is limited to users of the Claude Platform’s Managed Agents. These agents are a higher-level alternative to the standard messaging API. They are designed to run in managed infrastructure and are suitable for tasks that span several minutes or hours, often involving multiple agents working together.

The Importance of Dreaming for Long-Term AI Projects

One of the main benefits of dreaming is overcoming the limited context window of large language models. Normally, these models can only remember a certain amount of information at once, which can cause important details to be lost over time. Dreaming helps by reviewing past sessions and curating memories to keep the most relevant information accessible.

While many chat models use a process called compaction—analyzing conversations and removing less important data—this is usually limited to a single, ongoing chat. Dreaming, on the other hand, is a recurring process that examines multiple sessions across different agents. It identifies recurring mistakes, shared workflows, and team preferences to improve overall efficiency and accuracy.

Users can choose whether dreaming happens automatically or review the memory updates themselves. According to Anthropic, this process uncovers patterns that individual agents might miss, especially in long-running projects or multi-agent setups. It helps keep the AI’s memory focused on high-signal, important information as the project evolves.

Additional Updates and Improvements

Aside from dreaming, Anthropic announced that two other features—outcomes and multi-agent orchestration—are now more widely available. These features were previously in research preview and are now accessible to more developers. The company also announced that subscription plans for Claude, specifically Pro and Max, will see their five-hour usage limits doubled. This move responds to user frustrations as the infrastructure struggles to meet demand.

Overall, these updates aim to improve the AI’s ability to handle complex, long-term tasks and make multi-agent collaboration more effective. Dreaming is still in its early stages, but it shows promise for making AI agents smarter and more adaptable over time.

With these advancements, Anthropic continues to develop its Claude platform as a powerful tool for AI-driven projects that require memory management and coordinated efforts across multiple agents.

Inspired by

• https://arstechnica.com/ai/2026/05/anthropics-claude-can-now-dream-sort-of/