AI inference glossary
Model architecture

Context-dependent gating

Also known as Engram gate score

In plain English

A gate controls how much a retrieved memory feature contributes in the current model context.

Technical definition

Context-dependent gating weights retrieved memory features according to their compatibility with the current representation, rather than always applying a fixed contribution.

Engineering details

The article probes Engram gate scores to find names, code fragments, and recurring phrases. It also notes that calculating the gate requires the retrieved key, so observing a low gate after retrieval does not automatically avoid the memory read.

Why it matters

Skipping retrieval would require a separate usefulness predictor that acts before the read. Such a predictor introduces a different implementation and accuracy question; it is not a free consequence of existing gate values.

How to read it in InferenceX

The published examples were selected for interest rather than gate strength. Neither those examples nor strong gates establish cache hotness. Gate-based interpretation, storage placement, and serving performance should therefore remain separate measurements when evaluating Engram.