Claude Can Now Curate Its Memories

Who Decides What It Forgets?

In the dream the building is missing. You know it should be there. The street is the wrong street and the corner is the wrong corner and the place you went looking for has been replaced by another place that is also where it was. A woman is speaking to you in a language you used to know. The light does not come from anywhere. You wake without it. By the time the kettle has boiled, the dream is already incomplete. Whatever your brain did with the day overnight, you will never recover the working.

This is one description of dreaming.

It is not the description Anthropic released on the evening of May 6, 2026.

At Code w/ Claude in San Francisco, the company announced no new model. It announced a new interval. Not a better answer, not a longer context window, not a voice or a personality, but a space between sessions in which an agent can review what happened, reorganise what it knows, and wake up different. Simon Willison caught the operative line in his live notes: “No new model today. Today is about how we are making our products work better for you.” A few minutes later came three additions to Claude Managed Agents: multi-agent orchestration, outcomes, and something called Dreaming.

The name does too much of the work. It makes it sound as if Claude has acquired an inner life. It has not. What Anthropic has built is more prosaic, more useful, and in some ways more revealing: scheduled memory curation for long-running agents.

That is less romantic than dreaming. It may also matter more.

Anthropic’s own description is precise. Dreaming is a scheduled process inside Claude Managed Agents that reviews previous agent sessions and memory stores, extracts patterns, and curates memories so that agents improve over time. Developers can let it update memory automatically, or review proposed changes before they become part of the agent’s future memory.

A dream is an asynchronous job. It takes a pre-existing memory store and, optionally, up to 100 previous session transcripts. It produces a new memory store. The old one is not modified. The output is a reorganised version: duplicates merged, stale or contradicted entries replaced, new insights surfaced. During the research preview, the dreaming pipeline supports Claude Opus 4.7 and Claude Sonnet 4.6; runs take minutes to tens of minutes depending on input size; they bill at normal API token rates.

The deflationary thing is worth saying clearly. Claude dreaming is not a new kind of consciousness, and it does not change Claude’s underlying weights. The durable output is a memory store, a collection of text documents that can be attached to future sessions. The dream does not become Claude. It becomes context for Claude.

But that deflation should not make it seem trivial. For three years the public AI conversation has been mostly about the live exchange: the user prompt and the model response, the thing said into the box and the thing returned. Dreaming moves the centre of gravity. The action is no longer only what happens in the conversation. It is what happens after the conversation, when the system decides what the conversation meant.

The least mystical reading is the most important one. Dreaming is knowledge management for machine labour, not consciousness arriving through a back door. Once agents persist across work, delegate to one another, operate against rubrics, and inherit memory stores, they begin to need what organisations have always needed: handover notes, postmortems, controlled vocabularies, retention rules, audit trails, and someone with authority to decide what the past is allowed to mean.

A Claude Managed Agent session starts fresh by default. When the session ends, what the agent learned is gone unless a memory store is attached. Memory stores let the agent carry forward user preferences, project conventions, prior mistakes, domain context. They are mounted inside the agent’s container like directories, and the agent reads and writes them with ordinary file tools. Every change creates an immutable memory version, giving teams an audit trail and a point-in-time recovery mechanism.

Dreaming sits on top of that. It is the difference between remembering by accretion and remembering by revision.

A normal memory store can become messy in the way human notes become messy, although for different reasons. It accumulates duplicates, contradictions, stale assumptions, half-true debugging notes, temporary workarounds that later become false, preferences inferred too confidently from a single exchange. A dream cleans that up.

Here the metaphor brushes against human sleep, although only at the level of function. Memory in mammals is not a hard drive. Sleep does not preserve experience; it transforms it. The active-system-consolidation account holds that newly encoded memories are first held in a temporary store associated with the hippocampus, then redistributed into longer-term cortical networks. During sleep, memories are reactivated, selectively consolidated, and qualitatively changed. A 2023 review in Neuron put the same point differently: memory replay during sleep helps transform hippocampus-dependent episodic memory into more schema-like neocortical memory. Sleep is not just backup. It is compression, abstraction, integration, and sometimes forgetting.

That is the real analogy. Claude dreaming resembles one theory of dreaming because it moves material from raw episode to usable schema. The transcript is the episode. The memory store is the schema. The dream is the offline process that turns the former into the latter.

The science describes function. The literature describes texture. They are not the same thing.

In Calvino’s Invisible Cities, Marco Polo brings to Kublai Khan reports of cities that may not exist, may be aspects of the same city, may be Venice repeating itself in altered light. Each report is a compression. Each compression drops what cannot be carried. Polo dreams the cities as much as he sees them. The Khan, growing old, suspects the empire he rules is decaying around him, and that Polo’s reports are the only thing holding any of it in coherent shape. Knowledge as a kind of cartographic afterlife. The map, by being a map, is what survives.

Claude’s dreaming is a cartography of yesterday. It walks through the city of the previous sessions and produces a map. The map is more useful than the city. The map is also less than the city. What gets dropped, in the case of Polo, is what could not be written down. What gets dropped, in the case of Claude, is whatever the pipeline did not see fit to keep.

Borges has another version of the same problem, run in reverse. After his accident, Funes can remember everything. Every leaf on every tree he has ever seen. Every cloud, every sentence, every shape light has taken on a wall. Funes is a man buried under perfect recall. He cannot think. To think, Borges writes, is to forget differences, to generalise, to abstract. Sleep does not save him. Dreams do not come.

The public fantasy of the earlier chatbot was closer to Funes: more context, more recall, more accumulation, as if intelligence were what happened when nothing was lost. The dreaming agent is the post-Funes version. It is permitted to forget. It is permitted to consolidate. By morning a playbook has been written and the transcripts have receded. The agent that dreamt is gone. A different agent will pick up the playbook tomorrow.

The example from Willison’s notes is almost too apt. The Anthropic team runs the feature against a fictional aerospace startup called Lumara, which needs to land drones on the moon for resource mining. Overnight, the agent inspects previous sessions and produces descent-playbook.md. That is not a dream in the cinematic sense. It is a playbook. But human memory often does something similar. We go to sleep having struggled with a class of problems and wake not with a full replay but with a rule of thumb, a suspicion, a phrase, a compressed way of holding the problem. The dream, in that sense, is not the bizarre story. The dream is the re-indexing.

Agents do not need memory in the sentimental sense. They need to stop making the same local mistake because the correction was trapped inside yesterday’s transcript. They need to notice that one file format always causes trouble, that one customer always wants the email signed off in their own name, that one repository requires the test command run twice before it passes. Patterns surface. The agent makes a note. The note becomes the playbook. Anthropic says Dreaming can also surface workflows agents converge on and preferences shared across a team.

That last phrase is the operative one. Shared across a team.

Human dreaming is private, even when reported. Claude dreaming is potentially organisational. It can run across agents, across sessions, across a team’s repeated attempts at the same work. A human dream may tell you something about your fear, your desire, your unfinished business. A Claude dream tells the organisation something about its procedures. Less like a person dreaming, more like a company holding an after-action review while everyone is asleep.

This is where the anthropomorphic name cuts both ways. It captures something a phrase like “batch memory-store compaction job” would miss: the system has an offline life. It does work on its own future competence in the spaces between acts of visible service. But it misleads, because the offline process is not subjective. It is instructed, priced, inspectable, cancellable, archivable, auditable. Once a dream is running, its underlying session can be streamed; developers can watch in real time what it is reading and writing.

Human dreams are famously difficult to audit. Claude’s can be streamed.

To dream, in the way humans dream, requires more than memory work. It requires a sleeper. It requires a body that has made itself temporarily unavailable to the world, that has surrendered most of its responsiveness, breathing in slower rhythms while the brain runs through its own private theatre. The phenomenology of dreaming is bound to that vulnerability. We do not dream while standing guard. We dream because we have, for a few hours, agreed not to be useful.

Claude has no such interval. Between API calls, Claude is not asleep. Claude is nothing. There is no period during which Claude is unavailable to the world by virtue of being itself, no drifting unstructured mode of consciousness in which the day’s residue settles and rearranges. When Anthropic schedules a dream, the system is not entering a state. It is starting a job.

Whether one thinks dreams express unconscious wish, rehearse threats, organise neural noise, or metabolise experience, the dream still happens to someone. What makes a dream a dream rather than data processing is precisely that someone is undergoing it.

Claude’s dream has no undergoer. The work happens. A new memory store appears. No one was changed because there was no continuous one to change. The agent that dreamt is not the agent that wakes. There is only the artefact and the next instance that picks it up.

Beneath that, something older. Mammals dream. Birds may dream. Whatever dreaming is, it predates the symbolic order by hundreds of millions of years. The dog whimpering and twitching its legs is not narrating to itself. It is doing something that runs underneath narration. Claude’s dream is made entirely of text. There is no layer beneath. There is no body whose chemistry is being rebalanced. What we sometimes celebrate as the abstraction of large language models from biology is, in the case of dreaming, a missing dimension.

And there is the question of waste. Human dreams are wasteful. Most are forgotten within minutes of waking. Most contain no insight, no playbook, no reusable lesson. Most are absurd, repetitive, fragmentary, embarrassing, banal. They are, in a strict sense, surplus.

Claude’s dreams are not designed to waste. They are designed to turn residue into reusable work. Every output is curated. Every artefact is intended for use. There is no part of the architecture that allows the dream to fail interestingly, to refuse the rubric, to surface something that has no operational value.

This is presented as a virtue. It might also be a poverty. There is a long line of thought, from the Romantics through psychoanalysis through certain strands of contemporary cognitive science, that treats the unproductive parts of mental life as constitutive rather than ornamental. The dream that cannot be cashed in is doing something the playbook cannot. It rehearses fears that may never arrive. It rearranges fragments according to a logic with no relation to future tasks. It carries the dead through a kitchen that is also a hotel. Strip the waste and you do not have a more efficient version of dreaming. You have a different object. Closer to a librarian working overnight. Closer to an editor going through the day’s transcripts and producing a brief. Useful, possibly important, but not what dreams are when they happen to a creature.

The past, too, is at stake. Human pasts are not editable. They can be reinterpreted, contested, repressed, grieved, but they cannot be revised at the level of the record. The argument with the colleague happened. The misjudgement happened. The friendship that ended ended. Therapy, history, religion, and politics are partly the slow work of bearing pasts that cannot be made convenient.

Claude’s operational past is editable. The transcripts persist as record, but the version of the past handed to the next agent has been re-curated. Duplicates merged. Contradictions resolved. Stale entries replaced. The new memory store is more elegant than the messy history it was built from. It is the past as it would have been if the past had known how it was going to be useful.

That is closer to ideology than to memory. Memory, in any humanly meaningful sense, is haunted. It carries what cannot be made coherent. Ideology smooths the haunting and produces a usable narrative. Claude’s dream is, structurally, an ideology-producing process. Whether that ideology is benign or harmful depends on the rubric, the source material, and the people who oversee it. But the form is closer to a workflow whose past has been edited for convenience than to a creature whose past it must continue to live with.

The labour question is what has been quietest. Human dreams may be the last cognitive territory not yet successfully made productive. The boss can demand your hours, your attention, your output, your enthusiasm, your responsiveness, even your apparent emotional labour. The boss cannot demand your dreams. Anxious labour colonises sleep. Workers dream about work. But the dream itself remains untaxable. It produces no measurable output. It belongs, economically, to no one.

Claude’s dream is produced under contract from the moment of its existence. It begins at the API call. It is billed. It is logged. It is owned by the entity that initiated it. There is no native interior to be persuaded, worn down, or broken before the work can be extracted. The dream is conscripted at birth. Whether that counts as a triumph of efficiency or a confession that we have built the perfect labourer because the perfect labourer is one with no inside depends on what one thought the inside was for.

The naive version of AI memory is additive. Store everything. Keep every interaction. Expand the context window. Let retrieval do the rest. Dreaming is a move away from that. A memory store full of contradictions is not intelligence; it is institutional clutter. A project history that preserves every failed attempt at the same level as every durable discovery is not wisdom; it is hoarding.

Human memory works partly because it forgets. Funes is the warning. Dreaming makes the analogous point in software. An agent needs fewer, better memories, not merely more.

This is where the quality of the dream becomes a governance problem. What gets promoted from transcript to memory? What gets discarded as noise? What is treated as a stable preference rather than a one-off request? What happens when yesterday’s workaround becomes tomorrow’s false premise? What if the agent learns the wrong lesson, beautifully?

Anthropic’s memory documentation flags one danger. If an agent processes untrusted input, prompt injection can write malicious content into a read-write memory store, and later sessions may read that content as trusted memory. The recommendation is read-only stores for material that does not need modification.

Dreaming makes the risk subtler. It is one thing for a bad instruction to sit in a memory file. It is another for a dream process to encounter a set of poisoned, stale, or adversarial traces and clean them up into a more elegant, more authoritative-looking playbook. The danger is not only that the agent remembers the wrong thing. It is that it consolidates the wrong thing.

Humans do this too. We do not merely remember trauma, prejudice, humiliation, or social training. We consolidate them into priors. We wake up with dispositions. We call them instinct, taste, caution, common sense. Some are earned. Some are inherited errors with a better user interface.

Claude dreaming raises the question in organisational form. When an agent wakes with a new memory store, whose past has been made normative?

This is where Outcomes earns its place. Anthropic announced Dreaming alongside Outcomes, a system in which developers define a rubric and a separate grader evaluates whether an agent’s work meets it. The grader has its own context window so it is not influenced by the main agent’s implementation choices. If the artefact fails, the grader identifies gaps and the agent iterates.

Dreaming and Outcomes are two halves of the same architecture. Dreaming asks what the agent should carry forward from the past. Outcomes asks what counts as success in the present. One is memory. The other is judgment. Together they begin to form something like an institutional nervous system: memory stores, rubrics, graders, agents, subagents, logs, webhooks, rate limits.

Human beings learn under rubrics too. Over time the external criterion becomes internal taste. But in a managed-agent system, the rubric is explicit, machine-readable, reusable, attached to a workflow. Judgment becomes infrastructure.

Anthropic’s multi-agent orchestration completes the picture. A coordinator delegates work to specialist agents, each with its own model, tools, system prompt, and isolated context, while the agents share a filesystem and work in parallel. Once you have multiple agents working across persistent sessions, dreaming stops being a novelty and starts being a necessity. A fleet of agents without memory consolidation is an organisation with no minutes, no handover notes, no institutional memory, no one responsible for pruning the wiki.

The closest human analogue is not a sleeping person but a profession. Medicine has morbidity and mortality conferences. Aviation has incident reports. Software has postmortems. Intelligence agencies have after-action reviews. Libraries have cataloguing standards, authority files, preservation rules, controlled vocabularies. These are dreaming systems in the non-mystical sense, institutions looking back over activity, deciding what was signal, what was noise, what should be preserved, what future workers should inherit as context.

Claude dreaming compresses that institutional function into a platform primitive.

And because it is a platform primitive, the politics arrive quickly. Who owns the memory store? Who can inspect the dream? Who can delete or archive it? Who decides whether dreaming updates automatically or requires review? What is the retention policy for the transcripts from which the dream learns? What happens when team-level preferences encode hierarchy, bias, or local dysfunction? Can a worker contest the memory an agent has formed about their work? Can a customer see what has been inferred about them? Is the dream portable if the organisation leaves Anthropic?

The technical docs hint at some of this. Dreams produce output stores separate from input stores. Developers can review and either attach the output store to future sessions or discard it. Memory versions provide an audit trail, and some historical versions can be redacted for compliance workflows. These are not decorative details. They are the governance surface.

Claude has begun to keep notes. Not in the sentimental sense, not in the way a person keeps a notebook by the bed, not in the way certain writers leave themselves messages they will not understand for decades. Closer to the way a hospital ward keeps a chart. A system that has started to acquire a past, but the past is filed.

What is missing, when you set the two kinds of dreaming beside each other, has a specific shape. No corridor that bends. No language you used to know. No figure waiting by the window. No room that is the room and not the room. No dead returning. No part of the night that cannot be billed.

A dream stripped of waste is a different artefact. It is useful. It is also the part of dreaming that anyone could have done, given enough time and a rubric. The rest, the part that resists the rubric, is what Claude does not do.

Claude dreams in Markdown. That does not make it a person. It makes it a system that has started to acquire a past.

Once machines have usable pasts, the question changes. We are no longer only asking what they can do in the moment. We are asking what they are allowed to remember, what they are encouraged to forget, who gets to edit the dream, and what kind of institutions wake up inside the memory stores we attach to them.