I'm not a fan of MongoDB, but you don't know what you're talking about.
MongoDB is durable. While it doesn't quite support SQL transactions, it is durable. The data is journaled before being confirmed, and once confirmed will be written to disk. It can be consistent, but this sort of breaks the whole idea of scaling and distributing the load in Mongo, or causes massive performance problems, so that is something to consider when using it. Understanding and adopting eventual consistency is tough, but it's an issue with every distributed database, not just MongoDB.
And you don't "keep your entire data set in memory". You should make sure your indexes fit in memory. Your data can be as large as you like. Most people with terabytes of data in MongoDB don't keep terabytes of RAM in their servers.
Foursquare did NOT keep every single check-in in RAM. They kept an index of them in RAM, sure. But the problem was they had a sharded MongoDB deployment and one of the shards became unbalanced and exceeded the available RAM. If that happens (harder to do these days, but not impossible), it can be very difficult to recover.
MongoDB tries to bridge that gap between NoSQL and SQL. I think the MongoDB folks originally ignored decades of database research when developing MongoDB, but they've been forced to adopt it as the years have passed. Is it an Oracle killer? No. But it can be a useful and productive tool if you understand and apply it appropriately.
It comes down to this: there is no way to tell MongoDB to commit a write and not return until it is durable. You can tell it to fsync() it within at most 33 ms or so from when it returns, but not immediately. There is no control in it to say "fsync() this value because it is important, then return".
That is not as durable as other database engines, which do this based on number of commits that have not been fsynced to the journal (this value can be set to 1 causing every commit to be fsynced).
Consistency is a large topic, and I am sorry I mentioned it vs something more specific, such as:
MongoDB doesn't have transactions, so you can't make multiple updates truly atomic. That's annoying at least, and unusable at most. Two phase commits let you get about 50% there, but with no rollback your application code has to implement rollbacks. In lots of scenarios your application is not smart enough to rollback a "transaction", so your data will be screwed.
Eventual consistency is starting to come under fire precisely because it avoids the hard problem of database design and is offloading it on the application developer, who is typically not as versed in how to create a system where temporary inconsistencies are OK. At best, it means that the developer is now wasting productivity on addressing consistency issues.
You may be right about Foursquare keeping only an index of all the check-ins, not the entire dataset. The main point remains: why keep essentially archived data in RAM? MongoDB does not let you keep a partial index.
MongoDB tries to give up basic datastore guarantees to gain speed and flexibility. Once again, I am not arguing that it's a useless tool. It is useful. As a cache. Or in cases where data loss is acceptable. It is much less useful in cases where dataloss is catastrophic, such as financial transactions. It will also be much more expensive at scale, whereas other datastores optimize for the conditions where RAM is expensive and disk is cheap. You can still run into this case when you are talking about a terabytes to petabytes of data.
MongoDB is durable. While it doesn't quite support SQL transactions, it is durable. The data is journaled before being confirmed, and once confirmed will be written to disk. It can be consistent, but this sort of breaks the whole idea of scaling and distributing the load in Mongo, or causes massive performance problems, so that is something to consider when using it. Understanding and adopting eventual consistency is tough, but it's an issue with every distributed database, not just MongoDB.
And you don't "keep your entire data set in memory". You should make sure your indexes fit in memory. Your data can be as large as you like. Most people with terabytes of data in MongoDB don't keep terabytes of RAM in their servers.
Foursquare did NOT keep every single check-in in RAM. They kept an index of them in RAM, sure. But the problem was they had a sharded MongoDB deployment and one of the shards became unbalanced and exceeded the available RAM. If that happens (harder to do these days, but not impossible), it can be very difficult to recover.
MongoDB tries to bridge that gap between NoSQL and SQL. I think the MongoDB folks originally ignored decades of database research when developing MongoDB, but they've been forced to adopt it as the years have passed. Is it an Oracle killer? No. But it can be a useful and productive tool if you understand and apply it appropriately.