VibeMemBench (Frozen Verified Experience): leaderboard
VibeMemBench (Fan et al., Chinese Academy of Sciences and Alibaba, 2026) evaluates memory for coding agents on 111 repository coding targets from SWE-rebench V2. In this condition each target comes with its frozen verified experience, a lesson distilled from an earlier trajectory in the same repository and shown to help a reference solver, injected into the MiniSWEAgent context; percent of 444 target-seed runs whose declared tests pass, higher is better.
Metric: Resolved (%; the frozen verified experience for each target, distilled from earlier trajectories in the same repository and verified to help a reference solver, injected into the solver context; share of 444 target-seed runs (111 SWE-style repository coding targets from 90 SWE-rebench V2 repositories x 4 seeds) whose declared tests pass; MiniSWEAgent solver). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 80.4 | #137 |
| 2 | Qwen 3.8 Max | 80.2 | #55 |
| 3 | Kimi K2.7 Code | 74.3 | #180 |
| 4 | DeepSeek V4 Pro | 67.1 | #181 |
| 5 | GLM-5 | 57.4 | #256 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vibemembench-frozen-verified-experience · How It Works · Data refreshed daily, snapshot 2026-11-12.