AuthorsChenchen Lin, Wenhao Yuan, Xuehe Wang, Edith Cheuk Han Ngai
TL;DR
CAMA uses neuro symbolic latent evidence slots plus active independent evidence recovery to reach 87.9 EM on LongMemEval vs 86.8 for MADAM RAG.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CAMA identifies Memory Correlation Bias, where correlated memories are repeatedly counted and form a false majority, inflating support for a hypothesis.
In long term multi agent systems, this bias corrupts arbitration, writes wrong conclusions back into shared memory, and causes persistent, self reinforcing reasoning errors.
HOW IT WORKS
CAMA’s core mechanism chains Neuro-Symbolic Evidence Assignment, Effective Independent Evidence Estimation, Factor-Level Conflict Arbitration, and Active Independent-Evidence Recovery to reason over latent evidence slots instead of raw memories.
You can think of CAMA like a librarian clustering many copies of the same book into one catalog card, then deliberately searching for missing, independent books before deciding.
This design lets CAMA discount duplicated evidence that a plain context window would over count and actively pull in new, independent memories when support is too narrow.
DIAGRAM
This diagram shows how CAMA runs query-conditioned evidence decoupling, arbitration, and sequential recovery actions over time for a single query.
DIAGRAM
This diagram shows how CAMA is evaluated across benchmarks, baselines, and ablations with DeepSeek V4 Flash and Qwen3.6 27B.
PROCESS
Query-Conditioned Evidence Decoupling
CAMA runs Neuro-Symbolic Evidence Assignment over the retrieved slice and provenance graph to produce soft assignments Z(t) and overlap scores r(t)_ij.
Effective Independent Evidence Estimation
CAMA computes slot presence e(t)_j, normalizes to p(t)_j, and derives the Hill diversity based N(t)_eff as the effective independent evidence count.
Factor-Level Conflict Arbitration
CAMA scores memories s(t)_i(h), aggregates slot support β(t)_j(h), applies reliability weights ρ(t)_j, and forms arbitration logits ℓ(t)(h) and posterior P(t)(h|q,C(t)_q).
Active Independent-Evidence Recovery
CAMA’s recovery policy observes state S_t and chooses Expand(q′), Trace(m_i), or Stop, updating C(t)_q and Z(t) until evidence sufficiency or budget B is reached.
KEY CONTRIBUTIONS
Identification of Memory Correlation Bias
CAMA formalizes Memory Correlation Bias in long term multi agent memory, showing how correlated memories create false majorities that corrupt arbitration and propagate errors.
Correlation-Aware Memory Arbitration
CAMA introduces Neuro-Symbolic Evidence Assignment, Effective Independent Evidence Estimation, and Factor-Level Conflict Arbitration to aggregate at latent evidence slot level instead of raw entries.
Active Independent-Evidence Recovery
CAMA adds Active Independent-Evidence Recovery with Expand and Trace actions, achieving CMR 71.2 and RS 7.8 on MemoryAgentBench under DeepSeek V4 Flash.
RESULTS
EM
87.9
+1.1 over MADAM-RAG on LongMemEval with DeepSeek-V4-Flash
BERT
69.2
vs MADAM-RAG’s 62.4 on LoCoMo with DeepSeek-V4-Flash
CMR
71.2
+10.5 over MADAM-RAG on MemoryAgentBench correlation metrics
RS
7.8
down from MADAM-RAG’s 15.3 on MemoryAgentBench, halving replication sensitivity
On MemoryAgentBench, LongMemEval, and LoCoMo, which test long term memory reasoning and correlation robustness, CAMA consistently raises task scores and boosts Correct Minority Recovery while sharply lowering Replication Sensitivity.
BENCHMARK
On MemoryAgentBench, LongMemEval, and LoCoMo, which test long term memory reasoning and correlation robustness, CAMA consistently raises task scores and boosts Correct Minority Recovery while sharply lowering Replication Sensitivity.
BENCHMARK
Overall score on MemoryAgentBench with DeepSeek-V4-Flash.
BENCHMARK
Replication Sensitivity (RS) on MemoryAgentBench.
KEY INSIGHT
CAMA cuts Replication Sensitivity on MemoryAgentBench from MADAM-RAG’s 15.3 down to 7.8 while simultaneously increasing Correct Minority Recovery from 60.7 to 71.2.
This is surprising because adding a recovery policy and extra retrieval usually increases sensitivity to redundant evidence, yet CAMA’s latent slot reasoning makes more retrieval actually reduce bias.
WHY IT MATTERS
CAMA enables long term multi agent systems to arbitrate based on effective independent evidence rather than raw memory counts, even under heavy correlation.
Builders can now safely scale shared memories and multi agent pipelines without majority voting being hijacked by duplicated or biased sources, improving reliability for high stakes deployments.
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…