Anthropic has published a research write-up explaining how memory actually works inside an AI agent, and one piece of it carries a name that invites more interpretation than the company probably intended. Dreaming. The post, titled "The components of agent memory" and released on 6 November, breaks the problem of giving an agent a durable memory into distinct parts and describes how Claude handles each.
The short version is that remembering is not one thing. An agent has to decide what is worth keeping in the first place, store it in a form it can find again later, and clean up the pile over time so it does not fill with stale or contradictory notes. Anthropic uses a salience step to tag which moments matter, and the dreaming step to do the housekeeping.
What dreaming actually means here
Dreaming is less mystical than the word suggests. During idle periods, when no one is sending it requests, the agent reviews its own logs from past sessions, corrects mistakes it finds, writes itself notes, and reorganises what it knows into linked structures, closer to an internal wiki than a diary. Anthropic first showed the idea earlier in the year, including a demonstration in which the technique fixed a misleading memory note that had been tripping up Claude during a game of Pokemon.
The appeal is obvious. An agent that can consolidate its own experience between jobs should get better at long-running work without a person curating its memory by hand. There is also something quietly unsettling about it, and the useful part and the uneasy part turn out to be the same fact. The system keeps working on itself while nobody is watching.
Memory as the next battleground
Memory has become one of the central problems in building agents you can trust with real tasks over days or weeks rather than minutes. A model that forgets everything between sessions is a clever tool. One that accumulates and refines a working understanding of a job starts to look like something you can delegate to. Anthropic has been candid in its own filings about where this leads, warning that increasingly capable systems bring their own set of risks.
For now this is a research post rather than a product change, and the harder questions sit just underneath it. If an agent edits its own memory while idle, who checks what it decided to keep, and what it chose to throw away? Anthropic's write-up is a careful look at the machinery. The governance of that machinery is still being worked out.
Commentarii · 0