MemSyco-Bench (Xiamen U + Jilin U): New Benchmark Exposes 'Memory-Induced Sycophancy' — Agents Reach 52.3% Sycophancy on DeepSeek-V4-Flash, 61-62% of Errors Occur AFTER Memory Retrieval

Researchers from Xiamen and Jilin Universities introduce MemSyco-Bench, arguing existing memory benchmarks miss the real failure mode: 61-62% of errors in memory-enabled agents happen after relevant memory is retrieved. Adding incorrect user memories drives DeepSeek-V4-Flash sycophancy to 52.3% and drops Qwen3 accuracy from 49.12 to 26-36%. 'Memory caution' instructions help conflicts (+31.6%) but hurt personalization; confirmation prompts backfire, worsening personalization up to 46.3%.