Calculators that allow direct operations on data stored on persistent storage?

[Claudio L.]

David made a good point: if you need compact storage, you need to find the optimum format. However, if you need fast access, you have to work with the data in whatever format the CPU or the program uses, you can't afford to convert back and forth all the time.
There's always a compromise in between, and you'd be surprised how much data you can fit. I vaguely remember the data collector from Precision Solutions, LLC, it could store 7000 points (as in surveying coordinate points) in unused RAM areas of the 50g, you were able use the calculator normally (well, as a data collector) and all those points were "hidden" from view. It's all a matter of finding the proper format.

I think you want some kind of virtual memory management. With VMM you CAN have the whole thing in RAM, or at least you think you do, while the operating system swaps data to disk in/out as you use your data.
The storage format is still very relevant, though. For example, in a list of variable size objects (which is your preferred method of storage), you'd have to skip N objects to access item N, which implies reading the list from start to N for every access (swapping it in and out of memory). On the other hand, if you had fixed records you could compute the offset to object N and only read that one record from storage.
Helper routines would be the wiser way to go, I think, otherwise you'll pay the price in performance if you manage large lists.
You should probably start thinking of your "paging" solution before you run out of space.
One possible solution is for example to store 100 records per list, and store the lists in variables/files 'Lmmm', Where mmm is the higher digits of the record index: Index mmmNN is stored in list variable 'Lmmm', list index NN. It is simple that only takes you a few minutes to write the helpers, relatively fast, and you can move 'Lmmm' files to SD card when they become "cold" and are no longer modified.