AI

Claude Agents Get 'Dreaming' to Clean Up Memory Between Sessions

Anthropic has introduced 'dreaming,' a memory consolidation feature for Claude Managed Agents that mimics biological REM sleep. The tool reorganizes stored knowledge, removes outdated or contradictory entries, and improves task performance by 10%. Alongside this, Anthropic has made multi-agent orchestration and outcome-guided agents generally available, expanding the capabilities of its AI coding assistants.

Anthropic has added a new memory consolidation feature called 'dreaming' to its Claude Managed Agents, enabling AI agents to clean and restructure their stored knowledge between sessions. The feature is designed to address memory rot—a common issue where agents accumulate outdated, contradictory, or irrelevant information over time. Dreaming periodically reorganizes an agent's memory, converting relative dates to absolute timestamps, merging duplicates, and deleting stale entries. Early internal testing showed a 10% improvement in task success rates and higher-quality file generation when agents used the feature.

How Dreaming Works

Dreaming functions similarly to biological REM sleep, allowing Claude agents to review past interactions and refine their memory. During active sessions, agents collect debugging patterns, architecture decisions, and user preferences. Without consolidation, these notes degrade, leading to contradictions and inefficiencies. Dreaming runs periodically after a set number of sessions, systematically cleaning and restructuring the agent's knowledge base.

Key steps in the dreaming process include:

  1. Timestamp normalization: Converting relative dates (e.g., "three days ago") to absolute timestamps.
  2. Conflict resolution: Removing facts that contradict newer or more reliable information.
  3. Deduplication: Merging identical or near-identical entries to reduce clutter.
  4. Stale entry removal: Deleting outdated or irrelevant data, such as references to deleted files.

The feature is currently available as a research preview, requiring developer access. Anthropic has not disclosed a timeline for general availability.

Broader Updates to Claude Managed Agents

Alongside dreaming, Anthropic announced two other updates to its Managed Agents platform:

  • Outcome-guided agents: Developers can now set success criteria for tasks, with a built-in grader evaluating outputs and prompting revisions until the criteria are met.
  • Multi-agent orchestration: A lead agent can delegate tasks to specialist agents working in parallel, improving efficiency for complex workflows.

Both features were previously available as research previews but are now generally accessible. Multi-agent orchestration, in particular, has been in public beta since April 8.

Why Memory Consolidation Matters

Memory rot has been a persistent challenge for AI agents that retain context across sessions. As one technical analysis noted, "The biggest problem with AI coding assistants that remember context across sessions is not forgetting. It is remembering too much of the wrong things." Over time, agent memory files become cluttered with contradictory entries, references to deleted files, and outdated information, reducing efficiency and accuracy.

Dreaming addresses this by introducing a structured consolidation process, ensuring agents retain only the most relevant and accurate information. The feature is particularly useful for long-running projects where agents accumulate large volumes of context over multiple sessions.

How to Access Dreaming

Developers interested in testing dreaming can apply for research preview access through Anthropic's developer portal. The feature is compatible with all Claude Managed Agents and requires no additional setup beyond enabling the consolidation cycle. Anthropic has not yet announced pricing for general availability.

Tradeoffs and Limitations

While dreaming improves memory quality, it introduces a few tradeoffs:

  • Latency: Consolidation cycles add processing overhead, which may slightly delay task execution during active sessions.
  • Developer control: The feature runs automatically, with limited customization options for when or how often consolidation occurs.
  • Early-stage limitations: As a research preview, dreaming may lack some polish or edge-case handling present in mature features.

Bottom Line

Dreaming is a practical solution to a long-standing problem in AI agent development. By mimicking biological memory consolidation, Anthropic has created a tool that improves agent performance without requiring manual intervention. For developers working on long-term projects, the feature could reduce the need for manual memory management and improve overall reliability. Watch for broader availability and additional customization options as the feature matures.

Similar Articles

More articles like this

AI 4 min

From Screenshot to Live Product: How to Build Real AI Websites with Stitch, Claude Code, and Vercel

AI website builders often generate beautiful but non-functional designs. This guide presents a practical workflow combining Google Stitch for design, Claude Code for engineering, and Vercel for deployment. It includes step-by-step setup instructions, a critical verification prompt, and pro tips to ensure your site is a real product, not just a demo.

AI 3 min

Google Rules Out Liquid Glass for Android—Here’s What’s Next

Google has officially denied rumors that Android will adopt Apple’s Liquid Glass design, following a brief teaser that sparked speculation. Android ecosystem president Sameer Samat and other Google representatives dismissed the idea, reaffirming the company’s commitment to Material 3 Expressive. The upcoming Android Show on May 12 is expected to focus on other features, including a rumored Pixel phone notification LED system called "Pixel Glow."

AI 3 min

Gemini for Mac to Gain Autonomous Control—Rivaling Claude’s Agent

Google is preparing to expand its Gemini macOS app with agentic capabilities, allowing the AI to autonomously control a user’s computer—clicking, typing, and organizing files. The move follows Anthropic’s Claude Cowork, which already offers similar desktop automation for subscribers. While Google has not officially confirmed the feature, a teardown of the Gemini Mac app reveals preparations for screen access and accessibility permissions. The update could arrive as soon as Google I/O 2026, aligning with the company’s broader push into agentic AI.

AI 2 min

OpenAI, PwC partner to build AI agents for CFOs

OpenAI, PwC partner to build AI agents for CFOs CFO Dive

AI 2 min

Elon Musk's SpaceX Will Help Power Anthropic's Claude in Surprise AI Deal

Elon Musk's SpaceX Will Help Power Anthropic's Claude in Surprise AI Deal Decrypt

AI 3 min

Probe finds ChatGPT's model training violated Canada's federal, provincial privacy laws

Probe finds ChatGPT's model training violated Canada's federal, provincial privacy laws IAPP