MemTrace - Fresh: leaderboard
Long-term memory of a multi-session conversation history read directly in the model's context, asked early: questions on the current state, an earlier state and the change over time of user facts derived from HaluMem-Medium; MemTrace (Long et al., 2026). Gist accuracy in percent from a GPT-4o judge at the first two of eight checkpoints (the shortest histories), mean over the three question types; higher is better.
Metric: Gist accuracy (%; GPT-4o judge; the model answers from the visible multi-session history in its context; fresh history: checkpoints W1-W2, mean of current-state, earlier-state and change-over-time questions). Source: arxiv.org. Saturation forecast: Around December 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 35B A3B | 61.3 | #480 |
| 2 | GPT-5 Nano | 56.2 | #888 |
| 3 | Gemini 3 Flash | 52.3 | #159 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memtrace-fresh · How It Works · Data refreshed daily, snapshot 2026-11-12.