OWASP Agent Memory Guard: runtime screening for agent memory reads and writes

Source Summary

What It Contains

OWASP Agent Memory Guard is an incubator project and reference implementation for screening AI-agent memory reads and writes against prompt injection, protected-key tampering, secret leakage, size anomalies, and rapid-change patterns.

Extracted Claims

• Persistent agent memory, including RAG indexes, chat history, scratchpads, and vector stores, can become a privileged input if untrusted content is written into it.
• Memory poisoning differs from front-door prompt injection because the malicious content can survive across sessions and steer later tool use or data disclosure.
• The project wraps agent memory stores and applies policy actions such as allow, redact, quarantine, or block.
• It reports benchmark coverage across prompt injection, protected-key tampering, sensitive-data leakage, and size-anomaly categories, with local operation and no external API dependency.
• Integrations and recipes are documented for LangChain, OpenAI Agents SDK, AutoGen, mem0, and other memory-store-compatible frameworks.

Evidence Quality

Primary source — public OWASP GitHub repository. The repository was updated on 2026-05-31 and surfaced in a 2026-06-01 Hacker News Show HN item, but the control should be treated as defensive guidance and implementation evidence rather than proof of a specific exploited incident.

Follow-Up

• Track whether OWASP publishes additional guidance for ASI06 memory poisoning in agentic applications.
• Consider whether Armorer health checks should flag persistent-memory-enabled profiles that lack explicit memory scope, policy logging, or review gates.
• Monitor package releases and integration support before recommending production deployment of this specific library.