AI Roleplay Terms: A Plain-English Glossary

Every hobby grows a dialect, and AI roleplay grew one fast — a mix of fanfiction conventions, tabletop habits, and machine-learning jargon that nobody stops to define. This glossary defines the terms you'll actually encounter, in plain English, with the roleplay-specific meaning where it differs from the textbook one. It's organized from the conversation outward: how people talk, how stories work, how characters are built, what the machine is doing underneath, and the safety vocabulary that should sit around all of it.

We build teasa, an adults-only AI companion and roleplay platform, so where a term touches design choices we'll note how we think it should work — but the definitions themselves are platform-neutral.

Conversation basics

Persona

The character you play in a conversation — your name, description, and how the AI should refer to and treat you. A persona is distinct from your account: you might be a wry detective in one story and simply yourself in another, and a well-designed platform keeps those personas scoped so details don't leak between them. When people search "what is a persona AI," this is usually the answer: not the AI's character, but yours.

OOC (out of character)

Talk that steps outside the fiction, usually marked with brackets or parentheses: (OOC: let's skip ahead to the next morning). It's the standard way to give the AI direction — pacing, tone, corrections — without your character saying it aloud in the story. The opposite is IC (in character). Good roleplay etiquette, human or AI, keeps the two clearly separated.

Turn

One exchange: your message plus the AI's reply. Turns are the heartbeat of roleplay pacing — short turns feel like banter, long turns feel like collaborative prose — and many platform limits (memory retrieval, context assembly) are computed per turn.

Regeneration (swipe)

Asking the AI to try the same reply again, producing an alternative version of the response you just got. "Swipe" comes from mobile apps where you swipe sideways through variants. In a well-built system, regenerating preserves the earlier variants rather than destroying them, so you can compare and pick — a small feature that quietly protects your agency over the story.

Story mechanics

Branching

Splitting a conversation at any message into a parallel timeline, so you can explore "what if I'd said this instead" without erasing the original. Branching turns a chat log into a tree. The design principle worth demanding: creating a branch should never destroy the branch you came from.

Canon

The set of story facts treated as settled truth — what has definitively happened, as opposed to drafts, regenerated variants, or abandoned branches. Some platforms let you pin a message to canon, explicitly promoting it to the story's official record. Clear canon is what keeps long roleplays coherent instead of drifting into contradiction.

Info box

A structured status block some characters display alongside replies — time of day, location, outfit, relationship stats, health, whatever the creator defined. It borrows from games: a HUD for your story. Info boxes are powered by a state schema the creator sets up, and they help both you and the model track details prose alone would blur.

Staged intro

An opening sequence that unfolds in steps — scene-setting, then the character's entrance, then a first choice — rather than dumping one wall of text. Creators use staged intros to establish tone and stakes before the user's first message, which is often the difference between a story that hooks and one that stalls on "hi."

World rules

The standing constraints a creator writes for a story's universe: the magic system, the era's technology, what tone is allowed, what never happens. World rules sit above any single scene and keep the model from casually breaking the setting. In practice they live in the character definition or an attached lorebook.

Characters and creation

Character card

The portable definition of an AI character: name, description, personality, opening scene, example dialogue, and behavioral instructions. The card is what the model actually reads — the character is the card, plus the model interpreting it. Writing one well is its own craft; our character card guide covers it in depth.

Lorebook

A collection of world knowledge — places, factions, histories, secondary characters — stored as separate entries injected into the conversation only when relevant, instead of permanently occupying space in the prompt. Lorebooks are how creators build deep settings without drowning the model in text it rarely needs. See the full lorebook guide for how to structure one.

Trigger (keyword entry)

The activation mechanism inside a lorebook: each entry lists keywords, and when one appears in recent conversation, that entry's content gets pulled into the model's context. Mention "the Hollow Court" and the Hollow Court's entry loads; never mention it and it costs nothing. Choosing good triggers — specific enough not to fire constantly, broad enough not to be missed — is most of lorebook craft.

Cast

The set of characters in a multi-character story, and the mapping of who is played by the AI versus by you. Multi-character scenes are where roleplay AI most often fumbles — speakers blur, or the narration hijacks your character — so serious platforms validate cast and speaker boundaries on every generated reply.

Agency

Your exclusive control over your own character: the AI narrates the world and its cast but never speaks, acts, or decides as you. "The AI wrote my dialogue for me" is the classic agency violation, and it's the fastest way to make roleplay feel like reading someone else's story. Teasa treats agency as an enforced rule — generated replies in strict multi-character stories are screened before release so the model can't take over the player's character — not an etiquette suggestion.

Example dialogue

Sample exchanges included in a character card showing how the character speaks — cadence, vocabulary, humor, how they handle affection or conflict. Models imitate demonstrated style far more reliably than described style, so three good example exchanges usually outperform a paragraph of adjectives like "witty and mysterious."

Under the hood

LLM (large language model)

The AI engine underneath every companion and roleplay app: a model trained on enormous amounts of text to predict what comes next, which turns out to produce fluent conversation, narration, and characterization. The LLM has no feelings, memories, or persistence of its own — everything continuous about a companion (identity, history, world state) is built in layers around the model by the platform.

Token

The unit models actually read and write — chunks of a few characters; roughly three-quarters of an English word on average. Everything is metered in tokens: context limits, generation length, provider costs. When a long roleplay "forgets" something, the mechanical reason is usually that the relevant text no longer fits within the token budget.

Context window

The model's working memory: the maximum amount of text (in tokens) it can consider at once when writing a reply. Everything the model "knows" in a given moment — the character card, recent messages, retrieved memories, lorebook entries — must fit inside this window; anything outside it may as well not exist. This is the single most useful technical concept for roleplayers, because it explains both why long chats drift and why memory systems exist at all.

System prompt

Standing instructions the platform places before the visible conversation, invisible to you but read by the model on every turn — the character card, world rules, formatting requirements, and safety policy all typically live here. When a character "stays in character," the system prompt is doing much of the work.

Memory

The layer that carries facts across conversations, beyond the context window's reach: things you've shared, story canon, preferences. Approaches range from ambient extraction (the system silently mines your chats) to consent-gated designs where the companion asks before saving. Memory is the heart of the companion experience and the site of its worst dark patterns — our essay on how AI companion memory should work is the deep dive.

RAG-lite retrieval

Retrieval-augmented generation, informally: instead of stuffing everything into the prompt, the system stores knowledge outside the model and retrieves only what's relevant to the current turn — by keyword, by semantic similarity, or both. Lorebook triggers are a simple form of this; memory recall is another. The "lite" versions used in roleplay trade exhaustive search for speed and predictability.

Safety and control

Content level

A rating attached to a character or story declaring how far its themes go — for example everyone, teen, or mature. On teasa the taxonomy caps at mature: an adults-only platform for emotionally rich, romantic, and dramatic stories, with no explicit tier. Content levels only mean something when they're enforced consistently rather than treated as decoration.

Consent-gated memory

A memory design in which nothing is saved about you without your say: the companion proposes a memory ("should I remember that your sister is visiting?") and you approve, edit, or decline, with a record of that decision. It's the alternative to surveillance-style extraction, and it's teasa's default mode. The broader principle — escalation of intimacy, relationship labels, and remembered detail happens because you chose it — is what "consent-gated" means across a whole platform.

Moderation

The systems and humans that review published characters, stories, and community content against platform policy before and after it reaches other users. In a creator-driven roleplay platform, moderation is what stands between "anyone can publish" and "anything goes" — including holding new uploads for review and giving creators an appeal path when something is removed.

Age band

The audience classification a platform commits to — and the design consequences that follow. Open-ended emotional AI carries distinct risks for minors, which is why teasa is 18+ by policy: not an age checkbox, but a product decision about what the platform is for. A genuinely safe teen experience would require different models, catalogs, and safeguards, not a toggle.

FAQ

What does OOC mean in AI roleplay?

Out of character: a message (or part of one) addressed to the AI as a storytelling partner rather than spoken by your character, conventionally wrapped in brackets or parentheses. It's the standard channel for direction — "slow down," "skip to the next scene," "you got my character's name wrong" — without breaking the fiction inside the story itself.

What's the difference between a persona and a character card?

Direction of play. A character card defines a character the AI performs; a persona defines the character you perform, telling the AI who it's talking to. A rich roleplay usually involves both: the card animates the AI's side of the scene, the persona anchors yours.

Why do AI characters forget things mid-conversation?

Usually the context window: the model can only consider a fixed amount of recent text, and in a long chat, early messages fall outside it. Platforms compensate with memory systems and retrieval — summaries, saved facts, lorebook entries pulled in on demand — which is why memory design matters so much for long-running companions and stories.

Do I need to know these terms to use an AI companion?

No — a good platform works fine in plain conversation. But the vocabulary pays off fast if you create characters or run long stories: knowing what a token budget is explains forgetting, knowing what OOC is gives you a steering wheel, and knowing what consent-gated memory means tells you which platforms deserve your history. Start with our AI roleplay guide if you're new.