This problem applies to routing protocols.
On the Internet, packets are sent from any node to any other node, the routing relies heavily on the destination IP number structure (low significant bits are ignored, packets with the same high significant bits are routed to the same exit), this is very convenient because a routing table doesn't have to have information about each possible destination, but rather about different groups. The "view" of these groups changes from node to node. This has the inconvenience that the IP numbers (being limited) have to be carefully "supplied" by hierarchical organisations.
In a network where the identifier of a node has absolutely no relation with the identifier of its neighbours every address has to be remembered uniquely. But what about memory constrains?. The assumption in CryptoNet is that only a very small fraction of the total number of nodes can be remembered at a time.
The solution in CryptoNet is to remember information about an address only if this will help make profitable
routing decisions and therefore atract more packets. Memory is just another scarce resource.
Note that a vital assumption in CryptoNet is there will always be plenty of processing power to crunch smart decisions, and that the bottleneck will always be in the bandwith.
This applet ilustrates a packet being sent from wherever you click to the center. At the beguining the network has no information about anything; when a node receives a packet and it doesn't know anything about the destination it has to make a decision; when it knows anything at all about the destination it can make better decisions.
more ramblings later ...
Done by Diego Moriarty