Wouldn't this require retraining the model each time your "extended mind" needs more information? One benefit of RAG in my mind is that the "database" can be updated in real-time without needing any model retraining