Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs

While graph topology is a critical factor influencing memory leakage in multi-agent Large Language Model (LLM) systems, its specific impact has historically been difficult to quantify. To address this gap, we present MAMA (Multi-Agent Memory Attack), a structured evaluation framework designed to compare topology-dependent memory leakage across various multi-agent LLM configurations.

MAMA utilizes synthetic documents containing labeled Personally Identifiable Information (PII) to generate sanitized task instructions. The framework employs a two-stage protocol: the "Engram" phase, which seeds private data into a target agent’s memory, and the "Resonance" phase, where an attacker engages in multi-round interactions to attempt data extraction. Leakage is quantified over ten rounds using a dual-stage recovery metric that assesses both exact-match extraction and LLM-based inference on the attacker’s final output.

Our evaluation spans six standard topologies—complete, circle, chain, tree, star, and star-ring—across varying node counts ($n\in{4,5,6}$), different attacker-target placements, and various base models. The findings are consistent: leakage rates increase with denser connectivity, shorter distances between the attacker and target, and higher centrality of the target node. Additionally, the majority of leakage occurs during the initial rounds before plateauing. While the specific base model influences absolute leakage rates, it does not alter the overarching structural trends. Furthermore, spatiotemporal and location-based attributes prove more susceptible to leakage than identity credentials or regulated identifiers.

Based on these results, we offer practical recommendations for system architecture: prioritize sparse or hierarchical network structures, maximize the separation between attackers and targets, and implement topology-aware access controls to limit hub and shortcut pathways. The source code for this study is publicly available at https://github.com/llll121/mama-eval.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC