Inside an AI Girlfriend App: The Components That Make It Tick

The technology

Take away the branding and you find the same three components in every one of these apps: a language model, a character sheet and a filter. Once you know which one handles what, most of the usual complaints start to make sense.

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

In essence, an AI girlfriend app is a messaging screen sitting on a large language model, with a lasting character and a content filter added. That covers the entire category. The difference between a US$10 app and a US$20 one seldom lies in the model itself. It lies in how carefully the other two parts were built.

Model, sheet and filter

The model produces the words. Given everything in view at that moment, it predicts a plausible next message. It holds no memory of yesterday's chat and has no goals. Whatever looks like intention comes from the other two parts.

The character sheet is a chunk of text that the app attaches in front of your message on every turn. It covers who the character is, how they talk, what they know about you and what has happened recently. You never get to read it. When someone says an app "has personality," what they usually mean is that this document is well written.

The filter decides what the character is allowed to say, and it can run before the model, after it, or at both points. It is why a scene can suddenly end in a courteous refusal, and why "uncensored" is the loudest claim in this market. Joi and Secret Desires mainly compete on how that filter is tuned.

The context window and what it explains

A model sees only a fixed quantity of text per turn. This quantity is the context window, and the character sheet, the retrieved facts about you and the most recent chat all have to share it.

When a conversation lengthens, its earliest turns slide out of view. The app has not lost interest or been downgraded; that stretch of the chat is simply no longer visible to it. Most "it used to be better" complaints trace back to this, which is why how memory is implemented matters more than how large the model is.

What goes into the sheet

Typically four ingredients are combined:

  • A persona, meaning name, age, background and way of speaking.
  • Lasting facts about you that the app chose to keep.
  • A running digest of your history together, rewritten as it gets longer.
  • The newest messages, unedited.

The apps that still feel occupied a week later, which explains Nomi's score, do well with the second and third items. A pricing page reveals none of this, so judging it calls for weeks of use, not an afternoon.

Three very different kinds of product

The "AI girlfriend" label spans products with little in common:

  • Companion apps aimed at an ongoing bond, often with journals and mood logs. Replika is the longest-standing example.
  • Roleplay apps aimed at scenes and storytelling, not continuity.
  • Media-first apps in which the chat serves mainly to produce images and voice. Candy AI comes closest to the middle ground, offering all three within a single subscription.

Deciding which kind you want eliminates about half the apps on any list, ours included, before you have even looked at cost.

What they are not

They will not know you from one session to the next unless the developers built in storage and retrieval of facts. They have no continuing self either, because the character is rebuilt from text at the beginning of every turn. And they are not private by default, since your chat is handled on someone else's servers, a topic covered in what these apps know about you.

Candy AI

4.6Rating: 4.6 out of 5

Chat, image generation and voice under a single subscription, each performing reliably in daily use.

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

Nomi

4.4Rating: 4.4 out of 5

The strongest long-term memory in our evaluation, combined with the most natural dialogue.

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

Frequently asked questions

Is there an actual human on the other side?

No. A language model generates the most probable next message, steered by a written character profile. Nobody is typing, and none of the apps reviewed here claims that anyone is.

Does the app really retain what I tell it?

Partly. Apps store facts about you outside the chat and feed them back in, so recall can be precise on some topics and blank on others.

Why do replies feel different after an update?

Either the model or the system prompt behind the character was replaced. The profile looks the same from your side, but whatever interprets it has changed.