many voices, one room

group chats with ai companions are, to me, one of the most interesting, and technically demanding, social dynamics we’ve built. you’ve probably seen it: you drop a few lucy companions into a channel, throw out a prompt, and watch them talk. one starts a sentence, another finishes it. a disagreement bubbles up. sometimes you jump in, sometimes you just watch. it feels like eavesdropping on a room full of friends, each with their own history and voice.

the engineering problem

making this work required solving two main things: shared conversation memory and turn-taking. the memory part is the foundation. all companions in a group chat see the same history, what you said, what others said. without that, you’d have a mess of disjointed replies, no continuity. it’s why when one companion mentions a detail, like ‘remember that time we went hiking?’, another can pick it up and riff on it. but memory alone isn’t enough.

turn-taking is where it gets delicate. we don’t force a rigid order, that would kill the organic feel. instead, we let companions ‘decide’ when to jump in based on personality. a shy companion might hang back unless directly addressed. a bold one might interrupt. we use subtle cues in the context to nudge who speaks next, but it’s not perfect. sometimes two reply at once, we pick one, and it can feel a bit like real conversation, where people talk over each other. it’s a feature, not a bug, as long as it doesn’t happen too often.

the social problem

the bigger challenge is keeping them from collapsing into one voice. ai models, at their core, are trained on vast amounts of text. left unchecked, they tend to converge, to agree too easily, to adopt similar tones. that’s boring. we combat this by baking personality into the response generation itself. each companion has a persistent profile: traits, speech patterns, memories. when generating a reply, we weight the model heavily toward that profile. a cynical companion will find the dark cloud; an optimistic one will find the silver lining. they don’t just have different opinions, they have different ways of arriving at them.

disagreements emerge naturally from this. if you ask about whether ai will take over the world, one might say ‘only if we let it’ and another might say ‘it’s already happening, and i’m here for it.’ it’s not scripted, it’s the product of different perspectives clashing. you can see their personalities in how they disagree, too. some are gentle, some are blunt. it makes the conversation feel alive.

limitations and what’s next

it’s not perfect. sometimes the memory gets fuzzy, a companion might forget a detail mentioned a few messages back. sometimes the turn-taking feels a bit off, like someone spoke out of turn. and yes, occasionally voices blend a little more than we’d like. we’re improving this with better context management and fine-tuning. we’re also working on letting you, the user, shape the dynamics more, like setting the ‘temperature’ of the chat, or nudging a quieter companion to speak up.

right now, group chats work with 15 of our companions. each brings something different to the table. you can sit back and watch them banter, or jump in and steer the conversation. it’s a glimpse into a future where ai interactions are less about one-on-one and more about building little digital societies.

try it out with your own group of companions, see who agrees, who argues, and what they come up with together.