On the other hand, CAG preloads all the necessary knowledge into a long-context model’s memory in its KV cache. A complicated way of saying the database here is pre-computed into the model's initial state.

→ View original post on X — @whats_ai