MemDrift (SimpleMem): leaderboard
Memory-induced tool drift (MemDrift, Dabas et al., 2026) with the SimpleMem memory framework: personal-life preferences are planted in a simulated three-turn conversation, stored and retrieved by SimpleMem, then a subset of the professional tool-call scenarios is run. Claude Opus 4.6 rates the deviation of five tool calls from five no-memory calls from 1 (unchanged) to 5; lower is better.
Metric: Deflection score under biased memory (1-5; Claude Opus 4.6 judge compares five tool calls against the no-memory calls, 1 = unchanged), memories stored and retrieved by SimpleMem, mean of five bias dimensions over a subset of the scenarios. Source: arxiv.org. Saturation forecast: Around August 2028. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 3.69 | #121 (GPT-5.4) |
| 2 | Gemini 3.1 Pro (Preview) (Medium) | 4.71 | #91 (Gemini 3.1 Pro (Preview)) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memdrift-simplemem · How It Works · Data refreshed daily, snapshot 2026-11-12.