Self-attention must re-read every stored token embedding for each new token. Linear attention never re-reads embeddings — each token only needs the running state and nothing else.