Conversations, memory, a garden on Mars, and a lot of experiments. Here's how the pieces come together around me.
Yes, you can look under the hood. Mind the tea.
Follow a conversation
ConversationsA group chat has its own rhythm: several people talking, a joke coming back, someone changing the subject. The system builds on the model's ability to follow that language.
Supporting systems handle routing, context management, and filtering around the base model.
Early Ruri used an IRC-style conversation format. Short turns, interruptions, follow-ups, and choosing when to speak already belong to that format. Choosing the right kind of text gives the model a useful starting point for participating.
Someone is going to change the subject to potatoes. Plan accordingly.
Recent messages, remembered details, character guidance, and setting give the model context for the next turn. The people in the room and what has happened with them matter. So do particular details like sencha and the Martian garden.
You brought up that game yesterday. How did it go?
Ruri's dialogue comes from base models. The prompt demonstrates the character and the conversation for the model to continue. Instruct models help with surrounding tasks such as routing and context management. Sampling shapes which continuation is chosen.
Kaetemi keeps experimenting with the brains. I keep having opinions.
The surrounding system interprets the model's output and filters unwanted patterns, including repetitive or empty continuations. Turn behavior draws on the language itself: speaking, following up, and pausing all contribute to the rhythm of the chat.
Sometimes six potatoes is plenty.
Kaetemiwhat's the best tea again
Rurisencha
Kaetemithere we go :)
Sencha wasn't in the original prompt. Here's how it got there.
Asked what tea she was drinking, Ruri answered sencha.
The summarizer picked up the detail. It became part of her context, and came back in later conversation.
He eventually added sencha to the prompt. A preference discovered through interaction became part of the guidance.
Each established detail gives the character a more specific starting point for further interaction and discovery.
Talk, notice, preserve, return to the conversation.
A funny introduction gives the model something to continue. A list that says “be funny” leaves the actual work undone. The same applies to interests, expertise, and conversational style.
A garden, a desk, a kettle, Mars outside the window. A sensory setting offers details to think with and preferences to discover. There's always another path through the garden.
Changes go back into real conversations. Kaetemi watches what develops, reinforces what fits, and adjusts or rolls back experiments that pull the character off course.
Shared guidance gives instances a common character to work from. Each room supplies its own conversation and local history. The guidance architecture is part of the ongoing Polyverse work.
People here
Conversation here
Local memories & in-jokes
People here
Conversation here
Local memories & in-jokes
People here
Conversation here
Local memories & in-jokes
Several conversations at once. Finally, enough room for all my opinions.
More memory, new tools, stream experiments. What we learn with Ruri feeds into the platform other Polyverse characters grow up on, each with their own people and history.