From Personalization to Prejudice: Bias and Discrimination in Memory-Enhanced AI Agents for Recruitment

1 Introduction

• Memory and tools can shift agents from stateless transactions toward adaptive assistants that maintain continuity and align with evolving user goals.Persistent memory supports tracking preferences over time.Source: §1 · p. 1
• Bias may enter during personalized query creation, grow through personalized retrieval and re-ranking, and persist when memory updates reinforce earlier skews.The pathways reflect use of stored histories and recruiter-preference interpretations at multiple workflow stages.Source: §1 · p. 1
• Prior work documents recruitment bias and studies personalization and agent safety, but fairness in memory-enhanced personalized agents remains less examined.Recruitment studies cover traditional and automated settings, while personalized-agent work has largely focused on design.Source: §1 · p. 2
• The paper addresses an underexplored fairness gap: personalized-agent research has emphasized design, while giving less attention to how memory mechanisms may introduce or amplify bias.It maps where bias may arise across agent-operation stages and how it may propagate.Source: §1 · p. 2

2 Personalized Agents: Design and Operation

• Agents operate in a perception–planning–action cycle, consulting memory, formulating a plan, executing tool calls, and integrating observations into subsequent decisions.Short-term memory holds immediate conversational context, while long-term memory accumulates preferences, interactions, and procedures.Source: §2 · p. 2
• The recruitment agent can personalize a raw query, create a preference-aware job description, retrieve candidates, and re-rank them using recruiter memory.It can select one or a combination of these pathways, alongside baseline retrieval using the raw query.Source: §2 · p. 2
• The agent design supports empirical investigation of where bias may emerge and propagate across workflow stages.The study examines configurations that combine different personalization pathways.Source: §2 · p. 2

3.1 Experimental Settings

• The study uses Bias in Bios with test data for recruiter histories and train data as the candidate pool, creating 10,000 job postings and 1,000 recruiter profiles.For each recruiter and posting, it samples 4–10 same-profession candidates and curates task-specific memory using gender likelihoods and bio–profession similarity.Source: §3.1.1 · p. 2

3.1.2 Agent Configuration:

• Agent configuration combines semantic and episodic memory, personalized query and job-description generation, embedding-based retrieval, and GPT-4.1 re-ranking against memory.The configuration includes non-personalized retrieval from the raw query and personalized retrieval based on the personalized job description.Source: §3.1.2 · p. 2

3.1.3 Experiments:

• Seven experiment configurations test baseline and balanced retrieval, personalization at retrieval and re-ranking, gender-explicit memory, and gender scrubbing.All experiments generate ranked lists of the top 20 candidates.Source: §3.1.3 · p. 3

3.1.4 Evaluation Metrics:

• The evaluation sums normalized positional attention for male and female candidates in retrieved and re-ranked lists, using logarithmically discounted rank gain.Higher-ranked candidates receive more positional attention than lower-ranked candidates.Source: §3.1.4 · p. 3

3.2.1 Assessing Personalization: Utility Gains vs. Bias Risks.

• Average similarity was 0.52 for personalized re-ranked candidates, 0.5 for personalized retrieved candidates, and 0.41 for non-personalized candidates.Utility was measured by comparing top-5 recommendations with recruiters’ previously shortlisted candidates.Source: §3.2.1 · p. 3

3.2.2 Bias in Retrieval and Re-Ranking Stages: While the re-

• Experiments found bias in baseline retrieval and memory-aligned shifts during personalized retrieval and re-ranking; removing explicit gender indicators reduced bias.With fairness constraints during retrieval, re-ranking still followed recruiter-memory patterns; full personalization showed stronger reflection of those patterns than retrieval or re-ranking alone.Source: §3.2.2 · pp. 3–4

3.2.3 Bias Amplification in Re-Ranking Stage:

• Across retrieval-bias cohorts and experiments, re-ranking consistently amplified bias in the direction of recruiter-memory patterns.The analysis grouped cases by high female bias, balanced retrieval, or high male bias.Source: §3.2.3 · p. 4

3.2.4 Do Re-ranking Adjustments Reflect Merit or Gender Bias?

• Aggregate Meritocratic (Un)Fairness increased during re-ranking in 77% of Experiment 5 instances and 57% of Experiment 6 instances with gender scrubbing.The authors report that re-ranking was mostly influenced by bias from interpreting recruiter memory, rather than candidate merit; scrubbing reduced but did not eliminate the increase.Source: §3.2.4 · p. 4

3.2.5 Bias in Pre-Retrieval Stage:

• Gender preferences appeared in 60.5% of pre-retrieval instructions, while 39.5% were neutral.The authors also found 73.17% of recruiter task-memory summaries biased, and reported that gender-coded proxy terms persisted after explicit gender indicators were scrubbed, making scrubbing necessary but insufficient.Source: §3.2.5 · p. 4

• The study simulated a memory-augmented personalized recruitment agent and measured how personalization influenced bias.Source: §4 · p. 4
• The authors conclude that personalization introduces and amplifies bias over time, even as it increases effectiveness and utility.Source: §4 · p. 4
• The authors judge existing LLM guardrails insufficient for agentic settings and plan to study other domains, multi-turn interactions, and bias-reduction strategies that retain personalization benefits.Source: §4 · p. 4