retention, goals still reachable per world, after learning all three in order
gated (EFA)naive (no context mechanism)
forgettingworld-A skill lost by the naive net: —
the same gated net plays any world · pick one, reach the goal
Charlot Lab · Energy First Architecture · phase 3 spike (internal)
Phases 1–2 gave EFA a sparse-positive latent that holds a memory and is a plannable world model. Phase 3 adds the last mechanism: a dendritic context gate. EFA learned three worlds with different physics one after another, never revisiting the earlier ones. The world context arrives the way a dendrite gates a neuron: it selects a sparse sub-network, not as an input the network reads. The comparison is the textbook one: the same network, trained the same way, with no context mechanism. That naive net overwrites what it knew, learn world C and worlds A and B are gone. EFA's gate routes each world to disjoint synapses, so new learning never touches the old: continual learning on-device, no retraining trip to the cloud. (Trained offline; the page runs the trained network live.)