In the summer of 2026, self-evolving AI reached a historical milestone. While global NeoLabs are securing hundreds of millions in funding based on ambitious visions, Chinese research team EverMind, backed by Shanda Group, has quietly released three groundbreaking papers, delivering the industry's first full-stack roadmap for Self-Evolving AI. This achievement showcases China's rising capability in fundamental AI research.
Rooted in the research legacy of the Shanda Innovation Institute from the early internet era, Shanda has pivoted its strategic focus entirely to brain science and AI. As a key incubator, EverMind tackled the hardest but most foundational problem in AI: solving long-term memory and utilizing it to enable recursive self-improvement.
This comes amid a booming global NeoLab wave. In 2026, startups like Recursive Superintelligence (RSI)—co-founded by Meta's former scientist Yuandong Tian—raised $650 million at a $4.65 billion valuation. AlphaGo pioneer David Silver's Ineffable Intelligence secured a $5.1 billion valuation. These elite teams are all betting on Recursive Self-Improvement to build AI that adapts post-deployment.
However, most highly valued NeoLabs remain in conceptual stages. Engram focuses heavily on parametric weight memory, while Adaption Labs targets inference-time adaptation. In contrast, EverMind's trilogy of papers comprehensively covers the skill layer, the harness/scaffolding layer, and the model weight layer, forming a complete self-evolution technical loop.
This synergy was highlighted at the "Memory Origins" Hackathon, where RSI's Yuandong Tian and #EverMind's lead Yafeng Deng co-headlined a podcast. Deng emphasized that the core technical path of next-gen AI goes from long-term memory to continuous learning, and ultimately to self-evolution. EverMind's memory operating system, EverOS, and its self-evolving agent framework, Raven, serve as the practical manifestation of this philosophy.
True self-evolution is hard. Traditional LLMs are frozen post-training. Achieving continuous evolution requires overcoming massive hurdles, such as preventing overfitting in scaffold code modification and maintaining scalable long-term skill registries without catastrophic forgetting. EverMind's research directly addresses these core pain points.
[AgentUpdate Depth Analysis] As the marginal utility of raw foundation LLMs begins to plateau, the self-evolution capability of AI Agents has emerged as the definitive frontier for next-gen AI ecosystems. While highly-funded Western NeoLabs like RSI have dominated headlines, their methodologies often remain isolated in parametric fine-tuning or ephemeral runtime adaptations. In contrast, EverMind’s multi-layered approach—coupling the EverOS memory substrate with the Raven framework—presents a highly pragmatic and holistic blueprint. By simultaneously addressing scaffolding optimization and long-term skill persistence, EverMind offers a robust solution to overfitting and agent degradation in dynamic environments. Long-term, this dual-engine approach will transition AI Agents from static, version-locked software into continuously compounding entities, radically accelerating the commercial viability of fully autonomous workflows and adaptive physical systems.
