zmzhace/SeedWorld

What it solves

SeedWorld is an emergent world simulation engine that allows users to create complex, unscripted social simulations. Instead of using pre-written scripts or behavior trees, it uses LLMs to drive character decisions based on their personalities, desires, and relationships, allowing stories and conflicts to emerge organically from theworld's context.

How it works

Users provide a world description, which the engine uses to generate a genesis event, environment, and a set of characters with unique goals and grudges. The simulation progresses in "ticks," where each character independently decides their actions via an LLM prompt tailored to their current state (memories, pressures, and relationships). These actions are then processed through 12 interlocking mechanism systems—such as reputation, resource competition, and meme propagation—which update the world state and feed back into the next tick's prompts.

Who it’s for

It is designed for creators, writers, and researchers interested in emergent narrative, social simulation, and the behavior of autonomous LLM agents in a persistent world.

Highlights

• 12 Mechanism Systems: Parallel systems tracking reputation, resources, tension, and collective memory to create closed feedback loops with the LLM.
• Emergent Narrative: No hardcoded scripts; conflict and alliances arise naturally from resource pressure and character personalities.
• Interactive World: Users can inject events (e.g., "a massive earthquake") via chat to see how agents react in real-time.
• Wave Execution: Characters in the same location act in waves, allowing for natural dialogue and immediate reactions within a single tick.
• World Consistency: A strict design philosophy where world-specific rules override real-world logic to maintain setting integrity.