At the May 6 Code with Claude developer conference, Anthropic announced a new mechanism called Dreaming for Claude Managed Agents. This feature is still in research preview, designed for enterprise scenarios requiring long-running, multi-agent collaboration.
So-called dreaming, in essence, lets AI review recent sessions and memory storage when idle, screening out information worth long-term retention. Unlike traditional conversation compaction, which is limited to context management within a single session, Dreaming can identify patterns across sessions and agents, distilling scattered interaction experience into structured memory.
Anthropic notes that Dreaming can discover patterns that single agents cannot perceive — including recurring errors, converging workflow preferences, and behavior patterns shared across teams. This has direct value for long-term maintenance and knowledge transfer in complex projects.
From a technical perspective, this design aligns with Meta's memory mechanisms and OpenAI's memory features, but emphasizes collaborative memory sharing in multi-agent scenarios. Anthropic positions it as infrastructure upgrade for long-horizon work, not a simple feature overlay.
Personal view: agent memory is one of the main attack directions for vendors in the past two years, but most implementations still stay at remembering the current conversation. Cross-session, cross-agent persistent memory is the truly engineering challenge. If Anthropic can productize this capability, it would be an important step for enterprise AI application landing.