I've had both MongoDB and Cassandra perform nearly as well as Memcache when getting a single document/row when the document was in memory in MongoDB and the row was cached (row cache, not just key cache, so again: fully in memory) in Cassandra.

In memory operations are fast in many databases. Redis's default configuration (vm-enabled no) just only does operations in memory (with an occasional sync to disk). That's terrible durability but fantastic performance. Most databases, including Redis, can be configured for either that sort of high-performance/low-durability or the opposite. It's just that their default settings/behaviors vary widely.

Sorted sets. I mean, c'mon. That's worth the price of admission right there.