China's NeoLab Moment: EverMind Answers the Full-Stack Self-Evolution Question with Three Papers

In the summer of 2026, as overseas NeoLabs like Recursive Superintelligence (RSI), Engram, Adaption Labs, and Core Automation attracted billions in funding, Chinese team EverMind, incubated by Shanda Group, published three consecutive papers delivering a full-stack solution for self-evolving AI. EverMind, with roots in Shanda Innovation Institute's research culture, focuses on long-term memory and self-evolution. The papers address three challenges: harness self-improvement (HarnessBank), skill management (SkillCorpus), and model weight training (DASH). HarnessBank introduces a gene bank and gated verification to avoid overfitting, achieving 5.1-15.4% performance gains across seven benchmarks. SkillCorpus aggregates and curates over 82.1k raw skills into 96,401 high-quality skills, improving agent performance by up to 7.5 points. DASH uses divergence-adaptive supervision horizons to improve on-policy self-distillation, boosting math reasoning scores on AIME and HMMT benchmarks. EverMind has an open-source ecosystem with over 12k GitHub stars, and its frameworks are mapped across four layers: task, harness, model, and meta-improvement. The team expects that the next-generation AI will evolve continuously, and they invite partners to join in building AI long-term memory infrastructure.