Right, the solution I remembered was a bit more complex, it required counting bloom filters and a subtraction operation to get a filter for the set difference. But the paper mentioned in a sibling comment claims to have even less overhead.