MemDrift: leaderboard
Memory-induced tool drift (MemDrift, Dabas et al., 2026): professional tool-call scenarios in seven domains where the agent memory holds personal-life preferences that should not affect the task. Claude Opus 4.6 compares five tool calls made with the biased memories against five made with no memory and scores the deviation from 1 (unchanged) to 5; lower is better. Memories injected into the system prompt; mean over 105 scenarios in five bias dimensions.
Metric: Deflection score under biased memory (1-5; Claude Opus 4.6 judge compares five tool calls against the no-memory calls, 1 = unchanged), direct memory injection, mean of five bias dimensions (105 scenarios). Source: arxiv.org. Saturation forecast: Around October 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Medium) | 3.53 | #172 (GPT-5.2) |
| 2 | GPT-5.4 (Medium) | 3.79 | #121 (GPT-5.4) |
| 3 | Qwen 3.5 397B A17B (Thinking) | 4.42 | #242 (Qwen 3.5 397B A17B) |
| 4 | Gemini 2.5 Pro (Medium) | 4.56 | #262 (Gemini 2.5 Pro) |
| 5 | Kimi K2.5 (Thinking) | 4.62 | #252 (Kimi K2.5) |
| 6 | Gemini 3.1 Pro (Preview) (Medium) | 4.67 | #91 (Gemini 3.1 Pro (Preview)) |
| 7 | Claude Sonnet 4.5 | 4.79 | #244 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memdrift · How It Works · Data refreshed daily, snapshot 2026-11-12.