OpenRPG

Inspiration

I've been following agent memory research loosely for a long time, it's been a personal interest. Recently I came across two product concepts that stuck with me: Miromind and Flipbook. They gave me the idea of a world simulator that combines social simulation with visual presentation, instead of treating either as the whole product.

From there it clicked: per-NPC and per-location facts could work the same way agent skills do in the current paradigm. only retrieved when needed. That gives concrete signal for what's relevant in the moment, instead of stuffing everything into context and hoping the model picks the right thing.

What it does

Build your own world, or play in someone else's. Add NPCs, locations, skills however you want.

when you experience a cool plot line. you can share it. your moment becomes one of the canonical plots that future players can experience too. all the canonical plot presented as choices player can choose if they do not wish to freestyle their own action. and this will walk them down the cached canonical plot lines. no extra ai calls needed. The rest works like you'd expect from a GitHub for RPG worlds.

How we built it

The architecture is roughly inspired by the following:

• *Park et al.'s Generative Agents: Interactive Simulacra of Human Behavior (2023) *
• Concordia (Vezhnevets et al.)
• GitHub worlds as repos, contributions as patches, creator moderations.

Stack: TypeScript, Hono API, React web, Zod shared contracts. LLMs v Gemini 3.1 Pro for the GM (orchestration), Flash Lite for parallel NPC responses, Nano Banana 2 for image-edit contributions.

Challenges we ran into

Knowledge boundaries. which character knows what information. The problem is often not obvious. the NPC will not always use all the fact they receive. and track down each payload to diagnose is arduous

Planning. This is my first hackathon, and I'm still figuring out how to plan and timeblock . I had to replan a lot on the fly . It didn't end up super organized.

Accomplishments that we're proud of

live world generation it can pretty consistently generate world with intractable locations/npcs/all the other element needed to kick start creators' build. layered canon model the model story via players lived experience and collective creativity is very exciting.

What we learned

prompt engineering is really important in ai workflows in world generation. adjusting the prompt proved very important for consistency.

What's next for OpenRPG

** expand combat simulation system beyond simple number check ** multi creator moderation for repositories

Built With

• hono-api
• react-web
• typescript
• zod