Anthropic's Claude Managed Agents Can Now 'Dream' to Store Memories
During the Code with Claude developers' conference held in San Francisco, Anthropic unveiled a new capability named 'dreaming' for its Claude Managed Agents. This feature involves a scheduled review process where the agent examines recent occurrences and selects particular memories for future reference. Currently, dreaming is in a research preview phase and is exclusive to Managed Agents on the Claude Platform. These Managed Agents serve as a more advanced option compared to the Messages API, facilitating multi-agent tasks that can extend over several minutes or hours. Dreaming aims to overcome the limitations of context windows in LLMs, ensuring that crucial information is not lost during extended projects, akin to compaction in chat models that filters out unnecessary details.
Key facts
- Anthropic introduced 'dreaming' for Claude Managed Agents at the Code with Claude conference in San Francisco.
- Dreaming is a scheduled process that reviews sessions and memory stores to curate specific memories.
- The feature is currently in research preview and limited to Managed Agents on the Claude Platform.
- Managed Agents are a pre-built, configurable agent harness for multi-agent tasks over minutes or hours.
- Dreaming addresses the limitation of context windows in LLMs, preventing information loss.
- Compaction is a similar process used in chat models to remove irrelevant information from context windows.
- The conference was held in San Francisco.
- Anthropic describes dreaming as a way to store important information for future tasks and interactions.
Entities
Institutions
- Anthropic
- Claude Platform
Locations
- San Francisco
- United States