This paper provides a comprehensive review of knowledge mechanisms in Large Language Models (LLMs), focusing on knowledge utilization and evolution. The authors propose a novel taxonomy that includes knowledge utilization at a specific time and knowledge evolution across all periods of LLMs. They delve into the mechanisms of memorization, comprehension, application, and creation, as well as the dynamic progression of knowledge within individual and group LLMs. The paper discusses the limitations of parametric knowledge, such as fragility and hallucinations, and explores potential dark knowledge that remains unknown to both humans and models. The authors also propose strategies for constructing more efficient and trustworthy LLMs, including memory circuits, knowledge editing, and model merging. They highlight the importance of addressing conflicts and integration in knowledge evolution and discuss open questions and future directions in the field. The paper aims to provide insights for advancing the understanding and application of knowledge mechanisms in LLMs.This paper provides a comprehensive review of knowledge mechanisms in Large Language Models (LLMs), focusing on knowledge utilization and evolution. The authors propose a novel taxonomy that includes knowledge utilization at a specific time and knowledge evolution across all periods of LLMs. They delve into the mechanisms of memorization, comprehension, application, and creation, as well as the dynamic progression of knowledge within individual and group LLMs. The paper discusses the limitations of parametric knowledge, such as fragility and hallucinations, and explores potential dark knowledge that remains unknown to both humans and models. The authors also propose strategies for constructing more efficient and trustworthy LLMs, including memory circuits, knowledge editing, and model merging. They highlight the importance of addressing conflicts and integration in knowledge evolution and discuss open questions and future directions in the field. The paper aims to provide insights for advancing the understanding and application of knowledge mechanisms in LLMs.