XALIVE
Upload a character and it comes alive.
Give it a personality, a memory, a world, and the rules for how it relates to people. Everything underneath is handled for you.
Three stages, and the whole point of the shape is the last one. You make it in one place. It lives in many.
| Stage | What happens |
|---|---|
| Studio | Upload an image, a 2D sprite or a 3D model. Design personality, memory, world and relationship rules. |
| Runtime | Real time dialogue, voice, expression, memory recall, and openings it decides to make on its own. |
| Distribute | Connect it to an app, a game, a broadcast, a kiosk, or an enterprise service. |
Simple on the surface, six engines underneath
Characters are easy to build here. That is the result of the engines below being difficult, not of the problem being small.
| Engine | What it does |
|---|---|
| Sense | Reads signals Live voice in and out, presence detection, working out whether what was said was addressed to it, and signal freshness on a four step scale. |
| Emotion | Builds a state and leaves evidence Expression is decided with the response, not after it. Three axis coordinates, a separate hurt system, and a row per turn carrying grounds, confidence and which guardrail fired. |
| Memory | Judges rather than stores Four depths from passing to core. Wrong memories are invalidated, new facts supersede old ones with a history, and a supersession can be reversed. |
| Desire | Speaks first A curiosity signal produces a grounded question rather than silence. Every opening resolves to one of four reasons, and six gates decide whether it is spoken at all. |
| Values | Sets the basis for choices Personality is defined on five axes in words rather than numbers, and that identity is what assembles the prompt each turn. Changing it leaves an audit record. |
| Meaning | Connects its world Collection, refinement, human review, then admission into knowledge. Responses are bound to their grounds. On screen it is a map, not a graph. |
The research notes go into two of these in detail: how emotion is recorded and how it decides to speak first.
Not an AI that answers
The three things a character needs in order to last are an appearance, a relationship, and the ability to stand somewhere other than where it was made. Systems today tend to have one of the three.
It carries on from your yesterday, not from a context window that empties itself.
Where a character can stand today
This table is the honest one. Only the first row is what we would call working.
| Target | State |
|---|---|
| Web conversation | Working |
| Voice, broadcast overlay, website embed, exhibition kiosk, external integration, tool integration | In preparation |
| Game, mobile, desktop | Planned |
The kiosk target is running in real exhibition conditions and being verified repeatedly, which is the step before we would call it working. A target that is not ready cannot be switched on: the server refuses it rather than the interface hiding it.
Consent, rights and quiet
Once the system models the person as well as the character, this stops being an appendix.
- Personal identifying information is gated end to end and off by default.
- Quiet hours block every kind of first line, with no exception for camera triggered approaches.
- A daily limit on how often it may come to you, set by you.
- Changing one setting never overwrites another.
- Deletion exists as a plan with a reason recorded per system, separately for an account and for a character.
- A character without rights cannot be sent outside.
Keeping it the same character tomorrow
There are 965 automated checks, 59 of which are structural contracts. They exist to catch a specific class of failure: a new feature quietly changing how an existing character answers, or expression signals reaching only two of the three kinds of body, or a memory leaking into the wrong character.
A character that is different from yesterday cannot be a relationship. That is why this is not filed under quality assurance.
Creator beta · November 2026 Public launch · 2027 xalive.ai →