SillyTavern Group Chats: Who Talks, and What They Know (2026)

Two settings decide almost everything about a SillyTavern group chat. One controls who gets drafted to reply. The other controls whether your characters can read each other’s cards at all.

Most group chat problems — the same character dominating, nobody answering, four replies where you wanted one, characters who behave like strangers in a scene they have shared for an hour — come down to those two, and both are sitting on defaults that few people ever open.

The defaults are worth stating plainly, because neither is obvious: every character has a 50% chance to speak on every turn, and characters cannot see each other’s descriptions.

This guide covers how the draft actually works, what each mode changes about what your cast knows, and why the obvious fix for the second problem creates a new one. If you have not set SillyTavern up yet, start with the setup guide; if it is the memory rather than the cast that is failing, that is a different stack of systems. If you came here from Character.AI, its group chat works very differently, and it is covered in Character.AI group chat.

Making One

A group is created from the character management panel — a new group, then characters added into it from your library.

The thing to know before you start: a group chat is its own chat with its own history. It is not a room your existing one-to-one conversations move into. Starting a group with a character you have chatted with for weeks gives you a character who remembers none of it, because none of it is in this chat’s history. If that history matters, you have to carry it across yourself, which is what a summary or a lorebook is for.

Each member then gets a row in the group menu with three controls worth knowing immediately:

Mute — the struck-out speech bubble. A muted character is skipped by the draft.
Force Talk — the speech bubble. Triggers a reply from that character only, bypassing the reply order strategy entirely, and it works even if they are muted.
Remove — takes them out of the group.

Force Talk is the single most useful control in the panel and the one most people find last. It is the manual override for every draft problem below.

Who Talks: The Four Strategies

The reply order strategy decides how members are drafted. There are four, and the default is Natural Order.

Strategy	How it drafts	Best for
Natural Order	Name mentions, then a talkativeness roll per member, then a random fallback	Conversations that should feel unscripted
List Order	Strictly by position in the member list, no other rules	Round-table scenes, predictable rotation
Pooled Order	A random member who has not spoken since your last message; resets when all have	Making sure everyone gets a turn
Manual	Nobody replies automatically; you pick the speaker	Directed scenes, and writing-heavy play

Manual deserves a note because it generates a steady stream of “my group chat is not replying” reports. On Manual, user messages do not trigger replies at all. That is the documented behaviour, not a fault. You select the speaker from the menu or with the trigger command. Sending a generation with empty input picks a random unmuted member.

How Natural Order Actually Drafts

This is the default, so it is the one worth understanding properly. It runs in three stages.

Stage one: name mentions. The last message is split into words, and any member whose name contains one of those words as a whole word is activated. The documentation is explicit that only whole words count — a character named Misaka Mikoto activates on “Misaka” or “Mikoto”, never on “Misa” or “Railgun”.

This is the most useful practical fact in the whole feature. Typing a character’s name is the reliable way to make them speak. It happens before talkativeness is considered and it is deterministic. If you want a specific character to answer, name them.

It also explains a frustration that looks like a bug: characters with nicknames do not respond to their nicknames. If your cast calls someone “Doc” and the card says “Doctor Ellery Vance”, “Doc” matches nothing. The fix is to rename the character to the name you actually use in play.

Stage two: the talkativeness roll. Every remaining member is shuffled and then rolls. Each character has a Talkativeness value, defaulting to 0.5 for new characters, and activates if that value is at least as high as a fresh random number. The slider lives in the Advanced Definitions screen of the character editor — per character, not in the group panel, which is why so few people adjust it.

Because every member rolls independently, the number of speakers per turn is not fixed. In a four-character group at default talkativeness, you will average about two replies per turn, but a turn producing one reply and a turn producing all four are both perfectly ordinary. That randomness is the feature working as designed, and it is also the answer to “why did everyone answer at once”.

Stage three: the fallback. If nothing activated in either stage, one member is picked at random, preferring those with talkativeness above zero. So a turn almost never produces silence.

One more rule sits across all of this: the character who spoke last cannot be drafted again, unless you enable Allow Self Responses. That is what stops a single character talking to themselves indefinitely, and turning it on is how you allow a character to continue their own thought across two messages.

Tuning The Draft

Once the mechanism is clear, the fixes are direct:

One character dominates. Lower their Talkativeness in Advanced Definitions. Check that you are not naming them constantly, which activates them before the roll.
A character never speaks. Raise their Talkativeness, or name them. A character at 0% is a “shy” character who only ever speaks when mentioned — which is exactly right for a narrator or a background presence.
Too many replies per turn. Lower talkativeness across the cast, or switch to Pooled or List order, which draft one at a time.
You want a guaranteed rotation. List Order. It ignores mentions and talkativeness entirely.
You want this specific character, right now. Force Talk.

What They Know: Swap Versus Join

This is the second setting, and it changes the character of a group chat more than anything else in the panel.

One thing holds either way, and it is worth saying first: the chat history is always shared between all members, in both modes. Every character can see what has been said in front of them.

Swap character cards is the default. Each time a message is generated, only the active speaker’s card goes into the context. The model is told it is one specific person and nothing about anyone else.

Join character cards combines every member’s card into one prompt, in list order. The documentation lists exactly what gets merged: description, scenario (unless overridden for the chat), personality, message examples, and character notes or depth prompts.

The practical difference is a question about your cast:

In Swap, characters know only what has happened in front of them. If one character’s card says she is secretly the exiled heir and that has never come up in conversation, nobody else has any idea — including in ways you want, because that is how a secret works. If two characters are written as siblings and neither has said so in this chat, they will meet like strangers.

In Join, every character has read everyone’s card, including the secrets. Mutual awareness arrives immediately. So does omniscience.

The Warning On Join Is Real

SillyTavern’s own documentation does not hedge here. It says that because of how a typical character card is structured, Join mode can lead to unexpected behaviour including characters being confused about themselves, having merged personalities, and uncertain traits.

The reason is structural. A character card is written in the second person or about “you” — it is an instruction addressed to the model about who it is. Concatenate four of those and you have handed the model four conflicting identity statements at once and asked it to be one of them. Traits bleed. A reserved character picks up a talkative one’s speech patterns. Someone answers to the wrong name.

There is a mitigation built for exactly this: Join Prefix and Suffix. When Join is selected you can define text that wraps each character’s fields, so the prompt reads as four labelled blocks rather than one undifferentiated blob. Both support macros, and both replace the character’s name and a field-name placeholder, so you can produce headers along the lines of a name followed by which field this is. If you use Join at all, use these. The difference between a labelled Join and an unlabelled one is most of the difference between the mode working and the mode producing four-way personality soup.

Join also has two sub-modes, and you must pick one: include muted characters in the joint prompt always, or exclude muted ones unless they are the current speaker.

Which To Use

Situation	Mode
Characters should discover each other through play	Swap
A character has a secret the others must not know	Swap
The cast are established friends or family who obviously know each other	Join, with prefixes
Characters keep acting like strangers despite shared history	Join, with prefixes
Running on llama.cpp and prompt reprocessing is slow	Join — the context stops being rewritten per speaker
Characters are blurring into each other	Swap

The llama.cpp case is worth calling out because it is a performance argument rather than a creative one. In Swap mode the character card changes for every speaker, which invalidates the cached prompt each turn. Join keeps the context stable, which the documentation notes helps where altering large chunks of context is undesirable. On a local backend that can be the difference between a group chat that flows and one that stalls for thirty seconds per character.

Adding A Narrator

Nothing in SillyTavern marks a character as a narrator. You build one by making an ordinary group member whose card describes scenes instead of playing a person, and then controlling when they are drafted.

Two settings do the work:

Talkativeness at or near 0. The narrator then speaks only when named or forced — which is usually what you want, since narration on a random coin flip interrupts dialogue at arbitrary moments.
A short, distinctive name. Mentions are whole-word matches, so a one-word name is easy to summon deliberately and unlikely to fire by accident. Avoid a name that is also a common word, or every use of that word drags the narrator into the scene.

Then use Force Talk to call for narration exactly when you want it. That combination — near-zero talkativeness plus Force Talk — gives you a narrator on demand rather than a fifth voice competing for turns.

On the mode question, a narrator is one of the better arguments for Join: a narrator who has read the cast’s cards can describe them accurately, while a Swap-mode narrator knows only what has been said aloud.

Auto-Mode

Auto-mode keeps the group talking with no input from you. It follows the same reply order strategy, triggers generation automatically, and waits five seconds after the last drafted character finishes before starting the next turn. Typing in the message box switches it off, so stepping back in is easy.

It is genuinely good for watching a cast interact, and it is the fastest way to spend money on a paid API that SillyTavern offers, because every turn is a full generation for every drafted character and nothing waits for your approval. On Natural Order with four chatty characters, a few minutes of auto-mode is a lot of requests. Run it on a local model, or watch it.

Things That Look Broken And Are Not

Nothing generates when you send a message. Every member is muted or disabled, so your message is posted as-is with no reply. Check the group menu. Or you are on Manual, where this is the documented behaviour.
A deleted member was swiped. Swiping a message from a character who has since been removed produces a warning and no reply. Add them back to the group to swipe their message.
Replies arrive in an order you did not choose. Natural Order shuffles before rolling. It is not going down the list; List Order is the one that does.
The narrator keeps interrupting. Talkativeness is 50% by default, including for narrators. Set it to zero.
Characters ignore an established relationship. Swap mode. They have never read each other’s cards, only the history.
Characters have started blurring together. Join mode without prefixes. Add them, or go back to Swap.

Checklist

A group chat is a new chat with its own history; nothing carries over from solo chats.
Natural Order drafts by whole-word name mentions, then a per-character talkativeness roll at 50% by default, then a random fallback.
Naming a character is the deterministic way to make them speak. Nicknames that are not the character’s name match nothing.
The last speaker cannot reply again unless Allow Self Responses is on.
Talkativeness lives in the character editor’s Advanced Definitions, not in the group panel.
Swap, the default, shows each speaker only their own card. History is shared in every mode.
Join merges all cards and the documentation warns it merges personalities with them. Use Join Prefix and Suffix if you use it at all.
Force Talk overrides everything, including a mute. Auto-mode runs on a five-second delay and costs real money on a paid API.

Frequently Asked Questions

How do I make a group chat in SillyTavern? Open the character management panel and create a new group, then add characters to it from your library. A group is its own chat with its own history rather than a room your existing one-to-one chats move into, so starting a group does not carry over what happened in a solo chat with the same character. Once the group exists, each member gets a row in the group menu with controls for muting them, forcing them to talk, and removing them, and the group’s own settings panel holds the two choices that matter most: the reply order strategy and the generation handling mode.

How does SillyTavern decide which character replies? On the default Natural Order strategy it runs three steps. First it looks for character names mentioned in the last message, and only whole words count, so a character called Misaka Mikoto activates on Misaka or Mikoto but never on Misa. Second, every remaining member rolls against their Talkativeness value, which is 50% for new characters, so each one independently gets a coin flip on every turn. Third, if nobody was activated by either step, one member is picked at random. That is why the number of replies per turn varies and why a four-character group averages about two speakers but sometimes produces one and sometimes four.

Do characters in a SillyTavern group chat know about each other? Not by default. The default generation mode is Swap character cards, where only the active speaker’s card is included in the context each time a message is generated. The chat history is always shared no matter which mode you use, so characters know what has been said in front of them, but they do not know anything from another character’s description, personality or scenario. Everything a character knows about their companions has to have happened in the visible conversation. Switching to Join character cards puts every member’s card in every prompt, which fixes that and introduces a different problem.

What is the difference between Swap and Join character cards? Swap includes only the speaking character’s card, so each reply is generated by a model that has been told it is one specific person. Join combines the description, scenario, personality, message examples and depth prompts of every member into a single prompt in list order. Join is useful when you want the cast to be mutually aware, and it also helps with backends that benefit from a stable prompt, such as llama.cpp prompt caching, because the context is not being rewritten for every speaker. The documentation warns that Join can cause characters to become confused about themselves, develop merged personalities and show uncertain traits, which is a real risk rather than a theoretical one.

Why does the same character keep replying in my SillyTavern group chat? Check Talkativeness first, since it is per-character and set in the Advanced Definitions screen of the character editor rather than in the group panel. A character left at a high value will be drafted far more often than one set low, and because each member rolls independently a chatty character wins a lot of turns. The second cause is name mentions: if you keep writing one character’s name, you keep activating them explicitly, before talkativeness is even considered. The third is List Order strategy, which drafts strictly by position in the member list and ignores every other rule.

Why is my SillyTavern group chat not replying? If nothing happens at all when you send a message, the usual cause is that every member is muted or disabled, in which case your message is posted as-is and no generation runs. Check the group menu for muted members, marked with a struck-out speech bubble. If you are on the Manual strategy, that is working as designed: user messages do not trigger replies at all, and you pick the speaker yourself from the menu or with the trigger command. If you want a specific character to speak regardless of the strategy or their muted status, use Force Talk, the speech bubble icon beside their avatar.

Can you have a narrator in a SillyTavern group chat? Yes, and the usual approach is to add a narrator as an ordinary group member with a card written to describe scenes rather than to play a person. Set their Talkativeness deliberately: low if you want narration only when you call on them by name, higher if you want scene-setting between character turns. Because name mentions are matched on whole words, giving the narrator a short distinctive name makes them easy to summon on demand. Force Talk also works well here, letting you call for narration at exactly the moment you want it without changing any settings.

What is auto-mode in a SillyTavern group chat? Auto-mode makes the group keep talking without you. It follows the same reply order strategy and triggers generation with no user input, waiting five seconds after the last drafted character finishes before starting the next turn. It switches itself off as soon as you start typing in the message box, so you can step back in whenever you want. It is the closest thing to watching your cast hold a conversation on their own, and it is also the fastest way to spend money on a paid API, since every turn is a full generation and nothing is waiting for you to approve it.

Activation strategies, the talkativeness default, the self-response rule and the auto-mode delay were read from SillyTavern’s own source on 22 September 2026 and cross-checked against the project’s group chat documentation, which is also the source of the warning quoted about Join mode. SillyTavern is free, open-source software and Arcanum is an independent publication with no affiliation to the project or to any model provider or backend named here; all names are used nominatively.