This paper explores the development and potential of Personal LLM Agents, which are LLM-based agents deeply integrated with personal data and devices to provide intelligent assistance. The authors, from various institutions and companies, survey the current state of intelligent personal assistants (IPAs) and highlight the limitations of existing IPAs, particularly in terms of user intent understanding, task planning, and personal data management. They argue that the emergence of large language models (LLMs) offers new opportunities to enhance the capabilities of IPAs. The paper outlines the key components and design choices of Personal LLM Agents, including the foundation model, local resource management, user context and memory management, and skills. It also proposes a taxonomy of five intelligence levels for Personal LLM Agents, ranging from simple step-following to autonomous decision-making. Key challenges in implementing Personal LLM Agents are discussed, focusing on fundamental capabilities, efficiency, and security. The authors review existing solutions and propose potential improvements in areas such as task execution, context sensing, memorization, efficient inference, customization, and memory manipulation. They also address security concerns, including data confidentiality, decision reliability, and system integrity. The paper concludes by summarizing the contributions and future directions, emphasizing the need for further research to address the challenges and realize the full potential of Personal LLM Agents in the AI era.This paper explores the development and potential of Personal LLM Agents, which are LLM-based agents deeply integrated with personal data and devices to provide intelligent assistance. The authors, from various institutions and companies, survey the current state of intelligent personal assistants (IPAs) and highlight the limitations of existing IPAs, particularly in terms of user intent understanding, task planning, and personal data management. They argue that the emergence of large language models (LLMs) offers new opportunities to enhance the capabilities of IPAs. The paper outlines the key components and design choices of Personal LLM Agents, including the foundation model, local resource management, user context and memory management, and skills. It also proposes a taxonomy of five intelligence levels for Personal LLM Agents, ranging from simple step-following to autonomous decision-making. Key challenges in implementing Personal LLM Agents are discussed, focusing on fundamental capabilities, efficiency, and security. The authors review existing solutions and propose potential improvements in areas such as task execution, context sensing, memorization, efficient inference, customization, and memory manipulation. They also address security concerns, including data confidentiality, decision reliability, and system integrity. The paper concludes by summarizing the contributions and future directions, emphasizing the need for further research to address the challenges and realize the full potential of Personal LLM Agents in the AI era.