Large language models (LLMs) frequently hallucinate factually incorrect information. Developers traditionally address this by assuming missing knowledge and increasing model size or data. However, new research suggests LLMs often possess the facts but fail to recall them during generation. Frontier models encode a high percentage of facts, indicating recall is the primary bottleneck.The study proposes "knowledge profiling" to distinguish between encoded and known facts. Encoding means the model can reproduce a fact under specific conditions, while knowing means it can reliably answer questions about it across various phrasings. Encoding failures require pre-training interventions, while recall failures benefit from post-training techniques.A fact can be directly recalled, fail encoding, experience recall failure, be recalled with thinking, or be inferred without encoding. Experiments on numerous LLMs show frontier models encode most facts but struggle with direct recall. Inference-time thinking, like Chain-of-Thought, aids recall significantly, akin to human memory retrieval.Scaling models doesn't solve recall issues and can even exacerbate them by increasing inaccessible encoded facts. Recall is heavily influenced by query phrasing and context, with rare or reverse facts posing greater recall challenges. Developers should avoid treating all factual errors as retrieval problems and instead focus on improving recall.Selective use of inference-time reasoning and generate-then-verify pipelines can enhance factual accuracy. Testing semantic access beyond standard accuracy metrics reveals true model knowledge. Query reformulation and retries are also effective. While the WikiProfile benchmark focuses on encyclopedic facts, its methodology can be applied to domain-specific data, though encoding may be a more significant issue in specialized areas. This research shifts focus to post-training and inference-time optimizations, making improved AI factuality more accessible.