Chaulagain, Gaurav and Upadhyaya, Medhavee (2023) Persistent Memory Poisoning in Long-Lived AI Agents: A Threat Analysis and Trust-Aware Memory Defense Framework. American Journal of Engineering , Mechanics and Architecture (2993-2637), 1 (4). ISSN 2993-2637

Text AJEMA_Persistent+Memory+Poisoning.pdf Download (601kB)

Abstract

Many Artificial Intelligent Model (AI) based agents are using the persistent memory for maintaining the context information, remembering the user preferences and streamlining the performance of the tasks in multiple interactions with the user. This capability enables long-term personalization, but also creates a security risk because the agent's behavior could be affected by information from another interaction, albeit from an adversarial source. Persistent memory poisoning can be a more comprehensive form of attack in which the malicious information can still have a significant effect even though the original source is no longer in existence. In this paper, persistent memory poisoning in long-lived AI agents is explored, and a trust-aware defense framework called M-Guard is introduced, which incorporates provenance tracking, memory-write validation, trust-aware retrieval, temporal revalidation, and execution-time verification. The proposed research methodology is based on a series of controlled scenarios that include direct memory injection, misleading historical information, multi-record contamination, and delayed trigger-based behavior. A simulation-based evaluation is intended to compare the performance of the following: unrestricted memory, prompt-level filtering, write-time validation, retrieval-time screening, and the full system. The simulated "placeholders" in the example results provided in this draft are not empirical results. They show how to systematically measure the effectiveness of security as well as the level of security that can be achieved in completing a legitimate task, the false-positive rate, and how much memory is retained. It is the study's contention that persistent memory must be considered as being in a state of security, not automatically assumed to be a trusted repository of knowledge. The proposed framework provides a basis for reproducible experimentation and subsequent validation with real AI agents.

Item Type:	Article
Uncontrolled Keywords:	Persistent memory poisoning, artificial intelligent Model agents, AI security, long-term memory, prompt injection, trust-aware retrieval, provenance tracking, memory integrity
Subjects:	H Social Sciences
Depositing User:	admin eprints
Date Deposited:	29 Oct 2026 15:13
Last Modified:	29 Oct 2026 15:13
URI:	http://eprints.umsida.ac.id/id/eprint/17309

Actions (login required)

View Item