Try ZODB/ZEO, you will be able to access native python data structures (possibly cached in memory) from multiple python processes.
This does not solve the GIL problem - it reduces it just a little.
ZEO and ZODB are nice, but using them also does not solve the problem - there is still one serialization for every update and one de-serialization for every independent client/dirty-read. I love Zope and Plone and use them for a lot of situations (I won't think twice when someone wants a CMS for their intranet), but I sure would like to be able to set "zserver-threads" to 64, buy a Sun Niagara-based server and just live happy with it watching how all thread units get their fair share of usage.
[edit] Traditional DBMS architectures like that of postgres ,to my knowledge, rely on shared memory or memory mapped files (I'm sure this varies). They swap on disk pages in and out of memory and those pages retain their on-disk structure. There is little if any translation into in-memory data structures. Yes I could do that, but it's incredibly complex. Every single data structure basically has to be implemented on top of a byte array. I wouldn't be able to use any existing libraries for trees, lists, sets, hashtables, etc.