A mechanism where relevant information is retrieved from model parameters themselves rather than from external memory or attention, helping reduce computational bottlenecks.