Yes, that's a problem, if the serial numbers need to be maintained. If this is a one-time-per-dataset operation, and the serial numbers are there just while manipulating the data, it doesn't really matter.

Anyway, his problem seems to lie on the memory use more than the numbering scheme.

But the reason he's keeping all the old records around is to make sure he doesn't reuse a number. If he uses a unique identifier (the reference value is unique, automatically generated, and readily available), he doesn't have to keep all of the records around in memory.

The thing that bothered me was using grep to look for already-used phone numbers. What if they were the primary key of the hash? Then, it's a simple lookup to see if one's already used.