If the weights never change, is it really recursive self-improvement?

preview.redd.it/e0ydm43a55kh1.png This paper is using a much narrower definition of recursive self-improvement than the phrase usually suggests. AQuA stores validated evidence in a persistent research state that shapes later hypotheses. The underlying language model and evaluator remain fixed. I still find the narrower claim interesting, even if it sits closer to memory-augmented research automation than to a model rewriting itself. The paper does not establish any weight-level capability gain. Is persistent memory that improves later research decisions enough to call a system RSI, or should the term require changes to the system’s underlying capabilities?