Anthropic has published a detailed explanation of the components of agent memory, the knowledge Claude keeps from experience and reuses between sessions. In a post on its research site, the company describes memory as the context an agent can read later, separate from weight updates, in-conversation learning or the summaries used to free up context windows.
The team frames the problem with a simple tension: an agent must decide what is worth remembering before it knows what comes next. To handle that, Anthropic leans on a process it calls dreaming, where Claude reflects on its own logs during idle time to reorganize what it has stored, correct errors and produce new notes.
What agent memory means at Anthropic
Anthropic defines agent memory narrowly. It covers knowledge Claude gains from interactions with users and assigned tasks, saved as context it can read in a later session. It does not include changes to model weights, learning inside a single conversation, or compaction, the summaries that replace older turns so the model can keep working.
The company credits Letta and UC Berkeley researchers with proposing sleep-time compute, the idea that a model can use idle time between interactions to reorganize its existing context. Anthropic says the same approach can be applied to memory, letting an agent review logs to fix errors or generate new representations.
How dreaming and memory loops work
The design borrows from neuroscience. In the human brain, chemical signals mark certain experiences as important, and those memories are replayed more often during offline rest, helping them stabilize into long-term knowledge. Anthropic says Claude can do something similar by tagging salient events in a session log or writing them into a memory folder.
Memory, the post notes, is occasionally a prediction: writing down a verified fact assumes a future session will rely on it without checking again. Anthropic tested this with a Continual Learning Bench 1.0 task where Claude answers database questions, is told the right answer and can save a memory to help later.
- Claude Sonnet 4.6 produced scattered files full of untested guesses about why answers were wrong.
- Claude Opus 4.7 consolidated its notes into one reference document and marked checked versus unconfirmed facts.
