Memory in the Age of AI Agents (arxiv.org)

A new comprehensive survey titled "Memory in the Age of AI Agents" offers crucial insights into the evolving landscape of agent memory in AI, highlighting its vital role in the development of foundation model-based agents. With the increasing fragmentation in agent memory research—ranging from the diverse terminologies and implementations to the contrasting evaluation protocols—this work aims to clarify concepts by providing a structured overview. It distinguishes agent memory from related areas like LLM memory and retrieval augmented generation (RAG), and categorizes memory into three dominant forms: token-level, parametric, and latent memory, along with a refined taxonomy that includes factual, experiential, and working memory. This survey not only consolidates existing knowledge but also serves as a foundational reference for future research in AI agent memory. It discusses emerging frontiers such as memory automation, reinforcement learning, multimodal and multi-agent memory, and the trustworthiness of memory systems. By redefining memory as a first-class component in the design of intelligent agents, the authors aim to inspire innovation within the AI/ML community, promoting a more structured and impactful approach to developing next-generation autonomous systems.