The berkeley interfaces that come out of the box with e.g. Python don't have any of the advanced features that the sqlite interface has, e.g. concurrent access. Any advanced usage of Berkeley DB is far more complex than sqlite, e.g. try opening a database so that multiple users can read/write it. With sqlite, no extra steps are necessary if you have an occasional separate process that needs to access the database (obviously it's not built for efficient concurrent access).

Ad-hoc SQL queries on the sqlite database are also a huge win. Much better than defining your own data structures and then tools to read/write them.

I don't know you mean by "proper" transactions. Sqlite has transactions for DML.