ACL 2026 paper maps LLM agent memory: Storage → Reflection → Experience

A new survey paper accepted at ACL 2026 Findings introduces a structured framework for the evolution of memory mechanisms in LLM-based agents. Authored by Jinghao Luo, Yuchen Tian, and seven others, the work categorizes memory development into three progressive stages: Storage, Reflection, and Experience. In the Storage stage, agents simply preserve trajectories of past actions without modification. The Reflection stage introduces refinement—agents revisit and improve those trajectories. Finally, the Experience stage moves to abstraction, where agents extract generalizable knowledge from multiple trajectories, enabling proactive exploration and cross-trajectory learning. The survey identifies three key drivers behind this evolution: the need for long-range consistency in complex tasks, the challenges of operating in dynamic environments, and the ultimate goal of continual learning without catastrophic forgetting.

The paper goes beyond simple categorization to explore transformative mechanisms in the frontier Experience stage. Notably, proactive exploration allows agents to seek out new information intentionally, while cross-trajectory abstraction enables them to learn patterns across multiple episodes of interaction. By synthesizing insights from operating system engineering and cognitive science, the authors bridge a theoretical divide that has fragmented previous research. The survey provides concrete design principles and a clear roadmap for building next-generation LLM agents that can adapt and learn autonomously. This work is particularly relevant given the rapid adoption of agent-based systems in production environments, where memory reliability is a key bottleneck.