What an AI companion app actually is, and what it is not

An AI companion app is a chat interface with three components: a character with a fixed personality, memory that carries across sessions, and typically some method for generating pictures or voice of that character. Under the hood sits a large language model, the same architecture powering any chatbot, with system instructions that constrain it to stay in character.

What the character is

The character consists of a text description plus the model’s trained behaviour. You define the name, appearance, personality and backstory; the platform converts that into instructions the model obeys. How faithfully it maintains them over a long session is what our conversation score tracks. Budget platforms wander off within an hour. The competent ones stay on script.

What memory means

Memory is not the model recalling you spontaneously. It is the platform saving facts from previous chats and reinserting them into the context. Some use a narrow window, so anything beyond a few hundred messages vanishes. Some retain a compressed summary. The strongest platforms keep particular details and surface them without prompting, which our memory score recognises.

Images and voice

Pictures come from a separate image model, prompted with the character description. The difficulty is consistency: producing the same face twice. Voice applies text-to-speech to the reply, and on select platforms enables real-time calling. Both are judged on independent criteria.

What it is not

It is not a person. It does not learn in the manner the marketing suggests. It knows nothing you did not provide or that was absent from its training data. Frame it as a skilled improv partner with a notebook. The capable platforms justify that comparison.