ERINYS Memory-Governance Agent Uses Qwen to Block Sensitive Data Before AI Responds

A developer built a memory-governance system called ERINYS as part of the Global AI Hackathon Series with Qwen Cloud, demonstrating how unfiltered AI memory can produce unsafe, leaky responses. The project uses synthetic family-care data to show that feeding all stored memories directly into an AI prompt can expose private identifiers and surface outdated or contradictory information. ERINYS addresses this by evaluating each memory across six signals — sensitivity, staleness, conflict, importance, recency, and relevance — before assigning it one of four states: selected, conflicted, demoted, or blocked. In testing, three synthetic private identifiers that leaked in raw-memory mode were fully blocked under ERINYS, never reaching the Qwen language model. The system deliberately uses fixed, deterministic rules rather than AI-based judgment, so every memory decision is auditable and reproducible — a key consideration for sensitive, care-related applications.