I had the same issue today.
We are running two AMQ brokers in network of brokers, each one talking to the other one (the duplex is set to false). Under a high load after few hours I've noted that one of our queues started growing on both nodes, several thousand of pending messages in about 20 minutes (and the cursor memory usage was quite large). After I accidentally stopped the second node, the network bridge stopped on the first node and the queue size on the first node immediately became "-8".
So in my case this probably somehow related to the instability of network of brokers implementation, because if I disable the network connector the system runs stable for several days under the load.