In-memory is not exactly to do with persistence, you can still have non-volatile data be in an "in-memory" DB and take snapshots of it, but even so, what about when you decide you need to scale your compute but not your RAM and move the data structure to a networked box, or when you need to have more than one local process read the same data structure, or when you want to ensure your architecture will always support such a thing if you need to, or if your language doesn't have a decent implementation of a particular data structure, or etc.
Your examples become a problem if your graph can't fit in the RAM of one host. If not that, then you have then distinct graphs to handle, and so a paralellisable problem to be handled by multiple processes. All you have in that case is an abstraction which lets you have a unified interface to all of them. Notably that also isn't relevant as the one in the article doesn't do any of that. It is a graph library.
Seems like you might want that 100GB of user relationship graph to live somewhere else, so you don't (1) massively inflate your RAM requirements on the app, (2) have to fan out updates to hundreds of boxes instead of a couple, (3) deal with cold start problems during the relatively frequent app restarts, (4) etc.
I think most of the repliers are posting response to the blanket statement you made on _all_ classes of in-memory databases (think memcached, redis, voltdb, HANA etc)
If I wanted to do something in memory, I'd just use data structures in my process. Not so much a database if there is no persistence.