TRACE Bench - Short-Term Memory: leaderboard
Share of TRACE Bench cases in which the roleplay model correctly recalls a user-injected external fact probed later in the multi-turn dialogue (TRACE Bench, arXiv 2608.11236v1, Table 8). Percent over 200 cases; higher is better; thinking disabled for all targets.
Metric: Memory-probe completion rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 92 | #33 |
| 2 | MiMo-V2.5 | 91 | #196 |
| 3 | Gemma 4 31B (Non-reasoning) | 90.5 | #212 (Gemma 4 31B) |
| 4 | MiMo-V2.5-Pro | 90 | #167 |
| 5 | MiniMax-M2.7 | 88.5 | #270 |
| 6 | Kimi K2.5 (Non-reasoning) | 87.5 | #152 (Kimi K2.5) |
| 7 | DeepSeek V4 Flash | 87 | #155 |
| 8 | GLM-5 (Non-reasoning) | 87 | #149 (GLM-5) |
| 9 | Claude Sonnet 4.6 | 86 | #94 |
| 10 | Qwen 3 8B (Non-reasoning) | 85.5 | #678 (Qwen 3 8B) |
| 11 | Qwen 3.5 27B (Non-reasoning) | 85 | #192 (Qwen 3.5 27B) |
| 12 | DeepSeek V4 Pro (Non-reasoning) | 85 | #112 (DeepSeek V4 Pro) |
| 13 | Mistral Small 3.2 | 84.5 | #608 |
| 14 | Qwen 3 14B (Non-reasoning) | 84 | #542 (Qwen 3 14B) |
| 15 | Ministral 3 14B | 74 | #664 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trace-bench-short-term-memory · How It Works · Data refreshed daily, snapshot 2026-11-05.