Retrieval Unchanged, But 32 of 500 LongMemEval Results Flipped

przemarzec · reddit · 2026-08-21

Comparing Engrava 0.6.0 and 0.5.0 on LongMemEval revealed that while retrieval results were identical for 457 questions, 32 outcomes flipped (18 down, 14 up). This proves the score variance occurred post-retrieval in the reader or judge stage, not in the memory layer itself. The post advises diffing retrieved context IDs, not just final scores, when benchmarking memory layers.

Original post →

More from Research

Research channel →