Flybrain: an untrained, connectome-faithful simulation of the whole Drosophila brain and nerve cord, and what its memory system suggests for continual learning in AI

I am building a simulation of the complete adult fruit-fly central nervous system, brain and ventral nerve cord, about 167,000 neurons wired as in the MaleCNS v1.0 connectome. The synaptic weights are never trained as they come from the measured wiring, and the only quantities we change are a few global constants, each taken from a published fly measurement. If wiring and measured physiology are enough, the fly's behaviours should emerge without training; where they do not, the gap names the missing measurement. Every test is pre-registered and failures are reported. First result: the mushroom body, the fly's learning centre, now codes odours as the fly does: in fresh, pre-registered runs 2–5 % of Kenyon cells respond to an odour, specifically, and are nearly silent at rest. Several apparent brain problems turned out to be artefacts of simplifications. Vision, mushroom-body output, learning against fly data and sleep consolidation are not yet validated; an earlier sleep result was withdrawn after a bug was found. The fly's memory system learns on fixed wiring: a sparse key layer, a dopamine-gated output layer, several timescales, offline consolidation. We map this onto continual learning for frozen AI models, list the lessons so far, report a qualitative pilot on a frozen 9B language model, and disclose a set of brain-inspired AI memory designs as dated prior art (defensive publication).