Memory-World - Active Memory Instruction: leaderboard
Memory-intensive mobile GUI tasks from Memory-World (STAMP paper, Wang et al., 2026): in synthesized virtual apps, information shown on one screen leaves the visual context and must be recalled later, and the environment verifies whether the task was completed. This board covers general-purpose multimodal models given an appended instruction to write down information they may need later; the score is task accuracy on one attempt (pass@1), in percent; higher is better.
Metric: Task Accuracy (%; pass@1, general-purpose models with an explicit memory instruction). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 59.1 | #158 |
| 2 | GPT-5.4 | 52 | #121 |
| 3 | Gemini 3.1 Pro (Preview) | 49.6 | #91 |
| 4 | Qwen 3 VL 8B (Thinking) | 36.2 | #887 |
| 5 | Qwen 3 VL 8B Instruct | 16.5 | #982 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memory-world-active-memory-instruction · How It Works · Data refreshed daily, snapshot 2026-11-12.