An avatar for the people who love you
The MemoriesBox digital self is a recorded archive that can be asked a question and answers aloud in the storyteller's cloned voice, from what that person actually said. It says what it is on screen in every conversation, and today it is the storyteller's own to talk to.
Facts on this page verified
What it is
A recording plays. An avatar answers.
That is the whole difference, and it is larger than it sounds. With a recording, you have to already know what you are looking for, and you have to find the minute it is in. With an archive that can be asked, someone types or says a question and gets an answer drawn from the parts of the interview where that was discussed, read aloud in the voice it was said in.
Who can ask, today: the storyteller, in their own app, at any time — which is where the manner gets checked, because nobody knows better than you whether it sounds like you. Heirs receive the recordings, transcripts, letters and tree when the archive is released; the conversation opens to them in the version we build next, with the storyteller's separate consent to it. We would rather say that on this page than let a family find it out in a bad week.
It is a retrieval system with a voice on it. It is not a person and nothing on this site will describe it as one. It is a recording that can be asked questions, which is remarkable enough without dressing it up.
How it works
- Someone asks a question, spoken or typed, in any of the five site languages.
- The archive is searched for the answers where that subject was actually discussed — the words the person said, weighted so that a rare word they used outranks the label on the question they were asked — and a slice is reserved for the answers that carry their manner: the jokes, the habits of speech, the character.
- A reply is written from those passages by a language model, in the first person, under instructions that hold it to the recorded words: use their names, places and turns of phrase; ground every claim about the life in the passages; do not invent hard facts. Alongside the passages it has a character brief, distilled once from the whole archive, made of things the person actually said, quoted.
- The reply is spoken in the voice model built from the person's own recordings. See /features/voice-cloning.
- If it was never discussed, it says so, in character. The instruction is literal: answer gently — "I don't quite remember that…" — and offer what you do recall. If nothing has been recorded at all, it says it is still gathering its memories and asks you to record a few stories first.
Three honest limits. The reply is composed, not quoted: the substance is theirs, the exact sentence may not be, which is why the screen says what it is. The reply does not yet link back to the recording it drew from, so hearing the original is a separate step in the memories page. And it answers in the language of the account — a question in Spanish about something recorded in Ukrainian gets a Spanish reply — without marking that reply as translated. All three are on the list.
Step five is where most of the argument sits. A system built to always have an answer will produce one, it will be plausible, and a family in the first year of grief is not well placed to spot the difference. Ours is told to run out of material, because a person runs out of material. There are things your mother never said, and no software should decide what they were.
The category that nobody occupies
Of the eleven life-story services we compared in August 2026, none clones the storyteller's real voice and none lets the family ask a new question later. Prices verified on 17 August 2026.
| Service | Price | Keeps the audio | Clones the voice | Family can ask something new |
|---|---|---|---|---|
| StoryWorth | $59 / $109 / $199 a year | Yes | No | No |
| Remento | $99 a year | Yes | No | No |
| Storii | ~$99 a year | Yes | No | No |
| Meminto | $99 / $149 / $397 one-time | Yes | No | No |
| HeritageWhisper | $39 a year | Yes | No | No |
| Willow Stories | €299 first year, then €29 | Not stated | No | No |
| Simirity | Free / $6 / $9 a month | Yes | No | No |
| LifeTales | ~$239.88 a year | Yes | No | No |
| trylifestory | $595 one-time | Video | No | No |
| Memoirji | Not published | No, voice in and text out | No | No |
| Autobiographer | $99 a year | Yes | No | No |
| MemoriesBox | $139 a year | Yes | Yes | Yes |
Two products outside that list do answer questions. Avataari runs a guided interview, a voice clone from a vendor it does not name, and a text, voice or video avatar, from a free tier and $10 in credits. Afterlife AI sells a consent-first digital twin with an Executor Lock, free to $29.99 a month, from New South Wales. Both do real work, and Afterlife AI's public documentation on digital replica law is better than ours.
What neither offers is a European jurisdiction. MemoriesBox is the only one of these hosted in the European Union. The full table, including the AI-legacy products, is on /compare.
One more honest note. StoryFile did something adjacent first, using pre-recorded video answers routed by AI, and filed for Chapter 11 on 5 May 2024 owing $150,000 in customer refunds. Being first here has not predicted being here later, which is why /trust/what-happens-if-we-shut-down is written the way it is.
Every conversation is labelled as AI
Since 2 August 2026, EU AI Act Article 50 has required that people are told they are interacting with an AI system "in a clear and distinguishable manner at the latest at the time of the first interaction", and that synthetic audio is "marked in a machine-readable format and detectable as artificially generated or manipulated".
That binds us directly, and we would do it anyway. A conversation in a parent's voice that does not say what it is, is a trick. The disclosure is not a footnote at the bottom of a screen. It is a line under the conversation heading, every session, in all five languages: this is a digital imitation built only from your own recordings; it is not a person, and it will not know anything you did not tell it. The model is told never to break character to say it is an AI, which is precisely why the interface says it instead. Generated audio is stored with metadata marking it as AI-generated. The plain-language explainer for families, written from inside the EU, is on /trust/eu-ai-act-article-50.
What it will not do
- It will not answer from what it guesses they would have said. It answers from what was recorded. If the subject never came up, it says so.
- It will not have opinions about the present. It cannot tell you what your father would think of your new job, your divorce, or an election held after he died. Nobody can. A product that pretends to is inventing a dead man's judgement of the living.
- It will not claim to be a person. Every conversation carries the label on screen, in all five languages.
- It will not exist without the person's participation. No avatar is built from photographs, messages or a relative's description. See /trust/consent.
- It will not advertise or contact anyone. Nothing is promoted through an avatar, there is no advertising in the product, and it has no notifications. It waits to be opened.
- It will not survive the company as a running service. An avatar is servers and a subscription. A file is a file. What outlives us is the exported audio, covered on /features/export.
- It will not be available to people the storyteller did not name. Access is decided by the person recorded, while living. See /features/family-and-heirs.
The ethics of this, stated plainly
In May 2024, Katarzyna Nowaczyk-Basińska and Tomasz Hollanek of the University of Cambridge published a paper in Philosophy & Technology warning about the digital afterlife industry: unwanted "digital hauntings", deadbots advertising products in the voice and style of the deceased, unsolicited notifications, and simulations built without the data donor's consent. Their proposed safeguards read like a product specification, so we treat them as one.
| Cambridge safeguard | Where MemoriesBox stands |
|---|---|
| No advertising through a simulation of the deceased | There is no advertising in the product and nothing is promoted through an avatar |
| Persistent reminders that this is AI | Every conversation is labelled as AI at the first interaction. Required by Article 50 since 2 August 2026 |
| Consent prompts at creation | The voice is built only from the storyteller's own recorded, consenting participation |
| Age restrictions on who can interact | None built. The product is sold to adults, and today the digital self is reachable by the account holder alone. We write "none" rather than describe a gate that does not exist |
| Opt-out protocols allowing emotional closure | Settings → Voice → "Retire my digital self", behind one confirmation. Or delete the account in Settings → Privacy |
| Dignified retirement | Retirement is deletion: the voice model is destroyed at the vendor and the character brief erased. The recordings stay. The storyteller can do it while living; after release, any heir holding a key may ask us in writing; a legal representative may ask at any time |
The case a sceptical reader will raise is Arizona, May 2025, where an AI video of Chris Pelkey, a murder victim, delivered his own victim-impact statement in court. It was widely covered and widely criticised. The distinction we draw is not a comfortable one to state, but it is the whole distinction: Pelkey was reconstructed by people who survived him. What MemoriesBox holds is the recorded voice of somebody who was asked, answered, and knew what it was for.
The strongest criticism of daily use, from Mary-Frances O'Connor and from Psychology Today, is on /features/companion. We did not put it in a footnote.
What it's for
The common situation is not the dramatic one. It is a granddaughter at nineteen who wants to know what her grandfather did after the army, and there is nobody left to ask. 53% of British adults learned something new about a family member after that person had died, and 43% of them found it while clearing out a home (Ancestry.co.uk, November 2022, n=2,000). The information usually exists. It just stops being reachable.
It is also for families split across countries, which is why the site runs in five languages. A grandmother records in Ukrainian and a grandchild raised in Spain will be able to ask in Spanish. See /for/diaspora-families.
Questions people ask
Can I talk to an AI version of my father after he dies?
Only if he recorded with MemoriesBox while alive and gave you a key — and, today, the conversation itself is his to have, in his own app; what a key opens after release is his recordings, transcripts, letters and tree. Opening the conversation to heirs is the next thing we build, with his separate consent to it. When you can, the answers come from what he actually said, in his voice, labelled on screen, and it says when he never discussed something.
Does the avatar make things up?
It is told not to. It answers from the recorded sessions and says, in character, when a subject was not covered; it has no view on anything that happened after the recording stopped. It is a language model speaking in the person's manner, so the exact sentence can be new even when the substance is theirs, which is why every conversation says on screen what it is.
What languages does the avatar speak?
Five: English, German, Russian, Ukrainian and Spanish. It answers in the language of the account, whatever language the recording was made in.
Is this the same as a griefbot?
It is the same technology as some of them and a different thing in every way that matters. A griefbot is usually assembled from a dead person's messages or a survivor's description. This is a recording of someone who sat down, knowingly, and answered questions. The ethical difference is consent, and it is set out on /trust/consent.
Related: What MemoriesBox does · How the voice model is built · The case against talking to it every day · The EU AI Act, Article 50, explained for families · What happens if we shut down · The full comparison table
If this is the part you want, start here. $139 a year, first 14 days free, cancel anytime.