What is DeepSeek Engram...?

DeepSeek's proposed Engram is a new architectural block added to the transformer alongside attention and feed-forward networks. It acts as a conditional memory or lookup table that allows the model to instantly recognize multi-token phrases representing well-known entities. Instead of gradually composing meaning across multiple transformer layers, Engram lets the model directly retrieve and fuse the representation of a known entity (e.g., 'Diana, Princess of Wales') when it encounters a trigger token, reducing unnecessary compute.