Nous preprint demonstrates memory improvement certificates without source calibration

A preprint presents a method for certifying improvements in an agent's memory decisions without first estimating how reliable its information sources are. In one mathematical model family, learning and certifying useful decisions can require quadratically fewer records than calibrating sources as persistence declines. The researchers tested 45,000 held-out histories and 9,000 episodes across three MiniGrid memory environments with an introduced noisy-report interface. The new certificate accepted 9/9 improvements over a constant incumbent and 4/9 over a last-write-wins policy; the earlier certificate accepted none in MiniGrid. Strong established inference baselines remained competitive or better, limiting the result to learning and certifying decisions rather than establishing a universally superior memory algorithm.