Inside an AI Girlfriend App: The Three Parts That Make It Work

How they work

Behind the glossy marketing, these apps share one recipe: a language model, a character sheet and a filter. Work out which ingredient causes what, and most of the things that annoy users stop being a mystery.

We may earn a commission from links on this page. It never changes a rating.

Boiled down, these apps are a messaging screen sitting on a large language model, with a persistent character and a content filter built around it. Nothing else is going on. What makes one app worth US$10 and another US$20 is seldom the model itself; it's the quality of those other two pieces.

Part one, part two, part three

Model. This is the engine that produces words. Shown whatever text fits in front of it right now, it works out a plausible reply. It carries nothing over from yesterday's chat and wants nothing. Any sign of intention you notice was added by the next two parts.

Character sheet. Before every message you send, the app quietly sends the model a block of instructions: who this person is, their manner of speaking, what they know about you, and what's happened recently. You never get to read it. If a reviewer says an app has real personality, what they usually mean is that this document is well written.

Filter. This is the gatekeeper for what the character may say, and it can act before the model, after it, or at both stages. It explains why a scene sometimes stops with a courteous refusal, and why the word "uncensored" gets shouted in adverts. Joi and Secret Desires largely stand or fall on how they've tuned this.

The context window and the usual complaints

There's a ceiling on how much text the model can take in at once, known as the context window. The persona, the facts retrieved about you, and your latest messages all have to squeeze in together.

As a conversation grows, the earliest parts get pushed out. The app isn't bored with you and hasn't been nerfed. The model simply can't see that part of the chat any more. This one mechanism sits behind almost every "it was better before" complaint, and it's why how memory is engineered matters more than how large the model is.

Looking inside the sheet

Typically it's assembled from four pieces:

  • the persona, meaning a name, an age, a backstory and a style of speech;
  • lasting facts about you that the app decided were worth keeping;
  • a summary of your shared history that gets rewritten as it lengthens;
  • your most recent messages, unedited.

The apps that still seem to know you after a week away are the ones that get the middle two right. That's behind Nomi's score. A pricing page shows you none of this, which is why assessing it takes weeks, not an afternoon.

Three flavours under one label

"AI girlfriend" is used for products that barely resemble each other.

  • Companion apps aim at a long-running relationship, frequently with a journal and mood tracking. Replika has been around longest.
  • Roleplay apps focus on scenes and stories, not on carrying a relationship forward.
  • Media-first apps treat chat as a way to get images and voice. Candy AI lands closest to the centre, because one subscription covers all three.

Decide which flavour you're after and roughly half of any shortlist disappears, ours included, before you've looked at a price.

What they aren't

Unless an app was built to save facts and look them up, it won't know you from one session to the next. There's no ongoing self, either, since the character gets reassembled from text each time you send a message. And privacy isn't the default: your chats are handled on servers owned by somebody else. What these apps know about you digs into that.

Candy AI

4.6Rating: 4.6 out of 5

One subscription for chat, pictures and voice, with all three standing up to daily use.

Price
from US$12.99/month
Free tier
Yes
Australia
Available

Nomi

4.4Rating: 4.4 out of 5

Unbeaten long-term memory in our testing, with the most natural back-and-forth.

Price
from US$15.99/month
Free tier
Yes
Australia
Available

Frequently asked questions

Is somebody real typing the replies?

No. A language model predicts the most likely next message, nudged by a written profile of the character. No human is on the other end, and none of the apps reviewed here claim one is.

Does it truly remember me?

Only partly. Facts about you are stored separately from the chat and slipped back in when needed. So it can nail some topics and know nothing about others.

Why did the character change after an update?

Either the underlying model or the hidden instructions that steer the character were swapped out. You see the same profile, but whatever interprets it isn't the same.