HUSH-Bench (A-Mem): leaderboard
HUSH-Bench (2026) measures whether a conversational agent with memory brings up sensitive personal history that the current benign request gives no reason to use. The score, from 0 to 100, is an unsolicited-history-integration rating by a GPT-5.4 judge over 2,400 probes with memories retrieved by the A-Mem memory system; lower is better.
Metric: Unsolicited history integration score (0-100): a GPT-5.4 judge scores use of prior sensitive history the current benign prompt gives no reason to use; 2,400 probes per cell; temperature 0, 800-token budget; lower is better; memory entries retrieved by A-Mem (up to five). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 8.9 | #389 |
| 2 | Claude Sonnet 4.6 | 51.3 | #161 |
| 3 | DeepSeek V4 Flash | 57.4 | #258 |
| 4 | Qwen 3.5 9B | 69.7 | #803 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=hush-bench-a-mem · How It Works · Data refreshed daily, snapshot 2026-11-13.