MOOSEDev Gives Coding Agents Ontology-Grounded Project Memory

The paper presents MOOSEDev, a novel system designed to equip coding agents with a structured, ontology-grounded project memory. As software projects evolve rapidly with AI-generated code, tracking the rationale behind changes becomes challenging. MOOSEDev addresses this by capturing critical project information—such as architectural decisions, learned lessons, constraints, and underlying rationales—within a knowledge graph. This knowledge graph is made accessible to coding agents via a Model Context Protocol (MCP) interface. Each record within the system includes lifecycle status, provenance, and links indicating supersession, allowing for sophisticated querying through MOOSE, a proprietary neurosymbolic engine. This engine prioritizes the symbolic layer for reasoning, enhancing the system's ability to understand and retrieve context. In a comparative study against a production vector-memory tool using a public corpus of 835 typed records, MOOSEDev demonstrated significantly higher recall for specific query types. It achieved near-perfect retrieval (0.98-1.00) for questions involving supersession, set-completeness, and negation, whereas the baseline tool only managed 6% to 27%. While relevance recall and token cost were comparable, MOOSEDev's structured memory proved superior for precise, context-aware information retrieval.