RuleMem: Active Rule Memory for Long-Term Conversational Agents

Overview

RuleMem equips conversational agents with reusable rules learned from earlier interactions. It expresses these rules as natural-language Horn clauses and checks them using Rule Perplexity Consistency (RPC). The resulting memory helps retrieve evidence that may be distant in wording and provides a logical structure for producing answers.

The framework is evaluated on LoCoMo and LongMemEval_s*, with comparisons to existing conversational memory approaches. Its central goal is to make past interactions useful for both retrieval and reasoning over long dialogue histories.

Key Contributions

• A rule-based memory representation derived from conversation history.
• RPC-based validation of induced rules.
• Rule-guided evidence retrieval and answer generation, evaluated on two long-term conversational benchmarks.

Publication Details

• Venue: arXiv preprint
• Year: 2026
• Submitted: September 3, 2026
• arXiv: 2609.03915
• Paper: PDF

Authors

Xingyuan Zeng, Zuohan Wu, Yue Wang, Chen Zhang, Quanming Yao, Wei Liu, Jiuke Wang, Libin Zheng, Jian Yin

BibTeX

@misc{zeng2026rulemem,
  author        = {Xingyuan Zeng and
                   Zuohan Wu and
                   Yue Wang and
                   Chen Zhang and
                   Quanming Yao and
                   Wei Liu and
                   Jiuke Wang and
                   Libin Zheng and
                   Jian Yin},
  title         = {RuleMem: Active Rule Memory for Long-Term Conversational Agents},
  year          = {2026},
  eprint        = {2609.03915},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.03915}
}