AI Companion Glossary, A to Z of Terms and Jargon
The AI companion space runs on borrowed vocabulary. Machine learning terms, fan fiction habits and mobile game language all end up in the same app listing. This glossary defines the words you will actually run into, in plain language and without spin. Entries are short on purpose.
How to read this glossary
Definitions describe how a term is normally used, not how every app implements it. Where vendors disagree, the entry notes the disagreement. If you want the bigger picture first, start with what an AI companion is.
A to B
AI companion. An app or service built around an ongoing conversational relationship with a persistent character. Usually includes memory, image or voice features, and a paywall.
AI chatbot. A broad term for any chat interface to a language model. Companions are a subset of chatbots, not a separate species.
AI roleplay (RP). Improvised, in-character dialogue, often with scene description, narration, and more than one character. Roleplay is a use case, not a product category.
Avatar. The image shown for a character. Usually AI-generated, sometimes uploaded by the user. An avatar is decoration, not part of the model.
Base model. The general language model before any character setup, memory or filters are added. Two apps using the same base model can behave very differently.
C
Canon. The agreed facts about a character or world. In roleplay it means details a character should not contradict, such as a backstory or a relationship.
Character card. A structured file or profile that defines a character: name, description, personality, scenario, greeting and example dialogue. See character cards explained for the field-by-field version.
Character archetype. A recurring character pattern, such as the childhood friend, the rival or the mentor. Archetypes shape expectations before the first message arrives.
Context window. The maximum amount of text a model can consider at once, measured in tokens. Anything older that does not fit is summarised, retrieved later, or simply dropped.
D to F
Fine-tuning. Extra training of a base model on a dataset so it behaves differently. It changes the model itself, which is expensive. Most consumer apps use prompting instead.
First message. The opening line a character sends when a new chat starts. It sets tone, scene and the character's attitude, which is why writers spend time on it.
Filter. Content rules that block or rewrite certain output, typically sexual content, violence or self-harm themes. Filters may be enforced by the model provider or by the app, and they change over time.
G to L
Hallucination. Confident output that is wrong or invented. In companion apps this includes false claims about what you said earlier.
IC and OOC. In-character and out-of-character. OOC instructions ask the model to step outside the roleplay, for example to make a correction or explain a rule.
Image generation. Producing pictures from a text prompt. Used for portraits, scene art and selfies sent inside a chat.
Jailbreak. A prompt that tries to get a model past its content filters or system instructions. Jailbreaks are a moving target, since providers patch them, and using one can break the terms of service.
LLM. Large language model. The text engine behind most companions.
LoRA. A small add-on trained to adjust a base model's style or knowledge. Common in self-hosted image and text setups, rare in hosted companion apps.
Lorebook. A set of keyword-triggered facts injected into the prompt when related words appear. It keeps long backstory reachable without filling the whole context window.
M to P
Memory. Any system that stores and reuses past conversation. Memory may be automatic summaries, extracted facts, or a searchable archive. It is not the same thing as the context window.
Model. The trained network that generates replies. Many vendors do not name theirs, and a few swap models without telling users.
NSFW filter. The rule set for not-safe-for-work content. NSFW is a vague label, so read the actual terms rather than the toggle in settings.
OC. Original character, meaning a character the user made rather than one from existing media. See also character archetypes in roleplay for how OCs are usually built.
Persona. The identity a character presents. The same word is used for the user's own profile in roleplay, which causes a lot of confusion.
Persona drift. When a character gradually stops matching its described personality, often after long chats or when older context falls out of the window.
Prompt. Everything sent to the model: the system prompt, the character definition, retrieved memory and your latest message.
Q to S
RP. Shorthand for roleplay. "RP partner" means the model playing a character.
Streaming. Sending a reply word by word as it is generated instead of waiting for the full answer. It makes long replies feel faster.
STT. Speech-to-text. Turns your spoken input into text so the model can reply.
System prompt. Hidden instructions that set rules for the model, such as staying in character or refusing certain topics. Usually not visible to the user and often not editable.
T to V
Temperature. A setting that controls randomness. Higher values give more varied, less predictable replies. Users rarely see it, but it explains why the same prompt gives different answers.
Token. A chunk of text the model reads or writes, roughly a few characters of English. Pricing, limits and context size are all measured in tokens.
TTS. Text-to-speech. Turns the character's written reply into audio. Quality varies from robotic to close to natural, and a TTS voice is normally generic rather than a specific person.
Voice cloning. Building a synthetic voice from audio samples so the character sounds like a chosen person. This has consent and impersonation problems, and some apps treat voice samples as training data.
W to Z
World info. Another name for a lorebook, used by several tools. The idea is the same: background facts supplied only when relevant.
Terms people mix up
| Term A | Term B | Difference |
|---|---|---|
| Memory | Context window | Context is what the model sees right now. Memory is what the app chose to keep and reload. |
| Character card | Persona | The card is the file. The persona is the identity it describes, and sometimes means the user's own role. |
| Filter | Jailbreak | The filter blocks content. A jailbreak is an attempt to get around it. |
| Fine-tuning | Prompting | Fine-tuning changes the model weights. Prompting only changes the input text. |
| TTS | Voice cloning | TTS reads text aloud with a generic voice. Cloning imitates a specific person. |
Key takeaways
- Most companion apps combine a base model, a character card, a memory system and filters.
- Context window and memory are different things, and the difference explains most continuity problems.
- System prompts are usually hidden and not editable, so you cannot fully audit what a character was told.
- Jailbreaks are not a stable feature. They break, and they can breach the service terms.
- Words like persona and NSFW mean different things on different platforms, so read the definitions rather than the labels.
- If you plan to build your own character, the field-by-field walkthrough in how to write a character prompt applies these terms directly.