A virtual clone that thinks in your voice

Persoma builds an AI stand-in that mirrors one specific person: their personality, their phrasing, the way they weigh a decision. It travels in your pocket, and it opens only to the people you allow.

Request early access See what a clone does

Built on Microsoft Azure. The person being cloned holds the consent, the source material and the off switch.

What a virtual clone is

A model of one person, not a model of everyone

General assistants are trained to sound like anybody. A Persoma clone is shaped around a single individual's material until its answers are recognisably theirs.

It carries a personality, not just facts

Retrieval systems answer from documents. A clone also carries tone, hedging habits, pet phrases and the arguments its subject tends to reach for first.

It answers in real time

A clone holds a conversation: follow-up questions, clarifications, changes of subject. The point is a dialogue, not a generated document.

It stays under its subject's control

The person who was cloned decides what goes in, who may talk to it, and when it stops. Consent is a setting in the product, not a paragraph in a policy.

Where a clone earns its keep

Three patterns we think are the strongest fit.

  • Expertise that does not scale. One senior specialist, forty people who need ten minutes of their time each week.
  • Continuity through a handover. A founder, clinician or advisor steps back, and the reasoning behind years of decisions does not walk out with them.
  • Presence in more places than a calendar allows. A public figure or instructor who wants to answer the long tail of questions personally.
How it works

Four steps from consent to a working clone

  1. Consent and scope

    The subject signs off on what may be used and who may query the clone. Nothing is ingested before that record exists.

  2. Capture

    Written material, recordings, transcripts and a structured interview covering how the person reasons, not only what they know.

  3. Shape and review

    The persona is assembled, then the subject reads back real answers and corrects them. Voice is tuned by disagreement, not by settings.

  4. Publish and supervise

    The clone goes live to its approved audience. Conversations stay reviewable, and the subject can revise, restrict or revoke at any point.

Memory

A person is not the same in every room, and is still the same person

That sentence is the engineering problem. A clone has to move with a conversation without drifting away from whoever it is modelling. Persoma solves it with memory in layers, each on its own latency budget, and a personality that carries weight rather than a fixed script.

Persoma memory architecture Three memory layers stacked from fastest to deepest: immediate reflex measured in milliseconds, conversational memory covering the current exchange, and long-term memory covering everything the subject provided. A personality core runs vertically through all three layers, anchoring them. PERSONALITY CORE Immediate under 50 ms What just happened, this second. In-memory cache. The clone reacts rather than computes. Conversational 200 - 500 ms The thread you are in, held continuously. No resetting between turns. It remembers what it already said. Long-term up to 2 s The whole record the subject provided. Retrieved by relevance, never by recency alone.

Immediate

An in-memory cache holding the last few seconds, on a 50 ms budget. It exists so the clone can react to something that just happened rather than assemble a considered reply to it. Reflex, not thought.

Conversational

Continuous short-term memory across the exchange you are actually having, budgeted at 200 to 500 ms. It holds the thread, so the clone knows what you said six messages ago, what it has already told you, and what it has not yet answered.

Long-term

Everything the subject provided, plus the relationships between people, places and events. Budgeted at up to 2 seconds, because a deep retrieval is worth waiting for. Retrieved by relevance rather than recency, so an old position still surfaces when it matters.

Variable weight

Personality that moves, anchored to a core

Most systems hold personality as a fixed instruction, which produces a clone that is identical in a technical argument and a condolence call. Real people are not.

In Persoma the traits carry weights that shift with context - with accumulated sentiment, with what has been remembered, with what keeps happening. Under pressure the clone can become more direct; in a personal conversation it can soften.

The model itself never changes. Nothing is being trained and nothing is being learned. The weights are state, tethered to a core measured from the subject's own material, bounded by the range that person actually showed, and they settle back toward it. A clone can have a mood. It cannot have a personality change.

The test is not whether the clone stays still. It is whether it moves the way that person moved, within the range they actually showed, for the reasons that actually moved them, and whether it comes back.

Capabilities

What the platform gives you

Persona fidelity controls

Tune how closely the clone tracks its source voice, and how freely it may extrapolate beyond what the subject has actually said.

Grounded answers with provenance

Responses can cite the source passage they came from, so a reader can tell a remembered position from an inferred one.

Boundary rules

Define subjects the clone declines, and what it says instead. Medical, legal and financial guardrails are configured per clone.

Conversation review

The subject sees what was asked and answered, flags drift, and pushes corrections back into the persona.

Multilingual by default

English and French out of the box, with the persona's voice preserved across both.

Embed anywhere

A hosted page, a site widget, or an API for your own front end. Authentication rides on the identity you already use.

Built on Azure

A clone is personal data. We treat it that way.

Persoma runs on Microsoft Azure. That choice is about where the data sits, who can reach it, and what happens when someone asks for it back.

Data residency you can verify
Residency is set per deployment. We name the exact Azure regions your tenant runs in, in writing, before you commit.
Your content is not training data
Source material and conversations are used to serve the clone they belong to. They are not used to train foundation models.
Identity you already have
Sign-in uses Microsoft Entra ID, so access follows your existing directory, groups and conditional access policies.
Encryption in transit and at rest
TLS on every connection, platform-managed keys at rest, with customer-managed keys available for tenants that require them.
Revocation that actually deletes
A subject can withdraw consent, and the clone, its embeddings and its source material are removed within a stated window.
Audit trail
Administrative actions and clone conversations are logged for review and export.
Nothing is baked into a model
The underlying model is never altered by your material. A clone is a persona state and a retrievable record sitting on a model that stays exactly as it was. That is why deletion is deletion, and not a promise to try.

Compliance documentation is available on request during onboarding.

Questions

Frequently asked

Is this a deepfake?

No. A Persoma clone is built with the subject's participation and consent, it is labelled as an AI stand-in, and the subject can shut it down. A deepfake is defined by the absence of all three.

How much material does a clone need?

Enough to show how the person argues, not just what they know. In practice that means a body of their own writing or transcribed speech plus a structured interview.

What stops it from inventing opinions?

Two things: answers are grounded in the subject's own material with provenance available, and each clone has boundary rules for subjects it should decline rather than guess at. It will still be wrong sometimes, which is why the review loop exists.

Who owns the clone?

The subject owns their likeness and their source material, and can revoke both.

What happens to a clone if the subject dies?

What the subject decided, while they were able to decide it. Every clone records a standing instruction at onboarding: retire and delete, or continue for named people for a stated period. Where no instruction was recorded, the default is retire and delete. Silence is never read as permission.

Where a clone does continue, it never claims to be alive, never invents a life it did not have, and never expresses need designed to make leaving a conversation feel like abandonment. It speaks from the record, and it says so when the record is silent. Anyone in the named circle can ask us to stop it, not only whoever set it up.

The terms set this out in full.

Which model does it use?

Persoma runs on Azure OpenAI. The specific model is a deployment detail that changes as better ones ship, and is documented per tenant.

What does it cost?

Pricing is set per deployment during early access.

Early access

Request access to Persoma

Tell us who you would clone and what you would ask it. We are taking a small number of design partners.

A sentence or two is plenty.

We reply to every request. No newsletter, no sequence.

Prefer email? Write to hello@persoma.ai.