Aktualności10 października 2026 WikiSkill: agents log their own failures in a wiki, not the prompt
Google Research and Virginia Tech split an agent's failure memory from its prompt. The largest model tested gained 23.9 points of average accuracy, and skills transfer across model families.
