MiroFish

About

MiroFish is an open-source AI swarm intelligence engine that constructs fully simulated social environments from seed data — news articles, policy drafts, financial signals — and populates them with thousands of AI agents, each carrying distinct personalities and persistent memory. Instead of asking a single LLM what might happen next, MiroFish runs a parallel digital society and lets social dynamics emerge organically. The simulation unfolds in five stages: graph construction (extracting knowledge from seed materials via GraphRAG), environment setup (generating entity relationships and character profiles), simulation launch (dual-platform parallel execution with dynamic memory via Zep Cloud), report generation (a ReportAgent synthesizes findings from the post-simulation environment), and deep interaction (you can dialogue directly with simulated agents and the ReportAgent). Backed by Shanda Group and released under AGPL-3.0, MiroFish has gained over 32,000 GitHub stars since November 2025 — a growth rate that signals serious attention from researchers and applied AI teams. Documented case studies include predicting public opinion outcomes at Wuhan University and completing the lost ending of Dream of the Red Chamber through narrative simulation. The architecture supports any OpenAI-compatible LLM API (the team recommends Alibaba Qwen-plus) and requires Zep Cloud for agent memory management. Deployment is via Docker or direct Python setup (Python 3.11–3.12, Node.js 18+). Ideal use cases span political scenario modeling, policy impact analysis, brand sentiment forecasting, narrative prediction, and academic social science research. The core differentiator: most forecasting tools treat prediction as a calculation problem. MiroFish treats it as a simulation problem — and the results from complex social scenarios are significantly more nuanced than single-pass LLM predictions.

Key Features

• Thousands of AI agents with distinct personalities and persistent memory (via Zep Cloud)
• GraphRAG knowledge graph construction from seed documents and news articles
• Dual-platform parallel simulation with real-time variable injection
• Interactive dialogue with individual simulated agents post-simulation
• ReportAgent synthesizes detailed forecast reports from simulation output
• Docker deployment option for isolated, reproducible environments

Use Cases

• 1Public opinion forecasting from policy announcements or news events
• 2Scenario planning for corporate strategy or geopolitical analysis
• 3Social science research with large-scale agent-based modeling
• 4Narrative completion and story world simulation
• 5Brand sentiment simulation before product launches