Claude Managed Agents get a dreaming mode to refine memory
Anthropic documents dreaming, an idle-time step where Claude rewrites its own memory notes, and shows it lifting Opus 5.5 from 0.655 to 0.838 on CL-Bench.
Anthropic published an engineering post on 6 November describing dreaming, an idle-time process where Claude reviews its own transcripts to fix mistakes and reorganise its memory notes. It borrows the idea from “sleep-time compute” research at Letta and UC Berkeley. This is not a toggle inside Claude.ai: it is a documented workflow on the Claude Managed Agents platform, backed by a public cookbook and a GitHub quickstart for building a memory directory and a searchable wiki.
The numbers are specific. On the CL-Bench 1.0 database task, giving Opus 5.5 a memory store lifted its score from 0.655 to 0.838. Letting Claude dream over more sessions while building a wiki over a roughly 420-million-token internal corpus raised its score on a fixed set of questions from 13% to 48% as session count went from 2 to 128. In Harvey’s legal-brief benchmark, consulting the wiki scored 87% against 73% for search alone.
Why it matters: if you’re building an agent that needs to remember across sessions rather than re-read everything each time, this is a concrete, published architecture to start from rather than a research curiosity. The cookbook and quickstart are usable today.
The caveat: there’s no single dreaming=true flag. You build the memory directory and dreaming workflow yourself, and Anthropic’s own post admits the cadence and rules for what to keep are still open questions.