WikiSkill lifts agent scores by 18.6 points

Google Research’s WikiSkill lifted Gemini 3.5 Flash’s average score from 49.5% to 68.1% across five agent benchmarks, an 18.6-point gain. The paper separates raw execution traces, accumulated knowledge, and rollbackable skills; its reported results show smaller models with persistent skills can close the gap with larger models, so builders should test durable skill memory alongside cloud-session persistence economics.