AFAIK, with data.table it's all in-memory; whereas dplyr has the option of working with a database backend.

On the other hand, data.table has robust support for modify and update ops by reference, which can be a big performance saver.