The cause of forgetting was sharing. Dendrites remove it. In Numenta's Active Dendrites model, each unit has a dendrite that reads a context signal — which task am I in — and only the units whose dendrites respond most stay active; the rest switch off. A different context lights up a different, sparse, mostly non-overlapping sub-network. So task A trains one set of units and task B trains another, and because their sub-networks barely overlap, learning B leaves A almost untouched. On the hard version of this — a hundred tasks in a row — this holds around 81% accuracy where a standard network falls apart, and it was demonstrated on Meta-World robot-manipulation tasks, which is why it matters for embodied AI. Here you build the mechanism and watch task A survive. The dendrite's job here is not to compute the answer; it is to decide who computes it.