Overview
How MiroShark works: temporal knowledge graphs, grounded personas, 3-platform simulation loops, and cited reports.
MiroShark is a universal swarm-intelligence engine. Give it a scenario, URL, or document and it builds a temporal knowledge graph, grounds dozens of distinct personas, runs three platforms in lockstep, and writes a cited post-mortem report.
Two ways to run MiroShark
- Open-source engine (GitHub): self-host locally (
http://localhost:5001in dev) with one OpenRouter key, local Ollama, or Claude Code mode. - Hosted x402 API (
https://x402.miroshark.xyz): POST a seed, pay $1 USDC on Base, Solana, or Monad, and get a full simulation report in ~10 minutes with no API key.
Five-stage pipeline
document -> graph build -> persona grounding -> 3x platform loop -> report
| self-hosted only:
+- Director Mode (inject)
+- Fork (branch)
+- Persona Chat (DM)1. Graph build
Source text is parsed into a Neo4j temporal knowledge graph via batched UNWIND transactions, with entity resolution and contradiction adjudication.
2. Persona setup
Agents are generated from the graph and grounded in five context layers: node attributes, 1-hop relationships, semantic search, BFS neighbors, and optional live web research.
3. Simulation loop
Twitter, Reddit, and a Polymarket-style prediction market run concurrently each round via asyncio.gather. A market-media bridge connects trader prices and social posts while agent beliefs update every round.
4. Cited report
A ReACT report agent synthesizes the post-mortem from simulation_feed, market_state, graph search, and belief_trajectory. Every claim cites a real post or trade.
5. Interaction & exports
Self-hosted runs support mid-run news injection, counterfactual forks, and agent DMs. Every run exposes trajectory charts, signal.json, and provenance exports.
Core sections
Graph Memory
Temporal Neo4j ingestion, contradiction adjudication, and BGE-reranked hybrid retrieval.
Personas
Five-layer entity grounding, institutional vs. individual voices, and demographic anchoring.
Simulation Engine
Concurrent Twitter, Reddit, and Polymarket rounds with a live market-media bridge.
x402 Quickstart
Pay $1 USDC per simulation on Base, Solana, or Monad with no account or API key.