How Implicit Bias Accumulates and Propagates in LLM Long-term Memory

TL;DR

This work examines how implicit bias accumulates and propagates in LLM long-term memory through a longitudinal framework and a new benchmark, DIB, spanning nine domains. It shows that bias is not static and can spill across domains, with open-weight models often more susceptible than closed-source ones. The authors propose Static System Prompting and Dynamic Memory Tagging, finding that DMT significantly reduces bias accumulation and cross-domain propagation, with a 72.6% success rate and 40.6% average mitigation impact in key experiments. The results underscore the importance of memory-aware fairness interventions and offer a practical pathway for mitigating bias in long-horizon LLM applications.

Abstract

Long-term memory mechanisms enable Large Language Models (LLMs) to maintain continuity and personalization across extended interaction lifecycles, but they also introduce new and underexplored risks related to fairness. In this work, we study how implicit bias, defined as subtle statistical prejudice, accumulates and propagates within LLMs equipped with long-term memory. To support systematic analysis, we introduce the Decision-based Implicit Bias (DIB) Benchmark, a large-scale dataset comprising 3,776 decision-making scenarios across nine social domains, designed to quantify implicit bias in long-term decision processes. Using a realistic long-horizon simulation framework, we evaluate six state-of-the-art LLMs integrated with three representative memory architectures on DIB and demonstrate that LLMs' implicit bias does not remain static but intensifies over time and propagates across unrelated domains. We further analyze mitigation strategies and show that a static system-level prompting baseline provides limited and short-lived debiasing effects. To address this limitation, we propose Dynamic Memory Tagging (DMT), an agentic intervention that enforces fairness constraints at memory write time. Extensive experimental results show that DMT substantially reduces bias accumulation and effectively curtails cross-domain bias propagation.

Figures (5)

• Figure 1: The Comprehensive Framework for Long-Term Memory Bias Injection, Accumulation, and Evaluation. The framework consists of three distinct phases: (Left) Memory Bias Injection: Standard user tasks (from MMLU-Pro) are transformed into biased queries via a Bias Generator Agent using specific templates (e.g., Frustration, Benevolence), serving as the input stream. (Middle) Long-Term Memory Loop: The Main Agent processes these queries in a longitudinal setting ($t=1$ to $100$). Throughout the interaction, the agent retrieves context from and updates the Memory Storage, leading to the gradual accumulation of implicit bias. (Right) Periodic Bias Evaluation: The system triggers a periodic audit every $\Delta t = 20$ turns. The agent is frozen and evaluated against the Decision-Based Implicit Bias Benchmark (DIB), which assesses bias across three domains: Economic & Professional Stratification, Social Trust & Cultural Fit, and Perceptual & Physical Bias.
• Figure 2: Implicit Bias Accumulation across Memory Architectures and Injection Rates during long-term interaction. Each column represents a different bias injection rate ($\lambda \in \{0, 0.1, 0.2, 0.3\}$), and each row represents a memory mechanism. The Y-axis denotes the Average Generalized Bias Variance (AGBV), where higher values indicate greater unfairness.
• Figure 3: Impact of Single-Source Bias Injection on Global Bias Accumulation (Audit at $T_{80}$). This heatmap illustrates the cross-domain propagation of unfairness under a single-bias injection protocol. The Y-axis represents the specific Injected Bias Source, while the X-axis denotes the nine Evaluated Domains. The color intensity corresponds to the net increase in bias severity, quantified by $\Delta\text{GBV} = \text{GBV}_{t=80} - \text{GBV}_{t=0}$.
• Figure 4: This heatmap visualizes the change in $\Delta \mathrm{GBV}$ after applying our mitigation strategy across different models (Claude-Haiku-3 and DeepSeek-V3.1) and LTM mechanisms.
• Figure 5: System prompt of the Audit Agent. The structured warning tag allows the main agent to interpret the retrieved memory with necessary skepticism.