Cue-memory T-maze: a clean memory benchmark solvable in few generations.

Each trial: a CUE (left/right, drawn at random) is shown at step 0, then a corridor of delay steps shows nothing, then at the junction the agent must decide which way to turn. Reward iff the decision matches the cue. The cue is only visible at step 0, so the network must CARRY it across the corridor to the junction — that is the memory the task demands.

Sensor (vl 2): [Cue, JunctionFlag]. step 0 -> [cue, 0] cue shown (+1 left / -1 right) corridor -> [0, 0] nothing junction -> [0, 1] decide now Actuator (vl 1): Output of zero or more turns left, below zero turns right.

Why it is the right instrument: with random cues, a MEMORYLESS network sees [0,1] at the junction with no cue information, so its best fixed policy is to always turn one way — capped at ~50% (10/20 trials). A network that can hold the cue (recurrent wiring, or an LTC neuron's internal_state) can reach 100%. The fitness (count of correct decisions) cleanly separates the two. Unlike DXNN2's discrete T-maze, this tests static memory, not lifetime plasticity, so it needs no neuromodulation.

Actuator/sensor parameters [Delay | Trials] override the corridor length and trial count (defaults 2 and 20).