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

Researchers propose a unified evolutionary framework for LLM agent memory systems, categorizing development into three stages: Storage, Reflection, and Experience. The framework addresses fragmented research by synthesizing engineering and cognitive science perspectives, offering design principles for building more capable autonomous AI agents.

This arXiv paper addresses a critical architectural gap in LLM agent development by formalizing how memory systems should evolve. Rather than treating memory as a technical implementation detail, the authors position it as the foundational element that determines agent capability and autonomy. The three-stage framework—from raw trajectory preservation through refinement to abstraction—mirrors how biological systems and software engineering have historically approached information organization.

The research gains significance because current LLM agents struggle with consistency over extended interactions and fail to transfer knowledge across contexts. By framing memory evolution as a unified progression rather than isolated components, the survey provides theoretical scaffolding that practitioners lack. This matters beyond academic circles; developers building production AI agents face concrete challenges around context window limitations, hallucination, and inability to learn from past mistakes.

The two mechanisms highlighted in the frontier "Experience" stage—proactive exploration and cross-trajectory abstraction—suggest that next-generation agents will autonomously discover useful patterns and generalize insights across multiple problem instances. This capability leap would fundamentally change how AI agents scale from narrow task execution to general problem-solving.

For the AI industry, this framework establishes intellectual foundations that could guide the next development cycle. Organizations investing in agent infrastructure now have a theoretical roadmap for prioritizing which memory capabilities to implement first. The emphasis on continual learning as an end-goal signals that stateless, one-off LLM calls are becoming obsolete—future commercial systems will require persistent, adaptive memory architectures.

• →LLM agent memory evolves through three stages: Storage (preservation), Reflection (refinement), and Experience (abstraction).
• →Memory mechanisms address three core challenges: long-range consistency, dynamic environment adaptation, and continual learning.
• →Proactive exploration and cross-trajectory abstraction represent transformative mechanisms for next-generation agents.
• →The framework unifies fragmented research between engineering and cognitive science perspectives.
• →Persistent, adaptive memory architectures will become essential for commercial AI systems moving beyond stateless interactions.