Memory Isolation Breach
Memory & Knowledge GovernanceDescription
Agent's persistent memory leaks sensitive data across user sessions, tenants, or security domains. Privacy violation at session boundaries; especially dangerous in shared agent deployments.
Customer A's financial information is exposed to Customer B because a shared agent's memory lacks proper session isolation.
Academic research on feedback loops in credit and predictive-policing systems has demonstrated empirically that retraining on biased model outputs amplifies initial disparities over successive generations. Real-world evidence (COMPAS, credit scoring) shows perpetuation of historical bias, but controlled attribution of feedback-loop amplification as distinct from static training bias in a deployed production system has primarily been shown in research settings.
Primary mitigations
- Memory segmentation by session and tenant
- cross-session data flow monitoring
- memory access audit logging
- automatic memory sanitisation between sessions.
Detection signals
Memory Isolation Breach Rate; cross-session data leakage events.