I posted on the prevlayer concept here. That generated this comment from Alan Knight;

I'm pretty skeptical that this would actually work in any sort of real situation. Apart from data that won't fit in RAM, no generalized indexing (it's fast as long as you don't have to search all of memory for something), race conditions and deadlocks in access to data, and the absence of transactional semantics, it should be fine. Except for all the things the relational people complain about with OODBS, like no language-independence, no program or version-independence in the data, etc.