MiroFish

🇹🇭 ภาษาไทย

Next-generation AI prediction engine ที่ใช้ multi-agent simulation สร้าง “parallel digital world” จาก seed information แล้วให้ agent หลายพันตัววิ่ง social evolution เพื่อ predict อนาคต

GitHub: github.com/666ghj/MiroFish | Demo: 666ghj.github.io/mirofish-demo Backed by: Shanda Group | Engine: OASIS (CAMEL-AI)

“Rehearse the future in a digital sandbox, and win decisions after countless simulations.”

แนวคิดหลัก

MiroFish เชื่อว่า prediction ที่ดีที่สุดไม่ได้มาจาก statistical model แต่มาจาก Swarm Intelligence — emergent behavior ที่เกิดจาก agent จำนวนมากที่มี personality และ memory ของตัวเอง interact กัน ผลลัพธ์คือ collective emergence ที่ยากจะ derive จาก top-down model

Workflow

1. Graph Building    → Seed extraction, GraphRAG construction, memory injection
2. Environment Setup → Entity extraction, persona generation per agent
3. Simulation        → Dual-platform parallel sim, dynamic temporal memory (Zep Cloud)
4. Report Generation → ReportAgent with rich toolset
5. Deep Interaction  → Chat with any agent in the simulated world

Use Cases

ระดับตัวอย่าง
Macroทดสอบนโยบาย, simulation ข่าว public opinion, prediction การเงิน
Creativeต่อจบนิยาย, สำรวจ what-if scenarios

Tech Stack

ComponentTechnology
Simulation engineOASIS (CAMEL-AI)
Agent memoryZep Cloud
Graph/retrievalGraphRAG
LLM backendOpenAI-compatible API (แนะนำ Qwen-plus)
Frontend/BackendNode.js 18+ / Python 3.11–3.12
  • Swarm Intelligence — ปรัชญาหลักที่ MiroFish ยึด
  • Multi-Agent Simulation — framework ที่ MiroFish สร้างขึ้นบน
  • RAG vs Wiki — MiroFish ใช้ GraphRAG เพิ่ม dimension ที่สามใน spectrum

🇬🇧 English

Next-generation AI prediction engine that uses multi-agent simulation to construct a “parallel digital world” from seed information, then runs thousands of agents through social evolution to predict future outcomes.

GitHub: github.com/666ghj/MiroFish | Backed by: Shanda Group (incubation)

“Rehearse the future in a digital sandbox, and win decisions after countless simulations.”

Core Concept

MiroFish’s thesis: the best predictions don’t come from statistical models but from Swarm Intelligence — emergent behavior arising from many agents with individual personalities and memories interacting together. The collective emergence is something that cannot be derived from any top-down model.

Workflow

1. Graph Building    → Seed extraction from inputs, GraphRAG construction, memory injection
2. Environment Setup → Entity relationship extraction, persona generation per agent
3. Simulation        → Dual-platform parallel simulation, dynamic temporal memory (via Zep Cloud)
4. Report Generation → ReportAgent with rich toolset, deep interaction with post-sim environment
5. Deep Interaction  → Chat with any agent in the simulated world, follow-up with ReportAgent

Use Cases

LevelExamples
Macro / SeriousPolicy testing, news public opinion simulation, financial prediction
Micro / CreativeCompleting novel endings (e.g. Dream of the Red Chamber), exploring what-if scenarios

Technical Stack

ComponentTechnology
Simulation engineOASIS (CAMEL-AI open-source framework)
Agent memoryZep Cloud (long-term per-agent memory with temporal tracking)
Graph/retrievalGraphRAG
LLM backendOpenAI-compatible API (Alibaba Qwen-plus recommended)
FrontendNode.js 18+
BackendPython 3.11–3.12, uv

Unlike MemPalace which runs fully locally, MiroFish requires an LLM API key and a Zep Cloud account (free tier available).

Memory Approach Comparison

SystemMemory Approach
MemPalaceVerbatim conversation storage + semantic retrieval
LLM Wiki PatternCompiled synthesis in wiki pages
MiroFish / ZepPer-agent identity memory for simulation continuity