Research4h ago
RENDER Benchmark Tests LLM Memory Formatting
Matched-budget resolved packets outperformed recency-truncated raw dialogue by up to 72.6 points in new memory tests. The RENDER benchmark varies how history is presented—as summaries, typed records, or raw excerpts—across nine models. This proves that how a system formats retrieved memory impacts accuracy more than the content itself. Practitioners should optimize prompt templates over raw history.