Record a person's answers, their voice, and their stories while they are here — so the people who come after can still ask, and still hear.
Somewhere in a shoebox there is a cassette tape of your grandmother's voice. Nobody has a cassette player anymore.
That is the problem memoria solves. All of us are leaving behind less and less that lasts — our photos live in cloud accounts that will close, our messages in apps that will shut down. And the stories, the real ones, the ones someone would actually want to hear again in fifty years? Most of those are never recorded at all.
memoria is a small, quiet program that fixes this. A person sits down while they are alive, answers questions in their own words, lends their voice — and leaves behind a memorial that can still speak as them, long after.
It runs entirely on a small computer in your home. Nothing goes to any cloud. Nothing leaves the house.
They answer questions. Eighty gentle questions in eight chapters of a life — roots, growing up, love, work, wisdom, regrets, joy, and farewells. Where did you grow up, and what did it feel like? What made you laugh hardest in your whole life? What do you want the people you love to know that you have not said?
They can skip any question. Stop anytime and pick the thread back up next week. Every word is saved the moment it is spoken.
They lend their voice. A short reading script — a few sentences designed to capture every sound a voice makes. The program measures the voice: its pitch, its pace, its energy. Then it learns to speak in that voice's tone, so the memorial doesn't just say the words — it says them the way they said them.
They write letters. Sealed ones. To a granddaughter, to be opened on her 40th birthday. To a son, to be opened only when the family executor decides the time is right. The program keeps them sealed — truly sealed — until the person's own rule says otherwise.
A quiet page, reached by a QR code on a physical marker. They type a question:
"What was your earliest memory?"
And the memorial answers — in first person, in their grandmother's own words, in her voice's tone, with a small note showing exactly which of her recorded answers the reply came from.
If the archive doesn't hold an answer, the memorial says so, gently: "I never spoke about that, so I can't tell you — but I'm glad you asked."
The memorial never invents. This is the heart of the whole project, built into the code and checked by machine at every answer: it may only speak from what the person actually said. It will never guess, never fill in, never make up a memory. A memorial that invents is a lie about a person — that is the one thing memoria will not do.
Consent is real. Every answer is marked public, family, or private. Visitors see only what was meant for them. Private words stay private — even from the memorial itself.
Nothing leaks. Sealed letters stay sealed in every backup and every export. Nothing is sent to any cloud. The whole thing lives on your own hardware.
It outlives every format. Alongside everything else, memoria writes a "century file" — plain Markdown text of the person's answers. It needs no software, no accounts, nothing but the ability to read. If every computer dies, the century file is the gravestone.
- The interview — 80 questions, pause and resume, every word saved as it's spoken
- The archive — checksummed records that detect tampering, with JSON and Markdown exports
- The voice — a reading script, quality gates on every clip, a measured voice-print, and a speaking pipeline (Kokoro/Piper today; NeuTTS Air zero-shot cloning when its access gate opens)
- The mind — retrieval from the person's actual words (EmbeddingGemma 2), answered by a local LLM (Granite 4.0-3B), verified by a grounding checker that refuses rather than risks inventing
- The gates — sealed letters, consent tiers, a real executor key
- The keeper's tools — one-command backups, restore that verifies itself, and a migration runbook for the day the hardware dies
Designed for a Jetson Orin Nano 8GB — a $250 computer the size of a candy bar running the whole memorial on 8 gigabytes of memory, alongside a RAG embedding model and an agent, with room to spare.
python -m memoria interview --person "Ruth" # the guided interview
python -m memoria export --person <person_id> # the century file
python -m pytest tests -q # 155 tests, all green
uvicorn memoria.memorial_server:app --port 8093 # the memorial itselfThe full operations guide is in RUNBOOK.md; the promises in detail are in
ETHICS.md; the 20-step build plan with one commit per step is in
PROMPTS.md.
Nothing here is a simulation of a person. It is a record, given willingly in their own words — kept safely, and allowed to speak.
Light the candle. Sit down. Ask.