Anthropic · 2026-05-06 · major
Anthropic Managed Agents Get Dreams — Self-Improving Memory Curation Pipeline
Dreams is a research-preview pipeline that reads an agent's memory store plus past sessions and produces a deduplicated, reorganized memory store, surfacing recurring mistakes and shared team preferences.
An async pipeline that lets Claude reflect on past sessions and rewrite an agent's memory store without touching the original.
Key specs
| Sessions per dream | 100 |
|---|---|
| Models supported | claude-opus-4-7, claude-sonnet-4-6 |
What is it?
Dreams is a new research-preview feature for Claude Managed Agents that runs an offline pipeline over an existing memory store and up to 100 past session transcripts. It outputs a fresh memory store with duplicates merged, stale entries replaced, and new patterns surfaced — leaving the input store untouched so teams can review and discard the result if they don't like it.
How does it work?
A dream job is created via POST /v1/dreams with two beta headers (managed-agents-2026-04-01 and dreaming-2026-04-21) and an inputs array referencing one memory store plus optional session IDs. The selected model (Opus 4.7 or Sonnet 4.6) reads the transcripts, dedupes contradictions, and writes a new memory_store output. The job is asynchronous and exposes a session_id so callers can stream the underlying pipeline events.
Why does it matter?
Long-running agents accumulate duplicates, contradictions, and dead context that drag down quality across sessions. Dreaming gives teams a structured way to compress that history into a clean store on a schedule, and the human-review workflow keeps a safety net while the feature is still in preview.
Who is it for?
Teams running multi-agent Claude deployments with persistent memory.
Try it
Add 'anthropic-beta: managed-agents-2026-04-01,dreaming-2026-04-21' and POST to /v1/dreams.