There is obviously some overhead. But it is only so drastic for small files, it looks much better for larger files.
I tried in the past for some tests but stopped as it was much slower than tmpfs. It was even slower than SSD drive:) I guess all user space fuse translations + VFS layer + object based storage model (you can't modify one byte without re-uploading all file) makes it not what one would expect from RAM performance.
Not sure what do you need this for but if you need RAM disk - use tmpfs - which is normal block storage.
The memory backend is used mostly for testing. It should be fully functional and can be useful but if you are going to mount it then tmpfs will be better as it is implemented in kernel and is much more careful with memory management.
The memory backend is bucket based so think of it like an in memory S3 and you'll have an idea of its capabilities.