Memory Contagion: Cross-Temporal Propagation of Evaluator Bias via Agent Memory
A new failure mode threatens the reliability of LLM agents that rely on shared memory systems. Researchers have identified Memory Contagion, where bias embedded in evaluator feedback gets locked into consolidated memory and then propagates to downstream agents retrieving from the same store, even when memory consolidation itself is technically perfect. This reveals a structural vulnerability in multi-agent systems and long-horizon deployments where memory reuse is common, forcing practitioners to rethink how agent experiences are validated and stored.
Modelwire context
ExplainerThe critical detail the summary underplays is the 'even when consolidation is perfect' qualifier. This isn't a bug in how memories are written; it's a property of the architecture itself, meaning cleaner consolidation pipelines won't fix it.
This connects directly to two threads running through recent coverage. The 'Towards Root Memories' paper from the same day exposed how memory retrieval systems fail when they optimize for embedding similarity over logical relevance, and Memory Contagion adds a second layer: even if retrieval is logically sound, the content being retrieved may carry embedded evaluator bias that no retrieval strategy can filter out. Separately, the 'Uncertainty-based Debiasing and Unlearning' work highlighted how aggregate metrics obscure per-sample failures in evaluation pipelines. That same blind spot applies here: a multi-agent system could pass coarse quality checks on its memory store while silently propagating biased feedback across every downstream agent that queries it.
Watch whether any of the major agent memory frameworks (MemGPT, Zep, or comparable production systems) publish explicit mitigation guidance within the next two quarters. Adoption of provenance tagging at the memory-write layer would signal the field is treating this as an infrastructure problem rather than a research curiosity.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsLLM agents · Memory Contagion · evaluator bias · agent memory systems
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.