A new benchmark says today's AI agents still don't know how much to trust their own memories.
Researchers built MemCalib, a benchmark that tests whether large language models give each stored memory the right amount of influence over a response, rather than just recalling it correctly. Across frontier open-source and closed-source models, the results were poor: systems consistently over-relied on some memories while ignoring others, producing skewed answers. The team also tested standard post-training methods, including group relative policy optimization and on-policy self-distillation, and found a lopsided pattern - fine-tuning that fixed over-use tended to make under-use worse, and vice versa. To address that, they built MemCalib-RL, a training algorithm that separates over-use and under-use signals and traces credit back to specific response tokens through exact atom ablation.
Memory is the feature vendors pitch as making AI agents feel personal and persistent, but this research suggests the calibration underneath is shakier than the marketing implies. An agent that leans too hard on a stale preference, or ignores a fact it was explicitly given, produces the same failure users already distrust chatbots for: confident, wrong answers dressed up as personalization.
Tested on Qwen3-8B, Ministral-3-8B-Instruct, and Qwen3.5-35B-A3B, MemCalib-RL outperformed standard fine-tuning and the gains carried over to outside benchmarks - though holding up in a paper's own follow-up tests and holding up in a production agent are still two different claims.