MEME (Gold Facts): leaderboard
The reasoning ceiling of MEME (Jung et al., 2026): the model gets only the task-relevant gold facts from an episode's multi-session dialogue, with no filler, and answers exact-recall, aggregation, tracking, deletion, cascade and absence questions about fictitious personal or software-project facts. Mean of the six task accuracies in percent, GPT-4o judge; higher is better.
Metric: Accuracy (fraction x 100; only the task-relevant gold facts are given to the answering model, no filler and no memory system; GPT-4o judge; mean of the six task accuracies). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 91 | #75 |
| 2 | Claude Sonnet 4 | 87 | #397 |
| 3 | Claude Sonnet 4.6 | 84 | #158 |
| 4 | GPT-4.1 Mini | 70 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=meme-gold-facts · How It Works · Data refreshed daily, snapshot 2026-11-12.