MiroFish

One of MiroFish's demos feeds it the first 80 chapters of Dream of the Red Chamber and asks how the lost ending might go. The engine takes seed material like that, whether a data report or a novel, plus a prediction request written in plain language, and builds a parallel digital world from it: GraphRAG memory, generated personas and thousands of agents with their own memory and behavioral logic that interact over simulation rounds.

At the end a ReportAgent writes a prediction report, and users can keep talking to any agent in the simulated world. Running it from source needs Node.js 18+, Python 3.11 or 3.12, uv, a Zep Cloud key and an LLM API in the OpenAI SDK format, with Qwen-plus recommended. The project warns that simulations use the LLM API heavily and suggests starting with fewer than 40 rounds. The license is AGPL-3.0.