Radar
Top stories
Importance 2 of 5Covered by 1 source

Papr memory retrieval reports 96.0% on LongMemEval-S with official judge

A team ran all 500 LongMemEval-S questions through its Papr memory retrieval system using the official answer prompt and the unmodified GPT-4o judge, reporting 96.0% accuracy. Per-question answers and judge labels are public so results can be reproduced. They report retrieval recall is near saturated, with all evidence sessions in the top 5 for 97.9% of answerable questions.

Why it matters

Reproducible long-term memory benchmarks help teams building AI assistants and agents compare memory systems on equal terms.