If too many objects have been forced to remain in the shared pool by issuing the dbms_shared_pool.keep procedure, they may be consuming so much memory that Oracle simply may not be able to successfully manage what remains. Also, if a large object is not being kept in the shared pool, the instance has been running for a while, and then the object is referenced, when it is forced to be loaded, the memory may not be available. The key is to not casually keep objects in the shared pool. You should intelligently reduce and refine your pin list, as described earlier in the "Resolving Shared Pool Latch Contention" section.

Surely, 4031 errors are not the only reason to upgrade, but starting with Oracle Database 10g Release 2, Oracle began standardizing memory into 4KB chunks.6 While I would never recommend upgrading to this version only because of this improvement, it may be part of the rationale to upgrade.

Just as with standardizing on segment extent sizes, having standard memory chunk sizes increases the likelihood of quickly finding suitable memory. The quicker memory can be found, the fewer CPU cycles will be consumed, the less time the shared pool latch must be held, and the less likely a lot of wasted small chunks of memory will exist (increasing the likelihood of a 4031 error).