I'm not sure if my question was at all clear or not, so let me just dive in and give you the long version:
Someone said recently, when discussing high-volume web applications, that "disk is the new tape". Website administrators use huge clusters of memcache servers to make disk I/O-free round trips between the client and the server. In order to accomplish this, application developers are having to treat RDBMS's like generic data stores, tossing out valuable features like foreign key constraints, check constraints, and cascading UPDATES and DELETES.
But what if you could put your memcache cluster on the other side of your db interface, outside the realm of the application software (PHP, Ruby, Python...), and inside the RDBMS? Think of it as a large, distributed, in-memory database cache. To make it clear, I'm talking about having the type of memcache cluster that can store pretty much an entire database in memory, guaranteeing 100% cache hit rate when reading from the database, regardless of the parameters of your SELECT statement. Then, when writing application code, you can not only forget about memory caching, you can start dealing with the data store as if it were a Relational Database Management System again, and use normalization and JOINS the way RDBMS's traditionally encourage.
The performance gains, while maintaining/restoring the data integrity benefits of RDBMS's, might be worth looking into.
Does anyone know of a project where someone has done this already, perhaps as an open source project wherin they might modify an open source RDBMS such as PostGRE or MySQL? I haven't found anything ye, and I have no idea how these programs are structured or whether it would be possible to even implement such a storage engine.