From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms

Authors: Jinghao Luo, Yuchen Tian, Chuxue Cao et al.

arXiv · PDF

Summary

This survey organizes the scattered literature on LLM agent memory into a three-stage evolutionary framework: Storage (saving interaction traces), Reflection (refining and summarizing those traces), and Experience (abstracting general patterns from trajectories). The authors argue current work is fragmented between engineering and cognitive perspectives, and they propose design principles for building memory systems that enable long-range consistency, adaptation to dynamic environments, and continual learning. They highlight proactive exploration and cross-trajectory abstraction as transformative mechanisms in the Experience stage.

Main takeaways:

• Memory in LLM agents evolves through Storage → Reflection → Experience, moving from raw trajectory logs to abstracted reusable knowledge.
• Three core drivers push this evolution: need for long-range consistency, handling dynamic environments, and achieving continual learning.
• The frontier "Experience" stage involves proactive exploration (trying new things to learn) and cross-trajectory abstraction (generalizing across past episodes).
• Current research is fragmented; this framework offers a unified lens and design roadmap.
• The paper synthesizes engineering and cognitive science views into a coherent evolutionary perspective.

Relevance

Tangential but conceptually interesting for my persona/behavior installation work. If I think of persona markers or conditional behaviors as something the model needs to "remember" across turns, the Storage → Reflection → Experience framing could map onto how implanted behaviors generalize or leak—though the paper focuses on agent episodic memory rather than fine-tuned internal states.

Abstract

arXiv:2605.06716v1 Announce Type: new Abstract: Large Language Model (LLM)-based agents have fundamentally reshaped artificial intelligence by integrating external tools and planning capabilities. While memory mechanisms have emerged as the architectural cornerstone of these systems, current research remains fragmented, oscillating between operating system engineering and cognitive science. This theoretical divide prevents a unified view of technological synthesis and a coherent evolutionary perspective. To bridge this gap, this survey proposes a novel evolutionary framework for LLM agent memory mechanisms, formalizing the development process into three stages: Storage (trajectory preservation), Reflection (trajectory refinement), and Experience (trajectory abstraction). We first formally define these three stages before analyzing the three core drivers of this evolution: the necessity for long-range consistency, the challenges in dynamic environments, and the ultimate goal of continual learning. Furthermore, we specifically explore two transformative mechanisms in the frontier Experience stage: proactive exploration and cross-trajectory abstraction. By synthesizing these disparate views, this work offers robust design principles and a clear roadmap for the development of next-generation LLM agents.