A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle
arXiv:2604.16548v2 Announce Type: replace-cross Abstract: The emergence of writable, cross-session persistent memory in LLM agents introduces a qualitatively different threat landscape from conventional input-centric security concerns, characterized by three properties: persistence, statefulness, and propagation. To systematically characterize this landscape, we propose a Memory Lifecycle Framework that organizes attacks, defenses, and their cross-phase dependencies along two axes: six lifecycle phases (Write, Store, Retrieve, Execute, Share & Propagate, Forget & Rollback) and four security ob
The rapid advancement and deployment of LLM agents with persistent memory are highlighting new security vulnerabilities that require immediate attention and systematic analysis.
This survey identifies a critical, emerging threat landscape for AI agents, moving beyond conventional input-centric security to address persistent memory, statefulness, and propagation issues.
The focus of AI security shifts to a 'Memory Lifecycle Framework,' emphasizing the need for comprehensive defenses across different phases of memory interaction and away from purely input-output security.
- · AI cybersecurity firms
- · Developers of robust LLM agent security protocols
- · Entities investing in advanced AI governance frameworks
- · LLM agents with unaddressed memory vulnerabilities
- · Organizations relying on conventional cybersecurity methods for AI agents
Increased investment in research and development for LLM agent memory security.
New regulatory standards and compliance requirements emerge for AI agent deployment, focusing on memory integrity and data handling.
The development of 'memory firewalls' or 'memory-aware' operating systems specifically designed for autonomous AI agents.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL