The paper introduces UpliftMem, a method for learning memory retrieval in LLM agents. Instead of relying on outcome feedback from executed memory sets alone, it learns from the set-level execution uplift relative to the same executor operating without memory. This allows the system to compare the effect of different memory sets without requiring full rollouts for every alternative.
UpliftMem uses a probe selection strategy based on the expected value of sample information (EVSI) to allocate limited training rollouts. A closed-form solution under a correlated Gaussian model guides which memory sets to probe, making training efficient. The shared scorer is trained with a frozen executor and selects memory sets at test time without further probes.
The method is evaluated on ALFWorld, WebShop, and BigCodeBench, achieving the best success rates among compared baselines. Controlled experiments with fixed stores and matched probe budgets show that UpliftMem improves memory-use decisions and makes better use of execution feedback.