This is a test for prefrontal cortex (PFC) active maintenance mechanisms, which are a key part of the Deep layer model in Axon, and are essential for the Rubicon goal-driven motivation models.
The task requires encoding and maintenance of a stimulus (10 different stimuli by default, each a different random pattern) over N time steps (10 by default), with an error score on the final stimulus representation. It is also scored on predicting a simple localist time step representation that increases over the maintenance delay.
The SMaint self-maintenance mechanism for the PTMaintLayer is used by default, which simulates a population of interconnected PFC layer 5 neurons that mutually sustain each other via NMDA-gated recurrent excitation. The Config has an option to instead use direct recurrent connections within the layer, which requires a much stronger level of NMDA current for small-sized layers. Thus, the SMaint mechanism is an optimization allowing for smaller networks to be used.