@Sa4d_k1: Among the distinguished papers recently published in the field of AI Agents. The research involved 60 researchers from …

Among the distinguished papers recently published in the field of AI Agents. The research involved 60 researchers from prestigious institutions such as Stanford, Harvard, Google, Meta, and Salesforce, in addition to numerous other universities and research companies. It discusses one of the most important problems that has begun to emerge with the rapid and massive development in AI agents: How can the Agent work with you for days or weeks without forgetting what happened previously? For the Agent, during its operation, continuously gathers new information: conversations, decisions, previous errors, tool results, user preferences, and learning experiences from them. And placing all this information inside the context window (Context Window) is not a scalable solution, because the context size continues to grow with the length of the task and conversation—this is the problem the paper describes as Context Explosion. And here comes the importance of Agent Memory. The paper is titled: A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agents The paper attempts to organize the entire field of Agent Memory and explains memory from three main aspects: • Where is the memory stored? Inside the model itself (Internal Memory) or outside it in systems from which information can be retrieved (External Memory). • What type of information does it remember? Such as working memory (Working Memory), event and experience memory (Episodic Memory), semantic memory (Semantic Memory), and skill memory (Procedural Memory). • For whom is this memory? Is it for understanding the user and their preferences (User-Centric Memory)? Or for saving the Agent's own experiences, errors, and what it has learned from them (Agent-Centric Memory)? But the most striking idea in the research is that memory is no longer just a place for storing and retrieving information. Rather, it has become part of the way the Agent itself can evolve. It executes a task, saves the experience, extracts new knowledge or skill from it, then reuses it in the future to become better over time. And here we begin the transition from Agents that start almost from scratch in every task, to self-evolving agents (Self-Evolving Agents) capable of learning from their experiences and working for longer periods. The paper is extremely comprehensive and also covers memory management, Multi-Agent Memory, learning preservation and retrieval policies, Reinforcement Learning, Benchmarks, and even the idea of skills that Agents can learn, share, and reuse. For anyone interested in actually building AI Agents, this is one of the papers that provides an excellent overview of the entire field, not just one Memory approach. The Research: