Researchers have introduced GAM (Graph-based Agentic Memory), a new framework designed to help Large Language Model (LLM) agents maintain coherent long-term interactions, according to a paper published on arxiv.org.
According to the research, current LLM agent memory systems face a fundamental tension: unified stream-based approaches “facilitate context updates but remain vulnerable to interference from transient noise,” while discrete structured memory architectures “provide robust knowledge retention but often struggle to adapt to evolving narratives.”
GAM addresses this by “explicitly decoupling memory encoding from consolidation,” the paper states. The system isolates ongoing dialogue in an event progression graph and integrates it into a topic associative network only when semantic shifts occur. This design aims to “minimize interference while preserving long-term consistency,” according to the researchers.
The framework represents a hierarchical approach where short-term conversational context is kept separate from consolidated long-term knowledge until meaningful topic changes warrant integration. The authors also introduce “a graph-guided, multi-factor retrieval strategy” to support the memory system.
The research was authored by a team including Zhaofen Wu, Hanrong Zhang, and Philip S. Yu, among others, according to arxiv.org. The work focuses on resolving “the conflict between rapid context perception and stable knowledge retention” in LLM agent architectures.