Can Memory Make an AI Worse? My Benchmark with Qwen3.5-9B and Stellar Classification
(Antonio V. Franco) I ran 135 celestial object classification tasks using three memory approaches. The result was counterintuitive. It seems like common sense: if an artificial intelligence agent learns from its past experiences, it should become better over time. Each solved problem becomes a reference, a reusable pattern that accelerates and sharpens future decisions. This is precisely the intuition behind the ReasoningBank paper (Ouyang et al., ICLR 2026), a system that stores reasoning strategies in a memory bank and retrieves them when facing similar tasks. The promise is seductive and aligns with how human experts build expertise. ...