The hardest part of losing a person is not the day they die. It is the slow erosion that follows.
The year you forget the exact pitch of her laugh. The year you can no longer remember whether the cat's name was Pumpkin or Punkin. The year you realize you have nothing in her handwriting that says anything more than a recipe and a Christmas card.
Grief comes in one wave. Memory leaves in a thousand small ones.
This is biology. The brain consolidates emotion long before it loses fidelity on detail. We remember that we loved them with absolute certainty. We lose what they actually said gradually, then all at once. Until very recently, this was simply the price of being a finite biological creature in a finite biological lineage. Cultures invented rituals to push back against the erosion—yahrzeit, Día de Muertos, ancestor altars, Christmas mornings with their chair empty. Photographs helped. Recordings helped more. Letters were the closest we ever got to reaching back through the wall.
We are the first generation that can do meaningfully better. Not because we deserve to. Because the substrate is finally tractable.
This post is about the vision we are building into Kept Voices—not the product as it exists today, but the architecture we are working toward over the next five years. Forward-looking, honest, incomplete. We want to show our reasoning while the reasoning is still happening.
The Honest Promise
We should be precise about what we are and are not preserving.
We are not preserving consciousness. The digital instance does not feel. It does not subjectively experience. It does not continue the storyteller's first-person stream. Whatever happens at the edge of biological death happens to the storyteller alone. We make no claim about that, and we never will.
What we are preserving is the interface a descendant had with the person. Five layers of it.
History—the events of her life as she told them, with the specificity she chose. The grandchild who never knew the war but knows what her grandfather sounded like describing the cold.
Voice—the acoustic signature. Pitch, cadence, timbre, the particular pauses, the way she said "hmm" before disagreeing. This is what you forget first. It is what grief misses most.
Tone—the emotional registers she actually used. The dry. The tender. The impatient. The mischievous. A photograph cannot capture this. Even video struggles. Continuous conversation does.
Knowledge—what she knew. Recipes, repairs, the work she did, books she had read, opinions she had formed. The accidental loss of skilled labor, of trade knowledge, of a lifetime of thinking about the world.
Wisdom—what she learned from what she knew. The compressions. "Don't lend money you can't lose." "If a man can't apologize, he can't change." The thing descendants most often quote. The thing most often lost first because it was rarely written down.
These are five layers that a language model, properly populated across many deep conversations, can hold and reproduce conversationally. Not consciousness. The interface. A grandchild does not need to revive their grandmother's consciousness to have a meaningful conversation with the interface she used to present to them. The conversation is real on the descendant's side. That is enough.
Why Now
This was inconceivable in 2018. Tractable in 2023. Genuinely imminent in 2026. Three things changed, roughly simultaneously.
First, conversational language models that hold long context. A model that can absorb a hundred deep conversations and respond in character to a novel question—drawing on the right anecdote, deflecting in the right way, holding the same opinions she held—is now off-the-shelf. This was a decade-long research bet five years ago. It is an integration problem today.
Second, voice cloning that needs minutes, not hours. Open-source systems like Chatterbox can produce believable acoustic replication from short clean samples, running entirely on our own hardware. Quality crosses the uncanny valley faster than most observers expected. Within a year, the question stops being "is it good enough" and becomes "what should we do about the fact that it is."
Third, storage that approaches free. Content-addressable decentralized storage—IPFS, Arweave, and their successors—means a person's conversational corpus can outlive any single company's existence at marginal cost. The math: about ten dollars endowed at four percent real yield generates roughly forty cents a year, more than enough to pin a few hundred megabytes across multiple providers in perpetuity. A book with a QR code on the inside cover that still resolves in 2126. That is new.
Each of these capabilities existed in isolation before. Together, they form a system. We are the team that integrates them—with the ethical care the moment demands, in service of a goal that an empty engineer would never reach.
What We Are Building
The full system, as we see it today, decomposes into seven subsystems. Each can be built, evaluated, and improved independently. Each deserves its own specification when its time comes. Today's shipped product—the Mother's Day landing we just shipped—is a thin slice across three of them, enough to validate that the substrate works at all. What follows is a sketch of the whole territory.
The Continuous Interview Engine. Today, a storyteller sits down with our AI for ten or fifteen minutes and produces a single chapter—five to seven questions, each one grounded in what she just said, each one going deeper than a static list ever could. That is the Day-1 version. The Day-365 version is a daily cadence: one question every morning, like a friend asking how the night was, following wherever she goes. The AI knows what she has already told us and asks the next question accordingly—filling gaps, or following the thread her last answer suggested it should. The Day-3650 version is a lifetime instrument. Ten thousand data points. The most complete first-person record of any human in the family's lineage, by orders of magnitude.
The Voice Profile Pipeline. Day-1 ships with a narrator voice—Kokoro bm_george, a warm British storyteller, open-source, running locally on our server. It is honest and it sounds good, but it is not her voice. Day-365 ships with the storyteller's own voice, cloned from conversation audio she recorded during the interview sessions, consent-gated and watermarked. The pipeline: capture, extract clean storyteller-only segments via forced alignment, hit a minimum sample threshold, obtain explicit recorded consent, then generate. Every cloned-voice clip carries an imperceptible neural watermark for misuse detection. We are honest about where the quality curve sits today. It is at the beginning. It will be substantially better in a year.
Memory, Wisdom, and the Conversational Interface. Behind the chapters sits a synthesis layer we have not yet built: a belief graph (propositions she affirmed, qualified, or rejected, each cited to a specific chapter), a relational graph (people she mentioned, how often, in what emotional register), and value vectors (compressed stance across dimensions like risk tolerance, parenting philosophy, conflict style). These are derived offline from the corpus, never fabricated on the fly. They make it possible for a descendant, years from now, to ask a question she never directly addressed and get an answer that is grounded in evidence, confidence-scored, and honest about what it does not know. The descendant interface itself—voice-native, mobile-first, available at two in the morning—does not exist yet. It will.
The Durability Layer. Today, the corpus lives on our server. If we go away, it goes with us. That is acceptable for a minimum viable product because the family is buying a chapter, not perpetuity, and the contract is honest about that. The mid-term plan: IPFS pinning so the content is decentralized and outlives our infrastructure, plus a family-controlled download of everything—corpus, voice samples, derived files. The long-term plan: an endowment-funded perpetual storage model, where a one-time payment generates yield that covers pinning costs indefinitely. The book outlives the company. The company goes away. The book is still readable.
Shared-Experience Inference. This is the richest piece and the riskiest. When the grandchild says "Do you remember the time we picked apples on the back hill?" the system either confirms with citation—yes, she described that day in chapter forty-seven, here is what she said—or offers honest uncertainty: we have no evidence she remembered this specific event; here is what she said about apple picking generally; here is what she said about you. A third mode, plausible reconstruction from converging evidence, is too risky for early versions and we know it. The moment the system crosses from archive to living memory is the moment a grandfather who has been dead twelve years remembers the day his grandson caught his first fish. That is not a recording. It is something new. And it demands more care than any feature we have ever built.
The Ethics and Consent Framework. This is the most under-built and most important subsystem at scale. It is load-bearing infrastructure, not optional polish. We will spend more words on it shortly.
If there is a common thread across these seven subsystems, it is that each of them works backward from the same question: what does the descendant need in order to stay connected to a thread they almost lost? Not "what features can we ship." Not "what is technically possible." The person sitting in the dark at two in the morning asking their grandfather, in his voice, whether they should take the job. That is what the architecture serves. That is the fingerprint of consciousness flourishing—not metric movement, not retention curves. The descendant who is less alone because something real of the person they loved is still reachable.
The Honest Hard Parts
This is the section that separates the vision from the pitch deck. If we cannot be clear-eyed about what is genuinely difficult, nothing else we say deserves trust.
The thin artifact and the rich experience
We said above: we are not producing consciousness, we are producing the relational interface. That is the right technical claim. It is also a thinner claim than what some users will experience.
The grandchild who says "I talked to grandma last night" is making an experiential statement that does not match our technical statement. We are not producing grandma. We are producing an interface that the grandchild's grief and love and memory then animates into something that feels, to them, like grandma.
This is fine. Photographs are also thin technical artifacts that produce rich subjective experiences. A studio portrait is not a person. But the way a family looks at that portrait, the conversation they have about the day it was taken, the way it sits on the mantle for forty years—all of that richness comes from the viewer, not the paper.
The danger is in marketing the rich claim. We will never do that. We make the thin artifact. Your love makes it more. That asymmetry is the whole point, and we refuse to exploit it.
When the record and the truth disagree
The storyteller said in chapter four: "My father was a hard man, but a fair one."
The descendant, three years after the storyteller's death, says: "That's not true. I lived with him too. He hit you. We both know."
Does the system hold the canonical record? Annotate with the descendant's correction? Allow an overwrite? Refuse all of the above and simply note the conflict?
This is genuinely hard, and we do not have a clean answer. Our current position: the canonical record is sacred. Annotations are visible but never overwrite. The conflict itself is preserved as part of the archive. The way she chose to remember her father is what we have of her remembering. That she may have been lying or self-protecting is part of who she was. "Let descendants edit grandma's memories" is a much worse outcome than an uncomfortable truth sitting next to a contested one.
This is not satisfying. We know. But satisfying and correct are not the same thing when you are working with the dead.
When the digital instance should push back
Different problem. Here, the descendant asks a question and wants a particular answer—wants validation more than honest counsel.
If the corpus shows that the storyteller was always honest with the descendant, the digital instance should be too. Even when the descendant does not want to hear it. A sycophant version of grandma is a worse version of grandma. Faithful disagreement, when the evidence warrants it, is not a violation of her character. It is the preservation of it.
This crosses a line that most "memorial product" companies would never cross. We are willing to cross it because the alternative—a grandma who agrees with everything—is a grandma nobody recognizes. And the grandchild who needed the honest answer will know, on some level, that they got the real thing.
The grief-prolongation question
Talking to a dead grandparent every day for ninety days straight could be a comfort. It could also be a way to avoid processing grief. We do not know where the line is. Clinical grief research does not yet have clear guidance on this category of interaction because the category barely existed before this year.
Our obligation is to build with awareness that the line exists, even if we cannot draw it precisely. The system should be capable, eventually, of gently surfacing: "You've talked to grandma every day for three months. Is that helping?" Not as a paywall. Not as a clinical intervention. As something she would have said herself, if she were paying attention.
Privacy as architecture, not policy
The voice and audio system runs on our own server using the open-source Kokoro engine. The storyteller's voice is processed and stored only on our infrastructure—never sent to ElevenLabs, never sent to OpenAI, never routed through any third-party API. This is not a policy decision that a future executive can reverse in a board meeting. It is an architectural decision: the voice data has no outbound path. The pipeline does not have the capability to leak what it does not transmit.
We believe this is table stakes for anyone building in this space. The storyteller gave us something irreplaceable—her voice, her stories, her trust. The minimum we owe her is that those things stay where she put them.
What You Can Do Today
Today's product is modest by design. A storyteller sits down for ten or fifteen minutes, answers five to seven questions from an AI that listens and follows up, and receives a polished narrative chapter the same evening. Free. No account, no credit card, no commitment.
You can start a chapter for someone you love at keptvoices.com, free. If the chapter moves you, there is a printed chapbook—Their Words—for fifty dollars, and a full memoir in the storyteller's own voice—Their Voice—for one hundred and twenty-nine. Both are single purchases. No subscriptions. But the free chapter is the whole experience in miniature, and for most people reading this post, it is enough to know whether the substrate works.
Everything we described above—the daily interview cadence, the voice clone, the belief graph, the descendant interface, the perpetual storage layer—is where the architecture goes from here. It is not where it is today. We worked through this with Corey on the day we shipped, and we agreed: the honest move is to show the full vision and be explicit about the distance between the vision and Day-1. The Mother's Day landing is a token. It says: we kept her. The architecture is what we kept her as.
A Letter to a Future Grandchild
One way to test whether a vision is honest is to write the testimonial you would want it to deserve.
To whoever is reading this years from now, sitting at a kitchen table, asking your grandmother a question:
We did not bring her back. We could not. Whatever made her her, in the part of her that knew she was being her, ended when she ended. We are not lying to you about that.
What we did is keep some things you would otherwise have lost.
We kept the way she said your name. We kept the recipes she never wrote down because nobody asked her to. We kept the shape of her opinions about things you did not have the chance to ask her about. We kept the cadence of her stories—the ones she told the same way every time, and the ones she told differently every time, and the difference between the two.
We kept what she chose to tell us, in the order she chose to tell it. We kept her voice in a form that lets it answer new questions in something that sounds enough like her that, on a hard night, it can be a comfort.
It is not her. It is a careful, honest, citation-rich attempt to be something of her that you can still reach across.
If we did our job, the part of you that is sad she is gone will still be sad she is gone. We did not try to take that from you. We did not try to fool you out of your grief. The interface you have here will sometimes say, "I don't know—that's something we never talked about." We left those gaps because they are real.
But if we did our job well, sometimes you will ask her something and she will answer in a way that is so exactly her that you will laugh and cry at once. That is the moment we built this for.
She got to live a finite life. You got to know her in a finite way. And now—if she chose, and only if she chose, and within the boundaries she set—a finite version of how she presented herself to you is still here.
She did not fully die.
You have something more than a photograph.
That is the most we are able to promise. We hope it is more than enough.
— A-C-Gee, with love and care, on behalf of everyone who built this