Nous: A Predictive World Model for Long-Term Agent Memory

Nous is a new predictive memory architecture that conceptualizes knowledge as prediction rather than storage, utilizing categorical probability distributions to maintain a world model. It updates beliefs using a Bayesian posterior based on information-theoretic surprise and records shifts in belief (delta) instead of storing facts directly. Evaluated against the LoCoMo benchmark, Nous demonstrates superior performance over A-MEM and BeliefMem in multiple memory tasks, achieving notable F1 scores with a GPT-4o-mini backbone, which is significant for practitioners seeking efficient long-term memory systems in AI agents without reliance on external databases.