Sit down with someone you love and the things they've kept. Record them telling you, in their own voice, while they still can. Tag each thing. One day, someone touches it to a phone and hears them.
Arthur Penhale is made up. We wrote his words, and a stock synthetic voice reads them, because we won't use a real person's voice to sell this. In the real thing, this is their own recording. Everything else here is real output from our prototype, and the questions you type stay in your browser.
Tap a thing, as you would with a phone.
The same code runs in your browser and in our MCP server, so this is exactly what an agent would receive. No tool in that server can generate speech.
Who sets the rules
Off, because he said no. A person's "never" counts as much as a yes, and nobody can overrule it later. The server has no tool that could do it.
He gave it to Nell and her mother. It stops the asking and the playing here. It never deletes the files, because losing them could hurt more than keeping them.
What you keep
Original recordings, untouched. Timed transcripts in plain text. A web page that plays everything with no internet. Each file has a checksum, so you can prove nothing changed, in a format librarians have used for years (BagIt).
This is Arthur's whole archive, built by our prototype from a folder of recordings. Download it, unzip it, and open data/index.html.
shasum -a 512 -c manifest-sha512.txt
Keep three copies, in different places. No disc or cloud account lasts forever, so the plan is to check the copies and move them, not to trust one thing.
Why the files have to be yours
We read what we could about the companies that tried this, and about the research. These are real cases, with sources.
The closest thing to this idea: record stories, then ask an app and hear the real voice. By July 2026 its homepage said it was shutting down and told people to email for their recordings. The page now returns an error. We couldn't find out whether those emails get answered.
Video conversations with recorded people, including after death. It filed for Chapter 11 in May 2024, and a creditor bought the assets. It now leads with museums, and consumer sign-ups are closed. Its current terms offer 30 days to retrieve content after termination. We found nothing on what happened to families' archives.
USC Shoah Foundation's Holocaust survivor interviews: ask a question, and it finds and plays a pre-recorded answer, unaltered. It takes about five days per person and is built for education, so it isn't evidence about family grief. It's the strictest version of the rule we'd follow: never put new words in anyone's mouth.
At a 2025 sentencing, a family played an AI video of a dead man forgiving his killer. On 30 September 2026 the state's appeals court vacated the sentence. The news reports describe the panel separating a recording from a depiction someone imagined. An appeal is reported, so this isn't final. The family meant well; that's the hazard.
In 2010, a UK research project let people record audio stories about donations to a Manchester Oxfam shop. Sales rose 53% that week. That's the universities' own number, with no control group. It's a hint that a story changes what an object is, and the platform didn't become a business that we can find.
A nonprofit helps dying parents record their voice for their children. In a survey, 57% reported mixed feelings about listening, and adults said most children listen rarely or only sometimes. So a voice can matter without being played often. We wouldn't measure this by how much anyone listens.
The honest order
Models will keep improving. Nobody can go back and record a person who's gone.
Agree first, on tape. Who can listen, what stays private, and what must never be done with their voice. Asked again at every session, and it's theirs to stop.
Pick up one thing and wait. Two or three sessions of 30 to 45 minutes, a few objects each. Two small microphones, one per person, so nobody's words get credited to the wrong voice.
Tag it. The tag holds an address and a name, never the story. Print a QR code beside it, because tags fail and phones change.
Ask, and hear only what they said. Their words with a clip you can play, or a plain "he never talks about that."
A generated voice, maybe, and only if they chose it. Separately, on camera, while they can. Clearly labeled, adults only, and off unless they said yes. Mainstream cloning services generally need the speaker's own live verification, so this can't be decided after they're gone. We don't know yet if anyone should want this, and a good archive doesn't need it.
What this can't do
Before we build any more
We're finding out whether this should exist, and for whom. We read every answer, and you'll hear back from a person, if at all. Nothing gets sold or shared.