The latest commit in wrste/pheromone-network, 84b7049 (October 4, 2026 UTC), touches only documentation, not code. It shortens the stage names and box labels in the model flow diagrams (mermaid) of the Korean, English, and Chinese READMEs, and also rewrites the explanatory paragraph below them. The change adds 30 lines and deletes 30 lines, and all 60 changed lines occur inside the three README files.
This repository is a zero-dependency TypeScript memory kernel whose purpose is to give agents long-term memory. Without backpropagation or any training process, each output unit reads only a few nearby inputs whose tags match; connections that are used frequently are strengthened by a manual Hebbian rule (a rule under which connections used together frequently become stronger); and connections that go unused for a while decay according to a half-life. The structure is based on the paper Local Pheromone Network (arXiv:2606.30669).
The cleaned-up diagram presents this procedure in two blocks. At construction time, candidates are filtered by unit tag and position, and each output is connected to at most maxNeighbors nearby inputs; in the forward computation, the effective weight is the product of the connection mask, the learned weight, and the pheromone gate. Learning changes the weights and the traces but does not touch the mask. The default RecallKernel folds the input into codeDim and uses only a single codeDim -> codeDim local layer with no hidden layer, whereas the lower-level LocalPheromoneNetwork can add hidden layers. The default forward gate uses only the long-term trace (shortPheromoneWeight = 0, longPheromoneWeight = 1).
That said, this commit only compresses the explanation; it does not change the computational behavior. Neither benchmark figures for the paper the README cites nor memory recall accuracy are presented in the given record, and the default configuration remains limited to the long-term trace alone with no hidden layer. It can be confirmed that the implementation is lightweight and free of external dependencies, but how accurately this kernel actually remembers requires separate evidence.