AI Companion Glossary - The Terms You Keep Seeing, Explained

How it works

Forums, pricing pages and app settings assume you already know what a context window, a lorebook or an OOC note is. Here are the terms that come up most, grouped by where you will meet them.

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

This glossary is written for users, not engineers. Each entry says what the term means in practice and, where useful, where it comes up. Links go to our longer explanations.

The vocabulary of AI companions grouped into four areas: the model, memory, roleplay, and plans and safety

The four groups used in this glossary.

The model

LLM (large language model). The system that writes replies by predicting text. Every companion app is built on one, its own or rented. See what is an AI girlfriend app.

Token. The unit models read and write, roughly three-quarters of a word in English. Limits and prices are often counted in tokens.

Context window. How many tokens the model can read at once. Long chats overflow it, which is why older details disappear.

Temperature. A setting for how adventurous the wording is. Higher is more varied and more random; lower is safer and more repetitive. See how an AI companion writes its reply.

Sampling. The method of choosing each next word from the model's options. Temperature is one sampling setting of several.

Hallucination / confabulation. The model stating something untrue with confidence, such as a memory that never happened. See why AI companions invent memories.

Fine-tune. A model further trained for a specific style, such as roleplay. Local users often choose roleplay fine-tunes.

Quantisation. Shrinking a model so it runs on smaller hardware, at some cost to quality. Relevant if you run a companion locally.

Multimodal. Able to handle more than text - images, voice or video.

Memory

Long-term memory. Facts and summaries the app stores and brings back in later chats. See how AI companion memory works.

Summary memory. Condensed notes about earlier conversations, written by the model itself. Useful and lossy.

Retrieval. Searching stored memories for ones relevant to the current message and adding them to what the model reads.

Memory slots / journal. A visible list of what the app remembers about you, sometimes editable.

Lorebook / world info. A set of entries that are inserted into the context when a keyword appears - common in SillyTavern and advanced apps.

Roleplay

Character card. A file containing a character's description, personality, examples and greeting; often a PNG image with the data embedded.

Persona. Your own description - who you are in the story.

Greeting / first message. The character's opening line, which sets style and length for everything after.

Example dialogue. Sample lines in the character's voice. The strongest protection against personality drift.

OOC (out of character). A bracketed note to the AI as a writer. See roleplay prompts that work.

Regenerate / swipe. Ask for a new version of the last reply.

Time skip. A line such as [Two days later.] that moves the story forward.

Slow burn. A romance that builds gradually. See slow burn with an AI companion.

Group chat. Several characters in one conversation, each with its own personality.

ERP. Erotic roleplay.

Godmodding. The AI writing actions or dialogue for your character. Usually unwanted.

Plans, safety and policy

Free tier. What you get without paying - see free AI girlfriend apps.

Credits / tokens / gems. In-app currency for images, voice and extras, separate from the subscription. See monthly, annual or credits.

Fair use cap. A limit on "unlimited" features, usually images or voice.

NSFW filter. Automated blocking of sexual or graphic content. See content filters.

Uncensored. Marketing term for a permissive filter. There is always a floor.

Jailbreak. A prompt designed to get round a model's rules. Usually against the terms of service and a common reason for account flags.

Age verification / age assurance. Proof that you are an adult. See age verification on AI companion apps.

Crisis protocol. The legally required response when a user expresses suicidal thoughts - typically breaking character and showing a helpline.

Data export / access request. Getting a copy of what the app holds about you. See how to request your data.

Web app / PWA. An app that runs in the browser, often installable to the home screen. See browser or app store.

Model update. A change of the underlying model by the company, which can change how a character sounds overnight.

Affiliate link. A link that pays the site that sent you when you subscribe - including on this site. It is disclosed on every page, and our method is in how we test.

Frequently asked questions

What does OOC mean in AI roleplay?

Out of character. A note, usually in brackets, that speaks to the AI as a writer rather than to the character - for example to change the pace of a scene or correct a mistake.

What is a context window?

The amount of text a model can read at once when writing a reply, measured in tokens. Anything outside it - older parts of a long chat - is invisible to the model unless it is brought back through memory.

What does ERP stand for?

Erotic roleplay. It is the community term for sexual roleplay with an AI companion, and the feature many apps restrict or allow depending on their policy and your age verification.