GenericAgent Ships Self-Evolving Skill Tree From 3.3K-Line Seed — ArXiv Paper Shows 6x Token Reduction via Contextual Information Density Maximization

GitHub·medium signal

lsdefine/GenericAgent grows a self-evolving skill tree from 3,300 lines of seed code, achieving full system control while consuming 6x fewer tokens than baselines. Uses a five-layer memory architecture (L0 Meta Rules through L4 Session Archive) where solved tasks become reusable skills — the second invocation is a single-line call with no relearning. ArXiv paper published April 21 formalizes the approach as 'Contextual Information Density Maximization.' Gained 3,900 stars in its first trending week.