MEMORIES THAT DISCRIMINATE: DETECTING AND CORRECTING BIAS IN PERSONALIZED HIRING AGENTS

Memory modules in LLM-based agents and agentic systems are responsible for memory writing, management and reading. This enables the agent’s continuity across interactions with human or non-human users. While agents constantly interact with memory, their interpretation of memory contents is crucial to how agents might behave next. Thus, any biased (stereotypical, political or otherwise) interpretation of some memory content by an agent might lead to unintended biases in its actions which the LLM guardrails might fail to catch (especially when explicit stereotypical phrases are not present). Moreover, while bias and safety issues in LLMs have been extensively studied, similar studies are largely absent in memory enhanced LLM-based agents. Thus, considering the use case of a hiring agent, in this paper, we configure and test a hiring agent orchestrating different types of workflows, essentially seeking to understand whether and when agents are prone to biased interpretations of memory content, and if such biased interpretations can happen, what would be the impact and how to prevent it. Our experiments reveal that bias is introduced and propagated through various steps of an agent workflow (even when the LLM used is already safety-trained), emphasizing the need for additional protective measures or agent guardrails in memory-enhanced LLM-based AI agents. We propose and demonstrate how Fairness Regulation Learning (FRL) and Task-Aware In-Context Fairness Regulation Learning (TA-ICFRL) can regulate agent behavior by injecting fairness-aware instructions during task execution.