Anthropic just announced feature Dreams for Claude’s Managed Agents API, and it solves a problem every agent builder eventually hits: memory rot.
Agents write to memory as they work. Great. But over hundreds of sessions, that memory store turns into a mess. Duplicates. Contradictions. Stale entries that haven’t been true since three weeks ago. The agent keeps dragging this bloated context around, and nobody’s cleaning it up.
Dreams is how Claude cleans it up.
You trigger a dream as an async job. Claude reads an existing memory store alongside past session transcripts (up to 100 sessions), then produces a brand new, reorganized memory store: duplicates merged, contradictions resolved, stale entries replaced, and new patterns surfaced that weren’t explicitly written down anywhere.
The original store is never touched. You get a fresh output store to review, attach to future sessions, or discard if it doesn’t look right. Safer, auditable, reversible by design.
Three things builders should know:
It’s not a simple summarization pass. Claude is actively mining transcripts for patterns and insights, not just deduplicating keys. You can also pass in custom instructions to guide what it focuses on (e.g., “prioritize coding-style preferences; ignore one-off debugging notes”).
You can watch it think. While a dream is running, its session_id points to the live underlying session. You can stream the events in real time and observe what’s being read and written.
It costs what you’d expect. Billed at standard API token rates for whichever model you use (Opus or Sonnet). Usage scales with the number and length of input sessions, so start small.
The name is a deliberate metaphor, and it’s a good one. Biological memory consolidation mostly happens during sleep: the brain replays experiences, prunes weak connections, strengthens meaningful patterns. Claude’s dreaming mechanism is structurally the same idea, applied to agent memory.
But here’s what I think most people will miss: this introduces a second loop outside the agent loop.
The agent loop acts and writes. The dream loop reflects, reorganizes, and improves. That separation is what makes compounding performance over time actually possible. Before this, memory was basically an append-only log. You fixed noise with prompt engineering. Now memory becomes an evolving knowledge base with an offline learning mechanism sitting behind it.
Better agents don’t come from bigger context windows. They come from better memory lifecycle management. You can give an agent infinite context and it will still accumulate noise, contradiction, and lost signal. Dreaming is a bet that the answer is curation, not capacity.
We gave AI a voice, then a brain. Memory is the next layer. And now, apparently, it needs to sleep too.