WorldMemArena: leaderboard
WorldMemArena (Chen et al., 2026) evaluates multimodal agent memory over long multi-session trajectories through an action-world interaction loop, asking checkpoint questions whose supporting evidence has left the local context. This board reports QA Correct, the share of answers a GPT-5.4-mini judge grades correct against the gold answer and evidence, as a percentage where higher is better, for base models answering with full context and for agent harnesses that manage their own memory.
Metric: QA Correct (%; share of answers a GPT-5.4-mini judge marks correct against the gold answer and evidence, over long multi-session memory trajectories). Source: arxiv.org. Saturation forecast: Around May 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 58.27 | #389 |
| 2 | Gemini 3 Flash | 51.69 | #159 |
| 3 | Qwen 3.5 Plus | 51.05 | #227 |
| 4 | GPT-5.4 Nano | 48.31 | #633 |
| 5 | Claude Haiku 4.5 | 36.71 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=worldmemarena · How It Works · Data refreshed daily, snapshot 2026-11-12.