Memory mechanisms extend large language models beyond fixed limits
Learning to Remember: Advances in Memory Mechanisms for Large Language Models
Information Retrieval
Summary
Large language models usually can only remember and use information from a limited recent context, which restricts their ability to have long, continuous interactions or remember facts learned earlier. The authors survey different ways to add memory mechanisms to these models, allowing them to store, maintain, and retrieve information over time. They organize these by where the memory lives (inside or outside the model), how it changes over time, and how it helps agents learn from experience. This helps build smarter and more adaptable AI that can remember and learn continuously.
What this means in practice
- •For software developers: Design language model applications that remember user interactions over time by choosing appropriate memory mechanisms from this comprehensive memory framework.
- •For ai system architects: Build adaptive AI agents that persist and learn through experiences by incorporating memory techniques outlined for agent evolution and governance.
A survey. It maps existing work.
Authors
Chenye Han, Hongche Yu, Mengyue Wu, Dell Zhang, Jixiang Luo
Abstract
Memory has become a central mechanism for extending large language models (LLMs) beyond fixed context windows, stateless interaction, and static parametric knowledge. This survey organizes the literature around three complementary dimensions: the life cycle of memory, the storage substrate, and the system role that memory plays in LLM-based agents, while using cognitive-science-inspired memory concepts as a lightweight conceptual motivation rather than as a separate technical contribution or implementation family. On this basis, we distinguish three mechanism families: nonparametric memory, where information is stored outside model parameters in text, graph, multimodal, or database-like structures; parametric memory, where information is stored or accessed through activations, key-value caches, learned modules, adapter-based parameterizations, or model weights; and memory for agent evolution, where stored experiences support reinforcement learning, continual learning, procedural reuse, context engineering, and agent harness design. Evaluation is treated as a separate cross-cutting component rather than as a memory mechanism. Unlike surveys that focus primarily on retrieval-augmented generation or long-context modeling, this paper emphasizes the full memory life cycle: construction, maintenance, retrieval, consolidation, forgetting, governance, and evaluation. By aligning conceptual motivations, computational memory mechanisms, benchmarks, and deployment concerns, the survey provides a structured roadmap for designing memory-augmented LLM agents that are persistent, adaptive, auditable, and safe to update over time.