What it does Neural Redis is a loadable C module that turns Redis into a machine-learning engine. You create networks with NR.CREATE, feed them observations via NR.OBSERVE, and train them with NR.TRAIN — all inside the server. It handles regression and classification tasks, auto-normalizes your data, and keeps training in a background thread so queries don’t block.

The interesting bit The clever part is the zero-infrastructure pitch: no Python environments, no model serialization headaches, no separate inference service. Antirez wrote the whole thing — module, README, and all — in about two days, and it shows in both the brisk API design and the frank “alpha code” warning. Training happens on a copy of the network; weights merge back later, so you can query a model while it’s still learning.

Key highlights

• Native Redis commands: NR.CREATE, NR.OBSERVE, NR.TRAIN, NR.RUN, NR.CLASS
• Automatic data normalization and simple overfitting detection via AUTOSTOP
• Online training with RPROP; runs on a copy to avoid blocking reads
• Classification mode transparently handles one-hot encoding for you
• Only ~1000 lines of C, built against Redis unstable

Caveats

• Explicitly alpha code; author warns it may crash your Redis server
• AOF rewrite is unimplemented; only RDB persistence works
• Sigmoid activation and RMS loss only — no ReLU, no convolutions, no recurrence
• Auto-normalization can fail silently if your training data doesn’t span the full range you’ll later query

Verdict Worth a look if you want dead-simple ML for ranking, recommendations, or trend spotting without spinning up a separate pipeline. Skip it if you need computer vision, sequence models, or production-grade reliability.