As multi‑agent systems and large language models scale up, deployed AI assistants are tasked with reasoning over long, multi‑session dialogues. Existing benchmarks treat dialogue history as a retrieval or temporal reasoning problem and ignore a core interaction aspect—how something was said. VoiceLongMemEval (VLME) addresses this gap. Each question requires paralinguistic metadata—emotion tags, prosody descriptors, voice events—attached to conversational turns, which cannot be recovered from text alone. Every item passes a three‑stage adversarial gate, guaranteeing that a strong language model fails when only given the transcript. Evaluations of leading closed‑source and open‑weight models reveal a pervasive affect gap; adding text‑track paralinguistic metadata boosts accuracy by 0.09–0.38 (up to 0.61–0.69 with evidence hints), while standard ASR pipelines discard this signal. Audio‑native models extract cues directly from speech, achieving 0.354–0.412 versus 0.325 for blind baselines. Code and dataset will be released upon acceptance.
Review