At its Code with Claude developer conference last week, Anthropic announced a rather unusual new feature — letting Claude Managed Agents "dream."
Don't get the wrong idea — this isn't about AI gaining consciousness. "Dreaming" simply means the agent periodically reviews past conversations and operations, extracts information worth remembering long-term, and saves it to "memory" for future use.
Why make AI "dream"?
The problem is that LLM context windows have limits. An agent running for hours or days can easily lose early information. The existing solution is "compaction" — periodically pruning conversation history, discarding the unimportant while keeping what matters.
But compaction only works within a single conversation with a single agent. If you have multiple agents collaborating, each compacts its own information — no one sees the big picture.
"Dreaming" fills that gap. It analyzes past sessions and memories across agents, identifying patterns that no single agent can spot on its own. Like a recurring mistake, a workflow everyone converged on, or preferences the whole team shares.
As Anthropic puts it: "Dreaming surfaces patterns that a single agent cannot see on its own — recurring mistakes, workflows that agents converge on, and preferences shared across a team. It also restructures memory so it stays high-signal as it evolves."
Can you use it now?
Not yet. Dreaming is in research preview, limited to Managed Agents on Claude Platform, and requires applying for access.
The good news: two previously previewed features — outcomes and multi-agent orchestration — are now widely available. Anthropic also doubled the five-hour usage limits for Pro and Max subscribers, responding to user complaints about insufficient compute capacity.
Some thoughts
I find the "dreaming" name quite fitting. Unlike traditional database writes, this process is fuzzy, periodic, and somewhat "inspirational" — not unlike how human sleep consolidates memories. The brain replays the day's experiences and decides what to keep and what to discard.
At the end of the day, though, this is still a research preview. We'll know how well it works once more people get their hands on it.
