MemCalib: leaderboard
MemCalib (2026) tests whether an LLM gives each supplied memory proposition an appropriate degree of influence over its response, over 1,500 health, general-assistance and coding queries whose memory atoms are labelled ignore, bound or control. The Sample Calibration Score is a geometric-decay score over each response total memory over-use and under-use, rescaled to percent and averaged, with a DeepSeek-V4-Pro rubric judge and non-thinking generation over three seeds; higher is better.
Metric: Sample Calibration Score (%; geometric-decay score over per-response memory over-use and under-use totals, DeepSeek-V4-Pro rubric judge, mean over 3 seeds; non-thinking mode). Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Non-reasoning) | 46.25 | #29 (GPT-5.6 Sol) |
| 2 | Claude Sonnet 4.6 | 36.44 | #158 |
| 3 | Kimi K2.6 (Non-reasoning) | 35.54 | #173 (Kimi K2.6) |
| 4 | GLM-5.2 (Non-reasoning) | 34.4 | #137 (GLM-5.2) |
| 5 | DeepSeek V4 Flash (Non-reasoning) | 33.72 | #254 (DeepSeek V4 Flash) |
| 6 | Qwen 3 8B (Non-reasoning) | 31.17 | #1838 (Qwen 3 8B) |
| 7 | Qwen 3.5 35B A3B (Non-reasoning) | 26.54 | #480 (Qwen 3.5 35B A3B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memcalib · How It Works · Data refreshed daily, snapshot 2026-11-12.