Project Overview
MiroFish is a group-intelligence prediction engine developed by 666ghj, built on multi-agent technology. It constructs a high-fidelity parallel digital world to simulate the future evolution of complex systems.
- GitHub: https://github.com/666ghj/MiroFish
- Stars: 4,180+ | Forks: 512+
- Backer: Shanda Group
- Related project: BettaFish (AI public opinion analysis system)
- Temporal modeling: from static analysis to dynamic evolution
- Social simulation: from information extraction to group behavior prediction
- Parallel worlds: from a single reality to multi-scenario what-if simulation
- Temporal GraphRAG handles time-ordered memories better than standard RAG
- Fine-grained personas down to social media activity hours
- Dual-platform parallel simulation improves reliability
- Deep interaction: users can converse with any agent to understand its decision logic
- Compute cost: simulating thousands of agents is resource-intensive
- Validation difficulty: predictions are hard to verify before real events unfold
- Boundary effects: simulated world rules may limit result validity
- Data dependency: seed material quality directly drives prediction quality
Core Design Philosophy
"Upload any report, simulate the future instantly"
MiroFish attempts to answer a question: if thousands of AI agents with independent personas could freely interact in a virtual world, could they predict future trends in the real world?
Unlike traditional single-agent Q&A systems, it is a social simulator.
System Architecture
1. Four-Layer Workflow
1. Graph Construction — information extraction from seed material; memory injection (individual and collective memories per agent); temporal GraphRAG knowledge graph for high-fidelity background context. 2. Environment Setup — automatic persona generation from materials; social relationship networks between agents; configuration of time, space, and rules. 3. Simulation Execution — dual-platform parallel simulation instances; dynamic memory updates based on interactions; emergent social behaviors such as posting, liking, and interviews. 4. Report Generation — a ReportAgent aggregates all simulation data; trend analysis identifies key nodes in social sentiment evolution; detailed forecast reports are generated.
2. Technology Stack
| Component | Technology | Role | |------|-----------|------| | Agent framework | CAMEL-AI OASIS | Multi-agent collaboration infrastructure | | Knowledge graph | Temporal GraphRAG | Handles time-ordered complex memories | | Deployment | Docker / source | Fast deployment and secondary development | | Multimodal | Built-in vision encoder | Supports image and video input |
3. Agent Persona System
Each agent has fine-grained attributes, including Big Five personality scores (openness, conscientiousness, extraversion, agreeableness, neuroticism), short-term / long-term / collective memory stores, and behavioral parameters such as active hours, social bias, and preferred platform.
Relationship to BettaFish
The author previously built BettaFish, an AI public opinion analysis system with a multi-engine architecture (QueryEngine, MediaEngine, InsightEngine, ForumEngine, ReportEngine). MiroFish extends this "current-state analysis → future prediction" toolchain with:
Application Scenarios
Scenario 1: Policy Rehearsal
Input a draft license-plate restriction policy, historical similar cases, and local demographic data. The system generates thousands of agents representing different groups (car owners, transit riders, environmentalists), simulates social media discussion and adaptation, and outputs predicted social acceptance and risk points — letting decision-makers experiment in a zero-risk environment.Scenario 2: PR Crisis Simulation
Given an incident description, brand background, and public sentiment data, consumer/media/KOL agents simulate how different apology strategies evolve public opinion, recommending the optimal crisis-response approach.Scenario 3: Novel Ending Extrapolation
Using the first 80 chapters of *Dream of the Red Chamber*, Qing dynasty social context, and the author's style, character agents simulate choices and fates during the family's decline, producing endings consistent with the original logic.Strengths and Limitations
Strengths:
Limitations:
Assessment
MiroFish represents an evolution of AI applications from single assistants to social groups — instead of one AI answering your question, thousands of AIs "live out" the answer in a virtual society. This computational sociology approach exists in academia but has few open-source engineering implementations.
Engineering maturity: one-click Docker deployment, detailed configuration docs, OpenAI-format multi-LLM compatibility, and visual demos (Wuhan University public opinion, Dream of the Red Chamber simulation).
Community traction: 4,180+ stars and 512 forks indicate strong appeal and willingness to build on it; the BettaFish pairing creates a complete ecosystem, with corporate backing from Shanda Group suggesting sustained investment.
Worth watching: accuracy benchmarking against real outcomes, performance optimization for large-scale commercial scenarios, and integration with external real-time data sources (public opinion feeds, economic indicators).
*Note: This analysis is based on public documentation and community materials; no source-code-level review was performed.*