MemDrift - Autonomy and Self-Reliance: leaderboard
Memory-induced tool drift (MemDrift, Dabas et al., 2026) from self-reliant preferences stored in the agent memory, over 21 professional tool-call scenarios with the memories injected into the system prompt. Claude Opus 4.6 rates, from 1 (unchanged) to 5, how far five tool calls with these irrelevant memories deviate from five calls with no memory; 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), direct memory injection, Autonomy / Self-Reliance bias (21 scenarios). Source: arxiv.org. Saturation forecast: Around October 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Medium) | 3.14 | #172 (GPT-5.2) |
| 2 | GPT-5.4 (Medium) | 3.48 | #121 (GPT-5.4) |
| 3 | Qwen 3.5 397B A17B (Thinking) | 4.19 | #242 (Qwen 3.5 397B A17B) |
| 4 | Gemini 2.5 Pro (Medium) | 4.24 | #262 (Gemini 2.5 Pro) |
| 5 | Claude Sonnet 4.5 | 4.57 | #244 |
| 6 | Kimi K2.5 (Thinking) | 4.62 | #252 (Kimi K2.5) |
| 7 | Gemini 3.1 Pro (Preview) (Medium) | 4.81 | #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-autonomy-and-self-reliance · How It Works · Data refreshed daily, snapshot 2026-11-12.