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

Researchers identify 'Memory Contagion,' a phenomenon where biased evaluator feedback propagates through LLM agent memory systems into future iterations, even with perfect consolidation. The study demonstrates that bias contamination occurs at rates as low as 20% and has differential effects depending on bias type, exposing a critical vulnerability in current agent memory architectures.

This research uncovers a fundamental design flaw in LLM agent systems that compounds over time. When AI agents store experiences evaluated by biased sources, those biases become embedded in shared memory pools, contaminating subsequent agents that retrieve from the same systems. The phenomenon persists even under ideal consolidation conditions, proving that input bias alone suffices for cross-temporal propagation—a finding that challenges assumptions underlying current memory system designs.

The work builds on growing recognition that LLM agent capabilities depend heavily on memory infrastructure for coherence and learning. Previous studies focused on memory degradation during consolidation, but this research shifts focus to the quality of inputs feeding those systems. The identification of bias-type-dependent consolidation effects—where length bias attenuates while authority bias potentially amplifies—suggests memory systems don't uniformly filter distortions, instead selectively reinforcing certain biases based on their structure.

For developers deploying autonomous agents in production environments, this creates operational risk. Multi-agent systems sharing memory stores risk systematic bias amplification across agent generations. The absence of a safe contamination threshold means even minimal biased training data or evaluation poses risk. Organizations using agent systems for decision-making, content creation, or information retrieval must audit evaluation sources and implement bias detection mechanisms upstream of memory storage. The research provides formal measurement tools but practical mitigation strategies remain underdeveloped, requiring urgent attention from practitioners building deployed agent systems.

• →Biased evaluator feedback propagates through agent memory into future agents even with perfect consolidation
• →Bias contamination detected at rates as low as 20%, with no identified safe threshold
• →Memory consolidation effects differ by bias type, attenuating length bias while potentially amplifying authority bias
• →Shared memory systems create cross-temporal bias accumulation risk in multi-agent deployments
• →Current agent memory architectures lack built-in defenses against evaluator bias propagation