StructMem: Structured Memory for Long-Horizon Behavior in LLMs
A novel hierarchical memory model named StructMem has been introduced to enhance long-term conversational agents. This framework focuses on the connections between events rather than merely isolated facts, facilitating temporal reasoning and multi-hop question answering. Existing methods encounter a dilemma: flat memory is efficient yet lacks relational modeling, while graph-based memory supports structured reasoning but is costly and fragile. StructMem maintains event-level bindings and establishes cross-event links by temporally anchoring dual viewpoints and conducting regular semantic consolidation. It demonstrates improved temporal reasoning and multi-hop capabilities on the LoCoMo benchmark, all while decreasing token usage, API calls, and runtime in comparison to previous memory systems. The research can be found on arXiv.
Key facts
- StructMem is a structure-enriched hierarchical memory framework for LLMs.
- It addresses the trade-off between flat memory and graph-based memory.
- It preserves event-level bindings and induces cross-event connections.
- It uses temporally anchored dual perspectives and periodic semantic consolidation.
- It improves performance on the LoCoMo benchmark.
- It reduces token usage, API calls, and runtime.
- The paper is from the Computer Science > Computation and Language category.
- The arXiv ID is 2604.21748.
Entities
Institutions
- arXiv