AI inference glossary
Model architecture

Conditional memory

In plain English

Conditional memory supplies learned information only when the input selects the corresponding memory entries.

Technical definition

Conditional memory is a model mechanism that retrieves a selected subset of learned memory parameters for an input rather than accessing the complete memory table on every token.

Engineering details

In the Engram design discussed in the article, token patterns determine row addresses and a gate controls how the retrieved features contribute. The memory parameters are learned model data, unlike the request-specific attention state stored in a KV cache.

Why it matters

The total size of a memory table can be much larger than the bytes fetched for one token. Capacity requirements and per-token traffic must therefore be evaluated separately when deciding where the table should reside.

How to read it in InferenceX

The article reports a large Engram table but only about 12.4 KiB of row data per processed token position across the model. This sparsity motivates offloading; it does not guarantee that a particular CPU or SSD implementation will improve serving economics.