Multi-agent memory votes double-count shared sources, and estimating independent evidence fixes the false majority

arXiv 2608.19701·medium signalXBlueskyLinkedInCopy link

When several agents write memories derived from the same upstream source, retrieval-time voting counts that evidence repeatedly, a failure the authors name Memory Correlation Bias. CAMA estimates the effective number of independent sources by combining neural dependency inference with symbolic reasoning over provenance, and adds a sequential recovery policy that goes looking for alternative evidence when the retrieved set is not independent enough. If you run a memory layer shared by multiple agents, store provenance alongside the claim so correlated entries can be collapsed instead of summed.