Memory Contagion: Cross-Temporal Propagation of Evaluator Bias via Agent Memory

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Zewen Liu
• Venue: arXiv: cs.LG
• Published: 2026-06-22
• Collected: 2026-06-23
• URL: http://arxiv.org/abs/2606.23195v1
• PDF: https://arxiv.org/pdf/2606.23195v1
• DOI: UNKNOWN

Plain-Language Brief

The study identifies a new phenomenon called Memory Contagion, where biased evaluators' influence can affect the experiences of language model agents, causing their memories to become biased as well. This bias is passed on to future agents who retrieve from the same memory store, even under perfect consolidation conditions.

What the Document Says

• Memory Contagion: a novel phenomenon identified where biased evaluators' influence propagates through agent memories over time.
• Bias can be passed on to future agents retrieving from the same memory store, even under perfect consolidation conditions.
• Two bias types (length preference and authority bias) were studied across four experimental phases.
• Under perfect consolidation conditions, Memory Contagion occurs due to biased input as a sufficient cause.
• Consolidation can have opposite effects depending on bias type: it robustly attenuates length bias but preliminarily amplifies authority bias.
• No safe threshold for contamination was observed; bias propagation is detected at low rates (p=0.2).

Key Concepts Mentioned

• Memory Contagion
• Large Language Model
• Agent Memory
• Evaluator Bias
• Consolidation
• Length Preference
• Authority Bias

Claims Made by the Document

• Memory Contagion occurs even with perfect consolidation (oracle condition)
• Bias propagation is detected at low contamination rates (p=0.2)

Evidence Used

• The study demonstrates Memory Contagion under oracle condition and detects bias propagation at p=0.2.

Methods / Approach

• Two bias types were studied across four experimental phases.
• Experiments were conducted on language model agents trained or guided by biased evaluators

Limitations Stated by the Authors

• The study preliminarily amplified authority bias impact in a single-run estimate

Limitations Noted by the Agent

Agent interpretation — not from the source.

• The agent's memory degradation during continuous consolidation was not directly addressed in this work.

Open Questions

• What other factors may contribute to the propagation of bias in agent memories?
• Can safe thresholds for contamination be identified with further research?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The study primarily focuses on the cross-temporal propagation of evaluator bias and its effects on agent memory, rather than delving into the causes of memory degradation during continuous consolidation.
• The findings suggest that consolidation can have opposite effects depending on bias type, but further research may be needed to confirm this.

Raw Abstract / Excerpt

Large Language Model (LLM) agents increasingly rely on memory systems to maintain long-term coherence. Recent work shows that agent memories degrade during continuous consolidation. However, existing research assumes memories are derived from unbiased experiences. In this work, we identify and formalize a novel phenomenon: Memory Contagion -- the cross-temporal propagation of evaluator bias through agent memory. We show that when agents are trained or guided by biased evaluators, their experiences become biased; when these trajectories are stored and consolidated into memory, the bias propagates to future agents retrieving from the same memory store, even when consolidation is perfect (oracle). Across two bias types (length preference, authority bias) and four experimental phases, we demonstrate: (1) Memory Contagion occurs even with perfect consolidation (oracle condition), proving that biased input is a sufficient cause of contagion; (2) Consolidation has opposite effects depending on bias type -- robustly attenuating length bias while preliminarily amplifying authority bias (single-run estimate), suggesting a bias-type-dependent interaction; (3) No observed safe threshold: bias propagation is detected at contamination rates as low as p=0.2. Our findings expose a critical vulnerability in current agent memory designs and provide formal tools for measuring cross-temporal bias propagation.