Study Finds AI Assistants Recall Facts Well but Rarely Apply Them in Context

A July 2026 research study called InMind tested whether AI assistants can not only retrieve stored personal facts but also apply them appropriately to new, related requests. Using 125 synthetic tasks, the study found that one tested configuration achieved 100% accuracy when directly asked about a stored fact, yet applied that fact correctly in indirect scenarios only 9.6% of the time. A key example highlighted in the study involved an assistant that knew a user had a tree-nut allergy but still recommended an almond-flour recipe when asked about macarons. The researchers deliberately stripped obvious retrieval cues from indirect requests to stress-test this specific failure mode, where a stored constraint is relevant but not explicitly mentioned. The findings suggest that high recall scores on direct questions do not guarantee an AI will bring the right background knowledge to bear when it matters most.