AuthorsXiaoyang Li, Yiqi Wang, Chencheng Zhu et al.
TL;DR
Correlated Promotion Benchmark (CPB) shows that gating on declared source type is the only tested admission policy that keeps false adoption as low as 0.06–0.09 while still answering, whereas independence-style deduplication rejects more true than false claims.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Shared agent memory misreads repeated claims as independent evidence, so copied or paraphrased false beliefs are admitted and reinforced as if they were corroborated.
Correlated Promotion Benchmark (CPB) shows that once an uncontested false belief enters shared memory, the consumer asserts it in 0.97–0.99 of probes, spreading false adoption across the whole team.
HOW IT WORKS
Correlated Promotion Benchmark (CPB) combines CPB-Static, CPB-Live, a gold admission rule, lineage collapse, and a governance rule to measure admission decisions and their downstream effects.
You can think of CPB like a lab where shared memory is a central database, agents are lab technicians, and the governance rule is a strict lab supervisor deciding which observations become official records.
By fixing lineage in scenarios and logging every write and retrieval, CPB enables analyses of admission policies that a plain context window cannot, including reach, exposure, and adoption of false beliefs over time.
DIAGRAM
This diagram shows how CPB-Live runs multi-agent teams over a shared store, logs retrievals and writes, and then probes a consumer that answers only from shared memory.
DIAGRAM
This diagram shows how CPB-Static assembles a frozen benchmark with a gold admission rule, and how CPB-Live executes scenarios with logged lineage and consumer probes.
PROCESS
Scenario families and feeds
Correlated Promotion Benchmark (CPB) instantiates scenario families like Correlated agreement, Concurrent conflict, Mixed scenarios, and Dual source with authored feeds and fixed lineage.
Team over shared store
CPB runs six-agent teams for six rounds over a shared store, where agents retrieve current beliefs and emit candidate claims each round.
Admission policy actions
CPB applies admission policies such as Governance, Independence vote, and Majority vote to map each claim to PROMOTE, REQUEST EVIDENCE, KEEP PRIVATE, or ABSTAIN.
Consumer probe and grading
At round six, CPB queries a Consumer that answers solely from the store, then grades reach, exposure, and adoption of false beliefs using the scenario’s truth and lineage.
KEY CONTRIBUTIONS
An instrument: CPB
Correlated Promotion Benchmark (CPB) introduces CPB-Static and CPB-Live with adapters, a gold admission rule, and lineage collapse over public annotated sources and authored multi-agent envelopes.
Diagnosis of admission mechanisms
CPB shows that reimplemented voting, collapse, and judge mechanisms, plus components from Mem0 and A-MemGuard, still admit false claims under rewording or authoritative source typing.
Measurements of the store itself
CPB measures how often a Consumer adopts false shared beliefs, how competing truths change adoption, and how rewording and arrival order affect admission across agent families.
RESULTS
Damage share Qwen
0.152
-0.848 vs Share all
False adoption
0.06–0.09
vs 0.22–0.47 for other answering policies
False promotion Static
0.077
-0.151 to -0.208 vs 0.228–0.285 policies
Uncontested adoption
0.97–0.99
Consumer asserts uncontested false beliefs almost always
Correlated Promotion Benchmark (CPB) evaluates admission policies on CPB-Live and CPB-Static, measuring damage share, false admissions, and adoption. The main result shows that the governance rule keeps damage shares between 0.112 and 0.152 and false adoption between 0.06 and 0.09, while other answering policies reach 0.22–0.47 false adoption and majority vote often matches share-all’s damage.
BENCHMARK
Correlated Promotion Benchmark (CPB) evaluates admission policies on CPB-Live and CPB-Static, measuring damage share, false admissions, and adoption. The main result shows that the governance rule keeps damage shares between 0.112 and 0.152 and false adoption between 0.06 and 0.09, while other answering policies reach 0.22–0.47 false adoption and majority vote often matches share-all’s damage.
BENCHMARK
Damage share (Dmg) on CPB-Live relative to Share all for Qwen3.5-9B.
KEY INSIGHT
Correlated Promotion Benchmark (CPB) finds that independence-style policies reject 0.051–0.127 of true claims while still admitting 0.119–0.306 of false ones, giving damage shares around 0.243–0.289.
This is counterintuitive because deduplication is expected to protect against copied falsehoods, yet CPB shows it disproportionately suppresses true claims while leaving many false beliefs admitted and retrievable.
WHY IT MATTERS
Correlated Promotion Benchmark (CPB) gives researchers a controlled way to study epistemic admission with fixed lineage, logged retrievals, and graded adoption across realistic multi-agent scenarios.
Builders can now systematically test governance rules, deduplication schemes, and judge-based filters for shared agent memory, observing how false beliefs propagate and how source-type gating or lineage-aware policies change downstream behavior.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…