Anthropic Claude adds 'Dreaming' memory

AI agents are good at a burst of work. They are much worse at remembering what happened last Tuesday. That gap matters if you want an agent to handle projects that last hours or days instead of one prompt-response cycle. Anthropic’s new “dreaming” feature is basically an attempt to fix that. On May 6, the company added it to Claude Managed Agents as a research preview, alongside outcomes, multiagent orchestration, and webhooks. ### What is “dreaming,” exactly? It is not a model running in its sleep or inventing new ideas from scratch. Anthropic describes dreaming as a scheduled background process that reviews past agent sessions and memory stores, extracts patterns, and curates memory so future runs go better. The point is to improve the memory layer between sessions, not just stuff more text into the next prompt. (claude.com) ### Why does an agent need that? Because long-running agents keep hitting the same dumb wall — context windows end, sessions reset, and the next run can start half-blind. Anthropic has been pretty open about this problem in earlier engineering posts. It described agents that lose track of unfinished work, guess what happened in prior sessions, or even decide a project is done when it is not. Dreaming is a productized answer to that failure mode. (claude.com) ### So what changes in practice? Instead of relying only on whatever notes an agent happened to leave itself, dreaming goes back over prior work and tries to pull out the durable bits — repeated mistakes, useful workflows, and shared preferences across a team. Anthropic says developers can let those memory updates land automatically or review them before they stick. That matters for enterprise use, because memory is helpful only if it stays high-signal instead of turning into a junk drawer. (anthropic.com) ### Why launch it with “outcomes” too? Because memory alone does not tell an agent what “good” looks like. Outcomes lets developers write a rubric for success, then a separate grader checks the output in its own context window and sends the agent back for another pass if needed. Anthropic says this improved task success by up to 10 points in internal testing, with +8.4% on docx generation and +10.1% on pptx. Dreaming and outcomes fit together — one helps the agent remember, the other helps it judge. (claude.com) ### Where does Managed Agents fit in? Managed Agents is Anthropic’s hosted system for long-horizon work. The company launched it in April 2026 as a way to run agents through more stable interfaces while Anthropic keeps changing the harness underneath. That design matters here, because dreaming is not just a model trick. It is infrastructure around the model — session logs, memory stores, graders, orchestration, and background processes. (claude.com) ### Is this really “learning”? Sort of, but in the narrow product sense. The model is not being retrained every time it finishes a task. What changes is the agent’s working memory and the quality of the instructions and patterns it carries forward. Think of it less like teaching the brain new weights and more like cleaning up the notebook between shifts so the next worker does not repeat the same mistakes. That framing matches Anthropic’s own description of memory curation between sessions. (anthropic.com) ### What’s the catch? The catch is that memory can drift, overfit, or accumulate bad assumptions if the review loop is weak. Anthropic is clearly aware of that, which is why it gives developers control over whether dream-generated memory changes happen automatically or with review. And because dreaming is still labeled a research preview, this is closer to an early enterprise capability than a solved science. (claude.com) ### Bottom line? Anthropic is trying to make agent memory an actual product surface instead of a hack developers bolt on themselves. That is the real news here. If it works, Claude agents stop acting like talented amnesiacs and start looking a lot more usable for production workflows that stretch across many sessions. (claude.com)