Embed: each thought becomes a 384-number vector (). Similar meanings point in similar directions.
Topics: UMAP squashes the vectors to 5 dimensions, then HDBSCAN finds dense groups on its own and leaves loners as "unsorted". Names are each group's most distinctive words (c-TF-IDF).
Layout: a second UMAP squashes to 2D. It keeps local neighbourhoods but bends big distances, so the inspector's "nearest" list uses the full vectors, not the picture.
Try this: switch to with context and watch thoughts from one voice note snap together, even when they're about different things. That's why the map defaults to title only.