Frontier models store 95-98% of facts and fail to recall a third of them.

Google Research evaluated 13 LLMs across 4M+ responses on WikiProfile, 2,150 Wikipedia-derived facts tested from exact context completion through multiple choice (Google Research). Gemini-3-Pro and GPT-5 encode 95-98% yet fail to directly recall 26-34%, and extended thinking recovers 40-65% of those. Scaling Gemma3 from 1B to 27B cut encoding failures from 85% to 23% while the recall-failure share rose to a 40% peak without thinking. Scale fixes storage, not access.