Structured Episodic Event Memory
Zhengxuan Lu, Dongfang Li, Yukun Shi, Beilun Wang, Longyue Wang, Baotian Hu
2026-01-13
Summary
This paper introduces a new way for AI agents to remember and use past experiences, aiming to make them better at complex tasks and long-term interactions.
What's the problem?
Current AI systems, when trying to recall information, often grab bits and pieces without understanding how those pieces relate to each other. This is like trying to understand a story by only reading random sentences. For AI agents that need to interact with the world over time, this 'scattered retrieval' makes it hard to maintain a consistent understanding and reason effectively because they lack a structured way to organize memories and understand how events connect.
What's the solution?
The researchers developed a system called Structured Episodic Event Memory, or SEEM. It's designed like a layered memory system. One layer organizes facts and their relationships like a network, while another layer tracks events as they unfold in a story-like way. When the AI needs to remember something, it doesn't just pull up random facts; it reconstructs the context of the event, figuring out what happened and why, using 'provenance pointers' to trace information back to its source. They also created a way for the AI to connect fragmented pieces of information to build a complete picture.
Why it matters?
This research is important because it helps AI agents move beyond simply recalling facts to actually *understanding* experiences. By giving AI a more organized and dynamic memory, it can perform better on tasks that require reasoning, maintaining a consistent narrative, and learning from past interactions, ultimately making them more capable and reliable assistants.
Abstract
Current approaches to memory in Large Language Models (LLMs) predominantly rely on static Retrieval-Augmented Generation (RAG), which often results in scattered retrieval and fails to capture the structural dependencies required for complex reasoning. For autonomous agents, these passive and flat architectures lack the cognitive organization necessary to model the dynamic and associative nature of long-term interaction. To address this, we propose Structured Episodic Event Memory (SEEM), a hierarchical framework that synergizes a graph memory layer for relational facts with a dynamic episodic memory layer for narrative progression. Grounded in cognitive frame theory, SEEM transforms interaction streams into structured Episodic Event Frames (EEFs) anchored by precise provenance pointers. Furthermore, we introduce an agentic associative fusion and Reverse Provenance Expansion (RPE) mechanism to reconstruct coherent narrative contexts from fragmented evidence. Experimental results on the LoCoMo and LongMemEval benchmarks demonstrate that SEEM significantly outperforms baselines, enabling agents to maintain superior narrative coherence and logical consistency.