An AI memory method that weighs independent evidence rather than simply counting repeated memories reported the best displayed results among the methods tested in a new arXiv preprint. Called Correlation-Aware Memory Arbitration, or CAMA, it was designed for long-term multi-agent systems that must resolve conflicts between stored memories.
The paper calls the problem Memory Correlation Bias: a system may encounter many memory entries that are not effectively independent and mistake repetition for stronger support. Its research question is how to arbitrate such conflicts using independent evidence rather than apparent memory count.
CAMA’s central approach is to decouple correlated memories into effectively independent evidence sources, handle disagreements at the level of evidence factors, and recover missing independent evidence within a retrieval budget.
A memory count is not the same as independent evidence
For each query, the method represents relevant memories as latent, query-conditioned evidence slots and makes soft assignments between memories and those slots. Provenance information helps guide those assignments, but is not treated as a fixed label declaring which entries depend on one another.
It estimates whether evidence is genuinely present with a max-based calculation and a diversity measure called Hill diversity. The paper says that repeating one underlying factor does not increase the evidence mass.
Conflicting hypotheses are then compared by evidence factors rather than by individual memory entries, with each factor’s contribution added to the arbitration score.
When the available evidence is incomplete, a finite-horizon policy chooses among expanding the evidence, tracing dependencies and stopping. Actions that continue the search carry memory-access costs, so recovery is limited by a budget.
Tests on long-term memory benchmarks
The evaluation covered MemoryAgentBench, LongMemEval and LoCoMo, along with correlation-aware variants. Those variants added controlled correlated entries and distinct independent evidence while keeping the original queries and ground-truth answers.
Experiments used DeepSeek-V4-Flash and Qwen3.6-27B as backbone models. All methods shared the same backbone models, memory pools and retrieval settings; the setup used 10 retrieved memories, six latent slots and three recovery steps.
Across the displayed task metrics, CAMA had the best reported performance. On MemoryAgentBench, its two reported sub-scores, FC-SH and FC-MH, and its overall score were 78.9, 55.7 and 67.3 in the DeepSeek-V4-Flash block, compared with 76.5, 53.2 and 64.9 in the Qwen3.6-27B block.
To probe the effect of correlated memories, the study reported Correct Minority Recovery, Replication Sensitivity, Independent Evidence Gain and Evidence Resolution Rate. Under DeepSeek-V4-Flash on MemoryAgentBench, CAMA’s figures were 71.2, 7.8, 25.1 and 36.2 on those measures, respectively, and it had the best reported metric profile across the three datasets.
Under Qwen3.6-27B, the corresponding MemoryAgentBench figures were 68.1, 8.6, 23.4 and 34.0, with CAMA again listed as having the best reported profile across the three datasets.
The reported ablations showed lower task or correlation-aware performance for module-removal variants. The largest drop appeared when evidence decoupling was removed; removing Expand lowered Independent Evidence Gain and Evidence Resolution Rate, while removing Trace raised Replication Sensitivity.
Promising tables, narrow evidence
The authors describe a favorable accuracy-efficiency trade-off. On MemoryAgentBench with DeepSeek-V4-Flash, CAMA’s average latency was 6.7 in the reported units, with 4.2 LLM calls, a token cost of 14.6k, a context length of 6.8k and an accuracy improvement per kToken of 1.10. With Qwen3.6-27B, the corresponding figures were 5.8, 4.2, 14.6k, 6.8k and 1.14.
This is a computational methods evaluation based on benchmark data, not direct evidence from human studies or clinical outcomes. Its evidence comes from three benchmark families, controlled correlation-aware variants, the two listed backbones and the tested comparison methods.
The paper does not report the number of benchmark instances or constructed cases, train, validation or test sizes, repeated runs, confidence intervals or statistical significance tests. The reported tables therefore do not provide a measure of repeated-run variability or statistical reliability.
The correlation-aware variants were built with controlled paraphrasing, summarization, provenance links and added independent entries. The paper does not establish how closely those constructions match naturally occurring memory correlations; it also reports no separate numerical analysis for cases in which provenance is unavailable and the Trace action cannot be used.
The paper’s central argument is that memory reasoning should model dependencies and selectively recover missing independent evidence rather than simply accumulate more entries. Whether that pattern holds in naturally occurring multi-agent logs or transfers to additional models, domains and retrieval systems remains open.
Paper data and sources
Original title: Beyond Memory Majority: Latent-Source Reasoning for Multi-Agent Memory Arbitration
Authors: Chenchen Lin, Wenhao Yuan, Xuehe Wang, Edith Cheuk Han Ngai
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text