Human-inspired Perspectives: A Survey on AI Long-term Memory

TL;DR

This paper argues that AI long-term memory can be effectively designed by grounding it in human memory theory, offering a taxonomy that separates non-parametric (external) and parametric (internal) memory and mapping these to episodic, semantic, and procedural memory. It introduces SALM, a self-adaptive cognitive architecture with adapters to manage storage, retrieval, and forgetting, aiming to surpass the adaptability of current human-like memory systems. The work surveys existing AI memory approaches, identifies key challenges such as retrieval scale, coherence, and forgetting, and outlines concrete next steps, including task-driven measures and multi-modal applications like video understanding and cognitive simulations. By aligning AI memory design with cognitive principles, the paper provides a framework and practical guidance for developing robust, memory-driven AI that can learn, recall, and adapt over long horizons.

Abstract

With the rapid advancement of AI systems, their abilities to store, retrieve, and utilize information over the long term - referred to as long-term memory - have become increasingly significant. These capabilities are crucial for enhancing the performance of AI systems across a wide range of tasks. However, there is currently no comprehensive survey that systematically investigates AI's long-term memory capabilities, formulates a theoretical framework, and inspires the development of next-generation AI long-term memory systems. This paper begins by introducing the mechanisms of human long-term memory, then explores AI long-term memory mechanisms, establishing a mapping between the two. Based on the mapping relationships identified, we extend the current cognitive architectures and propose the Cognitive Architecture of Self-Adaptive Long-term Memory (SALM). SALM provides a theoretical framework for the practice of AI long-term memory and holds potential for guiding the creation of next-generation long-term memory driven AI systems. Finally, we delve into the future directions and application prospects of AI long-term memory.

Figures (8)

• Figure 1: Distributions of the research subjects, publishers and preprint servers of 37 papers related to AI memory Survey.
• Figure 2: Overview of Human Memory Hierarchy and processing. The Sensory Register receives and temporarily stores information; Working Memory stores critical information for immediate tasks; and Long-term Memory stores information relatively permanently.
• Figure 3: Diagram of Storage, Retrieval, and Forgetting of non-parametric memory and parametric memory. In the diagram, Relational Database Updating and Vector Database Updating within the Non-Parametric Memory module can be referenced in Sec. \ref{['sec:AI_memory:non_parametric:memory_storage']}; Rehearsal Enhancement Training, Distance-based Enhancement Training, Sub Network Enhancement Training, Dynamic Network Enhancement Training and Curriculum Enhancement Training within the Parametric Memory module can be referenced in Sec. \ref{['sec:AI_memory:parametric:memory_forgetting']}.
• Figure 4: Taxonomy of AI Long-term Memory and a collection of representative related works. AI long-term memory can be divided into non-parametric memory and parametric memory based on whether it is stored within model parameters. Both of these methods have specific mechanisms for storage, retrieval, and forgetting (Sec. \ref{['sec:founddation_ai_memory']}), and they are strongly related to human episodic, semantic, and procedural memory (Sec. \ref{['sec:hierarchy_ai_memory']}).
• Figure 5: The similarities between different types of AI long-term memory and human long-term memory. Types of Non-Parametric Memory refer to Sec. \ref{['sec:AI_memory:non_parametric:memory_storage']}, and types of Parametric Memory refer to Sec. \ref{['sec:AI_memory:parametric:memory_forgetting']}.