VibeMemBench (Memory Off): leaderboard
VibeMemBench (Fan et al., Chinese Academy of Sciences and Alibaba, 2026) builds 111 SWE-style repository coding targets from 90 SWE-rebench V2 repositories, keeping only targets where injected experience from earlier work in the same repository helps a reference solver. In this memory-off condition the MiniSWEAgent solver gets no prior experience; percent of 444 target-seed runs whose declared tests pass, higher is better.
Metric: Resolved (%; memory-off condition, no prior experience injected; 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 | Qwen 3.8 Max | 79.1 | #55 |
| 2 | GLM-5.2 | 76.8 | #137 |
| 3 | Kimi K2.7 Code | 69.8 | #180 |
| 4 | DeepSeek V4 Pro | 67.1 | #181 |
| 5 | GLM-5 | 55 | #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-memory-off · How It Works · Data refreshed daily, snapshot 2026-11-12.