I have used DBISAM and kbMemTable on different occasions.
What I like about DBISAM is that it has great features, and is usually very reliable. I have used it in large databases, full-text search, read-only mode, CGIs and many other situations.
It is fairly large compared to kbMemTable or SQLite based components, though. And you can't have a single file per database (or even table) - depending on the situation, that is a major disadvantage.
kbMemTable is tiny and it's great for small amounts of data. Since it runs in memory, it has to be a small amount of data, of course.
One other option I've taken on a couple of my desktop apps is dumping the data directly from/to my object hierarchy using TWriter/TReader. This is by far that smallest option, and is absurdly fast compared to using a database. The data files are tiny, too.
It has all kinds of drawbacks, though - you have to code versioning in if you might want to ever add/change fields, unless it's in-memory it is even more complicated, no multi-user support at all, etc.