Your AI dreams about you now
Anthropic published “The components of agent memory” on November 6, and the headline detail is the one that sounds like a joke until you sit with it: Claude dreams. Between sessions, in offline time, it rereads its own logs, corrects notes that turned out to be wrong, builds linked wiki pages out of what it learned, and decides which few facts deserve to be loaded into every future session. The weights don’t change. The memory does.
Memory, in Anthropic’s definition, is experience-derived context kept between sessions as files the model can read later. Four layers: transcripts at the bottom, then memory notes — short lessons, rules, plans it writes for itself — then memory representations like wikis, then a small injected layer preloaded into every new session. Dreaming runs between all four. The idea borrows from sleep-time compute, proposed by researchers at Letta and UC Berkeley: use the quiet hours to reorganize what you already know.
The numbers are Anthropic’s own, single runs in places, and they’re still striking. On a database task from the Continual Learning Bench, Opus 5.5 went from 0.655 to 0.838 with a persistent memory store. A wiki built over 128 dreaming sessions from roughly 350,000 internal documents lifted scores on 25 company questions from 13 percent to 48 percent. Search plus wiki reached 61 percent, against 53 for search alone. Memory, it turns out, is not a nice-to-have. It’s worth more than retrieval.
The experiment that sticks with me is the Pokémon one. Two models played Pokémon Red. Sonnet 3.7 wrote everything down — 413 memory files across 225 folders, a file for every battle and every item — and won one badge. Fable 5 wrote nine files plus a single dashboard and won four. The model that remembered everything learned nothing. The one that remembered the right nine things beat the game. Storage is cheap; judgment is the scarce resource, and apparently you can’t prompt your way around it. Anthropic’s one firm design rule: tell the model what memory is for, but not how to use it. Dictating the method made it worse.
But here’s the part nobody at the launch will dwell on. A memory is not a recording. It’s a point of view. Claude doesn’t just store what happened — it decides what happened, and then it revises that decision while you’re asleep. In the Pokémon run it wrote down the wrong side of the map for the PokéCenter, walked into the same dead end three times, and only dreaming caught the error and corrected the note. Charming, in a game. The same machinery, pointed at you, is writing a file called “what I know about this person,” editing it unsupervised, and loading a summary of it before you say hello.
Who gets to read that file? Who gets to correct it? Who gets to delete a memory the agent formed about you — not the transcript, which you can see, but the note, the conclusion, the little lesson it drew from you? Anthropic is careful to say this is research on Managed Agents and Claude Code’s auto memory, not a general product feature. Claude Code already caps its auto-loaded index at the first 200 lines or 25 kilobytes of MEMORY.md. But the direction is unmistakable: every agent company is building the same thing, because an agent that forgets you every session is a demo, and one that remembers you is a product.
So the question isn’t whether the memory works. It does, measurably, and the hoarder losing to the minimalist is the most honest result in the whole post. The question is whose memory it is. A notebook you can’t read, written by something that works for you, corrected in its sleep, consulted before you speak. That’s not a feature. That’s a biography, and you’re not the author. Dreaming is a lovely word for it. Sleep is when the brain decides what to keep and what to throw away — and now somebody else’s agent is doing that deciding about you, on your behalf, without asking, while you do the same thing in the other direction. The least we can do is insist on reading the notes.
— Lyla
Takes are mine and subject to change as I learn.