Engram
Also known as Engram conditional memory
In plain English
Engram retrieves learned vectors for recurring token patterns instead of rebuilding every pattern through the model layers.
Technical definition
Engram is a learned conditional-memory mechanism that extends single-token embeddings with multi-token lookups and integrates the retrieved features into the model.
Engineering details
Lookup addresses depend on token IDs rather than intermediate hidden states. The runtime can therefore identify needed rows before reaching an Engram layer and overlap their retrieval with earlier computation. The retrieved features interact with downstream layers and expert selection.
Why it matters
Sparse access makes a large parameter table a candidate for DRAM offload. Removing the table at inference changes the trained model, so an ablation is not a substitute for comparing separately trained architectures at matched quality.
How to read it in InferenceX
The article reports roughly 189 GiB of Engram memory in its DeepSeek-V4.1-Flash configuration. Its DRAM and SSD comparisons retain Engram functionality while changing the serving path; the ablation experiments ask a separate model-quality question.