Memories That Discriminate: Detecting and Correcting Bias in Personalized Hiring Agents

Himanshu Gharat, Himanshi Agrawal, Gourab K Patro
Proceedings of The Algorithmic Fairness Across Alignment Procedures and Agentic Systems, PMLR 344:34-71, 2026.

Abstract

Memory modules in LLM-based agents enable continuity across interactions with users, especially to deliver personalized experiences. While agents constantly interact with memory, their interpretation of memory contents is crucial to how agents might behave next. Thus, any one-sided interpretation (stereotypical or otherwise) 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. We seek to understand whether and when agents are prone to one-sided interpretations of memory content, and if it can happen, what would be the impact and how to prevent it. Our experiments reveal that one-sided memory misinterpretations can happen based on the context and potential fields (e.g., gender, race, or other attributes of interest), and as a result bias is introduced and propagated through an agent workflow (even when the LLM used is already safety-trained). This leads to bias and prejudice by agents even when it is not intended or preferred by the users, emphasizing the need for additional protective measures. Our proposed Fairness Regulation Learning (FRL) and Task-Aware In-Context Fairness Regulation Learning (TA-ICFRL) reduce such one-sided interpretations by injecting fairness-aware or balancing instructions during task execution.

Cite this Paper

BibTeX
@InProceedings{pmlr-v344-gharat26a, title = {Memories That Discriminate: Detecting and Correcting Bias in Personalized Hiring Agents}, author = {Gharat, Himanshu and Agrawal, Himanshi and Patro, Gourab K}, booktitle = {Proceedings of The Algorithmic Fairness Across Alignment Procedures and Agentic Systems}, pages = {34--71}, year = {2026}, editor = {Tang, Zeyu and Ganesh, Prakhar and Dieng, Awa and Rateike, Miriam and Watson-Daniels, Jamelle and Farnadi, Golnoosh}, volume = {344}, series = {Proceedings of Machine Learning Research}, month = {26 Apr}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v344/main/assets/gharat26a/gharat26a.pdf}, url = {https://proceedings.mlr.press/v344/gharat26a.html}, abstract = {Memory modules in LLM-based agents enable continuity across interactions with users, especially to deliver personalized experiences. While agents constantly interact with memory, their interpretation of memory contents is crucial to how agents might behave next. Thus, any one-sided interpretation (stereotypical or otherwise) 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. We seek to understand whether and when agents are prone to one-sided interpretations of memory content, and if it can happen, what would be the impact and how to prevent it. Our experiments reveal that one-sided memory misinterpretations can happen based on the context and potential fields (e.g., gender, race, or other attributes of interest), and as a result bias is introduced and propagated through an agent workflow (even when the LLM used is already safety-trained). This leads to bias and prejudice by agents even when it is not intended or preferred by the users, emphasizing the need for additional protective measures. Our proposed Fairness Regulation Learning (FRL) and Task-Aware In-Context Fairness Regulation Learning (TA-ICFRL) reduce such one-sided interpretations by injecting fairness-aware or balancing instructions during task execution.} }
Endnote
%0 Conference Paper %T Memories That Discriminate: Detecting and Correcting Bias in Personalized Hiring Agents %A Himanshu Gharat %A Himanshi Agrawal %A Gourab K Patro %B Proceedings of The Algorithmic Fairness Across Alignment Procedures and Agentic Systems %C Proceedings of Machine Learning Research %D 2026 %E Zeyu Tang %E Prakhar Ganesh %E Awa Dieng %E Miriam Rateike %E Jamelle Watson-Daniels %E Golnoosh Farnadi %F pmlr-v344-gharat26a %I PMLR %P 34--71 %U https://proceedings.mlr.press/v344/gharat26a.html %V 344 %X Memory modules in LLM-based agents enable continuity across interactions with users, especially to deliver personalized experiences. While agents constantly interact with memory, their interpretation of memory contents is crucial to how agents might behave next. Thus, any one-sided interpretation (stereotypical or otherwise) 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. We seek to understand whether and when agents are prone to one-sided interpretations of memory content, and if it can happen, what would be the impact and how to prevent it. Our experiments reveal that one-sided memory misinterpretations can happen based on the context and potential fields (e.g., gender, race, or other attributes of interest), and as a result bias is introduced and propagated through an agent workflow (even when the LLM used is already safety-trained). This leads to bias and prejudice by agents even when it is not intended or preferred by the users, emphasizing the need for additional protective measures. Our proposed Fairness Regulation Learning (FRL) and Task-Aware In-Context Fairness Regulation Learning (TA-ICFRL) reduce such one-sided interpretations by injecting fairness-aware or balancing instructions during task execution.
APA
Gharat, H., Agrawal, H. & Patro, G.K.. (2026). Memories That Discriminate: Detecting and Correcting Bias in Personalized Hiring Agents. Proceedings of The Algorithmic Fairness Across Alignment Procedures and Agentic Systems, in Proceedings of Machine Learning Research 344:34-71 Available from https://proceedings.mlr.press/v344/gharat26a.html.

Related Material