Hi,
My experience has taught me that any design that depends on  a fully memory based solution is asking for a lot of trouble. From what you have described, it looks like the full nodes are setup to make this assumption. Why not use Redis instead? My experience with Redis has been very good and yet is is better when backed up by SSD but not mandated. this way a linear dropoff is avoided. One of the disadvantages of a dispersed cluster solution is that individual node performance becomes paramount. if enough of the nodes are slow then you have other security problems to worry about. using Redis seems like a good compromise and it will keep your code simple.
regards
Adarsh