When Does Belief-Based Agent Memory Help? Reliability-Conditional Updating and Provenance-Capped Poisoning Defense
arXiv cs.CL 6 hours ago
Researchers investigated when belief-based memory improves language model agents using Nous, an architecture that represents entity-attribute pairs as probability distributions updated through Bayesian inference, finding that belief updating provides little benefit in existing benchmarks without contradictory evidence. A controlled contradiction benchmark showed belief updating with reliability-conditioned updates achieved 27.5 points higher performance on LLM-as-judge evaluation compared to token-F1 metrics, and provenance-capped updating successfully resisted memory-poisoning attacks. Probabilistic belief-based memory is most useful in environments with conflicting and differently trustworthy evidence rather than standard conversational recall.