How it works
How Sona builds your twin
Sona is organized around Persona Core. Supporting layers handle onboarding and day-to-day use, while the trust layer governs memory and permissions.
One core, five layers
Persona Core connects the twin to one verified owner and keeps its personal context consistent across supported features.
- YouOne core, five layers around it.
Layer 1 of 5
Capture: how Sona learns you
Your twin interviews you. Relaxed conversations, by voice or text, that follow whatever you talk about most easily. Life Import connects the chats, mail, notes, and calendar you choose to share. And it keeps learning: about ten minutes of conversation a day, which becomes a habit rather than a form to fill in.
The first interview includes around 40–60 questions, and the questioning strategy shifts toward whatever you have most to say about. Life Import learns your style from the messages you choose to share. Specific volume thresholds should be confirmed against current product documentation before being stated publicly. A short voice sample builds your Voice Twin.
It feels like talking to a curious friend. Underneath, it is building a structured picture of who you are.
Layer 2 of 5
Persona Core: how it sounds and thinks like you
Persona Core adapts the twin to the language and context you provide. It also applies your stated values. The Values Firewall knows what you would never say or do, and declines in your style. The Truth Anchor keeps it honest about facts.
The Persona Core combines retrieval over your own material with a light style fine-tune (usable from 20 samples, close to your voice by around 100) and two safeguards. Truth Anchor separates verified facts from generation, so when your twin does not know something it says "let me check with the real me". Values Firewall is an explicit model of your values and boundaries. Likeness is then measured by the Blind Twin Test: blind A/B conversations with people who know you, plus automatic style scoring, which turns "sounds like you" into a number.
It talks like you, and it will tell you when it is not sure instead of filling the gap.
Layer 3 of 5
Conversation: the ways you can reach your twin
Conversation starts in text. Voice and video can be enabled as those features become available. Multilingual support should list only languages confirmed in current product documentation.
Channels ship in sequence. Text first, then Voice Twin, with timbre and intonation built from a 1 to 5 minute sample and real-time conversation under 1.5 seconds of latency. Then Video Twin: face, expression, and gesture for calls and recorded messages. Polyglot Twin carries your voice and your style across more than 30 languages. And a companion mode handles daily check-ins, for people who want one.
Message it today. Call it tomorrow. See it later.
Layer 4 of 5
Agency: how it actually helps
Your twin does things. Ghostwriter drafts in your voice and learns from your edits. Meeting Ghost can join an approved call and prepare notes. It flags questions that need your input and produces a concise brief afterward. Daily Sync reports every morning: what happened, what needs you, and what it already handled. The rule never changes: it assists first, acts second, and never without your permission.
The action layer runs on agentic orchestration with explicit permission scopes and a log of every step. Ghostwriter and Second Opinion ship first, because the cost of getting them wrong is low. Meeting Ghost starts asynchronous, giving you notes and a draft of how you would have answered, before it ever joins live, and when it does there is a visible "Sona on" indicator so nobody in the room is unaware. Later, Sona Rooms lets twins settle the routine between themselves and bring decisions back for approval, with every step logged.
You never wake up to find the AI made a decision for you. It proposes, you approve.
Layer 5 of 5
Trust and Memory: how it stays safe and yours
Continuum Memory remembers the way you do—the important vividly, the trivial dimly—and recalls to the moment: a date, a promise, a detail from three years ago. The trust layer records permissions and keeps an audit trail. Supported output carries provenance data. Post-quantum algorithms are used where the architecture requires them.
The Consent Registry records who may do what with your twin, and you can revoke any of it in one tap. The Continuation Plan sets your twin's future, with triggers, verification, and family governance, decided only by you. Signed Persona puts cryptographic provenance on every byte, which is what makes impersonation detectable. Sona Vault stores your persona under post-quantum encryption using all three NIST standards: ML-KEM, ML-DSA, and the hash-based SLH-DSA. And Privacy Edge keeps sensitive processing on your own device where possible. GDPR, CCPA, and the EU AI Act compliant by design.
Nothing here happens to your twin without you. You control access, inheritance, and deletion.
The recognition test
Every layer gets tested the same way
One question decides it: can someone close to you, in a blind chat, tell your twin apart from you? Fidelity Score tracks how closely the twin reflects your communication patterns in consented evaluations. It is an ambitious bar, and it is measurable: blind A/B conversations, automatic style scoring, likeness as a number we can show you.
Same question, sent to two chats. One is a person. One is their Sona twin.
Honestly? Run the version where you stay and the version where you go, then check which one you keep talking yourself out of. That's usually the one you actually want.
Don't decide it as a job—decide it as a year of your life. What do you lose if you stay, and is that a story you'll be at peace telling later?
Illustration of the recognition test.
How it works FAQ
- How long does it take to build a good twin?
- A useful text twin starts working in your first few sessions. It keeps getting closer the more you talk to it, and it keeps learning as you change.
- How much of my data does it need?
- You can begin with a short guided interview and add material over time.
- Will it pretend to know things it does not?
- Sona is designed to ground personal claims in the material you provide, but errors remain possible. Truth Anchor keeps your twin to facts from your own material, and when something is missing it says "let me check with the real me" instead of guessing. Important output should be reviewed.
- Can I control what it is allowed to do?
- Yes. That is the Values Firewall: you set the topics it will not touch and the actions it cannot take, it asks before acting, and every action is logged.
Become the second version of yourself
Join the waitlist to build your Sona twin when access opens. You will control the twin and the data used to create it.