LongMemEval

Independent researcher proposes LSREP longitudinal memory benchmark

An undergraduate researcher posted an arXiv preprint introducing LSREP, a protocol that uses chronological replay, lifecycle schedules and repeated probes to test how conversational memory accumulates and is retrieved over time. The case system ICE v2 covered 1,985 turns, 219 probes and 52 checkpoints, but did not beat pure vector-RAG on a matched LongMemEval evaluation, and ordinary-density replay showed no quality advantage. Code and a frozen artifact are public; the author stresses the architecture is not fully validated.