Top-k Memory - LongMemEval (Gold): leaderboard
Top-k Retrieval Leaves Memory Value Undetermined v1, Appendix Table9: 300 LongMemEval oracle questions per hosted reader, gold evidence condition. Temperature0, 24000-character context; normalized exact match followed by same-model judge. Hosted endpoint names are retained; no snapshot equivalence is inferred. Pooling-cost differences and retrieval-identifiability diagnostics are not additional model scores. Reported metric: Accuracy (fraction).
Metric: Accuracy (fraction). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.6 Flash | 0.78 | #57 |
| 2 | Gemini 3.7 Flash | 0.76 | #24 |
| 3 | Gemini 2.5 Flash Lite | 0.61 | #432 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=top-k-memory-longmemeval-gold · How It Works · Data refreshed daily, snapshot 2026-11-05.