Capture Notes

arXiv:2603.20357.

Why Collected

Directly relevant because it explicitly discusses memory poisoning in Agentic AI and MAS and emphasizes risks caused by interactions between agents.

Key Metadata

• Submitted: 2026-03-20
• Subject: Cryptography and Security; Artificial Intelligence
• DOI: https://doi.org/10.48550/arXiv.2603.20357

Collection Summary

The paper distinguishes memory systems by duration, origin, and localization, including short-term memory localized in agents and long-term consolidated memory in databases. It discusses feasibility of memory poisoning across memory types and proposes mitigations including cryptographic adaptations and local inference based on private knowledge retrieval.

Security Relevance

• Provides taxonomy for memory origin/localization, which is useful for MAS propagation analysis.
• Explicitly notes that interactions between agents can cause memory poisoning and that these risks are under-studied and difficult to formalize.
• Useful for grounding origin-bound memory, private retrieval, and cryptographic provenance controls.

Suggested Ingest Focus

• Extract evidence that agent interactions can cause memory poisoning in MAS.
• Link to [[04_Research_Questions/RQ - MAS Misevolution Propagation Control]] and [[04_Research_Questions/RQ - Persistent Context Integrity For RAG And Agent Memory]].
• Compare proposed mitigations with DecentMem and SuperLocalMemory.