StructMem: Structured Memory for Long-Horizon Behavior in LLMs

TL;DR

StructMem introduces a hierarchical memory system for long-horizon LLMs that enhances temporal reasoning and multi-hop question answering while reducing resource usage.

Contribution

It presents a novel structure-enriched hierarchical memory framework that preserves event relationships and improves reasoning efficiency.

Findings

Improves temporal reasoning and multi-hop performance on LoCoMo.

Reduces token usage, API calls, and runtime compared to prior methods.

Supports structured reasoning with a scalable memory architecture.

Abstract

Long-term conversational agents need memory systems that capture relationships between events, not merely isolated facts, to support temporal reasoning and multi-hop question answering. Current approaches face a fundamental trade-off: flat memory is efficient but fails to model relational structure, while graph-based memory enables structured reasoning at the cost of expensive and fragile construction. To address these issues, we propose \textbf{StructMem}, a structure-enriched hierarchical memory framework that preserves event-level bindings and induces cross-event connections. By temporally anchoring dual perspectives and performing periodic semantic consolidation, StructMem improves temporal reasoning and multi-hop performance on \texttt{LoCoMo}, while substantially reducing token usage, API calls, and runtime compared to prior memory systems, see https://github.com/zjunlp/LightMem .