“Engram's continual learning approach can reduce token inference consumption by up to two orders of magnitude (100x) by internalizing context that would otherwise require large system prompts.”