Claude adds “dreaming” offline memory‑consolidation and doubles context limits

Anthropic is trying to fix two very different AI problems at once. One lives inside the model — agents forget badly, or remember too much junk. The other lives outside it — even good models feel cramped when compute is scarce and usage caps bite. On Wednesday, May 6, Anthropic announced a new memory feature called “dreaming” for Claude Managed Agents and, separately, a SpaceX compute deal that it says will raise Claude usage limits starting now. (platform.claude.com) ### What is “dreaming” here? Not consciousness. Not a hidden background mind. Basically, it is an offline cleanup pass for an agent’s memory. Anthropic’s docs describe it as a process that reads an existing memory store plus past session transcripts, then produces a new memory store with duplicates merged, stale facts replaced, contradictions resolved, (platform.claude.com)inspect the rewritten version and keep it or throw it away. (platform.claude.com) ### Why did Anthropic need this? Because persistent memory gets messy fast. Managed Agents can already carry information across sessions using memory stores — little text files that hold preferences, conventions, prior mistakes, and project context. But the normal write path is incremental. An agent keeps adding notes as it works. Over time, that turns(platform.claude.com)es. Dreaming is Anthropic’s answer to that sprawl. (platform.claude.com) ### How does it actually run? As an asynchronous job. Anthropic says a dream can take an existing memory store and optionally up to 100 prior sessions, then run for minutes to tens of minutes depending on input size. In research preview, it supports Claude Opus 4.7 and Claude Sonnet 4.6, and it requires a separate beta header on top of the Managed Agen(platform.claude.com)is still gated and experimental. (platform.claude.com) ### Why does that matter for agents? Because agent failures are often memory failures wearing a different costume. An agent loops because it remembers the wrong intermediate rule. It gives inconsistent answers because yesterday’s preference and today’s preference both survived. It drags irrelevant context forward because nothing cleaned house. Dreaming(platform.claude.com)nt behavior less brittle by compressing memory into something more coherent. That is the real promise here. (platform.claude.com) ### What changed on the compute side? Anthropic also said it signed an agreement with SpaceX to use all of the compute capacity at the Colossus 1 data center in Memphis. The company says that adds more than 300 megawatts of capacity — over 220,000 Nvidia GPUs — within the month. Anthropic tied that capacity directly to product changes effective May 6: (platform.claude.com)terprise plans, removing peak-hours limit reductions for Pro and Max, and raising API rate limits for Claude Opus models. (anthropic.com) ### Why is SpaceX part of this story? Because compute is now strategy, not plumbing. The odd wrinkle is that SpaceX is tied to Elon Musk, who also merged SpaceX with rival AI company xAI earlier this year. So Anthropic is effectively buying room to grow from infrastructure linked to a competitor. That tells you how tight the market still is — if a major lab can get 300 megawatts and 220,000 GPUs quickly, it takes the deal. (cnbc.com) ### So what is the bigger pattern? Anthropic is moving on both software efficiency and raw capacity. Dreaming makes agent memory denser and cleaner. The SpaceX deal makes Claude less rationed. Put those together and the company is trying to make agents that can both remember more usefully and run more often. In plain English — Claude is getting less forgetful and less throttled at the same time. (platform.claude.com) ### Bottom line? This is not one flashy feature. It is infrastructure plus behavior. Anthropic is betting that better AI agents need both — cleaner memory after the session, and more compute during the next one. (platform.claude.com)