llamaIndex - 💡(How to fix) Fix [Feature Request]: Memory poisoning defense (OWASP ASI06) integration for agent memory stores

Feature Description

OWASP recently published the Top 10 for Agentic Applications, which includes ASI06: Memory Poisoning — a threat specific to AI agents that persist memory across sessions.

LlamaIndex's agent memory stores (ChatMemoryBuffer, VectorMemory, SimpleComposableMemory) are a prime target for this attack. A malicious user can craft inputs that get stored as "trusted" memories, which then influence future agent behavior — leaking data, taking unauthorized actions, or hijacking tool calls across sessions.

The OWASP reference implementation for ASI06 is agent-memory-guard (Apache-2.0, pure Python, zero external dependencies):

pip install agent-memory-guard

Proposed integration: A GuardedMemory wrapper that implements LlamaIndex's BaseMemory interface, screening all put() and get() calls through the OWASP defense pipeline (prompt injection detection, integrity verification, policy enforcement).

Reason

LlamaIndex currently has no built-in defense for memory poisoning attacks. As agent use cases grow (especially multi-session, multi-user deployments), this becomes a critical security gap. The OWASP ASI06 reference implementation provides a drop-in solution.

Value of Feature

Provides OWASP-compliant memory security for LlamaIndex agents. Addresses a formally recognized threat category (ASI06) with a production-ready, zero-dependency implementation. Enables enterprise adoption of LlamaIndex agents where security compliance is required.