ZJU-LLMs/OpenStory

What it solves

OpenStory is a multi-agent simulation framework designed to create interactive, LLM-driven story worlds. It allows users to generate complex narratives where autonomous agents interact based on their personalities and environments, moving beyond static scripts to create dynamic, emergent plotlines.

How it works

Built on the Agent-Kernel architecture, the system uses a plugin-based mechanism to manage the full lifecycle of agents—including perception, planning, execution, and reflection. It utilizes YAML files for flexible configuration of environments, actions, and agent settings. The framework supports dynamic addition and removal of agents during a simulation and integrates with Redis for data busing and caching. It also includes a "WorldKernel" feature that can automatically decompose a single natural language prompt into roles, locations, paths, and rules to instantiate a new simulated world.

Who it’s for

• Storytellers and writers who want to simulate parallel universes or expand on existing classics.
• Developers interested in multi-agent system (MAS) research and interactive narrative generation.
• Gamers looking for an interactive experience where they can observe or participate in AI-driven story simulations.

Highlights

• WorldKernel: Generate a fully functional interactive story world from a single sentence.
• Dynamic Agent Management: Ability to add or remove agents on the fly during the simulation process.
• Visual Interface: Includes 1:1 simulation maps (e.g., for the Dream of the Red Chamber) to track agent trajectories and status in real-time.
• Plugin Architecture: Highly configurable system for agent perception and decision-making.
• Pre-built Scenarios: Includes detailed simulations for Dream of the Red Chamber and Westworld.