MiroSharkMiroShark

Personas

How MiroShark grounds simulated agents in 5 graph context layers, live web research, and Nemotron demographics.

Each simulated agent is generated from the knowledge graph, grounded in five context layers, and anchored to realistic demographics.

Five-layer grounding

entity -+- attrs          (Neo4j properties)
        +- relationships  (1-hop edges)
        +- semantic       (vector neighbors)
        +- BFS context    (multi-hop)
        +- web enrichment (public figures / thin context)
                  v
        persona prompt  ->  Wonderwall agent loop
Layer / FeatureHow it works
Graph contextCombines node attributes, 1-hop edges, semantic vector neighbors, and multi-hop BFS context.
Individual vs. institutionalOntology rules switch the persona schema: individuals speak in first person; institutions speak in official press-release voice.
Live web enrichmentSet WEB_SEARCH_MODEL=perplexity/sonar-pro (or self-hosted MIROSHARK_SEARXNG_BASE_URL + MIROSHARK_FIRECRAWL_BASE_URL) for citation-backed context on public figures.
Demographic anchoringNon-named agents draw age, region, education, profession, and platform habits from a Nemotron-aligned population sample.

Graph Memory

Inside MiroShark's Neo4j temporal knowledge graph, entity resolution, hybrid retrieval, and BGE reranking.

Simulation Engine

How Twitter, Reddit, and a Polymarket prediction market run in lockstep with belief updates and sliding-window memory.

On this page

Five-layer grounding