Anthropic’s Claude Managed Agents Get a “Dreaming” Memory Pass

For Claude and Claude Code developers, this is one of those updates that sounds whimsical but is actually pretty practical. Anthropic is pushing harder on long-running, multi-agent workflows, and “dreaming” is really about making memory less brittle when agents work for hours instead of minutes.

Key Points

Anthropic introduced “dreaming” for Claude Managed Agents at its Code with Claude conference.
In this context, dreaming is a scheduled review process that scans recent sessions and memory stores to find useful information to keep for future tasks.
The feature is in research preview and is limited to Managed Agents on the Claude Platform.
Managed Agents are Anthropic’s higher-level, pre-built agent harness that runs in managed infrastructure, aimed at longer-running or multi-agent work.
Anthropic says dreaming can surface:
- recurring mistakes
- workflows agents converge on
- preferences shared across a team
Users can choose automatic memory updates or review memory changes directly.
This is distinct from ordinary chat compaction, which mostly helps a single conversation stay within context limits.
Anthropic also said outcomes and multi-agent orchestration are moving from research preview to wider availability.
Anthropic will double five-hour usage limits for Pro and Max users of Claude Code because of demand and infrastructure strain.

My Take

What strikes me is that Anthropic is treating memory as an operational system, not just a chat feature. I think that’s the right direction: once you have multiple agents doing real work over time, the hard problem stops being “can the model answer?” and becomes “can the system remember the right things, at the right level of abstraction, without getting noisy?”

The “dreaming” framing is a little cute, maybe even a bit over-branded, but the underlying idea is solid. If it actually helps agents notice repeated failures or team-level preferences, that’s genuinely useful. I’d be curious whether the memory curation stays understandable in practice, because the moment an agent starts rewriting its own memory, trust becomes the whole game.

What I’d actually try, if I were building with this, is a narrow workflow with clear success criteria: maybe a support triage agent, a code review helper, or a multi-step research pipeline. Those are the kinds of tasks where recurring patterns matter and where memory drift would be easy to spot. I think that’s the right way to evaluate this kind of feature—less “wow, it dreams,” more “does it make the second and third week of usage better than the first?”

The usage-limit increase is also worth noting because it’s a very practical admission: demand is outpacing compute. That’s not glamorous, but it matters more to real users than another demo feature. If you’re building on Claude Code, these operational improvements may end up feeling more valuable than the headline feature itself.

Bottom line: this is an interesting step toward agents that can learn across time, not just within a single prompt. The concept is promising, but the real test will be whether the memory system stays useful, editable, and predictable when developers push it beyond the demo.