A closer look at the memory matrix during training of the London Underground problem