MemEye (Full Context) - Open-Ended: leaderboard
Long-term multimodal agent memory: questions about eight life-scenario conversation histories that require keeping fine-grained visual evidence and tracking visual states that change across sessions; MemEye (Guo et al., 2026). The model reads the whole native multimodal history (a 128k-token budget, oldest turns dropped first) with no memory system; a GPT-5.2 judge scores the 371 mirrored open-ended answers from 0 to 1, reported in percent and averaged over the benchmark's 12 visual-granularity by memory-reasoning cells. Higher is better.
Metric: LLM-judge score (%; GPT-5.2 semantic-correctness score in 0-1 on the 371 mirrored open-ended questions, macro-averaged over the 12 visual-granularity x memory-reasoning cells; full native multimodal history in context, no memory system). Source: arxiv.org. Saturation forecast: Around May 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 43.91 | #389 |
| 2 | Gemini 2.5 Flash Lite | 38.62 | #958 |
| 3 | GPT-4.1 Nano | 35.08 | #1992 |
| 4 | Qwen 3 VL 8B Instruct | 21.12 | #982 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memeye-full-context-open-ended · How It Works · Data refreshed daily, snapshot 2026-11-12.