model of a CPU/program accessing memory, except that the CPU/program is learned via a neural net
memory operations are all made to be differentiable via: reads/writes are not to a single memory location but rather linearly distributed via an 'attention' vector
memory ops:
3 kinds of reads:
vanilla
content-associative
follow episodic association chain
write
erase (relative to a 'how-much-this-location-is-in-use' per-location activation level, which is: ?set during write using a formula learned by the network?)