AMemGym (Off-Policy): leaderboard
Metric: Normalized memory score (times 100): overall question-answering accuracy rescaled between a random-guess baseline (0) and the same model's upper bound when given the ground-truth user states (100), averaged over evaluation periods, on the AMemGym base configuration: 20 synthetic user profiles, each with 10 multiple-choice personalization questions (4 to 7 options) asked after each of 11 interaction periods while an LLM-simulated user (GPT-4.1) reveals evolving user states over about 47 turns; the native LLM keeps the whole history in context; temperature 0; off-policy: the model answers from GPT-4.1's recorded on-policy conversation instead of its own; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 33.9 | #194 |
| 2 | Gemini 2.5 Flash | 31.7 | #237 |
| 3 | Gemini 2.0 Flash | 21.4 | #331 |
| 4 | Gemini 2.5 Flash Lite | 20.4 | #413 |
| 5 | GPT-4.1 Mini | 19.8 | #346 |
| 6 | DeepSeek V3 | 16.5 | #312 |
| 7 | GPT-4o Mini | 16.4 | #588 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=amemgym-off-policy · How It Works · Data refreshed daily, snapshot 2026-10-11.