XMPP breaks a lot of things down into extensions, even for things like images, file uploads and voice. I think the idea was to break it all up and let people implement the parts they want, but it also makes the entire protocol kinda broken. History doesn't really get stored and can be different on every client if they're connected or not. Matrix tries to incorporate everything and it's starting to bloat like a web engine; getting more difficult to implement an entire server with all the features.

Their Synapse server kinda sucks at scale, but Dendrite does look promising, and ejabberd has also started adding Matrix support into their XMPP server.