This paper studies how false claims enter and spread in shared , where repeated statements can be mistaken for independent evidence. The authors introduce the Correlated Promotion (CPB), which has two parts: CPB-Static, a frozen test split from public annotated sources, and CPB-Live, which runs multi-agent teams over a shared store with logged writes and retrievals. They test eight admission policies across four agent families. They find that policies which deduplicate sources reject many true claims along with false ones, while policies that preserve answer coverage admit nearly as many false claims as unrestricted sharing. A policy that gates on declared source type reduces false adoption to 0.06–0.09, compared with 0.22–0.47 for other policies. However, once an uncontested false belief enters memory, a consumer relying only on that memory asserts it in 0.97–0.99 of probes. No non-oracle policy consistently rejects false claims across verbatim copies, paraphrases, and paraphrases declared authoritative. The work is interesting because shared agent memory is increasingly used in multi-agent systems, and false beliefs can propagate through copying, making admission control critical.
Shared agent memories confuse repeated claims with independent evidence
Across eight admission policies and four agent families, the authors find that filtering correlated claims creates a persistent trade-off between retaining true information and blocking falsehoods.
Top University
University of Southern Queensland · University of New South Wales · Aikaier · Nanjing University · Facebook
Research Digest··2 min read
In one line
Gating on declared source type reduces false adoption in shared agent memory, but no policy consistently rejects false claims.
Before you read this
Three ideas this paper builds onagent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.241 papers →benchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.366 papers →multi-agent coordinationThe collaborative operation of multiple AI agents to solve tasks more effectively or with greater capability.91 papers →
Read it as
Li and colleagues introduce the Correlated Promotion Benchmark, which tests whether claims should enter a memory shared by multiple AI agents.
The authors built two complementary evaluations.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.