Memory plays a crucial role in AI systems based on LLM, supporting sustained and consistent interactions over time. Although previous surveys have explored memory on LLM, they often lack attention to the fundamental operations governing memory functions. Key components such as memory storage, recovery and generation based on memory have been studied in isolation, but a unified framework which systematically integrates these processes remains underdeveloped. Although some recent efforts have offered operational views of memory to classify existing work, the field still lacks cohesive memory architectures which clearly define how these atomic operations interact.
In addition, existing surveys tend to approach only specific sub-themes in the wider memory landscape, such as long-term manipulation, long-term memory, personalization or knowledge editing. These fragmented approaches often miss essential operations such as indexing and do not offer complete overviews of memory dynamics. In addition, most previous works do not establish a clear research range or do not provide structured references and tool coverage, limiting their practical value to guide future progress in memory for AI systems.
Researchers from the Chinese University, the University of Edinburgh, Hkust and Lab from Huawei UK R&D Ltd. present a detailed investigation into memory in AI systems. They classify memory in parametric types, contextual and contextual structured, distinguishing between short -term and long -term memory inspired by cognitive psychology. Six fundamental operations – consolidation, update, indexing, oblivion, recovery and compression – are defined and mapped in key research areas, in particular long -term memory, long -context modeling, parametric modification and multi -source integration. Based on an analysis of more than 30,000 articles using the relative quotation index, the survey also describes the tools, benchmarks and future orientations.
The researchers first develop a taxonomy in three parts of the IA memory – parametric (weight of the model), contextual structured (for example, indexed dialogue stories) and non -structured contextual (raw or integral text) – and distinguish short -term stresses. They then define six basic memory operations: consolidation (storage of new information), update (modification of existing inputs), indexing (organization for quick access), forget (delete outdated data), recovery (recover the relevant content) and compression (distillation memory). To found this framework, they extracted more than 30,000 high-level AI articles (2022-2025), classified them by a relative quotation index, and the work in high impact cluster in four themes-long-term memory, long-detext modeling, parametric publishing and multiple multiple integration and the tools of cartography and active memory and active research works and Surviving keys and tools.
The study describes an ecosystem in layers of AI systems focused on memory that supports long -term context management, user modeling, knowledge retention and adaptive behavior. This ecosystem is structured on four levels: fundamental components (such as vector stores, large-language models such as Llama and GPT-4, and recovery mechanisms such as Faiss and BM25), frameworks for memory operations (for example, Langchain and Llamaindex), memory layer systems for orchestration and persistence (such as Memary and Memose), and orchestration products (such as Memary and Memose), and orchestration products (notably Memary and Memose), and orchestration products (notably Memary and Memose), and orchestration products (notably Memary and Memose), and orchestration products (especially Memary and Memose), and orchestration products and Memose), and Fayer Facerser (including the Bot. Chatgpt). These tools provide an infrastructure for the integration of memory, activation of capacities such as earthing, searching for similarities, understanding of the long -term context and personalized AI interactions.
The survey also examines open challenges and future research orientations in AI memory. It highlights the importance of spatio-temporal memory, which balances the historical context with real-time updates for adaptive reasoning. The main challenges include parametric recovery of memory, lifelong learning and effective knowledge of knowledge between types of memory. In addition, the article is inspired by models of biological memory, emphasizing double -memory architectures and hierarchical memory structures. Future work should focus on the unification of memory representations, the management of multi-agent memory systems and the fight against safety problems, in particular memory safety and malicious attacks in automatic learning techniques.
Discover the Paper. Also, don't forget to follow us Twitter And join our Telegram And Linkedin Group. Don't forget to join our 90K + ML Subdreddit. For promotion and partnerships, Please talk to us.
Sana Hassan, consulting trainee at Marktechpost and double -degree student at Iit Madras, is passionate about the application of technology and AI to meet the challenges of the real world. With a great interest in solving practical problems, it brings a new perspective to the intersection of AI and real life solutions.
