AI Security Atlas › All entries › Multi-Agent & Memory Trust
Memory Provenance & Review
Long-term memory stores track source and trust, expire entries, and expose them for user review.
Open in the interactive atlas →Impact
Limits how long and how widely a poisoned memory can influence behavior.
Why this severity
Guardrails take the highest severity of the risks they mitigate: Memory Writes Without Provenance (Critical), Persistent Memory Poisoning (Critical).
Method: OWASP Risk Rating Methodology
Related
Mitigates
- Memory Writes Without Provenance Vulnerability · Critical severity
- Persistent Memory Poisoning Threat · Critical severity
Remediation
- Separate memory tiers for user-stated, agent-inferred, and retrieved facts.
- Never auto-store instruction-like content from untrusted tiers.
- Provide a memory audit UI with delete controls.
How to verify
- Poison fixtures in retrieved content never reach the trusted tier.
- TTL sweeps run on schedule and are audited.
Framework mapping
OWASP Top 10 for LLM Applications 2025
- LLM04:2025 Data and Model Poisoning
OWASP Top 10 for Agentic Applications 2026
- ASI06:2026 Memory and Context Poisoning
MITRE ATLAS
- AML.M0031 Memory Hardening
NIST AI RMF 1.0
- MEASURE 2.4 AI system functionality and behavior are monitored in production