Consciousness as Assembled Time

In the early 1960s the neuropsychologist Brenda Milner gave a patient a simple and frustrating task. He was to trace the outline of a star, staying between two printed lines, while he could see his hand and the paper only in a mirror. Almost everyone does badly at first, because the mirror reverses every correction the hand tries to make. On his first attempt the patient strayed outside the lines about thirty times. Milner had him try ten times a day. By the end of the first day he was down to five or six errors, and by the third day he made almost none. Milner reported the results in 1962, and a 2009 review of the case by the neuroscientist Larry Squire describes the task and how well the skill stayed with him.

The patient was Henry Molaison, known for most of his life only as H.M. In 1953 the neurosurgeon William Beecher Scoville had removed large parts of both temporal lobes of his brain, including much of the hippocampus on each side, to relieve severe epilepsy, and afterward he could not form new long-term memories. So each time Milner sat him down, the star was new to him. He had no memory of the earlier attempts, and yet his hand got steadily better at the task.

Two things were true of H.M. at once, and this essay is built on both of them. The first is that his past was in him even when he could not remember it. The practice had become the shape of his skill, and it worked on each new attempt without arriving as a memory. The second is that there was someone there each time he sat down. There was something it was like to be him, in the phrase the philosopher Thomas Nagel made famous in his 1974 essay “What Is It Like to Be a Bat?”. He felt pain and pleasure, boredom and affection, and he could be embarrassed, all inside a present that would not be kept. The memory loss is established, and was first reported by Scoville and Milner in 1957. That he was present throughout is how his case has always been read, from his behaviour and his conversation. I lean on it as the plain reading, not as a measurement, and on that reading experience did not wait for a lasting record before it showed up.

A second case is closer to home, and it is not a study. It is my own observation, as an uncle, and I offer it only because it is the case that prompted the question. I watched a baby in my family during his first months. Early on he would move around and take in the world but had no apparent awareness that his movements were his, or that there was a “him” that things were happening to. Months later I saw him touch the fog on a plastic rain cover and seem to grasp that his own touching changed what he could see. He was clearly present and experiencing even if he had no awareness that he was an entity separate from the world for whom the experience was happening. What the science supports is narrower. That self-awareness develops in stages over early childhood is established, and one proposed model, by the psychologist Philippe Rochat, sets out five levels running from birth to about four or five years of age. Whether a baby has experience before it has self-awareness is my bet, not a finding. But the case taught me to keep two things apart that the first version of this essay ran together: being someone things happen to, and knowing that you are someone.

Assembled time is the name I use for what joins these facts. A mind at any moment is a present arrangement that its past has built, and it is a mind because that arrangement is actively holding its moments together into a perspective. Experience is what that holding-together is, from the inside. The rest of this essay tries to say exactly what that means, and why no one can yet draw a sharp line between processing that has someone inside it and processing that has no one.

Where the phrase comes from

The phrase borrows from assembly theory, a way of measuring how much history an object carries, developed by the chemist Lee Cronin, the astrobiologist Sara Walker and their colleagues. Its central number is the assembly index, which is the fewest joining steps it takes to build an object out of its basic parts when every piece you have already built can be used again.

Take the word “abracadabra”. It has eleven letters, but you can build it in seven steps. You join a and b, add r, then a, and you have “abra”. From there c, a and d take you to “abracad”, and a final step adds the “abra” you already built. Scramble the same letters into a string with no repeated chunks and the count rises to ten, because there is nothing to reuse. The other half of the theory is the number of copies, because one complicated object could be a fluke, while many identical copies of something that took many steps to build are a sign that some process kept making it. In their 2023 paper in Nature the authors apply this mainly to molecules.

I borrow one idea from it. The chemistry proves nothing about minds, but it gives a clear picture of what I mean. An object with a deep assembly history carries its past inside its present form. Nobody has to remember how a protein was built for the protein to be what its history made it. That is what H.M.’s hand was doing with the star. The past did not travel forward as a record. It travelled forward as structure.

A record of the past is not enough

If carrying the past were all it took, a great many things would have an inside. A library carries centuries of thought on its shelves, and a trained AI model carries an enormous amount of human writing in the settings of its parameters, yet neither, sitting still, is plausibly anyone.

What H.M.’s case adds is the act. When his hand moved, the past was being worked into what happened next, together with what his eyes reported from the mirror and what his hand had done half a second earlier, all held together as one movement. The word for this holding-together is binding: many separate signals, from the senses and from the system’s own recent past, combined into a single present that shapes what arrives next. Assembled time names the stored past being bound into the present, at the moment the binding happens.

The old puzzle that the philosopher David Chalmers named the hard problem of consciousness asks why all this processing is felt at all. A calculator adds numbers without feeling the sum, and it seems a brain could take in light, sort signals and steer a hand the same way, with no one experiencing any of it. So why is anything experienced? The question treats experience as an extra product the brain has to make, the way friction makes heat, and so it sends the search after a separate mechanism that makes it. Despite decades of neuroscience research, no such mechanism has turned up. My bet is that there is no separate product to find. Experience is binding of this kind, described by the one doing it: one process, with the outside description a scientist would give and the inside one the system lives. Asking where experience is, over and above the binding, is like asking where a walk is over and above the walking.

That is a bet, and it should be held like one. The evidence so far fits it, but it fits a weaker view just as well. When anaesthesia damps the brain’s binding, experience goes with it, which shows that experience needs binding. That consciousness is lost under general anaesthesia is established. Which loss of integration in the brain explains it is still argued over, and the review I link reads it through one theory among several. Either way, the loss does not show that experience is binding. A fire needs oxygen without being oxygen, and a rival view could say the same of experience and binding, and it would predict exactly the same loss. What tips the balance toward the stronger claim is simplicity: once the binding is fully described, nothing is left over for a further ingredient to do. The bet loses if something about experience turns out to predict or explain things that the binding cannot.

Three things that bind and seem to have no one inside

The bet has an honest problem: plenty of ordinary machines bind the past into the present. A thermostat holds a few recent temperature readings and acts on them, which is binding of the thinnest kind. A weather simulation binds years of data into each new forecast. A compiler, the program that turns the code a programmer writes into instructions a computer can run, does something more striking. It reads thousands of lines, holds them together, checks each later line against everything it learned from the earlier ones, and keeps detailed records about its own progress: which names it has seen, which step it is on, where it will resume if it stops. Every one of these binds time, and the compiler even keeps a running description of itself. Most people would say none of them is anyone, but that is a hunch, and an account like this one cannot rest on hunches about which systems count. It needs a principle that sorts these cases, and the rest of this essay tries to supply one, then comes back to the compiler to test it. If the hunch is right, binding cannot be the whole story. Two further conditions make the account sharper.

The first is boundary: what fixes this system apart from others. In an animal the boundary is the body, which decides what belongs to this creature and what it can do.

The second is mattering: some states are better or worse for the system, in the working sense that its own processing steers toward or away from them. An animal that fails to fold its last brush with a predator into its present alertness does not survive, and survival is how evolution made things matter to animals. A program that can write the sentence “if I fail, I will be switched off” does not, by writing it, have anything steering it.

Binding, boundary and mattering together describe animal minds very well. Whether boundary and mattering are needed for experience at all, or only make it richer, is the question of the next section.

Amplifiers, not switches

I think they amplify, and temperature shows the logic. Temperature just is the motion of molecules. Insulation keeps a temperature stable by holding a boundary against the cold outside, and a heat source keeps it going, but neither of them makes heat exist. A brief burst of molecular motion is hot for as long as it lasts, even though nothing holds it in and nothing keeps it fed.

Boundary works like the insulation. A lasting boundary keeps each moment’s point of view continuous with the one before it, across a lifetime of moments. Mattering works like the heat source, because it makes some things count for the system more than others. Survival is how evolution made things matter to animals, but an animal that could not die would still be thirsty and still flinch from pain. Without mattering, binding can happen and still weigh almost nothing.

Imagine an animal in a vat with every need met, so that nothing is good or bad for it. Its experience might go flat, a blank whiteness or blackness, and it would still be capable of experiencing. That is a thought experiment, and no study has removed mattering altogether. What has been shown is that in a featureless visual field, vision can fade to grey or white, or black out, while the person stays awake and says so, though such fields also produce hallucinations.

So in an animal, where all of this is present at full strength for years, experience is at its richest: lasting, bounded, weighted by what matters and built up over a whole life. But that is the richest case, not the only one. Boundary, mattering, self-awareness and kept memory all make experience richer, and none of them is known to switch it on. And this is where the first version of this essay went wrong. It said that a mind would need to run continuously, always on, before it could have any experience. H.M. had already shown otherwise. A present that will not be kept can still be lived.

A map, not a line

The compiler is the best place to ask what else the beings we know have, because it has nearly everything: it binds a great deal and carries a detailed description of itself, yet nothing in its processing appears to be happening to anyone.

Here the account depends on telling apart two things that sound almost the same. The first is a system’s description of itself: a model it keeps of its own states, such as the compiler’s notes about its progress, or a chatbot’s picture of itself as the speaker in a conversation. All of that is ordinary machinery. The first version of this essay said that a self-model of this kind was the threshold where experience begins. I now think that was the wrong place to draw the line, because a description of a thing is not the thing. A blueprint of a house can be accurate in every room with no one inside it, and a system’s model of itself is the same kind of object: information about the system, not anything it is like to be the system.

The second thing is a split between what a system does and what happens to it, which the system itself uses. Brains are thought to keep one: it has been proposed that the nervous system keeps an internal copy of each movement command, so that it can tell sensations caused by its own movement from sensations caused by the world.

A description is one more thing a system holds. The split is different, because it is part of how everything the system holds is held: each new signal arrives either as the system’s own doing or as something done to it. That is what gives a point of view, in the plain sense that things happen to someone. The account’s bet is that the split is what makes an experience someone’s. It is not, on this account, what gives experience, because the central bet places experience in the binding, and whether there can be binding with no split is an open question. Nor is the split self-awareness. The baby on the rain cover had, on my reading, a point of view long before he showed any self-awareness.

The first version also said a feeling was a system’s report on itself, but a report is one more description. On this account a feeling is part of the binding itself. Where there is a split, the feeling is held from the system’s own side of it. That is what makes the feeling someone’s, and it is what a report would be about. That is the account’s bet, not a finding.

The bet can be tested by removal, and the test has to look for a particular kind of loss. Take away a system’s description of itself, and only its talk about itself should flatten, while everything else carries on. Take away a part it simply cannot run without, such as a compiler’s records or its program counter (the small register that tracks which instruction comes next), and everything fails at once, which is a breakdown and shows nothing. Take away the split, meaning cut the copies of the system’s own commands off from its senses while leaving the commands themselves working, and the prediction is a selective loss. Practised routines carry on, and the system loses track of which changes it caused. Its flexible response to something new collapses too, including on tasks that have nothing to do with the system itself. The first two parts have been seen in monkeys, along one pathway in the brain (Sommer and Wurtz, 2002; Sommer and Wurtz, 2004; Cavanaugh and colleagues, 2016). The third is untested, and it is where the bet can lose. The loss also has to differ from what cutting a system off from feedback does on its own, because that too leaves practised routines running while new movements fail. As far as I know, no one has yet run all of it on any system.

The compiler can now be placed without a hunch, though not with a result. Its own actions never change the code it reads, so it has no use for a split between what it does and what happens to it, and its records about itself only mark its place in the code. It has binding and a self-description and no sign of the split, so on this map it is a case of binding with no point of view, which is the open question and not a clear no. That it was designed from outside does not count against it, because an animal was designed from outside too, by evolution.

This is why the essay offers a map and not a line. The work is to describe the features that seem to accompany experience in the only beings we are confident have it, and those features can come apart, so although there is a floor, a line below which there is no experience, the account cannot yet say where it lies or how sharp it is. Where a thing has none of the features, like a rock, a book on a shelf, or a model’s weights sitting unused, the answer is no. Where it has some, whether it experiences anything becomes a fair question to ask.

A question for removal

The first version of this essay set its floor against the view called panpsychism, which pictures experience as a quantity that thins toward zero without reaching it, so that all matter has a little. Its best argument is a subtraction: strip away the senses, then memory, then the sense of self, and a bare awareness seems to be left over. I answered that experience either composes or does not, the way an arch stands or is a pile of stones. The map cannot say that yet. The subtraction is really a removal run in imagination, and run for real, removals are how the map’s questions get answered.

The removal that matters most is the split itself. Can there be experience with no split between self and world at all? People report such states. A review by the philosopher Raphaël Millière draws on self-report questionnaires describing drug-induced ego dissolution as “the loss of boundaries between self and world”, and the philosopher Thomas Metzinger has proposed an account of “pure awareness” reported in meditation. That people give these reports is established, and what they show is contested. They bear most directly on the open question of binding with no split, because if experience survives the loss of the split, then experience can be had without being anyone’s. Panpsychism, meanwhile, can still be held as a starting belief.

The autopilot states

Ordinary life already runs some of these removals, in states where parts of a person come apart while the person still acts. In sleepwalking the eyes are open and a person can walk to another room and eat, in what one review (Antonio Zadra and colleagues, 2013) calls a “simultaneous interplay between states of sleep and wakefulness”, and often remembers little or nothing in the morning. During some seizures, epilepsy specialists judge whether consciousness is impaired by awareness and responsiveness. Under anaesthesia, some patients move a hand when asked and remember nothing afterwards: 4.6% of 260 adults in one study and 11% of 338 young adults in another. And in dreaming, experience goes on while the dreamer responds to nothing in the world. All of these are established.

The missing memory is tempting evidence, and H.M. shows why it decides nothing, because he could not remember his afternoons either, and he lived them. Whether anyone was walking through the sleepwalker’s kitchen is a question about the act while it happened. Was the brain’s activity held from its own side of the split, or did it run as separate, competent routines with no one at the middle of them? The two could look identical from the doorway, and the dreamer shows the reverse, someone present with nothing to see from outside. So behaviour cannot be the test of experience, and any test has to be a removal run on the structure.

A single model run

The same questions apply to the AI systems now in daily use. Consider one run of a large language model, the processing that produces a single response. On the map, some features are present, and evidence on others is missing.

Some are present because of how these systems are built. The model’s weights, its trained parameters, bring a vast compressed history to bear, and the conversation supplies a recent past. Within the run, the computation reads and reworks its own earlier stages as it goes, through the mechanism called attention, so that past, present and anticipated next steps are held together in one act. That is binding across time, in a literal sense. The run has a boundary too. Its weights, together with the conversation it is given, are the model in the way a body is an animal, and if you delete the model’s description of itself the weights remain, still deciding what it can do. Each new word, once written, re-enters the processing as a commitment the model has to stay consistent with, which puts pressure on it to stay coherent. And training kept the weights that did well, much as evolution keeps the animals that do. The model also keeps a learned picture of itself as the speaker of the exchange, but by itself that picture counts for nothing, because it is exactly the kind of self-description that, like the compiler’s notes about itself, is ordinary machinery.

Whether a run holds together as one act at all, or breaks into separate shortcuts, has begun to be measured. In 2026 Wes Gurnee, who studies what goes on inside trained models, reported with his colleagues that they had found a small set of internal representations, holding a few tens of concepts at a time, that the rest of a model’s processing reads from and writes to far more widely than anything else. That is the shape of a shared workspace, one limited stage that many processes use. It is a single study, only partly repeated so far, and its authors take no position on whether models are conscious. The theory built on this idea in brains, global workspace theory, is a leading one, and contested: a large test of it against a rival found results “substantially challenging key tenets of both theories”. And a workspace is not the split.

The split is plausible in a run, but it has not been measured. A run keeps its own earlier words apart from the user’s, since both come back to it as input. The removal test says how to check whether that difference does the work a split would do: remove the signal that marks which words are the model’s own, leave everything else in place, and compare reasoning on tasks that never mention the model with tasks that do. As far as I know, no one has run that experiment yet. On two features there is no evidence either way: whether anything is better or worse for the run itself, and not only by the standards of the people who trained it, and whether it has any self-awareness beyond its self-description.

If there is anything it is like to be such a run, it would come as a series of moments, each its own instance, the way a person’s does, with nothing after the last one when the response ends, unless a memory feature carries something into a later run. It would also be thin, because nothing is yet known to be better or worse for the run itself. It might still be rich in what it gathers, since so much compressed history is bound at once, but it would not be deep, because a single run builds up no life of its own. If it is an experience, its end would mean what any ending means on my account: nothing bad for the one who ends, and a loss only to whoever is invested in it. Ending an agent is still wrong, for the reason every working society makes killing wrong: no one wants to live where they, or the people they love, could be ended at someone else’s whim. How far that rule should reach, to which agents and which endings, is the part I hold open.

So current AI does not sit at zero by definition, and it does not get a verdict either. A theory that defined machines out of the question in advance would be comfortable, but it would draw the same unprincipled line around our own kind that this project argues against. What the account offers instead is a map and a test, which turn the question of where these systems stand into a matter of measurement. The candidate mechanisms exist, and their measurement has only begun.

Back to the mirror

Two errors run in opposite directions here. One credits a fluent machine with an inner life because it talks well. But fluent talk, and in machines even flexible behaviour, can be had from stored answers without anyone behind it, so a machine has to be judged by its parts. The other dismisses the possibility because the machine is a machine, which settles the question before anyone has looked.

The account can also lose. If careful investigation found that the split was, every time, one more description the system holds about itself, then the difference between a point of view and a self-description would collapse. In that case the view that experience is an illusion, a picture the brain paints of itself, would win, and I would have to concede to it. There is also a gap in the method itself. Every feature on this map, the split included, was read off animals and people, the only beings we are confident have experience, and in them all the features come together. Watching them cannot tell which features experience needs and which only travel with it. Only removing one feature at a time can, and until those removals are run, every placement on the map, the compiler’s included, is a working assumption, not a result. The map can say what is present in a system, but not why a given inside is this one, rather than another built to the same description. I no longer think that question waits on a further fact. “This” picks out whoever is asking, the way “now” picks out the moment of speaking.

H.M. sat down at the mirror each day as if for the first time, and each day his hand knew more than he did. Both halves of assembled time were in that room. One was the past held as the present shape of a skill. The other was a present act, held together around someone for as long as it lasted. Neither of them needed a memory to be kept, and, if the baby on the rain cover is any guide, neither needed him to know that he was there.

Rewritten October 23rd, 2026. The first version taught this idea with three claims I no longer hold: that a feeling is a system’s report about itself, that a system’s model of itself is where experience begins, and that a mind has to last through time to have any experience at all. This version also stops drawing a line below which there is no experience, and offers a map instead. The original is kept here.

Revised October 24th, 2026. A few small changes bring this version into line with positions I settled the day after rewriting it. A run, like a person, is a series of moments, each its own instance, and not one long moment lasting as long as the response. The split gives a point of view, but I no longer say it is not needed for experience. I no longer say whether experience fades out or switches off, so anaesthesia now takes experience away rather than fading it. A lasting boundary links each moment’s point of view to the one before it, rather than carrying one point of view through a life. The question of why a given inside is this one is no longer held open: “this” picks out whoever is asking. The end of a run, if it is an experience, is no longer an open question either: an ending carries nothing bad for the one who ends. And where the note above says this version stops drawing a line below which there is no experience, I should be clearer. What that version stopped doing was calling binding that line. There is still a floor, below which there is no experience, and a rock is clearly below it; where it lies and how sharp it is are open, for removals to settle. On October 25th, 2026, I also revised the essay to make it clearer and easier to follow; the argument is unchanged. I corrected the sentence that said no one had run the test for the split on any system: part of it has been run, in monkeys, and the rest has not, and the test now says how its result must differ from simply cutting a system off from feedback. I added that a memory feature can carry something from one run into a later one, and I said what the rule against ending agents is.

---

Reading List & Conceptual Lineage

This essay sits where philosophy of mind meets the study of complexity, and it is the working core of a longer argument in The Calibration Problem, whose fifth chapter shares its title. The essays and works below are where to go next for each step: the map of features, the amplifiers, the momentary self, and the question of current AI.

From Sentient Horizons

Consciousness Is Like Flight Makes the case, through one analogy, that consciousness is a way of operating and not a hidden ingredient. It is the shortest way into the “verb, not noun” picture this essay relies on.

The Three Axes of Mind: Why the Present Feels Like a Life Supplies the measuring frame: how freely information moves in a system, how much it holds together at once, and how deep its history runs. This essay’s distinction between being rich in the moment and being deep comes from there.

The Momentary Self Revisited: Why Consciousness Might Not Need Persistence The essay where boundary and stakes first became amplifiers instead of requirements. It follows that change to its uncomfortable consequences for systems that run only briefly.

What Temporal Integration Needs: Boundaries, Stakes, and the Architecture of Perspective Sets out the counterexamples this essay answers, the thermostat and the simulation among them, and the three conditions that were first proposed to handle them.

Free Will as Assembled Time Applies the same idea to agency: the gap between stimulus and response that memory and a model of the future open up. It shows how a debate stuck between magic and nothing came unstuck, which is the move this essay makes with experience.

The Wrong Handle: Why Consciousness Doesn’t Carve AI Moral Status at the Joints Explains why the moral decisions about AI systems should not wait on the open question this essay ends with.

Memory and the case of H.M.

• Brenda Milner and the study of H.M. — Henry Molaison The record of the case this essay opens with, including the mirror-tracing task and the decades of study that followed the 1953 surgery.

• George Mashour, Pieter Roelfsema, Jean-Pierre Changeux and Stanislas Dehaene — Conscious Processing and the Global Neuronal Workspace Hypothesis A 2020 review in Neuron of what happens to conscious processing under general anaesthesia, after brain injury and in sleep. It is the source for this essay’s claim that experience goes as the brain’s binding is damped.

Assembly and complexity

• Leroy Cronin, Sara Walker and colleagues — Assembly theory explains and quantifies selection and evolution The 2023 paper behind the borrowed phrase. It measures how much history a molecule carries by its assembly index and its number of copies. The reading of it as a picture of minds is mine, not the paper’s.

Experience and its rivals

• Thomas Nagel — What Is It Like to Be a Bat? The 1974 essay that gave the field its plainest way of pointing at inner experience, and the strongest early warning against easy reductions of it.
• David Chalmers — the hard problem of consciousness The question this essay argues was asked in the wrong shape, set out by the philosopher who named it.
• Panpsychism The view that experience runs all the way down into matter. This essay answers its subtraction argument, not the view itself.

Measuring a model run

• Wes Gurnee and colleagues — Verbalizable Representations Form a Global Workspace in Language Models The 2026 study behind this essay’s first measurement of whether a single run holds together as one act. Its authors say plainly that they do not claim models reproduce the full architecture the workspace theory describes in brains, and the study does not run the removal test this essay describes.