VentureBeat Covers Memento-Skills: Self-Designing Agent Framework Achieves 116% Accuracy Improvement Without Retraining Underlying Model

VentureBeat·low signal

VentureBeat highlighted the Memento-Skills framework, where agents autonomously construct, adapt, and improve task-specific sub-agents through a Read-Write Reflective Learning loop that stores evolving skills as structured markdown files. Starting from elementary skills (web search, terminal), the agent continuously rewrites its own executable tools, achieving 26.2% and 116.2% relative accuracy improvements on GAIA and Humanity's Last Exam benchmarks respectively — all without updating LLM parameters. For production agent builders, this validates externalized skill evolution as a viable alternative to fine-tuning.