A new paper proposes swapping vector search for a knowledge graph so coding agents stop losing track of why the code looks the way it does.
The system, called MOOSEDev, stores architectural decisions, constraints, and rationales in a graph database rather than as embeddings. Each record carries a lifecycle status, a provenance trail, and links showing what superseded what, and agents query it through a Model Context Protocol interface backed by a symbolic reasoning engine called MOOSE. The researchers tested it against a production vector-memory tool on a shared corpus of 835 typed records. On questions that required tracking supersession, completeness of a set, or the absence of something, MOOSEDev returned essentially the full expected answer set (0.98 to 1.00), while the vector-based baseline surfaced only 6% to 27%. On plain relevance recall and token cost, the two systems were roughly tied.
That gap is the real story. Vector search is good at finding things that resemble a query, which is exactly the wrong tool for a question like has this decision been overridden or list everything we ruled out. As coding agents write a growing share of a project's code, the ability to answer those structural questions is what keeps a codebase's history legible instead of turning into a pile of orphaned commits.
Worth noting: the comparison is self-run, the corpus is modest, and MOOSE itself is a proprietary engine, so the numbers deserve outside replication before anyone rebuilds their agent memory stack around it.