interpretabilityfactneutralPersonalization memory in LLMs reshapes reasoning trajectories on open-ended questions, not just final answersArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02374v1