Cache hotness
Also known as hot embedding rows, row access frequency
In plain English
A hot row is accessed often or recently enough that keeping it in a faster tier may save repeated retrievals.
Technical definition
Cache hotness describes the observed access frequency or recency of data within a workload and time window, informing which data may benefit from faster storage.
Engineering details
Recommendation systems commonly retain frequently or recently used embedding rows in faster memory while colder rows remain on SSD. Engram gate scores answer a different question: how strongly retrieved features contribute in a given context.
Why it matters
A strongly weighted row is not necessarily frequently requested. Cache policy needs access traces and resource constraints rather than a ranking of interesting examples or large gate activations alone.
How to read it in InferenceX
The article explicitly warns that strong gates do not identify cache-hot rows. Its example scan also does not measure table capacity by content category, so it cannot justify a cache allocation or prove that certain learned content wastes memory.