Yes, RDBMDSes have built-in cache, but they are not as fast as memcached. So why the difference? I haven't really checked, but you have to populate and invalidate memcached explicitly, and it doesn't honor database isolation (if you invalidate cache after writing an object, some instances may read stale data after you already updated the db, etc). MongoDB's cache may cut corners in a similar way, I'd love to know.

They might be referring to Mongo writing data into memory rather than requiring that it be written to disk before responding success (in its default configuration) or that Mongo uses memory mapped files (and tries to uses as much memory as it can) to hold the active records in memory.