EverMemBench (Full Context): leaderboard
EverMemBench (Hu et al., 2026) tests long-horizon memory over simulated multi-party workplace chat: five year-long projects, about 845,000 tokens of group dialogue each, and 2,400 questions, here with the whole dialogue history in the model's context. Accuracy in percent averaged over nine sub-tasks of fine-grained recall, memory awareness and profile understanding; higher is better.
Metric: Accuracy (%; mean of the nine sub-tasks, whole multi-party dialogue history in context, open-ended recall judged by an LLM). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 72.61 | #159 |
| 2 | Llama 4 Scout Instruct | 40.18 | #1532 |
| 3 | GPT-4.1 Mini | 37.44 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evermembench-full-context · How It Works · Data refreshed daily, snapshot 2026-11-12.