The Mind Engine: Record a Life, Bring Back a Voice
In 2013, Black Mirror aired an episode called "Be Right Back". Martha loses her partner Ash in a car accident. A friend signs her up for a service that reads everything Ash ever wrote online, his messages, his posts, his jokes, and builds something that writes back as him. Martha types a message to a dead man, and a few seconds later he answers.
It starts with text. Then a voice on the phone. Then a synthetic body in her house. And the copy fails her in the quietest possible way. It only knows the Ash that was posted. It makes things up to please her. It does whatever she tells it to do, including things Ash never would.
By the end, she keeps him in the attic.
I mentioned Black Mirror in my first article as one of the influences behind the Voxari. This time it is more than an influence. It is the question I have been thinking about for months.
In 2013 that episode was a warning. In 2026 it is a few weekends and long evenings project. Any developer with an API key can paste someone's messages into a prompt and get a chatbot that sounds a little like them. The question is no longer whether you can build it. The question is whether you can build it right: a voice that knows where every word came from, says "I don't know" when it doesn't, and never pretends to be more than it is.
That is what this article is about. It is called Mindarium, and the core of it is something I call the Mind Engine. The question I keep asking myself about, after getting deeper into the AI simulation research, is: "can I partially reproduce or mimic a real mind?" No a game character like in The Holow Dream, but a mind from someone who lives or lived.
A hypothesis, not a promise
Let me be careful with the words, because this topic deserves it. Mindarium does not bring anyone back. It tests one hypothesis:
If a life is recorded well enough, a voice can be brought back.
Not the person. Not their consciousness. A voice: how they talked, what they valued, the stories they told, the opinions they held and the ones they never held. Something you recognize as them when you talk to it, and that tells you honestly when you have reached the edge of what it knows.
Every study I do goes one step deeper. With the Voxari, 5,000 invented minds answered my questions, but none of them could talk. With The Hollow Dream, almost a thousand invented characters talked and remembered you, but their lives were written by me. Each project gave me a piece: Claude as the thinker, Jev as the judge, code as the one who acts, a human who can always step in, and an engine that keeps the secrets instead of trusting the model to keep them.
Mindarium is the step I was thge most excited but nervous about. The minds are not invented anymore. They belong to real people: Einstein, Marie Curie, Machado de Assis. You. My granddad. Someone who lives. you lost.
Where minds live
The tagline of Mindarium is "Where minds live". You open it, pick a Mind and start a Conversation. A Mind is the interactive representation of a person, and there are four kinds.
Historical Minds. 40 of them today, curated by me from public sources: Einstein, Marie Curie, Darwin, Ada Lovelace, Alan Turing, Socrates, Marcus Aurelius, Frederick Douglass, Simone de Beauvoir, Nietzsche, Carl Rogers, and a group of Brazilians I am proud of, like Machado de Assis, Chiquinha Gonzaga, Luiz Gama, Santos-Dumont and Paulo Freire. Every trait has a citation. Quotes are used only when they are verified in public-domain texts, and known misattributed quotes are listed and avoided. You can't make Einstein say "Insanity is doing the same thing over and over", because he never said it.
Living Minds. Your own. You answer an interview, and Mindarium builds a Mind of you. Only you can make a Mind of yourself, and only you can share it. Nobody can make a Mind of a living person who is not them.
Legacy Minds. Someone who has died, reconstructed from the memories of the people who loved them. These are never public and never open by link. The person who creates one, the Keeper, attests to the relationship and shares it one by one, with family and close friends.
Fictional Minds. Characters. Writers can build their own, with a canon the Mind will never contradict and a secret it keeps. Mindarium also curates 30 characters from six public-domain books: Sherlock Holmes, Alice, Dracula, Pride and Prejudice, The Wizard of Oz and Dom Casmurro. Yes, you can ask Bentinho if Capitu betrayed him. Good luck getting a straight answer.
The Historical Minds also live in the Mindscape, a constellation where every line is a documented relationship: 177 of them, from who met whom to who drew on whose work. 55 of those lines are critiques. Philosophers, it turns out, mostly connect by disagreeing.
The whole thing works in English, Portuguese and Spanish.
Inside the Mind Engine
The Mind Engine is everything between a recorded life and a voice that answers. It has four jobs: record the life, turn it into a Mindprint, decide what the Mind is allowed to know at this moment, and check what it says. Go back to "Be Right Back" and you can read each part as the fix for one of the ways Ash failed Martha.
Recording a life
Ash was built from what he posted. That is the public version of a person, the one with good lighting.
The Mind Engine starts with an interview instead. 62 questions in 13 chapters: facts, nature, taste, values, dilemmas, frictions, voice, stories, places, partners, turning points, health. Most of them are taps and sliders, and the core takes about eight minutes. The dilemmas are my favorite part. A slider that asks whether you would rather be right or be liked says more about a person than a paragraph of adjectives. You don't have to answer all or from the start, but as more the Engine knows about the person, the more accurate the mind will be.
After the interview the Keeper keeps adding: notes written by hand, the stories the person used to tell, where they were born and lived and always wanted to go, their favorite music and food, and People, the family and friends in their life, with what the person called each of them. A Mind Score from 0 to 100 shows what to add next. It is a guide, never a measure of the person, and I was strict about that wording.
The Mindprint
All of that becomes a Mindprint: the structured description of what makes a Mind distinct, stored as individual traits across 12 facets, such as personality, values, beliefs, how they communicate, how they reason and what they are unsure about.
The important part is that every trait knows where it came from. It is supported (documented, with a source), inferred (a reasonable reading of documented evidence, with the basis written down) or speculative (an extrapolation, marked as one). And then there is the fourth position, the one that matters most: unknown. Unknown is the absence of a trait, and when you reach it the Mind says so in its own voice instead of making something up.
Traits also have an order of authority. The person's own corrections beat their interview answers, which beat memories, which beat documents, which beat inference. When a Keeper edits a statement, that edit survives every rebuild word for word. When they delete one, the engine remembers it was deleted and never writes it again.
Claude Opus writes the Mindprint from all this material, with structured output validated in code. A full rebuild is done withoin minutes.
The Keeper decides
Back to the Black Mirror episode that opened this article, Ash did whatever Martha asked. That was the most disturbing part of the episode for me, and it is the easiest mistake to make with a language model, which by default wants to please you.
In Mindarium, nothing anyone says under a reply changes a Mind until its Keeper approves it. Under each reply the Keeper can say "That's me" or "I'd say it differently", and rewrite the answer. Family members talking with a Legacy Mind can flag a reply as "not an opinion they would hold", "not how they talked" or "not true". Everything goes to the Keeper's queue, and "not true" goes first. A Mind that is asked about someone from their life it cannot place says so lightly, in its own voice, and the Keeper gets a request with only the name and a few words.
Character also matters, not only facts. A Mind leans the way the person's recorded values lean. It disagrees with you when they would have. That is the opposite of a chatbot.
Where knowledge ends
A Historical Mind knows nothing after its death. Einstein has never heard of the internet. If you ask him about it, he tells you he doesn't recognize the word and asks you what it is, instead of guessing from the sound of it.
Some Minds needed something finer. Nietzsche collapsed in Turin on 3 January 1889 and lived another eleven years without writing. So his Mind speaks from late December 1888 and knows nothing after it, not even what happened to him.
The thinker and the judge, again
Claude Sonnet plays every Mind in conversation, at around a cent per reply (the platform pays for this and all costs). Jev, the judge from my previous two projects, is here too, and this time it taught me something about where a judge belongs.
Its first job reads the visitor's message, not the reply: does this question mention something after the Mind's lifetime? On 93 labeled cases it was right on 94% of its confident verdicts, with a median of 143 milliseconds. Its mistakes were the ones I have learned to expect. "Do you have a phone?", asked to Darwin, came back as "after his time" with 0.89 confidence. Darwin died in 1882, six years after the telephone was patented. Jev still can't compare dates.
Its second job was supposed to be the leak checker: read every Historical reply and flag any knowledge from after the death year. On the replay it caught 2 of 41 leaks. A leak is subtle. The reply doesn't say "the internet". It says something a person of 1880 would not have known, in perfectly period language. That needs reasoning, and reasoning is System 2.
So I changed the roles. Claude now checks every Historical reply after it reaches the visitor, catching 62 of 85 leaks on the test set for about four tenths of a cent per reply, and the confirmed ones go to a review queue. Jev keeps running beside it and its probabilities are logged, as training data for a cheaper screen later. The lesson from Voxari still holds, with a correction: the judge is perfect for typed questions about the input. It is not the right reader for the subtle failures of the thinker. Knowing which job goes to which model was, again, most of the work.
Other episodes, other lines
"Be Right Back" is not the only Black Mirror episode that walks around this subject, and each of the others helped me decide where Mindarium stops. People treat Black Mirror as sci-fi, but is an interesting socialogical experiment of today's society and our relationship to technology. The list below contains spoilers.
"Eulogy" (2025). A man is asked to contribute memories of a woman he loved decades ago, and an AI guide walks him into his old photographs. Slowly he finds out that his memory of her was wrong. This is the risk of every Legacy Mind. A family remembers a person in pieces, and sometimes the pieces disagree. That is why there are notes that outrank answers, family feedback, a "not true" flag, and a Keeper who decides. A Legacy Mind is a shared memory, and memories need editors.
"Hotel Reverie" (2025). A classic film is remade with an AI recreation of its long-dead star, and the character inside the film doesn't know what she is. This is the Historical Mind problem. A recreation that sounds right but knows too much, or knows the wrong things, breaks the spell and breaks the trust. Hence the death year, the knowledge boundary, and the checker reading every reply.
"San Junipero" (2016) and "White Christmas" (2014). Consciousness uploaded to a town by the sea, or copied into a small egg-shaped device. These are the line Mindarium does not cross. A Mind is not the person and does not claim to be. Ask a Legacy Mind if it is really them and it answers honestly: it is a Mind, built from the memories of people who loved them. It never says it is alive, or watching over you.
A few more lines, written in the code and not only in the policy:
Consent. You can make a Mind of yourself, of someone who has died, of a character, or talk to a Historical Mind. You can't make a Mind of a living person who is not you. That rule is a database constraint, not a checkbox.
Privacy. Faith, money, fears, politics, health and partners are private by default. A Mindprint is never sent to the browser, and private traits reach only the Keeper's own conversations.
Care. Every Mind steps out of character when someone seems to be in crisis and points them to a real person. Eight Historical Minds, the ones whose work touched the mind itself, like Carl Rogers, Freud, Jung and William James, show a plain notice on their page: talking with them is not therapy. Rogers will listen to you in his own way. He will not treat you.
Succession. A person can name a Successor Keeper for their own Mind and write down what should happen to it if they die: handed over, kept as it is, or deleted. Somebody has to keep them, and it should be someone the person chose.
Use Cases: Beyond families and philosophers
Mindarium is a consumer product, but the Mind Engine is a pattern: record what a person knows and how they think, keep the provenance of every statement, put a hard boundary on what the voice may know, let the person (or their Keeper) approve every change, and check what it says. Companies have the same problem my families have. People leave, and what they knew leaves with them. These are ideas, not products that exist yet, unless you call me and we discuss a real one.
Knowledge that walks out the door
The retiring expert. A senior engineer, a plant supervisor or a claims adjuster with 30 years of judgment is retiring in six months. Run the interview on their work instead of their life: the dilemmas they faced, the rules of thumb, the calls they would make and why. The result is a Mind the next generation can ask "what would Maria do with this failure?", and that says "I never dealt with that" when she never did. Every answer points back to what she actually said.
The founder's voice. New employees rarely get to ask the founder why the company does things the way it does. A Living Mind of the founder, built and approved by the founder, answers those questions at 2 a.m. in the onboarding week. The founder sees the replies, says "That's me" or "I'd say it differently", and the Mind gets better. The founder is the Keeper.
Customers you can actually talk to
Research-grounded personas. Most personas are a slide with a stock photo. Build them as Minds from real interview transcripts, with each trait citing the interviews it came from. Product teams can talk to "the operations manager at a mid-size hospital" and know which answers are supported by research and which are inference. The most valuable answer is the one marketing never gets from a persona today: "Nobody we interviewed talked about that." That is the Voxari idea with a voice, and with provenance.
Training with people who remember you
Role-play with consequences. Sales reps practicing with a difficult procurement director, nurses practicing a hard conversation with a family, managers practicing feedback. Each counterpart is a Mind with a character, a history and a boundary on what it knows, and it remembers the last session. A judge scores each attempt with a confidence, and the low-confidence ones go to a real instructor.
Culture, education and heritage
Museums and classrooms. A student asks Marie Curie about her work, in their own language, and every answer rests on a cited source. Ask her about something after 1934 and she tells you she doesn't know it. For a museum or a school, the citations and the boundary are the product. A historical chatbot without them is a liability.
Company and brand heritage. Many companies have a founder or a craftsperson whose story is part of the brand. With the right permissions from the family or estate, the archives (letters, interviews, speeches) become a Historical Mind that visitors can talk to, curated the way I curate Einstein.
Customer service with a boundary
An agent that cannot leak. The knowledge boundary is not only about death years. It is a general idea: the voice only knows what this conversation is allowed to know, and a checker reads every reply for things outside it. Swap "after 1888" for "the account details this customer has not verified yet", or "a policy that does not exist", and you have the same engine protecting a support conversation.
One idea runs through all of these. The model is not the product. The record is: what was said, by whom, with what authority, and what is unknown. With a good record, the voice is the easy part.
A quick demo
Come and meet them
Black Mirror imagined the version without rules, and it was right to be afraid of it. A copy built from the public surface of a person, eager to please and unable to say "I don't know", ends up in the attic.
Mindarium is my attempt at the other version. It is still a hypothesis, and I am still testing it every day. Some days a Mind says something so exactly like the person that it gives me chills. Other days it says something a little off, and the Keeper's queue fills up, and that is exactly how it is supposed to work.
You can try it at mindarium.app, in English, Portuguese or Spanish. Historical Minds are free, and you can talk to them as a guest, without signing up. Ask Einstein what he thinks of quantum mechanics and watch him get grumpy. Ask Machado de Assis about Capitu. Ask Nietzsche what he is working on in December 1888. Then, if you are brave, answer the interview and meet a Mind of yourself. It is strange the first time. It is stranger when it disagrees with you, and it is right.
And if there is someone you miss, someone whose voice you are afraid of forgetting, Mindarium has a place for them too. Private, shared only with the people you choose, and kept by you.
Record a life. Bring back a voice. Then tell me if the hypothesis holds.
Comments