You don't parse self referential nested trees with simple REs. You parse them with a parser (typically recursive descent) and then work over the nodes.

Regardless of that the code I supplied does the same as the regexes you were trying to use in your original post except a lot more efficiently. You seem to have missed what it does. Also you seem happy to blame a memory leak in the RE engine. Besides the fact that your REs are really rather badly written I pointed out you have an extra set of braces:

while(<>) {
   {   # surplus to requirements
[download]

this may well be creating a closure.