By Dave Fuller, Managing and Technical Director, Accent
Anthropic has given its AI agents a feature called dreaming. Strip away the borrowed poetry and what you find is something far more familiar to anyone who has maintained a codebase for a decade: housekeeping. That does not make it unimportant. It makes it worth understanding properly.
The memories belong to someone else
The most unsettling scene in Blade Runner is not the shootout or the rain-soaked monologue. It is the quiet moment when Rachael discovers her memories are not hers. The spider outside her window, the boy who tried to take it, the recollections she has built a self around: all of it implanted, curated by Tyrell from someone else's life, chosen for effect.
Philip K. Dick's 1968 novel Do Androids Dream of Electric Sheep? asks whether an artificial mind can possess an authentic inner life. The title has become shorthand for AI consciousness, wheeled out whenever a technology company reaches for a word from the human interior. Anthropic has now reached for one of the biggest.
But Dick's real subject was never the dreaming. It was the curation. Who decides what a mind holds, what it discards, and whether it can tell the difference. That question turns out to be the useful one here, because it is exactly what Anthropic's new feature does.
What Anthropic actually shipped
On 6 May 2026, at its second annual Code with Claude developer conference in San Francisco, Anthropic launched dreaming in Claude Managed Agents as a research preview.
The mechanism is unglamorous. An AI agent doing real work writes notes to a memory store as it goes: conventions it has learned, mistakes it has made, quirks of the environment it is operating in. Over many sessions that store accumulates duplicates, contradictions, and entries that were true in March and are nonsense by August. Dreaming is a scheduled process that reads the memory store alongside transcripts of prior sessions, extracts patterns across them, and produces a curated version.
Three design decisions are worth noting, because they tell you what Anthropic thinks the risks are.
First, the input memory store is not modified. The process produces a separate output store, which makes it a copy-on-write operation rather than a destructive edit. Second, you choose the level of control: memory can update automatically, or changes can sit for review before they land. Third, it shipped as a research preview with gated access rather than general availability.
Those are the choices of a company that knows the thing being edited is the agent's understanding of your project, and that getting it wrong is expensive.
Harvey, the legal AI company, has reported substantially improved task completion after adopting it, with agents retaining filetype workarounds and tool-specific patterns between sessions. Take vendor-supplied figures for what they are: encouraging, unaudited, and drawn from one particular kind of work.
A note on what you will read elsewhere
Before going further, a caveat that matters more than it should.
Search for this feature and you will find a great many confident articles describing "Auto Dream" in Claude Code, complete with slash commands, trigger thresholds and four-phase pipelines. Several are dated weeks before the May announcement. Some cite each other. At least one developer has reported the documented command simply returning an unknown skill error.
I checked Anthropic's official Claude Code memory documentation on 8 August 2026. It describes two mechanisms in detail, CLAUDE.md files that you write and auto memory that Claude writes for itself, stored locally per repository. It does not mention dreaming, Auto Dream, or a dream command anywhere.
The feature Anthropic announced lives in Claude Managed Agents, its cloud platform for long-running agent work. Whether an equivalent has quietly reached the Claude Code CLI is, at the time of writing, something the secondary coverage asserts and the primary documentation does not support.
I mention this not to be pedantic but because it is the whole problem in miniature. A system designed to resolve contradictions between accumulated notes is being described by a body of accumulated notes that badly needs the same treatment. If you are making decisions about tooling, read the vendor documentation first.
Why the word does real work
It would be easy to dismiss the naming as marketing, and I do not think that is fair.
The dominant theory of what happens during REM sleep is memory consolidation. The brain replays the day's raw input, strengthens connections that proved useful, discards what did not, and reorganises the remainder into something durable. That is a genuinely accurate description of what this process does to a memory store. The metaphor is doing explanatory work, not just atmospheric work.
Anthropic could have called it compaction, or reconciliation, or garbage collection, and nobody outside the engineering team would have paid attention. They chose a word that made people look. Given how many organisations are about to hand persistent memory to autonomous systems, people looking is not the worst outcome.
Where the metaphor breaks
It breaks completely, and the break is the interesting part.
There is no experience here. Nothing is happening to anyone. A dream cycle is a scheduled job that reads files, applies judgement about which facts supersede which, and writes new files. The agent does not wake with a sense of having processed something. There is no self for the memories to belong to, no continuity of perspective that curation might threaten. Dick's replicants feared their memories being revealed as false because their identity depended on them. An agent has no such stake.
So the answer to the question in the title is no, and the question was never really the point.
The point is Rachael. Her memories were curated by someone else, according to criteria she did not set, in service of purposes she was not told about. When a dream cycle resolves a contradiction in an agent's memory, something has decided which version of the truth survives. The heuristics are reasonable: newer entries generally supersede older ones, explicit corrections outrank inference. But reasonable heuristics still make judgement calls, and in ambiguous cases they will sometimes pick wrong.
What survives is what the agent believes about your project. That is worth reviewing.
What this looks like from year twenty-seven
I have worked at Accent since 1999. The agency itself has been building for clients since 1991. In that time I have watched a specific kind of failure repeat itself with total reliability, in every organisation, on every stack, regardless of anyone's good intentions.
Documentation rots.
The wiki page describing the deployment process was accurate in 2019. The README references a build step that was removed two refactors ago. There is a comment at the top of a config file warning you not to change a value, and the person who wrote it left in 2016, and nobody now knows whether the warning still applies. Three separate places document the same convention, two of them wrong, and there is no way to tell from the inside which is which.
This is not a documentation problem. It is an entropy problem. Written knowledge decays at a rate proportional to how fast the system it describes changes, and the only thing that has ever worked against it is a human being deliberately going back and doing the tedious work of reconciliation. That work is unglamorous, unbillable, and consequently almost never done.
Which is why I find this feature more interesting than the framing around it deserves. Not because agents can dream, but because someone has finally automated the bit everyone skips.
The caution is equally clear from those years. Automated reconciliation inherits every weakness of the source material. If your project's accumulated notes contain a confident, repeated, wrong belief, consolidation will not correct it. It will promote it, merge the duplicates into one authoritative entry, and hand you a tidier version of the same mistake. Clean memory and correct memory are different properties, and the process optimises for the first.
For a marketing site with a two-year life, this hardly matters. For a bespoke system we will still be maintaining in 2031, it matters a great deal.
What to do about it
If you are running Claude Code today, you already have auto memory whether you have thought about it or not. It is enabled by default and writing notes about your projects to your machine right now. Run /memory in a session and read what it has saved. That five minutes is worth more than anything else in this article.
Beyond that:
- Put anything that must never change in CLAUDE.md, not in auto memory. Instructions you write are treated differently from notes the model keeps for itself. Consolidation treats fresher evidence as authoritative, which is the correct default and exactly wrong for a standing rule.
- Review before you trust. Where the option exists to inspect curated memory before it takes effect, take it, at least for the first several cycles. You are looking for confident wrong answers, not messy right ones.
- Treat memory as project data. If an agent is accumulating knowledge about a client's system, that store has the same confidentiality status as the codebase. It needs the same thinking about retention, access and location.
- Do not confuse this with understanding. A tidy memory store makes an agent faster to orient. It does not make it right, and it does not transfer accountability for the output away from you.
We are early in evaluating this properly at Accent, and I would rather say so than pretend to conclusions I have not earned. What I will say is that the industry has spent three years asking whether AI can think, and comparatively little time asking who curates what it remembers. Dick was asking the second question all along.
Sources
- Anthropic, New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration: https://claude.com/blog/new-in-claude-managed-agents
- Claude Code documentation, How Claude remembers your project: https://code.claude.com/docs/en/memory
- VentureBeat, coverage of the Code with Claude announcements, 8 May 2026
- The New Stack, coverage of Managed Agents dreaming, 6 May 2026
- Philip K. Dick, Do Androids Dream of Electric Sheep?, Doubleday, 1968
A note on how this article was produced
The argument, structure and editorial position of this article originated with Dave Fuller, Managing Director at Accent. Source material included a full transcript of the referenced podcast episode. AI tools were used for research summarisation. All claims were verified and the final text was reviewed and edited by a person, who takes responsibility for it.
Dave Fuller is Managing and Technical Director of Accent Design Group Ltd, the Norwich digital agency that has designed, built, hosted and maintained websites and web-based applications for over 27 years. He holds qualifications in web accessibility and design.