Like SQLite, it's not necessarily in-memory, it has an on-disk format as well. I haven't given it a massive workout for bigger-than-memory workloads but in theory they are possible.

With that said, the resident memory size of stuff-in-parquet/arrow/duckdb is a lot less in practice than stuff-in-pandas (i know) and stuff-in-sqlite (i believe), so it still enables more in-memory workloads than you might otherwise be able to do.