I second the suggestion for Berkeley DB. I used it before it was bought by Oracle. It is not a full relational database, but just stores key-value pairs. We switched to that after writing our own paging B-Tree implementation. It was a good learning experience, but we kept adding features until is was just a (poorly) implemented version of BDB.
If you want to do it yourself, here is an outline of what we did. We used mmap to map pages into memory. The structure of each page was index based, so with the page start address you could access any element on the page. Then we mapped and unmapped pages as necessary. We were indexing multi GB text files, back when 1 GB of main memory was considered a lot.
