MEME (In-Context) - Tracking: leaderboard
Long-term conversational memory with the whole episode in context: about 35K tokens of multi-session dialogue about fictitious personal or software-project facts, then questions; MEME (Jung et al., 2026). Tracking asks for the full revision history of a value in order; accuracy in percent from a GPT-4o judge, higher is better.
Metric: Accuracy (fraction x 100; the whole ~35K-token episode transcript, 32K tokens of filler around the evidence sessions, fed to the answering model with no memory system; GPT-4o judge; tracking: the full revision history of an entity in order). Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 Mini | 69 | #775 |
| 2 | Claude Sonnet 4.6 | 58 | #158 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=meme-in-context-tracking · How It Works · Data refreshed daily, snapshot 2026-11-12.