Presentation Information
[4Yin-A-62]Memory Governance for Cross-Session Prohibition Retention
〇Yu Amano1 (1. Activarch Consulting)
Keywords:
Memory governance,Memory-augmented dialogue evaluation,Adjacent-session retention,Naive RAG baseline comparison
Large language models often lose prohibition compliance across sessions. This study evaluates whether memory governance improves prohibition retention using a LoCoMo-inspired Japanese dataset. In the directly comparable P1 setting, retrieval-based augmentation reduced violations versus baseline, and MCP-governed memory showed the lowest observed violation rate. In P2 with distractors, non-MCP and MCP settings were analyzed separately because implementations differ; thus, the main evidence comes from P1, while P2 is reported as supplementary exploratory analysis of trends. Future work focuses on unified-condition attribution and broader validity checks.