Search the Atlas

Search risks, controls, and glossary terms

CriticalAgenticMemory PoisoningRealized

Persistent Memory Corruption

Memory & Knowledge Governance

Description

Malicious or erroneous data injected into agent memory stores persists across sessions, corrupting future behaviour. Long-term behavioural manipulation through persistent memory; poisoned memory citations self-reinforce.

Example scenario

Adversary injects false safety approval records into agent memory; agent subsequently bypasses safety checks believing prior approval exists.

Real-world evidenceRealized

Multiple confirmed production incidents establish systematic demographic bias in AI-assisted lending and underwriting: the Apple Card / Goldman Sachs gender-bias investigation by the NY Department of Financial Services (2019), the HUD complaint against Facebook's ad-targeting algorithm for housing discrimination, and regulatory actions against multiple fintech lenders confirm this risk has materialized in deployed systems.

Primary mitigations

  • Memory integrity validation
  • write-access controls on persistent memory
  • anomaly detection in memory update patterns
  • memory source verification.

Detection signals

Memory Integrity Score; memory poisoning detection rate; unexpected behavioural shifts post-memory-update.

Mitigating controls

5
Dual coverage

Related risks in Memory & Knowledge Governance