RAG is like constantly reading your notes instead of integrating experiences into your processes.

Not really. For example we still can’t get coding agents to work reliably, and I think it’s a memory problem, not a capabilities problem.

On the other hand, test-time weight updates would make model interpretability much harder.