MemCredit preprint reports agent gains by selecting memories for decision utility

Researchers report in a preprint that MemCredit improves agent task performance by selecting memories for their contribution to decisions. On ALFWorld, it achieves 86.57 SR@50 under a 500-token historical-memory budget, compared with 65.67% for BM25 retrieval. The framework uses controlled counterfactual interventions to estimate how individual memories affect actions, then trains a lightweight estimator to predict their utility. During inference, semantic relevance generates candidates, while predicted utility determines which memories fit within the token budget. This leaves the underlying agent policy unchanged and requires no additional environment interaction. The evaluations cover ALFWorld and WebShop; the reported numerical comparison concerns ALFWorld.