The idea is that two elements x and y despite having the same fingerprint, will be stored into two different buckets. So you can have several time the same fingerprint in a Cuckoo filter, which you can't in a Bloom Filter.
When you want to remove x, you remove one of the two fingerprints, it does not matter which since they're the same. Next when you query y you will be sure to still have a positive answer.
Note that however the deletion won't necessarily make the filter forget about x. It will just clear some place in the structure.