Skip to main content
Back

Renmin University's EvoOntology Stops Data Agents From Relearning SQL Context Every Task

Renmin University researchers release a plugin for Claude Code and Codex that builds and evolves a semantic layer over your data, grounded in actual agent interactions.

  • EvoOntology is a self-evolving semantic layer plugin for Claude Code and Codex, from Renmin University of China.
  • Exposes ontology to agents via MCP tools browse_semantics and resolve_semantics instead of stuffing full context.
  • Three-layer design: Content graph (Terms, Mappings, Constraints, Evidence), Schema rules, and Tool runtime.
  • Updates ship only if a Candidate ontology beats its Parent under paired evaluation, making evolution reproducible.
  • Installs through plugin marketplaces, no clone or pip needed, MIT licensed.
  • Ships adapters for BIRD, DDR-10K, and InsightBench benchmarks with a documented extension contract.

Researchers at Renmin University of China have released EvoOntology, an MIT-licensed plugin for Claude Code and Codex. It builds a semantic layer across tables, files, and databases, then proposes updates based on the agent’s execution history and measured performance.

Data agents can inspect schemas and file paths, but those structures rarely explain what a metric means, which columns represent the same entity, or which business rules constrain a query. Agents often reconstruct that context for every task, leading to incorrect joins, filters, and SQL. EvoOntology stores the missing context as versioned agent state and retrieves relevant fragments through Model Context Protocol, or MCP, tools.

Comparison of a data agent operating with and without the EvoOntology semantic layer
EvoOntology inserts a persistent semantic layer between an agent and its data sources.

Why agents keep relearning semantics

Hand-authored semantic systems such as dbt metrics, LookML models, and analyst-curated knowledge graphs require ongoing maintenance as schemas and business definitions change. Loading a complete glossary into every prompt also consumes context that the agent could use for data, code, and intermediate results.

EvoOntology exposes semantics on demand through MCP, allowing an agent to browse or resolve concepts during a task. Its evolution process analyzes recorded tool calls, outputs, and outcomes, then creates localized patches that must pass a controlled evaluation before publication.