Most methods that get strong multi-hop answers (GraphRAG, HippoRAG, RAPTOR, trained retrievers) build a knowledge graph or fine-tune a retriever over the corpus. That's fine until the data changes — then you re-extract / rebuild / retrain before the new facts are usable. For a corpus that updates daily, that's a real cost. MOTHRAG does the multi-hop reasoning at query time over a plain dense index instead. An update is just embed + append (one embedding call) — no graph reconstruction, no re...