|Welcome to the Monastery|
I'd draw your attention to the first word of both of the sentences you quoted; and also to both the id est; and contraction that follows it.
If the OPs circumstances do not comply with either of those two criteria; then *I* wouldn't use this approach.
But, his records might only be 80 characters in size (ie.<1GB of data); and if I were purchasing my next machine right now, I wouldn't consider anything with less than 8GB, preferably 16; and I'd also be looking at putting in a SSD configured to hold my swap partition effectively giving me 64GB (or 128GB or 256GB) of extended memory that is a couple of orders of magnitude faster than disk.
So then you are trading 2x O(N Log N) processes + merge at disk speeds; against a single O(N2) process at ram speed. Without the OP clarifying the actual volumes of data involved; there is no way to make a valid assessment of the trade-offs.
Also, if they are free-format text records -- ie. the key is not in a fixed position; or there might be multiple or no keys per record -- sorting them may not even be an option.
Equally, the OP mentioned 'patterns'; if they are patterns in the regex sense of the word, that would exclude using a hash. And, if you had to search the records to locate the embedded keys in order to build a hash, you've done 90% of the work of the in-memory method, before you've started to actually use the hash.
The bottom line is, I offered just one more alternative that might make sense -- or not -- given the OPs actual data; and it is up to them to decide which best fits.
With the rise and rise of 'Social' network sites: 'Computers are making people easier to use everyday'
Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.
"Science is about questioning the status quo. Questioning authority".
In the absence of evidence, opinion is indistinguishable from prejudice.