Developer Documents 8 Months and ~200 Failed Experiments Seeking a Non-Neural AI Memory System

A developer has published the second installment of a research series detailing an eight-month effort to build a non-neural, CPU-only system capable of accumulating and applying experience without retraining a large language model. The project, documented through roughly 200 failed experiments, aimed to find a knowledge substrate that could change future behavior after experience, survive restarts, and generalize to unseen but related cases. The core challenge identified was not data storage — which proved straightforward — but what the researcher calls 'transferable causal transition': preserving the logic of when a learned condition-action-consequence rule should and should not be applied. Numerous candidate knowledge carriers were tested, including memory graphs, typed edges, graded vectors, and topology fields, but each preserved only one aspect of consequence while failing to generalize correctly. Surviving code from the project has been published in an open repository called AuraSDK, and the series continues to document what mechanisms held up and where the precise limits were found.