RUMA: Routed Updatable Memory Architecture for Versioned Document Updates

Keeping language systems current is expensive, and standard retrieval layers often do not distinguish cleanly between superseded guidance and genuinely conflicting active sources.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 On a harder answer-quality pass over the frozen external corpus, superseded updates reach compositional exact match 1.0, while the base-only system remains at 0.0, naive append reaches 0.4, and one conflict-heavy synthesis case still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.