Research2h ago
RENDER Benchmark Tests LLM Memory Formatting
Matched-budget resolved packets beat recency-truncated raw dialogue by up to 72.6 points in the new RENDER benchmark. Researchers tested nine models using 500 LongMemEval questions to see how input formatting affects recall. The results show that structured memory templates significantly outperform raw conversation. Practitioners should optimize how they render history for RAG systems.