Neuro-Symbolic-SNN demos continual learning on MNIST

Most of the apparent stability here comes from the training policy around the SNN, not from any magical new neuron design. That makes it interesting as a prototype, but the evidence is nowhere near strong enough to call it a general continual-learning breakthrough.

• –The core network is a standard LIF SNN with surrogate gradients and LayerNorm, so the novelty is mostly in orchestration.
• –Replay weighting and the decaying plasticity schedule are sensible anti-forgetting tricks, and they likely explain much of the observed stability.
• –The LLM is acting as a curriculum planner and poisoning gate through a local Ollama model, which is clever but prompt-sensitive and easy to overtrust.
• –The repo is still very early, basically a single-file demo with no releases, so the MNIST result should be read as anecdotal rather than benchmark-grade.