Papr reports 96.0% on LongMemEval memory benchmark

Papr says it answered 480 of 500 LongMemEval questions correctly (96.0%), which it calls the highest published score it has found, using the benchmark's own answer prompt and grader. Its 95.4% configuration sends the answer model 16.5k tokens per question, about a third of what Hindsight sends; with gpt-5-mini it reaches 94.4%. Retrieval places all evidence sessions in the top 5 for 97.9% of answerable questions. Papr says every answer and grading decision is public, and that it tuned prompts on half the questions and checked on the other half. It notes there is no further held-out set, which it says is true of most published LongMemEval results. The embedding model, papr-embed, is available through the Papr Memory API.

Generated from reports · updated 14 min ago

Not enough continuous observations yet to draw a trend.