// the find
antirez/neural-redis
Neural networks module for Redis
Neural Redis is a Redis loadable module from antirez (Redis's creator) that trains small feed-forward neural networks directly inside Redis, using an RPROP learning algorithm. The pitch is that you collect data, train, and run inference all through Redis commands — no separate ML stack. It targets simple regression and classification problems like ad targeting or user preference prediction, not vision or NLP.
The API is genuinely simple: three commands (NR.CREATE, NR.OBSERVE, NR.TRAIN) get you a working model without touching Python or a separate service. Training runs in a background thread and the live network stays available while a copy trains, so you don't block reads. Auto-normalization means you don't have to scale your inputs to [-1, 1] yourself. The BACKTRACK option for overfitting detection is a thoughtful addition that saves the best checkpoint rather than just stopping.
Last commit was 2018 — this is effectively dead software, and Redis's module API has changed enough since then that it likely won't compile against a current Redis without work. AOF rewrite is explicitly unimplemented, so you can lose all your trained networks and datasets on restart if you're not using RDB snapshots. Sigmoid-only activation and RMS loss only; no dropout, no batch norm, no GPU path — fine for the stated toy problems but you'll hit a ceiling fast. The alpha warning isn't just boilerplate: a 1000-line C module written in two days and abandoned for six years is a real risk in any production context.