edrtech/working-memory-depth-recurrence

This repository contains a small Python demonstration of an alternative approach to building AI systems. The author, who describes themselves as not being a data scientist or ML engineer, built it out of frustration with how large language models work. The core engine is about 60 lines of Python and requires no third-party libraries to run. The underlying idea is called working memory depth recurrence. Instead of using backpropagation, gradients, or any global learning signal, the system learns using only local rules operating on a single shared graph. The author argues this setup more closely mirrors how biological brains actually process information. The repository includes three small demos. The first teaches the system to understand numbers by counting piles of objects (characters, words, or anything countable), after which it can perform addition even though no addition examples were ever shown. The second learns what each type of card shuffle does to a 52-card deck and can then predict the final order of any deck after any number of shuffles, training in under two seconds from scratch. The third first trains the system on bad examples so it only memorizes, then trains it properly on top of the same memory, resulting in real generalization without any forgetting. The author is transparent about possibly being over-confident in the significance of this work, and explicitly invites people with genuine ML expertise to check whether it holds up. The code is intentionally short and readable so anyone can verify each demo's results by hand in a few minutes. This is a research-stage prototype and a personal project, not a production library. There are no external dependencies. You clone the repository and run it directly.