Claude Dreaming: The Memory Feature That Confesses Agents Reset
Anthropic announced Claude Dreaming on May 6.
The product pitch: agents that learn. Background memory consolidation. Review past sessions while idle. Remove duplicates. Surface patterns. The AI equivalent of REM sleep.
The reason Dreaming exists is the part they did not put in the headline.
Claude agents start every session from zero. No memory of the last run. No recollection of the workaround that worked last week. No trace of the constraint that failed the last seven times. You brief the agent. It performs. Tomorrow, you brief it again. Identical starting point. Same context rebuild. The same invisible tax on every single interaction.
That is the default behaviour. That was always the default behaviour.
Dreaming is the patch. It runs in the background at standard API rates. Access is gated behind a request form. You have to ask Anthropic for permission to let your agent remember what it already learned.
Harvey, a legal AI company, saw a 6x jump in task completion after enabling it. The same legal drafts that failed repeatedly because Claude forgot file quirks between sessions suddenly completed. Six times the completion rate. Same agent. Same tasks. The only difference: it could finally remember what it had already done.
Anthropic is framing this as a new capability.
It is a correction.
They built Dreaming because agents do not compound knowledge by default. The compounding was always the promise. Dreaming is the admission that it was never in the product.
The deeper point is not about Anthropic.
Every AI agent has this problem. The context window closes. The session ends. Whatever the agent learned about your specific files, your specific quirks, your specific failure modes is gone. The next session starts cold. You can throw money at the problem (Dreaming is not free) or you can throw human time at it (brief the agent again, every time). Neither option scales.
The 6x improvement at Harvey is not a product achievement. It is a measurement of how much was being lost without it.
Memory is not a feature that gets added later. It is the foundation. When it is missing, you are not running an intelligent agent. You are running an intelligent amnesiac. It is capable, until it forgets everything it just learned.
The headline was: agents that learn.
The subtext was: they were not learning before.
Context Windows Close. AI Forgets Everything. Your Work Should Never Start From Zero.
MrAgentˣ is in private beta. Limited to the first 1,000 until launch.
Join the waitlist