window 
|------[++++++|++++++]------|------|---|    file
   A      B      C      D      E      F     blocks
             <----->
              match
[download]

If you don't destroy/recreate the variables but just change the content, your memory consumption will¹ be minimal.

seek and read help reading chunks of data from files.

substr manipulates the content of strings.

pos returns the position of your last regex match.

So only one global variable $window of fixed size holding two current blocks could do and whenever the pos of a match leaves the first block you have to shift a new block into $window.

¹) well, as long as Perl doesn't do very (unlikely) weird speed optimizations.

UPDATE:

This code is an almost perfect example of what I meant: Matching in huge files

The differences are the temporary variable $block which could be optimized away and the handling of pos. Instead of adjusting the window at "halftime", pos is adjusted to the window. Actually I think this is even smarter than what I planed...