We present AION, a 2.2B-parameter language model architecture that decouples factual recall from reasoning computation through a learned Hard Bypass mechanism. When an associative memory module (Engram) retrieves a stored fact with confidence exceeding 0.90, the model skips all transformer layers entirely. Gate training achieves 99.6% confidence on factual queries and 0.0% on generative queries after 12 epochs.