// the find
666ghj/MiroFish
A Simple and Universal Swarm Intelligence Engine, Predicting Anything. 简洁通用的群体智能引擎，预测万物
MiroFish spins up an LLM-agent swarm simulation (built on CAMEL-AI's OASIS) over a GraphRAG-constructed world model, then asks a ReportAgent to narrate what the simulated crowd 'predicted' — pitched for public-opinion and financial-forecasting scenarios. It's for people experimenting with multi-agent social simulation as a storytelling/scenario tool, not for anyone needing an actual forecasting system.
It reuses OASIS for the actual agent-simulation mechanics instead of reinventing that from scratch, and offloads long-term/temporal agent memory to Zep rather than hand-rolling a graph memory layer — both are reasonable choices to avoid solving already-hard problems twice. There's a real backend test suite (17+ files) covering Zep paging, retries, lifecycle, and ontology generation, which is more test discipline than most agent-demo repos bother with. LLM backend is swappable via any OpenAI-compatible API, so it's not locked to one vendor despite the README defaulting to Qwen-plus.
Calling this 'predicting anything' is overreach — there's no eval harness, backtest, or accuracy metric anywhere in the repo; it's LLM agents role-playing a scenario and a report agent summarizing the role-play, dressed up as forecasting. The dependency chain to even try it is heavy: an LLM API key, a Zep Cloud account, Node + Python + Docker, and the README itself warns simulations are 'high consumption' and to stay under 40 rounds — this gets slow and expensive fast for what is fundamentally a scripted multi-agent chat loop. Hard dependency on Zep Cloud (not just optional) means there's no fully self-hostable path if you don't want your simulation data living in a third-party memory service. README leans on demo videos and screenshots over documenting the actual ontology/ReportAgent tool APIs, so understanding what the simulation is really doing requires reading source, not docs.