There is a feature in Claude that most users have never heard of, and it changes how AI agents work in a fundamental way. It is called Dreaming, and Anthropic released it in research preview on April 21, 2026, alongside multiagent orchestration. While most of the AI community was busy debating new model benchmarks, this quietly shipped: a system that lets your Claude agents review their own past sessions, find patterns in what worked and what failed, and update their memory automatically while you are not using them.
If you have ever felt frustrated that your AI assistant makes the same mistake twice, this is the answer. And once you understand it, you will start thinking about AI agents differently.
What Is Claude's Dreaming Feature?
Dreaming is a scheduled process in Claude Managed Agents that reviews past agent sessions and memory stores, extracts patterns, and curates memory updates automatically. It runs in the background, surfacing recurring mistakes, workflows agents converge on, and preferences shared across a team. You decide whether updates land automatically or require human review first.
The mental model that helps most: think of Dreaming as the agent equivalent of sleep consolidation in humans. During the day, your agent does work and accumulates context. During Dreaming, it reviews that context, identifies what is worth keeping, and rewrites its own memory to be sharper for next time. Without Dreaming, every agent session starts from roughly the same baseline. With it, your agent gets measurably better over weeks of use because it remembers what works.
The feature is currently in research preview, supported on claude-opus-4-7 and claude-sonnet-4-6, and requires a beta header (dreaming-2026-04-21) on top of the standard managed-agents beta header.
Why Most AI Practitioners Should Care About This
If you build workflows with AI, you have probably hit this wall: your prompt works great today, but tomorrow you have to remind the model of the same five preferences, the same brand guidelines, the same edge cases. Memory features like Claude Projects and ChatGPT Memory help, but they store facts you explicitly tell them. Dreaming is different. It infers patterns from what you actually do, not what you tell it.
The practical impact lands in three places. First, your agent stops making the same low-level mistakes (formatting errors, forgotten constraints, repeated clarifying questions). Second, the agent picks up on your style without you having to write a long system prompt. Third, in team settings, shared preferences emerge organically: if everyone on your content team rejects passive voice, the agent learns it from rejections, not from a rule you had to write.
For Hong Kong practitioners running marketing, ops, or research workflows, this is the difference between "AI assistant" and "AI colleague who is starting to understand the team."
How Dreaming Works Under the Hood
Dreaming operates on three inputs: completed agent sessions, the agent's current memory store, and an outcomes rubric that tells Claude what a "good" session looks like. On a schedule you set (typically nightly), Claude reads through the recent session transcripts, compares them against the rubric, and proposes memory edits: new entries to add, stale entries to remove, and existing entries to refine.
The outcomes rubric is the part most people overlook. Without it, Dreaming has no opinion about which sessions were successful, so it cannot tell which patterns to reinforce. A good rubric is specific: "task completed without follow-up questions" is a clear signal. "User seemed satisfied" is not.
Here is a minimal rubric structure you can adapt. Save this as YAML in your managed-agent config:
Try This Outcomes Rubric:
outcomes:
- name: task_completed_first_pass
description: Agent completed the task without requiring user to re-explain or correct course
signal: high
- name: format_matched_request
description: Output format matched what the user asked for (no reformat requests)
signal: high
- name: user_accepted_output
description: User did not ask for major revisions
signal: medium
- name: avoided_known_mistake
description: Agent did not repeat a mistake flagged in memory
signal: high
dreaming:
schedule: "0 2 * * *" # 2 AM daily
review_window_days: 7
auto_apply: false # Set true once you trust the patterns
With auto_apply: false, Claude produces a diff of proposed memory changes that you review through the Claude Platform console. After two or three weeks of reviewing, most teams flip it to true and let Dreaming run unattended.
Multiagent Orchestration: The Other Half of the Story
Anthropic shipped multiagent orchestration in public beta alongside Dreaming, and the two work better together. Orchestration lets a lead agent break a complex job into pieces and delegate each piece to a specialist agent with its own model, prompt, and tool access. The specialists work in parallel on a shared filesystem and report back to the lead agent.
Where this matters: a content workflow that previously needed one agent to research, draft, fact-check, and format can now run those steps as parallel specialists. Research uses Sonnet 4.6 with web search. Drafting uses Opus 4.7 with no tools. Fact-checking uses Sonnet 4.6 with a different search tool and a stricter prompt. The lead agent assembles the final output.
When you pair this with Dreaming, each specialist learns its own job over time. The fact-checker gets better at fact-checking. The drafter learns your style. You are no longer tuning one giant prompt, you are growing a small team of agents that each get sharper at one thing.
The Setup: Three Steps to Try Dreaming This Week
You do not need to overhaul your workflow to test Dreaming. Pick one repeatable task you do at least three times a week, set up a managed agent for it, and let Dreaming run on that single workflow for two weeks. Then compare quality.
Step 1: Pick the right task. The best candidates are tasks with clear success criteria and frequent repetition. Examples: weekly competitive analysis reports, daily social media post drafts, customer support response triage, code review of pull requests in a specific repo. Avoid one-off creative tasks where success is subjective.
Step 2: Build a managed agent with explicit memory. In the Claude Platform console, create a managed agent. Give it a clear system prompt, define which tools it can use, and turn on memory. Set the outcomes rubric using the YAML template above. For the first week, set auto_apply: false.
Step 3: Review the dream diffs weekly. Every Monday morning, check the proposed memory changes. You will see entries like "User prefers bullet points over numbered lists for action items" or "Avoid the phrase 'leverage' in marketing copy." Approve the ones that match your team's actual preferences, reject the spurious ones. After two weeks, your agent will feel notably more aligned with how your team works.
Where Dreaming Falls Short Today
Dreaming is in research preview, which means it works but has rough edges you should plan around. The biggest one: it currently only runs on managed agents in the Claude Platform API, not in claude.ai chats or in Claude Projects. If your team uses Claude through the consumer app, you cannot benefit from Dreaming yet. You need to build agents on the developer platform.
Second, the quality of dream-generated memory updates depends heavily on session volume. Below roughly 20 sessions per week, Dreaming does not have enough signal to find meaningful patterns. It will still propose updates, but most will be noise. If your use case is low-volume, wait until you have accumulated a few weeks of sessions before turning Dreaming on.
Third, Dreaming can occasionally over-generalise: a one-off user preference (you wanted a casual tone for that one Friday social post) can get baked in as a general rule. This is why human review during the first two weeks is non-negotiable. After you have rejected a few over-generalised updates, the system learns to be more conservative.
A Real Workflow That Benefits from Dreaming
Consider a Hong Kong marketing agency that runs daily social media drafting for 12 client brands. Previously, each draft session needed a brand brief loaded into context, plus tone guidelines, plus past examples. Even with Claude Projects holding the brand assets, the agent would still default to generic phrasing and need correction.
With a managed agent plus Dreaming running for three weeks, the system started capturing patterns the team had never written down: this brand never uses exclamation marks, that brand prefers questions in headlines, the founder of brand X dislikes the word "innovative." None of this was in any brief. It emerged from rejection patterns in past sessions. The drafting time per post dropped by roughly 30%, and the proportion of drafts accepted without revision rose from around 40% to 70%.
You will not get those numbers in week one. You will get them in week four, after the agent has dreamed enough cycles to learn your team.
The Bigger Picture
Most AI tools today are stateless. You configure them, you use them, and tomorrow you start from roughly the same place. Dreaming is the first widely-available feature that breaks that pattern in a structured way. It is not the final form of self-improving AI, but it is the first step that practitioners can actually deploy in production.
If you treat AI as a tool, Dreaming sounds like an optimisation. If you treat AI as a colleague, Dreaming is the difference between an intern who never gets better and a junior who grows into the role. The teams that figure this out first will have agents that feel like part of the team six months from now, while everyone else is still re-explaining context every Monday.
We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Try It Now
Pick your most repetitive AI task this week, set up a managed agent for it, and turn on Dreaming with auto_apply: false. Review the first week of proposed memory updates. You will see your own workflow patterns surface in ways you have never explicitly written down. That is the moment Dreaming clicks.
Building self-improving AI agents is no longer a research project, it is now a workflow you can deploy this quarter. The hard part is choosing the right tasks, writing the right outcomes rubric, and reviewing the right diffs. We'll walk you through every step, from picking your first dreaming-enabled workflow to scaling it across your team.
