The AI Hippocampus: How Far are We From Human Memory?

Memory plays a foundational role in augmenting the reasoning, adaptability, and contextual fidelity of modern Large Language Models (LLMs) and Multi-Modal LLMs (MLLMs). As these models transition from static predictors to interactive systems capable of continual learning and personalized inference, the incorporation of memory mechanisms has emerged as a central theme in their architectural and functional evolution. This survey presents a comprehensive and structured synthesis of memory in LLMs and MLLMs, organizing the literature into a cohesive taxonomy comprising implicit, explicit, and agentic memory paradigms. Specifically, the survey delineates three primary memory frameworks. Implicit memory refers to the knowledge embedded within the internal parameters of pre-trained transformers, encompassing their capacity for memorization, associative retrieval, and contextual reasoning. Explicit memory involves external storage and retrieval components designed to augment model outputs with dynamic, queryable knowledge representations—such as textual corpora, dense vectors, and graph-based structures—thereby enabling scalable and updatable interaction with information sources. Agentic memory introduces persistent, temporally extended memory structures within autonomous agents, facilitating long-term planning, self-consistency, and collaborative behavior in multi-agent systems, with relevance to embodied and interactive AI.

The architecture of memory in modern (M)LLMs is increasingly analogous to the synergistic relationship between different human brain systems, particularly the neocortex, the hippocampus, and the prefrontal cortex. This brain-inspired framework, which echoes the principles of Complementary Learning Systems theory, provides a powerful lens through which to understand the different memory paradigms evolving in AI. Implicit Memory: The Neocortex. We conceptualize the model's internal parameters as its digital neocortex. In the brain, the neocortex is the primary repository for long-term semantic knowledge, skills, and consolidated memories, which are learned slowly and stored in a distributed manner. Similarly, a transformer's weights embody the implicit memory of the model—the foundational "world knowledge" acquired during pre-training. Explicit Memory: The Hippocampal System. To access specific, real-time, or episodic information, an AI system requires a mechanism analogous to the hippocampus. The hippocampus is critical for the rapid encoding of new episodic memories (i.e., specific events and their context) and acts as an index that binds together disparate elements of an experience stored across the neocortex. Explicit memory systems in AI, such as Retrieval-Augmented Generation (RAG), mimic this function. Agentic Memory: The Prefrontal Cortex. The functionality of agentic memory is best analogized to the prefrontal cortex (PFC), the brain's executive control center. The PFC is responsible for working memory, goal-directed planning, and integrating information from both long-term stores and recent episodic memories to guide behavior. Agentic memory systems similarly maintain a persistent state across interactions, manage working memory (e.g., a scratchpad), and orchestrate the strategic retrieval and use of both implicit and explicit memory to formulate plans and execute complex tasks.

In this report, we present a narrative review of three distinct types of memory integrated into large language models: implicit memory, which is embedded within model parameters; explicit memory, which involves external storage and retrieval mechanisms; and agent memory, which captures persistent interactions with environments. Additionally, we systematically examine memory mechanisms specifically designed for and utilized by multimodal LLMs. Our survey is meticulously structured to trace the developmental trajectory of memory mechanisms, spanning from foundational concepts to the most recent advancements. Based on this comprehensive review, there is a critical need to advance our understanding of the internal mechanisms of Transformer architectures and to develop more effective frameworks for implicit memory modeling. Enhancing the long-context processing capabilities of LLMs, either through extended context windows or RAG, is essential; however, each approach presents trade-offs in terms of computational efficiency and scalability. Dynamic memory adaptation, inspired by human learning strategies such as recursive retrieval and experience reflection, holds promise for improving reasoning and communication in agent-based systems.