Claude Managed Agents Gain 'Dreaming' Feature to Preserve Memories Across Sessions
Key Takeaways
- ▸Anthropic introduced 'dreaming,' a new feature for Claude Managed Agents that preserves memories across sessions by identifying patterns and important information
- ▸Unlike existing compaction processes, dreaming works across multiple agents and sessions, surfacing insights no single agent could identify independently
- ▸The feature is in research preview with limited availability; developers can request access and choose between automatic or manual memory curation
Summary
At its Code with Claude developers' conference, Anthropic announced a new feature called "dreaming" for Claude Managed Agents. Dreaming is a scheduled process that reviews recent sessions and memory stores to identify and preserve important information for future tasks and interactions. Unlike traditional compaction processes that work within a single conversation, dreaming analyzes patterns across multiple agents and sessions, identifying recurring mistakes, converging workflows, and team preferences.
The dreaming process is currently available as a research preview and is limited to Managed Agents on the Claude Platform. Developers can request access to the feature, which addresses a key challenge in long-running projects: the limited context windows of large language models mean important information can be lost over time. Users will have the option to choose between an automatic process or manually reviewing changes to memory.
Alongside the dreaming announcement, Anthropic revealed that two previously announced research preview features—outcomes and multiagent orchestration—are now more widely available. The company also responded to user demand by doubling rate limits for subscribers to its Pro and Max subscription plans, addressing infrastructure constraints that have frustrated users.
- Anthropic also made outcomes and multiagent orchestration features more widely available and doubled rate limits for Pro/Max subscribers
