Overwriting a minimal number of pointers ensures that the time behavior of the algorithm remains tractable, but adds to the complexity of programming.
Second, to avoid overwriting, previous training sets must be interleaved with the current set.
Overwriting the thunk for 1 with a black hole immediately it is entered solves this space leak, because a black hole retains no pointers.
The term (1/m) could represent a "theoretical maximum memory span," given a memory system subject to overwriting.
It is also possible to obtain both these advantages in a slightly more subtle way, without performing the black-hole overwriting operation.
I can identify basically three: overwriting, perturbation, and contextual distinctiveness.
Because overwriting is a stateful operation, everything must be in some appropriate monad.
Overwriting models of this sort seem possible for the description of memory for sequential stimuli but not simultaneous arrays.
