Logseq does this too - or used to. They found that for the rich operations they want to support (live collaboration, richer data), persisting the data in a graph structure makes more sense to them. It probably adds some conveniences, but the file model is just unbeatable for simplicity.