RENDER finds a 72.6-point memory gap

RENDER, a benchmark for LLM memory and RAG, held the conversation fixed while changing how the model saw its history: across 500 LongMemEval questions and nine models, resolved packets beat recency-truncated raw dialogue by 42.4 to 72.6 points. The paper’s best-to-worst spread across deployed-style templates was 24.6 to 48.8 points, a warning that teams should benchmark the rendered memory artifact—not just retrieval recall—alongside the archive’s analysis of agent skill-library retrieval, according to the RENDER paper.