Hard-coded thresholds kept misrouting memory lookups,
so we trained a router to learn the pattern
Fixed thresholds kept choosing the wrong memory mode for an AI assistant's queries, and tuning them by hand did not hold up as traffic changed.
Task
SENET’s memory engine can answer from a short-term cache or a deeper sparse memory store, and picking the wrong one either wastes time or misses context. A hand-tuned threshold made that choice and was wrong often enough to notice.
What we did
We replaced the threshold with a real Kolmogorov-Arnold network, written in plain numpy with analytical backpropagation, that learns which memory mode a query needs. It retrains on the assistant’s own telemetry as usage patterns shift. If its accuracy ever degrades, it falls back to the old thresholds automatically, so a bad retrain can never take the system down.
Result
Measured against 108 real routing decisions, the trained router reached 0.91 accuracy, with a safety net under it the whole time. The router ships inside the product, not as a separate service a buyer has to maintain.
Full story: SENET: a sellable AI assistant with its own memory engine and a self-training router. Want the same? Get a plan.