Besides that a CPU can only see data in memory, there are useful things like buffering layers etc. Of course, I am aware of that - sorry for not being precise enough.

What I meant was if I could have regular expressions act directly on a file without having to explicitly load parts of that file to a variable in memory, or more precisely, without having to load parts that are dependent on the expected size of the match or such things. Such dependencies would mean that I generally would have to load the complete file to memory because I don't know anything about the size of possible matches in advance.

Regarding the streaming XML parsers: Several years ago, I have tried some of them for another project. They all have been a fine example for the very same problem:

They were streaming only in the sense that they read the source file line-by-line (but what if the source file does not contain any line breaks which is perfectly acceptable according to the XML standard?) or that they broke the file in chunks at syntactical markers, e.g. tags (but what if there were 100 GB text before the next marker?). I didn't see any streaming XML parser which didn't rely on such mechanisms. I admit that I have tested only a few of these parsers, so I might have missed the ultimate one.