ATM-Bench: leaderboard
Agentic long-term memory benchmark (atmbench.github.io): models answer questions over a large accumulated interaction history. Oracle track scores each answer model with the full context provided; QS is an LLM-judged question score in percent.
Metric: Oracle QS (%). Source: atmbench.github.io. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 86 | #87 |
| 2 | GPT-5 | 85.3 | #105 |
| 3 | GPT-5.4 | 85.1 | #83 |
| 4 | Qwen 3.6 27B | 84.5 | #182 |
| 5 | GPT-5.2 | 84.12 | #115 |
| 6 | MiniMax-M3 | 83.6 | #151 |
| 7 | Claude Sonnet 4.5 | 83.3 | #148 |
| 8 | MiMo-V2.5 | 82.6 | #196 |
| 9 | Gemini 2.5 Pro | 78.6 | #159 |
| 10 | Qwen 3 VL 8B Instruct | 78.19 | #419 |
| 11 | Gemini 2.5 Flash | 76.9 | #255 |
| 12 | MiMo-V2.5-Pro | 71.7 | #167 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=atm-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.