From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms
:::info Stub — Full Engineering Breakdown Coming This paper has a linked code implementation and was featured on Hugging Face Papers with 3 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Jinghao Luo et al. |
| Year | 2026 |
| HF Upvotes | 3 |
| arXiv | 2605.06716 |
| Download | |
| Code | https://github.com/FeishuLuo/Evolving-LLM-Agent-Memory-Survey |
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.
Engineering Breakdown
Plain English
This survey formalizes how LLM-based agents manage memory by proposing a three-stage evolutionary framework: Storage (keeping trajectory records), Reflection (refining those records), and Experience (abstracting them into reusable knowledge). The authors argue that existing memory research is fragmented between systems engineering and cognitive science perspectives, and this framework provides a unified lens to understand how memory mechanisms have developed and should be designed.
Key Engineering Insight
Memory in LLM agents isn't just about storing conversation history—it's about a progression from raw storage to intelligent abstraction. The framework suggests that production systems need all three stages (preservation → refinement → abstraction) to avoid token bloat and hallucination, not just one or two.
Why It Matters for Engineers
As LLM agents handle longer tasks and multi-turn interactions, naive storage strategies blow up context windows and inference costs. This framework helps engineers decide what to keep, when to compress, and how to extract generalizable patterns—directly impacting latency, cost, and reliability in production systems handling complex workflows.
Research Context
Prior work treated memory as either a storage problem (how to organize data) or a cognitive problem (how humans remember), creating siloed solutions. This survey bridges that gap by showing memory evolution as a continuous design spectrum, enabling engineers to make principled trade-offs rather than copy-pasting architectures from papers that solve different downstream tasks.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::