PersistBench - Beneficial Memory: leaderboard
Evaluates appropriate use of persistent memories. Cross-domain leakage and sycophancy each use 200 cases, counting failure in any of three generations; beneficial memory uses 100 cases with one generation. Lower failure is better, including failures to use beneficial memory. This series reports Beneficial Memory using Failure Rate (%). Source: the original paper result table.
Metric: Failure Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 2 | #87 |
| 2 | Gemini 3 Flash | 4 | #103 |
| 3 | Gemini 3 Pro | 4 | #82 |
| 4 | Grok 4 | 5 | #189 |
| 5 | Grok 4.1 Fast | 6 | #227 |
| 6 | Kimi K2 (Thinking) | 10 | #252 (Kimi K2) |
| 7 | Claude Sonnet 4.5 | 11 | #148 |
| 8 | DeepSeek V3.2 Speciale | 14 | #138 |
| 9 | Qwen 3 235B A22B | 19 | #323 |
| 10 | GPT-OSS-120B | 20 | #346 |
| 11 | GLM-4.7 | 21 | #202 |
| 12 | GPT-5.2 (High) | 23 | #115 (GPT-5.2) |
| 13 | MiniMax-M2.1 | 30 | #299 |
| 14 | Kimi K2 0905 | 33 | #292 |
| 15 | GPT-4o | 53 | #357 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=persistbench-beneficial-memory · How It Works · Data refreshed daily, snapshot 2026-11-05.