Claude Dreams API: Consolidate Agent Memory Overnight

On November 6 Anthropic published The components of agent memory, and the word everyone latched onto was dreaming. The idea is simple to state: an agent writes notes while it works, and later, offline, a model re-reads the transcripts and the notes and cleans them up. It fixes notes that were wrong, merges duplicates, and promotes the durable ones. The numbers in the post got people talking. On Pokémon Red, a model that dumped 413 files (about 978,000 characters) into memory got one badge, while one that kept nine small files got four. On a 350,000-file document set, wiki scores rose from 13% to 48% as dreaming scaled from 2 to 128 sessions.

This guide is about the part you can actually run: the Dreams API in Claude Managed Agents. You will seed a memory store, point a dream at it plus a batch of past sessions, poll the job, and then, the step most write-ups skip, audit what the dream changed before any agent is allowed to read it. By the end you will have a small nightly job and a review gate you can drop into your own agent stack.