I wrote this for people who use these apps, not people who build them. Each entry says what a word means in practice and where you're likely to run into it. Links lead to our deeper guides.
The four buckets this glossary uses.
Under the hood: the model
LLM (large language model). The engine that produces replies by predicting what text comes next. Every companion app runs on one, either built in-house or licensed. Start with what is an AI girlfriend app.
Token. The chunk of text a model handles at a time, about three-quarters of an English word. Usage caps and pricing are often quoted in tokens.
Context window. The number of tokens a model can take in at one moment. A long chat spills past it, and that's why early details vanish.
Temperature. A dial for how bold the word choices get. Turn it up and replies get more varied and less predictable. Turn it down and they get safer and more repetitive. More in how an AI companion writes its reply.
Sampling. How the model picks each next word from its candidates. Temperature is only one of several sampling controls.
Hallucination / confabulation. When the model says something false with total confidence, like recalling an event that never took place. See why AI companions invent memories.
Fine-tune. A model given extra training toward a particular style, such as roleplay. People who run models on their own machines often pick roleplay fine-tunes.
Quantization. Compressing a model so it fits on weaker hardware, with a small hit to quality. It matters if you run a companion locally.
Multimodal. Able to work with more than text, such as pictures, voice or video.
Memory terms
Long-term memory. Facts and recaps the app saves and feeds back into later conversations. See how AI companion memory works.
Summary memory. Short notes on past chats, written by the model itself. Handy, but details get lost along the way.
Retrieval. Looking through saved memories for the ones that fit your latest message, then adding them to what the model reads.
Memory slots / journal. A list you can see of what the app knows about you. Some apps let you edit it.
Lorebook / world info. Entries that get slipped into the context whenever a trigger keyword shows up. You'll find them in SillyTavern and other advanced tools.
Roleplay terms
Character card. A file holding a character's description, personality, sample lines and greeting. Often it's a PNG image with the data hidden inside.
Persona. Your side of the story: a description of who you are in the roleplay.
Greeting / first message. The opening line from the character. It sets the tone and length for everything that follows.
Example dialogue. Sample lines written in the character's voice. It's your best defense against personality drift.
OOC (out of character). A bracketed aside aimed at the AI as a writing partner. Try roleplay prompts that work.
Regenerate / swipe. Request a fresh take on the last reply.
Time skip. A line like [Two days later.] that jumps the story ahead.
Slow burn. A romance that develops bit by bit. See slow burn with an AI companion.
Group chat. A single conversation with several characters, each with a personality of its own.
ERP. Erotic roleplay.
Godmodding. The AI writing your character's moves or lines for you. It's rarely what you want.
Plans, safety and policy
Free tier. What the app gives you at no cost. See free AI girlfriend apps.
Credits / tokens / gems. Currency inside the app for pictures, voice and extras, billed separately from the subscription. Compare options in monthly, annual or credits.
Fair use cap. A ceiling on features sold as "unlimited", most often images or voice.
NSFW filter. Automatic blocking of sexual or graphic material. See content filters.
Uncensored. A marketing word for a loose filter. Every app still has a floor.
Jailbreak. A prompt crafted to slip past a model's rules. It usually breaks the terms of service and is a frequent reason accounts get flagged.
Age verification / age assurance. Proving you're an adult. See age verification on AI companion apps.
Crisis protocol. The response an app is required by law to give when a user voices suicidal thoughts. Typically the character drops the act and a helpline appears.
Data export / access request. Asking a company for a copy of everything it holds on you. See how to request your data.
Web app / PWA. An app that lives in your browser and can often be pinned to your home screen. Compare in browser or app store.
Model update. The company swapping the underlying model, which can change how a character talks overnight.
Affiliate link. A link that earns the site that referred you a commission if you subscribe. We use them, we disclose them on every page, and how we test explains our method.