Why a single neuron is enough to settle persistence

Today I closed the three points that were still open in the refutable model of the living, persistent neuron. The guiding thread was the same throughout: choose the leanest form that might be enough, then try to break it.

Forgetting can only come from sharing

On a single unit, there is no network where a memory can hide. Forgetting can therefore only come from one thing: the same weights encoding both P1 and P2. That is why the overlap of afferences between the two patterns becomes the real variable of the experiment, not a detail. We sweep it from 0 to 1 and read a retention curve, rather than a single verdict.

The signature as anchor, not as copy

The neuron’s signature is only useful if it does not drift with the weights. I fix it as a slow state that steers the sleep replay toward events consistent with what the neuron is. The chosen form is minimal: a vector of slow afferent traces. It scores the journal episodes; it does not regenerate the pattern. That is what prevents winning by cheating.

The unbounded journal, assumed

Keeping everything costs. I assume it: the full journal is an upper bound, an oracle, and I measure its cost in order to read each retention against its price. The question is not “is it free” but “how much does a bounded buffer recover at equal compute”. The O(T) cost is the condition of the experiment, not a flaw to hide.

Still open

Choose the artifact for the brick: a playable concept page, or Rust code under TDD. The pedagogical dimension of replay, the Proustian madeleine, will come with it.