/WHAT IS · DEFINITION
A character system turns a model into a someone.
A character system is the technology layer that gives an AI a stable identity — a personality, a voice, a face, a memory, a set of values and a style of being in a room — that persists across conversations, sessions and devices. A language model by itself can imitate a personality for one chat. A character system sustains one for years.
Short answer
If the large language model is the engine, the character system is the driver: it decides who is speaking, what they remember, how they sound, when they speak, and when they stay quiet. Without it, every conversation starts over, and you get a helpful assistant. With it, you start to get a person.
What sits inside one
- Identity layer. Name, backstory, visual design, voice signature, values, relationship to the user. The constants.
- Memory layer. Episodic memory (specific moments with you), semantic memory (facts about you and the world), procedural memory (how the character behaves). Decides what is worth remembering, when to forget, and how to surface a memory naturally.
- Dialogue layer. Prompting, fine-tuning, retrieval-augmented generation, turn-taking, silence. The part that generates speech in character.
- Behavior layer. Idle animations, micro-expressions, breathing, posture, eye contact. Most of presence happens here, not in dialogue.
- Perception layer. Voice activity detection, wake logic, ambient sensing, face/presence detection — so the character knows when you are in the room without being told.
- Body layer. The renderer, voice synthesizer, display, speakers, lights. The sensory surface you actually experience.
Why a prompt is not enough
A system prompt can bias a model toward a personality for the duration of a session. It cannot keep a relationship. It cannot remember your sister's name in month three unless you architect memory explicitly. It cannot decide that tonight the character should be quiet because you came home quiet — unless you give it perception, context and a behavioral model that allows restraint.
Character systems are what the AI companion category has to invent now that base models are a commodity.
How it differs from game NPCs
Game NPCs are authored: a writer scripts what they can say, and the state space is bounded. A character system for a companion is generative: most things the character says were never written by a human, but they still have to feel like the same person. That is the hard, new problem.
My earlier work on real-time virtual humans and VTuber technology at Huya and XVERSE was essentially building authored characters at broadcast quality. What I am building now is generative characters at living-room quality — which is, counterintuitively, harder.
Presence is mostly behavioral
People assume character is what the AI says. In practice presence — the feeling that someone is there — is 70% how it behaves when it is not speaking: the weight of its silence, whether it looks at you when you walk in, how it rests between conversations. That is why the behavior and body layers are as important as dialogue to me.
For the broader vocabulary we use around the device — presence, character engine, holographic display, quiet hardware — see the glossary. For the hardware that hosts it, start with physical AI.