Reproduce “Ontology-Grounded Project Memory for Coding Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
James Adam
Trivyn introduces MOOSEDev, a system that provides coding agents with structured, ontology-grounded project memory, addressing "comprehension debt" in agent-driven software development. It demonstrates that agents can achieve significantly enhanced structural reasoning capabilities, such as set completeness and supersession traversal, with high accuracy compared to vector-based memory and flat-file documentation.
Key insights
- Project memory for coding agents is fundamentally a "modeling problem" requiring formal representation via ontologies, rather than just a "note-keeping problem."
- The structural information within a knowledge graph is critical for enabling complex structural reasoning (e.g., set completeness, negation, supersession traversal) that is beyond the capabilities of vector-based or keyword search methods.
- Guaranteed currency of information can be achieved by construction through explicit modeling of record lifecycles, ensuring agents always receive the most up-to-date guidance by excluding superseded records.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.