@Steven Lord said it exactly correct. The problem is not that a tool like ismember rounds anything, nor that ismembertol would help you. The problem is that as stored in MATLAB as a double, the numbers are already quantized to live in only 52 bits for a mantissa. And that means IF you need to compare numbers where the differences are relatively smaller than a double will allow, then you need not to store or compute the numbers as a double at all. And that means working entirely in a higher precision. Tools like syms, or my own HPF (available for download from the file exchange) would allow you to do those things more precisely. However, they will also be VASTLY slower, so your next question will be in how to make your code run more quickly.
And that begs the question of whether you really need to be doing this computation to such a high degree of accuracy anyway. Do you KNOW those numbers EXACTLY to that many digits? Was the data that went into them known to 20+ decimal digits? I would argue that few things you can possibly measure are known that accurately.
Hmm. Let me see. An atomic clock can meaure time to something like 10^-18 seconds (maybe 10-20, as I saw in another link). And the fine structure constant, while known pretty accurately, has way worse precision, only something like 85 parts per trillion.
