Anthropic introduces dreaming for agents

AI agents are good at doing one session of work. They are much worse at staying sharp across many sessions without turning their own memory into a junk drawer. That is the gap Anthropic is trying to close with “dreaming,” a new Claude Managed Agents feature it showed on May 6 in San Francisco. Basically, the agent pauses between jobs, re-reads what happened, and rewrites its memory into something cleaner and more useful. (platform.claude.com) ### What is “dreaming,” actually? It is not a model free-associating in the background. It is an asynchronous job in Anthropic’s Managed Agents system that takes an existing memory store, optionally looks through past session transcripts, and produces a new memory store with duplicates merged, stale entries replaced, and fresh patterns surfaced. Anthropic has it in research preview, not general release. (platform.claude.com) ### Why do agents need this? Because long-running agents accumulate messy state. They write little notes to themselves as they work, but over time those notes can conflict, repeat, or stick around after they stop being true. Anthropic’s own docs frame dreaming as cleanup plus synthesis — less like adding raw memory, more like taking a scattered notebook and turning it into an edited briefing. (platform.claude.com) ### What does it do between sessions? The key move is that dreaming happens after work, not during the main task loop. That matters because the agent can step back and look for patterns across sessions instead of only reacting in the moment. Anthropic says it can update preference files and other context, which means the next run starts with a tidier understanding of how the user likes things done and what has already been learned. (platform.claude.com) ### How much history can it use? Up to 100 sessions in a single dream run. Anthropic’s docs also say the output memory store is separate from the input store, so developers can inspect the rewritten version and throw it away if it looks wrong. That separation is a quiet but important safeguard — the system is editing memory, but not destructively. (platform.claude.com)Managed Agents? Because Managed Agents is Anthropic’s bigger pitch to developers: stop hand-building all the brittle scaffolding around long-horizon agents and let Anthropic host the harness. The company has been arguing for weeks that agent infrastructure should survive model changes, and dreaming fits that story. If the harness handles persistence, (platform.claude.com)o babysit. (anthropic.com) ### Is this only about memory? No — it is also about reliability. Anthropic announced dreaming alongside broader Managed Agents upgrades, including outcomes and multiagent orchestration. So the idea is not just “remember more.” It is “run longer, delegate better, and improve over time without a human constantly re-tuning the workflow.” That is much closer to a production agent story than a chatbot feature story. (claude.com) ### What is the catch? The catch is that memory editing can help and still go wrong. A system that merges duplicates and replaces stale facts is making judgment calls about what matters now. Anthropic seems aware of that — hence research preview status, explicit developer instructions, limited model support, and the choice to generate a separate output store instead of silently overwriting the old one. (platform.cl([claude.com)ocs/en/managed-agents/dreams)) ### Why does this matter now? Because AI companies are racing to make agents feel less like demos and more like coworkers that can pick up where they left off. Anthropic is leaning especially hard into enterprise use cases — coding, finance, and other long-running work where context persistence is the whole game. Dreaming is a small weird-sounding feature name for a very pr(platform.claude.com) time. (tech.yahoo.com)