GenericAgent is a compact autonomous agent framework from Fudan University researchers, introduced in April 2026 via a GitHub release and an arXiv paper (2604.17091). At just 3,300 lines of core code, it achieves full local system control while consuming six times fewer tokens than comparable frameworks. The agent evolves by automatically crystallizing solved task execution paths into reusable skills stored in its L3 skill tree memory layer — meaning every completed task makes future similar tasks faster and cheaper. The framework implements four tightly integrated subsystems: a minimal nine-tool atomic set (browser, terminal, file ops, keyboard/mouse, vision, ADB), a hierarchical five-layer on-demand memory (L0 meta rules through L4 session archives), a self-evolution mechanism that converts verified trajectories into Standard Operating Procedures, and a context truncation layer that keeps active context under 30K tokens even on complex multi-step workflows. As of April 2026, GenericAgent supports Claude, Gemini, Kimi, and MiniMax as reasoning backends, ships a million-scale public Skill Library, integrates with WeChat, QQ, Feishu, and DingTalk bot frontends, and includes L4 session archive memory for long-horizon recall across days or weeks. The project surged to nearly 8,000 GitHub stars within weeks of its arXiv paper publication on April 21, 2026.