How AI companion apps actually work

AI companions can feel uncanny, but the machinery is understandable. There are five pieces, and once you know which piece does what, most of the app's behaviour — including the parts that feel surprisingly human and the parts that break — becomes predictable.

1. The language model

Every reply is generated by a large language model: the same class of technology behind ChatGPT and similar tools. The model itself has no personality and no knowledge of you. It takes a block of text and continues it plausibly. That's it.

Most companion apps don't train their own model. They use a general-purpose model — commercial or open-weight — and do all the character work in the layer above. This is why companions from different apps can feel oddly similar in their rhythms, and why an app can feel completely different overnight when it switches model.

2. The system prompt — where the persona lives

Before your message reaches the model, the app prepends instructions: who this character is, how she speaks, her backstory, her appearance, what she must never do. That block of text is the persona. It is why one companion reads as warm and shy and another as bold and chaotic — same model, different instructions.

Two consequences worth knowing. A persona is much shallower than it feels: change a paragraph and the character changes completely. And the persona is competing with your conversation for the model's attention — which is why long conversations sometimes drift out of character.

3. The context window — why she forgets mid-conversation

The model can only see a limited amount of text at once. Everything has to fit: the persona instructions, the retrieved memories, and the recent conversation. When a conversation gets long, the oldest turns fall out of that window.

This is the single best explanation for the most common complaint about these apps — that she contradicted something said forty messages ago, or forgot a detail from earlier in the same session. She didn't forget. It was no longer in front of her.

4. The memory store — why she remembers last month

Because the context window is too small to hold your whole history, apps store memories separately and retrieve the relevant ones. In practice: important facts get written to a database (usually a vector database that searches by meaning rather than keyword), and when you send a message the app searches it, pulls the few most relevant items, and pastes them into the prompt.

So "she remembered my sister's name" means: your message mentioned something semantically close to a stored note, the note was retrieved, and it was in the prompt when the reply was generated. Full detail in how AI companion memory works.

5. Media generation and moderation

Selfies come from a separate image model, conditioned to keep the same face each time — which is a genuinely hard technical problem and the main reason image quality varies so much between apps. Voice notes come from a text-to-speech model with a fixed voice per character. Both are far more expensive to run than text, which is why they're metered or gated behind higher tiers.

Moderation sits on top, in two places: instructions in the prompt that the model is told to obey, and automated classifiers checking output before it's sent. The first is soft — prompts can be talked around. The second is the one that actually holds.

What this means when choosing an app

• Memory quality is the main differentiator, and it's an architecture problem rather than a model one. Apps that feel like they know you have invested in retrieval.
• Image consistency is the second. Keeping one face across hundreds of generations is hard; a free trial shows you immediately whether an app has solved it.
• "Advanced AI" in marketing copy means nothing. Nearly everyone is using the same handful of models. The differences are in the prompt, the memory and the media.