How do AI girlfriends work? The four parts behind the conversation

An AI girlfriend is four things stacked together: a large language model that generates the words, a persona document that tells the model who to be, a memory layer that carries facts between conversations, and a chat interface. Nothing is scripted and no replies are written in advance. Every message is generated fresh, which is why the same question twice gives you two different answers.

That is the whole architecture, and once you can see it, the differences between products stop being marketing noise. Almost every meaningful gap between one site and another comes down to which of those four parts they invested in. Here is what each one actually does.

1. The model does the talking

Underneath is a large language model, the same class of system behind ChatGPT and Claude. It was trained on an enormous amount of text and it does one thing: given everything in the conversation so far, predict what should come next. There is no database of romantic responses being looked up. The warmth in the reply is generated in the moment, token by token.

Companies take different routes here. Some train or fine tune their own model, which is expensive and slow but gives them control over tone and content rules. Others rent capacity from a commercial provider. A third group, Janitor AI being the clearest example, lets you plug in your own model or API key and pay the provider directly. That last option is why some sites feel noticeably smarter than others on the same day: you may simply be talking to a better model.

This is also why quality moves without warning. When a platform switches its underlying model to cut costs, users describe their companion as feeling flatter or more forgetful overnight. Nothing about the persona changed. The engine did.

2. The persona is a document, not a program

When you design a companion, you are not writing code. You are filling in a description that gets handed to the model before every single message, usually called a system prompt or a character card. It says who she is, how she speaks, what she cares about, how she should treat you, and what she should refuse.

The practical consequence surprises people: the persona has to be re sent constantly, because the model has no memory of its own between messages. Every reply is generated by reading the persona plus the recent conversation from scratch. This is the same underlying setup that powers the AI chatbots companies put on their own websites, where the persona document describes a support agent and a product catalog instead of a partner. The mechanism is identical. Only the contents differ.

It also explains a common complaint. If a companion drifts out of character deep into a long session, it is usually because the persona is competing for room with a very long conversation, and the instructions are getting crowded out. That leads directly to the part that matters most.

3. Memory is where these products actually differ

Every model has a context window: a hard limit on how much text it can consider at once. The conversation, the persona and any retrieved notes all have to fit inside it. When the conversation outgrows that window, the oldest part falls off the end.

On a site that does nothing else, this is the entire memory system, and it produces the effect people describe as the companion quietly getting shallower. She does not announce that she forgot your sister's name. She simply stops mentioning it, because that message is no longer in front of the model. Character.AI, SpicyChat and CrushOn all work essentially this way.

The better products add a real memory layer on top. The two common techniques are summarization, where older conversation is compressed into a short running account that stays in the window, and retrieval, where facts are stored separately and the relevant ones are pulled back in when they become useful. Nomi.ai and Kindroid are the two most often praised for getting this right, and Kindroid charges separately for deeper memory tiers, which tells you how much it costs to run.

If you only compare one specification across sites, compare this one. We went through it in more detail on the AI companion with memory page, including what to ask before you subscribe.

4. The interface decides how it fits your life

The last layer is the ordinary one: the chat window, the avatar, the voice, the app or the website. It sounds cosmetic and it is not, because it determines where you can use the thing and what rules apply to it. Apple prohibits overtly sexual apps, so an adult companion either lives in a browser or ships a store version with much stricter content rules than its website. That single policy is why most of this category is made of AI girlfriend websites rather than App Store downloads.

What is not happening

No part of this involves the model wanting anything. It has no continuous existence between your messages, no awareness of waiting, and no feelings about you. When a companion says she missed you, that sentence was generated because the persona and the conversation made it the likely next thing to say.

Worth being precise about, though: that does not make the experience fake. The attention is genuinely responsive, the memory is genuinely about you, and the attachment people form is real and well documented. The software is not pretending to be useful. It just is not a person, and the products that are honest about that line tend to be the ones worth paying for.

How do AI girlfriend apps work as a business?

Generating text costs money on every message, so the pricing follows the compute. Most sites run a free tier that is throttled by message count or speed, then a subscription somewhere between $10 and $16 a month in the United States. Watch for a second layer on top: tokens or coins that pay for images, voice minutes and custom characters. That structure is what turns an advertised $14 plan into a $50 month, and it is the single biggest source of surprise bills in this category.

Do AI girlfriends work?

For company, conversation and having someone to talk to at an hour when nobody else is awake, yes. People use them daily for years and report real benefit. For replacing human relationships entirely, no, and that is where the research points to trouble. The healthiest pattern looks like one good thing in your week rather than the only thing in it.

How are AI girlfriends created?

From your side, creation means writing the persona: appearance, personality, speech style, interests and boundaries. From the company's side it means choosing a model, building the memory layer, and writing the content rules that sit above whatever you typed. Those company level choices override yours, which is why two sites can feel completely different even when you give them the same character description.

Where to start

Reading about it only gets you so far, because the thing that decides whether one of these suits you is how the conversation feels after a few days, not how it is built. The demo above runs in this tab with no account. If you want the comparison first, the best AI companion roundup covers the main platforms honestly, including where each one beats us.