Keeping language systems current is expensive, and standard retrieval layers often do not distinguish cleanly between superseded guidance and genuinely conflicting active sources. This work presents RUMA, a routed updatable memory architecture that stores chunk-level records outside dense weights, applies append-first updates with lineage and provenance metadata, and answers through a citation-first grounded retrieval layer. The current implementation targets controlled changing-document settings rather than general-purpose generation. Across synthetic update tasks, held-out real-text chunk tasks, an expanded versioned-manual corpus, a frozen external corpus built from official-source Python ecosystem documents, and new routing and retrieval benchmarks, explicit same-lineage supersession consistently outperforms no-update and naive append baselines while preserving retained guidance. On a harder answer-quality pass over the frozen external corpus, superseded updates reach compositional exact match 1.0, while the base-only system remains at 0.0, naive append reaches 0.4, and one conflict-heavy synthesis case still fails. These results do not establish a universal replacement for dense language model architectures, but they do support a narrower claim: version-aware routed external memory is a promising path for low-cost document updates with inspectable evidence and controlled conflict handling on the current controlled corpora.
Vatsal Chavda (Tue,) studied this question.