Current AI agents are remarkably capable at solving individual tasks, yet they often suffer from a memory problem where they fail to retain long-term expertise. Every time an agent encounters a new scenario, it essentially starts from scratch, frequently rediscovering solutions it has already mastered in different contexts. To solve this, researchers from Zhejiang University have unveiled SkillNet, an open-source infrastructure designed to serve as a permanent library for capabilities. By formalizing these skills as durable, reusable assets, the platform prevents systems from starting over and enables them to build upon past successes.

The core of SkillNet is a massive repository containing over 200,000 skills organized through a structured system of categories and relationships (a unified ontology). This setup allows systems to fetch specific strategies for tasks ranging from scientific problem-solving to digital shopping. Unlike traditional methods that treat every interaction as a one-off event, SkillNet allows models to connect disparate pieces of knowledge. This connectivity ensures that a strategy learned in one environment can be efficiently adapted to another, mirroring how humans accumulate expertise over a lifetime.

The technical impact is significant, with testing on benchmarks like ALFWorld and WebShop showing a 40% increase in success rates. Beyond pure performance, SkillNet emphasizes responsible deployment by evaluating skills across five dimensions, including safety, cost-awareness, and maintainability. By providing a Python toolkit and an interactive platform, the team aims to transform development from training isolated models into cultivating an ecosystem of evolving, modular intelligence.

Current AI agents are remarkably capable at solving individual tasks, yet they often suffer from a memory problem where they fail to retain long-term expertise. Every time an agent encounters a new scenario, it essentially starts from scratch, frequently rediscovering solutions it has already mastered in different contexts. To solve this, researchers from Zhejiang University have unveiled SkillNet, an open-source infrastructure designed to serve as a permanent library for capabilities. By formalizing these skills as durable, reusable assets, the platform prevents systems from starting over and enables them to build upon past successes.

The core of SkillNet is a massive repository containing over 200,000 skills organized through a structured system of categories and relationships (a unified ontology). This setup allows systems to fetch specific strategies for tasks ranging from scientific problem-solving to digital shopping. Unlike traditional methods that treat every interaction as a one-off event, SkillNet allows models to connect disparate pieces of knowledge. This connectivity ensures that a strategy learned in one environment can be efficiently adapted to another, mirroring how humans accumulate expertise over a lifetime.

The technical impact is significant, with testing on benchmarks like ALFWorld and WebShop showing a 40% increase in success rates. Beyond pure performance, SkillNet emphasizes responsible deployment by evaluating skills across five dimensions, including safety, cost-awareness, and maintainability. By providing a Python toolkit and an interactive platform, the team aims to transform development from training isolated models into cultivating an ecosystem of evolving, modular intelligence.