Anthropic released a new “Dreaming” capability for managed agents this week.

Dreaming is a scheduled review process that analyzes past agent sessions and memory to identify patterns (things like recurring mistakes, converged workflows, team preferences). It restructures memory to keep it high-signal as agents accumulate experience, enabling self-improvement across sessions.

You control whether dreaming updates memory automatically or requires your approval first.

Catch up on this and more AI updates for product teams on YouTube: youtube.com/watch?v=_6a…