AI Companion & Roleplay Glossary: 19 Terms, Defined

AI companion

Also known as: Virtual companion, AI partner

An AI companion is a conversational AI character you keep an ongoing relationship with, distinguished from a general-purpose chatbot by three things it has and an assistant doesn't: a persona it stays in, memory that survives between separate sessions, and no task it is trying to complete.

The distinction is architectural rather than cosmetic. An assistant is tuned to resolve your request and end the exchange; a companion is tuned to continue it, and that changes what the system optimises for at every layer — the prompt holds a character in voice, a memory layer carries details forward, and reply length tracks yours instead of front-loading an answer. It is also why companion apps live or die on their memory rather than their model, since most run the same base models everyone else does. See the AI girlfriend hub for the most common form of one.

AI roleplay

Also known as: RP, Interactive fiction

AI roleplay is collaborative interactive fiction with a language model, in which you and the model each drive characters inside a shared scene — as opposed to a question-and-answer exchange, where one side asks and the other resolves.

A roleplay turn usually mixes dialogue with narration, often set off with asterisks or italics, and the model is expected to advance the scene rather than summarise it. That makes it unusually sensitive to prompt shape: a first message written in past tense pulls every later reply into past tense, and a two-line greeting teaches the model that two-line replies are correct. Most complaints about a character feeling flat trace back to the setup rather than the model behind it. We ranked platforms on this in best AI chatbot for roleplay.

Persona

Also known as: User persona, User description

A persona, in AI chat, is the description of you that gets injected into the prompt so the character knows who it is talking to — the mirror image of the character's own profile, and the reason a character can use your name without being reminded.

A persona typically carries your name, how you want to be addressed, a short appearance sketch, and anything the character should treat as already known. It is the fix for the most common complaint in long roleplay — a character contradicting an established fact about you — because anything in the persona is re-sent every turn, while a fact mentioned once in conversation eventually falls out of the context window. Keep it short and factual; a persona longer than the character's own profile starts competing with it.

Full guide: Writing a good user persona

Persistent memory

Also known as: RAG memory, Long-term memory

Persistent memory is the external store that lets an AI character recall details from earlier sessions, and it exists because the language model itself is stateless — it retains nothing between turns and re-reads its entire world on every single message.

The usual implementation embeds each message into a vector store, then retrieves the most relevant past fragments and re-injects them into the prompt before the model replies — a pattern called retrieval-augmented generation, or RAG. Some apps add a pinning layer for details you explicitly mark as important, which survives regardless of what retrieval happens to surface.

Full guide: How AI memory works in practice

Character card

Also known as: Bot card, V2 card

A character card is a portable file carrying a character's full definition — name, personality, scenario, greeting, and example dialogue — usually as a PNG with the JSON embedded in its metadata, so the image doubles as both avatar and data.

The embedded-PNG convention is what makes cards shareable: drag the image into a supporting app and the character arrives intact. Raw JSON works identically without the picture. Field names follow a community-maintained spec, so a card exported from one front-end generally imports into another without editing.

Full guide: Character cards explained, PNG vs JSON

NSFW

Also known as: Not safe for work, Adult mode

NSFW stands for not safe for work, and in AI chat it describes a policy setting on the model and the platform rather than a genre of content — which is why the same label means materially different things from one app to the next.

Some platforms permit explicit text but not explicit images. Some permit both and still refuse specific themes. Some advertise NSFW support while routing replies through a filter that intervenes mid-scene, which is the version people complain about most. The label tells you a platform allows adult content; it does not tell you where that platform draws its line, and the line is what actually determines whether a scene survives to the end. Worth separating from uncensored, which describes the model's own tuning rather than the platform's policy.

Full guide: Where one major platform draws the line

Context window

Also known as: Context length, Token limit

A context window is the fixed budget of tokens a language model can read in a single turn, covering the system prompt, the character definition, retrieved memory, and recent messages combined — which is the direct mechanical reason a character starts forgetting once a conversation outgrows it.

Everything the model knows on a given turn has to fit inside that budget, and when a conversation exceeds it the oldest messages stop being shown first. Nothing is deleted from your transcript; the model simply no longer sees the earlier part of it. A larger window pushes the problem further out without solving it, which is why persistent memory exists — retrieval spends a small slice of the budget re-showing a handful of relevant older fragments instead of trying to fit the whole history.

Full guide: The conversation engine, in depth

Token

A token is the unit of text a language model reads and bills in — roughly four characters of English, or about three-quarters of an average word — and it is not the same thing as an in-app credit, coin, or message allowance.

Tokenisation is why model limits are quoted in numbers that map cleanly to neither words nor characters. Common words are usually a single token, while rare words, names, and non-English text split into several, so two messages of identical length can cost noticeably different amounts. The confusion worth avoiding: a platform's purchasable currency is a billing abstraction that platform invented, whereas a token is a fixed property of how the model reads text.

System prompt

Also known as: Character definition, System message

A system prompt is the always-on instruction block that tells the model who a character is and how it should speak, sent along with every single message and never shown in the visible transcript.

It is the difference between a model and a character: the same base model produces entirely different output depending on what sits in this block. Most apps keep the character definition and the system prompt as separate fields that get concatenated — the definition supplies who, the prompt supplies how. Both are re-sent every turn, so both consume context window budget every turn, and an overlong system prompt eats the room otherwise available for memory and conversation.

Full guide: How to write a character bot

Greeting

Also known as: First message, Intro message

A greeting, also called the first message, is a character's opening turn — written by a human rather than generated by the model — and it silently sets the tense, point of view, and prose length that every later reply imitates.

This is the highest-leverage field in a character definition and the most commonly wasted one. The model reads the greeting as an example of correct output, so a greeting in third-person past tense produces a scene narrated in third-person past tense indefinitely, and a one-line greeting teaches the model that one-line replies are what you want. Writing the greeting in the register you actually want back works far better than any instruction telling the model to write that way.

Full guide: Images and formatting in a first message

OOC (out of character)

Also known as: Out of character

OOC stands for out of character, a bracketed side-channel for directing a scene without the instruction becoming part of the story — written inline as something like [OOC: skip ahead a week] so the model reads it as stage direction rather than dialogue.

The convention exists because roleplay has nowhere else to put author-level notes. Without it, asking a character to change the pacing gets absorbed into the fiction and answered in character, which is rarely what you meant. Most models handle the bracketed form reliably because it is well represented in roleplay training data. Its counterpart, IC or in character, occasionally marks the return to the scene, though most people simply resume.

The {{user}} and {{char}} macros

Also known as: Placeholders, Template variables

The macros {{user}} and {{char}} are placeholders in a character definition that the app substitutes at runtime — {{user}} becomes your persona name and {{char}} becomes the character's — so one definition works for any player without hardcoding names.

They matter most for portability. A card that hardcodes a player name reads as broken to everyone else who imports it, while the same card written with macros adapts silently. Support is near-universal across roleplay front-ends, though the exact set of available macros varies. A definition written with {{user}} also survives a persona rename, because substitution happens fresh on every turn rather than being baked in when the card was saved.

Full guide: Macros and persona setup

Large language model (LLM)

Also known as: LLM, Language model

A large language model, or LLM, is a system trained to predict the next piece of text given everything before it — the component that actually generates a character's words, and one that holds no state, no beliefs, and no memory of you on its own.

Everything that feels persistent in an AI companion is built around the model rather than inside it. The model receives a block of text and returns a continuation; between one message and the next it retains nothing at all. Persona consistency comes from re-sending the character definition every turn, and recall comes from a retrieval layer re-sending relevant history. Understanding that collapses most confusion about why characters forget, why they contradict themselves, and why memory is a product feature rather than a model capability.

Full guide: Which models these apps actually run

Reasoning model

Also known as: Thinking model

A reasoning model, sometimes called a thinking model, generates a hidden block of deliberation before producing its visible reply — spending extra computation on planning the answer rather than emitting it directly.

In roleplay the trade is concrete: the hidden pass buys better scene continuity, more consistent character logic, and fewer contradictions across a long session, at the cost of noticeably slower replies. Whether that is worth it depends on what you are doing. It helps most in plot-heavy, multi-character scenes and least in fast casual chat, where latency is the thing you actually feel. Some platforms expose the choice directly; others pick per conversation without surfacing it.

Full guide: Which models these apps actually run

ERP (erotic roleplay)

Also known as: Erotic roleplay

ERP stands for erotic roleplay — sustained adult roleplay with a character, distinguished from one-off explicit text or image generation by continuity, since the scene, the relationship, and the history all carry forward between sessions.

The distinction is technical as well as cultural. One-off generation needs only a good prompt, whereas ERP needs the same infrastructure as any long roleplay: a persona, a memory layer, and a model that will not break character partway through. It is also the use case most affected by filtering, because an intervention mid-scene is far more disruptive than a single refused request.

Full guide: AI sex chat roleplay, compared

Uncensored

Also known as: Unfiltered, No filter

Uncensored, or unfiltered, describes a model that has not been refusal-tuned to decline adult themes — meaning it engages with explicit material rather than deflecting, though it does not mean the platform around it applies no rules at all.

Two separate systems are usually at work: the model's own tuning, and a moderation layer sitting between you and it. A platform can run a genuinely uncensored model and still block specific categories at that layer, which is what every legitimate operator does. Content involving minors, real people's likenesses without consent, and non-consent themes is refused everywhere regardless of marketing. Uncensored is a claim about refusal behaviour on adult themes, not an absence of limits.

Full guide: Uncensored AI chat, compared

LoRA

Also known as: Low-rank adaptation, Character LoRA

A LoRA, short for low-rank adaptation, is a small trained adapter applied on top of a base image model so the same face and body render consistently across every generation, instead of producing a new stranger each time.

Character consistency is the hard problem in companion image generation, because a base model given the same prompt twice returns two different people. A LoRA encodes one specific character into a file small enough to train quickly and load on demand. IP-Adapter approaches solve the same problem by conditioning on a reference image rather than training an adapter, and some pipelines use both together.

Full guide: How companion images are generated

Image-to-video

Also known as: I2V, Animate image

Image-to-video is the generation mode that animates an existing still frame into a clip, as opposed to text-to-video, which builds a clip from a description alone and re-rolls the character's appearance in the process.

For a companion this is usually the only usable path, because the first frame anchors identity — the character in the clip is the character in the image you fed it. Text-to-video is stronger for scenes where no specific person has to persist. The practical workflow is therefore two-stage: generate a still you are happy with, confirm the likeness, then animate that exact frame.

Full guide: AI video generation

dreamcoin

A dreamcoin is the single unit of currency on ourdream.ai, spent across chat, image generation, video generation, and voice — one balance covering every feature rather than a separate quota per tool.

Cost scales with what the work actually takes: a longer video at a higher resolution costs proportionally more than a short one, and image batches cost per image. This is the one entry on this page that is specific to a single platform rather than industry-wide vocabulary — every companion app invents its own currency, and the units are not comparable between them.

Full guide: Plans and dreamcoins