AI大家说 | 忘掉《Her》吧，《记忆碎片》才是LLM Agent的必修课

Core Viewpoint - The article discusses the evolution of AI from chatbots to AI Agents, emphasizing the importance of context engineering in enabling these agents to perform complex tasks effectively [3][5][6]. Group 1: AI Evolution - The narrative of the AI industry has shifted from chatbots to AI Agents, focusing on task decomposition, tool invocation, and autonomous planning by 2025 [3]. - The film "Memento" is suggested as a metaphor for the new Agent era, illustrating how a system can operate in an incomplete information environment to achieve a goal [3][4]. Group 2: Context Engineering - Context engineering is defined as a comprehensive technology stack designed to manage information input and output around the limited attention span of large language models (LLMs) [5][6]. - The success of an AI Agent hinges on providing the right information at each decision point, which is crucial for avoiding chaos [6]. Group 3: Memory Systems in Agents - The protagonist Leonard in "Memento" exemplifies an agent with a clear goal (revenge) and the use of tools (camera, notes) to navigate a complex reality [4][5]. - Leonard's memory system serves as a metaphor for the challenges faced by AI Agents, particularly the need to execute long-term tasks with limited short-term memory [8][9]. Group 4: Three Pillars of Context Engineering - The first pillar is an external knowledge management system, akin to Leonard's use of photographs to capture critical information, which corresponds to retrieval-augmented generation (RAG) in AI [12][14]. - The second pillar involves context extraction and structuring, where information is distilled and organized for efficient retrieval [16][18]. - The third pillar is a layered memory management system, ensuring that agents maintain focus on their core mission while adapting to new information [19][20]. Group 5: Vulnerabilities in Agent Design - The article highlights two critical vulnerabilities in agent design: external poisoning, where agents are fed misleading information, and internal contamination, where agents may misinterpret their own notes [23][24]. - The lack of a verification and reflection mechanism in agents can lead to a cycle of errors, emphasizing the need for systems that can learn from past actions and adjust accordingly [27].