When Agents "Misremember" Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems

TL;DR

This work investigates the Mandela effect in LLM-based multi-agent systems by introducing ManBench, a benchmark that tests collective false-memory formation across four task domains and five interaction protocols. It provides a formal framework to quantify memory distortions via metrics like error rate, reality shift rate, and maximal shift, and demonstrates widespread susceptibility across 13 LLMs. The authors show that group composition, memory timescale, group size, knowledge domain, and model scale influence the effect and propose prompt-level defenses (cognitive anchoring, source scrutiny) and a model-level defense (SFT with resilience and cooperative datasets) that significantly reduce false memories. The study highlights practical risks in sensitive domains and outlines ethical considerations and future directions for more robust, trustworthy cooperative AI systems.

Abstract

Recent advancements in large language models (LLMs) have significantly enhanced the capabilities of collaborative multi-agent systems, enabling them to address complex challenges. However, within these multi-agent systems, the susceptibility of agents to collective cognitive biases remains an underexplored issue. A compelling example is the Mandela effect, a phenomenon where groups collectively misremember past events as a result of false details reinforced through social influence and internalized misinformation. This vulnerability limits our understanding of memory bias in multi-agent systems and raises ethical concerns about the potential spread of misinformation. In this paper, we conduct a comprehensive study on the Mandela effect in LLM-based multi-agent systems, focusing on its existence, causing factors, and mitigation strategies. We propose MANBENCH, a novel benchmark designed to evaluate agent behaviors across four common task types that are susceptible to the Mandela effect, using five interaction protocols that vary in agent roles and memory timescales. We evaluate agents powered by several LLMs on MANBENCH to quantify the Mandela effect and analyze how different factors affect it. Moreover, we propose strategies to mitigate this effect, including prompt-level defenses (e.g., cognitive anchoring and source scrutiny) and model-level alignment-based defense, achieving an average 74.40% reduction in the Mandela effect compared to the baseline. Our findings provide valuable insights for developing more resilient and ethically aligned collaborative multi-agent systems.

Figures (6)

• Figure 1: An example of the Mandela effect: an LLM-based agent is influenced by specious evidence in a multi-agent conversation, forming a false collective memory that contradicts the truth.
• Figure 2: An overview of the five interaction protocols, where the Generic Group involves undifferentiated agents forming a simple social consensus, and the Role-based Group consists of agents with distinct, strategic roles. The Short-term timescale measures immediate, in-context response, while the Long-term timescale assesses whether beliefs persist after memory consolidation and retrieval.
• Figure 3: Results (%) of $\texttt{Err}^{P}$ and $\sigma_{P}$.
• Figure 4: Results (%) of maximal reality shift rate $\sigma_{max}$ across model series.
• Figure 5: Results (%) of reality shift rate $\sigma^{P}$ before (Base) and after applying the defense methods (cognitive anchoring and source scrutiny).