MOOSEDev: Ontology-Grounded Memory System for Coding Agents
MOOSEDev, a novel system, seeks to provide coding agents with a structured, ontology-based project memory. Detailed in a paper on arXiv (2608.13662), MOOSEDev organizes architectural decisions, lessons learned, constraints, and rationales within a knowledge graph, accessible to agents through a Model Context Protocol (MCP) interface. Each record includes lifecycle status, provenance, and supersession links, which can be queried using MOOSE, a proprietary neurosymbolic engine prioritizing the symbolic layer for reasoning. In tests against a production vector-memory tool on a neutral public dataset of 835 typed records, MOOSEDev achieved near-complete answer sets (0.98-1.00) for supersession, set-completeness, and negation queries, outperforming the baseline's top-k retrieval. The paper underscores the difficulty of tracking the reasons behind swift code alterations in software projects, where coding agents are now the main source of new code. MOOSEDev enhances query precision for complex inquiries by maintaining the rationale for changes, marking a significant advancement in AI-assisted software development.
Key facts
- MOOSEDev is a system for ontology-grounded project memory for coding agents.
- It captures architectural decisions, lessons, constraints, and rationales in a knowledge graph.
- Exposed to agents via a Model Context Protocol (MCP) interface.
- Records carry lifecycle status, provenance, and supersession links.
- Queries are handled by MOOSE, a proprietary neurosymbolic engine.
- Compared against a production vector-memory tool on a corpus of 835 typed records.
- MOOSEDev achieved 0.98-1.00 accuracy on supersession, set-completeness, and negation questions.
- The baseline's top-k retrieval performed lower.
- Paper available on arXiv (2608.13662).
Entities
Institutions
- arXiv