Detection and security caveats
Shared by Pavel · Updated 2026-11-16
The guide proposes SOC monitoring ideas for agent memory, but describes them as untested and dependent on logging memory writes and their sources.
Detecting attacks
- Log each memory write with its source so investigators can trace suspicious entries.
- Watch for cross-user or cross-agent leakage through tenant isolation, branch-scoped sharing, and logged reads of shared branches.
- Enforce memory scope in retrieval at read time; do not rely on the model to respect permissions. Agent identity matters because one user’s agents may have different access.
- Treat suspicious memory content and retrieval as attack signals: memory poisoning plants false or malicious content for later retrieval and action, while indirect prompt injection can hide instructions in retrieved pages, email, or memory.
Shared-memory limits and privacy risks
- Shared memory can let one agent read information it should not; cross-agent and cross-user leakage is also called AgentLeak.
- Shared-memory access control is not guaranteed by organizing data into groups: Graphiti’s reference servers allow callers to read, change, or delete data in any group.
- Multi-agent shared memory is a leakage channel; AgentLeak’s coordinator-worker tests found internal messages leaked far more than final outputs.
- Conversational memory contains personal data, so deletion and retention obligations apply to both raw and derived copies.
- A memory store can expose information to anyone able to phrase the right query, making access control and scoping as important as deletion.