Long-term conversational agents store past dialogue as facts, but retrieval based on surface similarity misses evidence that is logically relevant yet semantically distant. RuleMem instead induces reusable logical rules, represented as natural-language Horn clauses, from historical interactions and validates them with a Rule Perplexity Consistency (RPC) mechanism. At query time the rules act as retrieval cues and as explicit logical premises for reasoning.
RuleMem was evaluated on LoCoMo and LongMemEval_s*. Against 14 baselines on LoCoMo it achieved the highest accuracy, exceeding the baseline average by 27.47 points, a 54.3% relative improvement. This is interesting because it shifts memory from passive storage to active reasoning support, addressing both retrieval gaps and reasoning failures.