Anthropic explains how agent memory is organised and repaired
Anthropic's new engineering article examines stored context and the offline work needed to keep it useful.
Giving an agent a memory also gives it something to tidy up.
What happened
Anthropic published an engineering article on 6 November explaining components of agent memory, including notes, wikis and session transcripts. It describes an offline process called dreaming, which reviews past interactions and reorganises or corrects stored notes.
Why it matters
For an agent working across sessions, retrieving more information is only part of the problem. Stale notes, mistakes and contradictions can make remembered context less useful. Memory therefore needs maintenance as well as storage.
The bigger picture
This is a new technical explanation, rather than the first launch of dreaming, which had already been introduced. It also concerns information stored for future use, rather than a claim that the underlying model continually retrains its weights. Developers can examine the design choices without assuming universal product availability or guaranteed improvements.