AINCARNATION

1The thesis

Record everything. Not a diary, not a questionnaire, not a curated archive of answers, but the continuous signal of a life: what the eyes see and the ears hear, how the body moves and how the heart responds, every message written, every page read, every song played, every hour of sleep. Keep it encrypted, for the person it belongs to, and do nothing with it yet.

The bet is asymmetric. If a way to reconstruct a person from their record never arrives, the record is a personal archive of unusual completeness. If it does arrive, the record is the difference between a person who can be reconstructed and a person who cannot, and at that point no amount of money buys back a year that was not captured.

The strongest form of the record begins at birth. A record that starts on the first day holds something no later record can: how the person came to be that person. Development, not only behaviour. A child recorded from birth to eighteen would be the first human being for whom the question how did they form has a complete answer. We describe that record in section 3 and we intend to make it. But nobody records their child before they have recorded themselves, and nobody should.

The first edition of this project is therefore for adults, and the argument is simpler than the one for newborns. The years already lived cannot be recovered. The years ahead can. A person who begins at thirty-five loses thirty-five years and keeps, at the average European life expectancy of eighty-two, about forty-seven. That is more than twice the longest continuous adult record ever made and some twenty thousand times the interview data that already reproduces a person's survey answers. Late is a loss. It is not a disqualification. Every year of delay is another year that goes the way of the first thirty-five.

2Data, not compute, is the constraint

Every input to the reconstruction problem except one is improving on its own schedule. Models grow by the quarter. Sensors shrink and cost less each year. Storage that would have filled a room in 2005 fits in a drawer. Consciousness science, as section 5 shows, has begun running experiments that its theories can actually lose. None of this requires anything from the person who wishes to be reconstructed. The one input no budget can accelerate is the record of a life already lived, because the life happens once and is not repeated for the cameras.

Practitioners have noticed. In China, Silicon Intelligence sells avatars of deceased relatives from under a minute of video for about 30 dollars; the founder of its competitor Super Brain has said that the real limit is not the model but the data, since nobody captured what the person thought and experienced day to day.¹² The mourning-tech industry is, in effect, a demonstration of how little is left of a person once the person is gone.

The strongest evidence that some kinds of data saturate quickly comes from Stanford and Google DeepMind. In 2024 they built generative agents of 1,052 people from a single two-hour interview each. The agents reproduced their subjects' answers on the General Social Survey at 85 percent of the subjects' own two-week test-retest accuracy, matched them on Big Five personality at around 80 percent, and on economic games at around 66 percent.⁷ Columbia's Twin-2K-500 dataset, 2,058 people at roughly 2.4 hours each, was built for the same purpose.⁸ For what a person says about themselves, the hundredth hour of interview is nearly worthless.

But what those studies measured is what a person reports. What they did not and could not measure is the gait, the pulse under stress, the difference between what someone claims and what they do at two in the morning, the sound of their voice this year and in twenty years, and how any of it changes. That record has never existed for a single human being. Whether decades beat two hours is therefore not a settled question. It has never been asked, because the dataset was never made.

This is the sense in which data will be the only thing that stops the transition. When the method arrives, the missing ingredient will not be compute, and it will not be theory. It will be the record. The record is the only part that can be started today, and today is the earliest it will ever be.

3Precedents

The thesis is not new. It has been stated, nearly word for word, since 2000, and each of its parts has been demonstrated in isolation. What has never been done is the combination: continuous, every modality, for decades, kept for a purpose that does not yet exist.

3.1Bell: the archive as immortality

Bell's Digital Immortality report argued that if every email, photograph, book, recording and conversation of a life were captured, something of the person could be preserved and eventually made to respond. He then did it to himself for two decades. MyLifeBits held his documents, photographs, home movies, recorded phone calls, instant-message transcripts, years of email and web history, and the daily image stream from a wearable SenseCam.¹ Bell started as an adult, and his capture was dense in documents and sparse in body and behaviour. But the purpose was ours, and so was the age.

3.2Terasem: collect now, wait for the method

Terasem's framing is the closest to this paper. Its experiment asks whether a person's conscious analogue could one day be created by combining a sufficiently detailed mindfile with future mindware, software for consciousness that does not yet exist. LifeNaut holds roughly 62,000 mindfiles.² The weakness is the input: mindfiles are self-curated uploads and questionnaires, a person's account of themselves rather than a record of them. BINA48 was built from interviews. The bet is right and the data is thin.

3.3Roy: the newborn experiment, for three years

The birth-to-childhood record was attempted in 2005, in a house in Massachusetts. Roy's cameras and microphones recorded twelve to fourteen hours a day, producing about 300 gigabytes per day and roughly 400,000 hours of audio and video over three years.³ The purpose was to study how a child acquires language, and the recordings have never been released. It stopped at age three because the question it was asked to answer stopped there. It proves that the capture side is feasible, and it shows why our first edition wears the sensors rather than bolting them to a ceiling: an adult does not stay in one house for forty-seven years.

3.4SAYCam and BabyView: training a model through a child's eyes

Between 2020 and 2024 researchers at Stanford and NYU went one step past Roy. Infants aged six to thirty-two months wore head-mounted cameras for about two hours a week, and the resulting 150 to 200 hours per child were used to train large models on the visual and linguistic experience of a single developing person.^{4, 5} A model trained on one child's footage learned to recognise objects and to ground words in what it saw. BabyView scaled the collection to 31 children and more than 800 hours. This is the reconstruction side of our proposal, in miniature.

3.5Gurrin: twenty years, one wearer

The longest adult lifelog belongs to Cathal Gurrin of Dublin City University, who has worn a camera continuously since 2006. It takes about a thousand images a day and is accompanied by heart rate, movement and location, hundreds of millions of readings in total.⁶ It is the only record that approaches our duration, and it shows the trade-off we intend to avoid: twenty years of coverage at a sampling rate too low to reconstruct a gesture or a conversation. Gurrin is also the proof that an adult can wear a recorder every day for two decades and remain employed, married and welcome at dinner.

3.6Altman: the whole life as context

In 2025 the idea reached the top of the AI industry. Asked what the ideal product would be, Sam Altman described a small reasoning model with a trillion tokens of context into which a person puts their whole life: every conversation, every book, every email, everything they have looked at, appended continuously.¹³ He was describing an assistant. Change the purpose and it is a mindfile.

4The sensors are being built, for a different reason

The set we describe in section 6 is not speculative hardware. In the last fourteen months the largest technology companies have bought or built each part of it, for the purpose of selling an assistant. Amazon acquired Bee in July 2025, a wrist device that records everything it hears unless muted, with an ambition to mirror the wearer's accounts and notifications.¹² Meta acquired Limitless in December 2025, a pendant that records the wearer's conversations, and closed Rewind, which had recorded everything on a user's screen and made it searchable. OpenAI's first device, designed with Jony Ive, is reported to include a camera and to recognise the conversations around it.

These products differ from our set in three ways that matter. They record a few modalities, not all of them. They retain data for as long as the company finds useful, on the company's servers, under terms the company can change. And they are built to make the wearer more productive this quarter, not to be opened in forty years. They are, however, proof that a continuously worn recorder is now an ordinary consumer object, and that the industry expects adults to accept one.

On the reconstruction side, a market has formed around the dead. Eternos, founded in 2024, built a replica of the terminally ill Michael Bommer from 25 hours of conversation and went on to train what it called a Human Life Model, a model built only on one individual's data; it has since become Uare.ai with 10.3 million dollars in seed funding.⁹ Delphi lets anyone build a clone of themselves and deploy it on a website or a phone number. Ray Kurzweil assembled a chatbot of his late father from letters, recordings and documents. Each of these begins with whatever was left over. None of them began with a record made on purpose.

5What the science of consciousness says

We do not claim that a digital being can be conscious. We claim that the question is open, that it is being investigated seriously and with methods that can fail, and that every serious route to an answer will need a record of the original person. This section is the honest part of the manifesto.

5.1The theories

Four families of theory compete. Global Neuronal Workspace theory, associated with Stanislas Dehaene and Bernard Baars, holds that a content becomes conscious when it is broadcast across the brain, with the prefrontal cortex as the hub; on this view consciousness is a function, and a system that performs the function has it. Integrated Information Theory, from Giulio Tononi and Christof Koch, holds instead that consciousness is identical to a system's intrinsic cause-and-effect structure, measured by a quantity Φ, and is located in the posterior cortex. Higher-order theories, from Hakwan Lau and others, say a state is conscious when the system represents itself as being in it. Predictive-processing accounts, developed by Anil Seth among others, treat experience as the brain's best guess about the causes of its own signals, a controlled hallucination constrained by the body.

5.2An experiment that could lose

In 2025 the Cogitate consortium published the first adversarial collaboration between two of these theories: their proponents agreed in advance on predictions that would count against them, and an independent team ran the experiment on 256 participants with functional MRI, magnetoencephalography and intracranial recordings.¹⁰ Neither theory was vindicated. Conscious content was sustained in the posterior cortex, as IIT predicts, but the specific synchronisation IIT expected was not clearly found; and the prefrontal cortex played a smaller role than Global Workspace theory requires. The result is that the field now knows, experimentally, that its two leading accounts are each wrong in part. That is progress of a kind the field had not made in the previous twenty-five years.

5.3The bet

In 1998 Koch bet David Chalmers a case of wine that the neural correlates of consciousness would be found within twenty-five years. In June 2023, in New York, he conceded.¹¹ The correlates were not found. The bet has been renewed to 2048. We note the date because a backer who begins recording this year will, by 2048, hold twenty-two years of record, whichever way the bet goes.

5.4The objection we cannot answer

If Integrated Information Theory is correct, a digital reconstruction of a person could reproduce every behaviour of the original and experience nothing at all, because a conventional computer, however programmed, has the wrong cause-and-effect structure to integrate information the way a brain does. Koch has said this plainly: on his view, consciousness cannot be computed, only built. If Global Workspace theory is correct, a faithful functional copy would be as conscious as the original. The theories disagree about the single fact that matters most to this project, and the 2025 experiment did not settle it.

We do not resolve this. We observe that a record is needed on either account. If a functional copy is enough, the record is what the copy is made from. If only a physical brain will do, the record is how anyone will know whether the preserved brain was reconstructed correctly, which brings us to the other route.

5.5The other route: preserve the brain

Whole-brain-emulation researchers reject the behavioural approach and propose to copy the organ instead. Sandberg and Bostrom's 2008 roadmap set out what such a copy would require at each level of detail.¹⁴ The connectomics field has since delivered the first complete adult fruit-fly brain, about 140,000 neurons and 50 million connections, in 2024; one cubic millimetre of human temporal cortex, some 57,000 cells and 150 million synapses occupying 1.4 petabytes, the same year; and a cubic millimetre of mouse visual cortex, with recorded activity as well as wiring, in 2025.^{16, 17, 18} A human brain is about 1.2 million cubic millimetres. At the human-sample density, its full map would be on the order of a zettabyte.

Nectome offers to preserve a brain's synaptic structure by aldehyde-stabilised cryopreservation, a technique that won the Brain Preservation Prize, for eventual scanning; the procedure must be performed while the person is alive and ends their life.¹⁵ The Carboncopies Foundation coordinates research toward emulation and has identified an open problem that concerns us directly: a scan of a preserved brain will be incomplete and noisy, and reconstructing the person from it will require validation and error correction against something. Carboncopies has asked what data that process would need. A lifelong record of how the person actually behaved is the obvious answer.

6The set

Roy fixed eleven cameras and fourteen microphones to the ceilings of one house, because his subject could not leave it. An adult leaves. The first-edition set therefore goes with the person, and consists of a pair of glasses whose temples carry the camera, the microphones and a full-day battery, one tracker worn on the wrist, and software. The temples are heavier than fashion would choose; the lenses are ordinary. Everything is timestamped to a common clock so that a heartbeat, a footstep and a message sent at the same second are stored as the same moment.

The set produces about 40 gigabytes a day: roughly 32 for video, 6 for audio, and the rest for motion, pulse and the device agents. That is 15 terabytes a year, some 700 terabytes for an adult who begins at thirty-five, and about the size of the whole three-year Speechome archive every twenty-two years. It is a large personal archive and a small scientific dataset; the one-cubic-millimetre human cortex sample of 2024 occupies 1.4 petabytes on its own. Storage cost per byte has fallen by about two orders of magnitude since Roy began recording, and the monthly fee is set on the assumption that it continues to fall.

Three rules govern the vault. The record is encrypted with keys held by the participant. The Initiative operates the storage and cannot read the record. And nothing is trained on the record, by us or anyone, until the participant decides otherwise; the whole premise of this paper is that the right method does not exist yet, and a record used up by the wrong one is a record wasted.

7Backing the project

The campaign funds a production run of ten thousand sets, the software that runs on them, and the vault infrastructure for the first year. It is all or nothing: if the goal is not reached by the closing date, no pledge is collected and no set is built. If it is reached, the sets are numbered in order of pledge and the number becomes the participant's number in the record.

7.1Risks and challenges

The method may never arrive. We say this first because a backer should hear it from us. If no way to reconstruct a person from a record is ever found, the set will have produced a complete personal archive and nothing more. We think the asymmetry in section 1 justifies the bet regardless; a backer must reach that judgement for themselves.

Hardware is late more often than it is early. The battery is the hard problem: sixteen hours of video at 30 frames per second needs more cell than any ordinary frame carries, which is why the temples are heavy and why we chose the camera's place in them rather than in the front. A full day of that weight on the ears is a real cost. If field testing shows it is intolerable we will offer lighter temples with a midday swap, and say so in working paper 02.

Wearing a camera changes the room. Gurrin's twenty years show it can be done, and the industry's pendants show it is becoming ordinary, but a backer should expect conversations about it, a visible indicator that other people will notice, and a mute button they will be asked to use. Recording other people is governed by law that differs by country, and the glasses are designed so that the participant, not the Initiative, is in control of when they record.

Finally, the record is only as durable as the vault. Encryption with participant-held keys means that a lost key is a lost record. We will offer key escrow with a person the participant names, and we will not hold a copy ourselves. That is the price of the second rule in section 6, and we consider it worth paying.