Memory-First Fact-Checking: A Knowledge-Graph-Grounded Multi-Agent System for Misinformation Detection

Amelia Petrenciuc; Alexandru Lecu; Adrian Groza

Source-linked AI summary

Memory-First Fact-Checking: A Knowledge-Graph-Grounded Multi-Agent System for Misinformation Detection

Amelia Petrenciuc, Alexandru Lecu, Adrian Groza

arXiv:2608.29617v1Aug 30, 2026cs.CLcs.AI

TL;DR

Misinformation fact-checking needs evidence-grounded verification that can handle claims missing from static Knowledge Graphs while limiting unsupported language-model outputs. The paper proposes a memory-first KG system with web fallback and adversarial evidence deliberation, achieving higher reported performance than the Llama 3.3 70B baseline on the evaluated benchmark.

Problem
Existing classifiers and language models lack independently inspectable evidence, while preconstructed Knowledge Graphs cannot verify claims outside their stored coverage.
Method
The framework retrieves claims from dual-index KG memory, applies NLI and graph-aware confidence, invokes trusted web evidence and support–contradiction–judging agents when needed, then stores verified triples.
Results
97.4% accuracy on resolved claims versus 87.7% accuracy at 97.3% coverage for the Llama 3.3 70B baseline on Dataset 2.
Takeaways & Limitations
Verified web evidence can be converted into reusable structured knowledge, while unresolved claims can remain principled abstentions when confidence thresholds are unmet.

Abstract

from arXiv · show

This paper introduces a hybrid fact-checking framework that integrates Knowledge Graph-based semantic memory with adversarial multi-agent reasoning for explainable misinformation detection. The proposed system follows a memory-first, web-fallback architecture, in which input claims are initially evaluated against a dual-index Knowledge Graph through Sentence-BERT-based semantic retrieval and Natural Language Inference. When the evidence retrieved from the graph is insufficient to support a reliable decision, the framework collects information from trusted web sources and assesses it using an adversarial tribunal composed of support, contradiction, and judging agents. A graph-aware confidence mechanism combines semantic similarity, NLI confidence, and structural graph evidence to determine whether internal knowledge is sufficient, thereby reducing unnecessary web retrieval. Following verification, validated information is transformed into structured triples and incorporated into the Knowledge Graph, supporting the incremental expansion of the system's semantic memory. Experimental evaluation on a curated COVID-19 misinformation benchmark demonstrates that the proposed framework achieves an accuracy of 97.4\% and a macro-averaged F1-score of 92.6% on resolved claims, outperforming a Llama~3.3~70B baseline, which obtains an accuracy of 87.7% and a macro-averaged F1-score of 86.3%.

I. INTRODUCTION

The paper addresses scalable misinformation verification by combining inspectable evidence, Knowledge Graph memory, and adversarial reasoning in a hybrid architecture. Its memory-first design uses internal graph evidence before invoking web retrieval when coverage or confidence is insufficient.

Misinformation’s rapid circulation, especially during COVID-19, increases the need for automated fact-checking that operates at scale.
Content-based classifiers and transformer models can classify claims but do not inherently retrieve independently inspectable external evidence.
Large language models add reliability concerns because fluent outputs may contain plausible but unsupported information.
Knowledge Graph verification is reusable but cannot cover claims absent from the stored graph, particularly emerging or domain-specific claims.
The proposed framework combines dual-index KG retrieval, graph-aware confidence scoring, adversarial support–contradiction–judging agents, and persistent integration of verified triples.

B. Semantic Retrieval and Stance-Aware Inference

The framework separates semantic evidence retrieval from stance assessment, using NLI and graph safeguards to decide whether KG evidence is reliable enough. Specialized adversarial agents then evaluate competing interpretations of retrieved evidence.

SBERT and RAG identify semantically related evidence, but semantic relevance alone cannot distinguish support, refutation, and neutrality.
NLI models entailment, contradiction, and neutrality, although negation and numerical reasoning remain vulnerable points for fact verification.
Claim-level and atomic-fact indexes retrieve candidates, while NLI assesses their logical relationship and graph safeguards determine whether external verification is required.
ProgramFC demonstrates interpretable task decomposition by breaking complex claims into simpler subtasks executed through specialized functions.
The proposed tribunal assigns support, contradiction, and judging roles, grounding its final decision in cited passages rather than unconstrained interaction.
Lightweight entity resolution allows newly verified facts and aliases to be reused during incremental graph updates.

III. SYSTEM ARCHITECTURE

The system first tests claims against persistent Knowledge Graph memory and routes unresolved cases to web evidence and an adversarial tribunal. Verified external conclusions are stored as structured facts, while explanations cannot alter established verdicts.

The memory-first workflow queries a persistent KG, evaluates candidates with semantic similarity, NLI, and graph-aware confidence, then routes unreliable cases to web investigation.
Externally verified conclusions become Subject–Predicate–Object facts persisted in Neo4j, allowing later claims to benefit from accumulated graph memory.
Verdicts are established by the deterministic KG layer or evidence-grounded tribunal, while downstream explanations only summarize associated evidence.

B. Dual-Index Knowledge Graph

The Knowledge Graph uses complementary MACRO and MICRO indexes to preserve either claim context or atomic-fact structure. Retrieval eligibility and deterministic arbitration govern whether graph memory resolves a claim or triggers external verification.

The MACRO index stores complete Claim nodes, whereas the MICRO index stores atomic VerifiedFact and ClaimedFact Subject–Predicate–Object statements.
Both indexes use 768-dimensional all-mpnet-base-v2 embeddings and cosine similarity for retrieval.
A graph node is eligible according to its similarity to the input claim; if neither index produces an eligible candidate, the claim goes directly to external verification.
When both indexes qualify, deterministic arbitration selects the representation level using claim length and similarity-score differences.

C. NLI and Graph-Aware Decision Layer

The decision layer combines graph retrieval with NLI to interpret claims relative to stored truth status, then uses a composite confidence score and safeguards to decide whether internal evidence is sufficient.

The BART NLI model compares the retrieved graph statement as premise with the input claim as hypothesis, estimating entailment, contradiction, and neutral probabilities.
NLI relations are interpreted jointly with the graph node’s stored truth status rather than treated as factual verdicts alone.Entailment with a true node supports the claim, while entailment with a false node indicates reproduced misinformation.
C = 0.60*S + 0.30*N + 0.10*G combines semantic similarity, directional NLI confidence, and structural graph support.The weights are fixed empirically, with semantic similarity dominant and graph structure providing secondary corroboration.
The composite score routes claims to internal resolution or external verification rather than serving as a calibrated probability.Neutral NLI outcomes require stronger semantic and structural evidence before a graph verdict is accepted.
Numeric, temporal, regional, path-specific, and high-neutrality safeguards treat unreliable graph matches as unresolved and transfer them to web verification.A corrective relation is accepted only when corrective evidence is available in the graph neighborhood.

D. Web Evidence Pipeline

The web evidence pipeline retrieves and diversifies trusted-source passages before adversarial analysis, while preserving NEUTRAL when evidence is insufficient or execution fails technically.

The web branch activates when graph candidates or confidence are insufficient, or when an NLI safeguard rejects the retrieved relation.Evidence collection finishes before any deliberative agent is invoked.
An LLM optimizer preserves claim entities, relations, polarity, and numerical information in a concise DuckDuckGo query, whose results are filtered through a curated trusted-source whitelist.DuckDuckGo returns up to 40 candidate results.
Up to five pages from distinct trusted domains are cleaned, split into overlapping 100-word passages, and ranked against the claim using all-mpnet-base-v2 embeddings.Explicit fact-checking or refutation language receives small relevance adjustments.
The final evidence bundle contains up to six diversified passages selected from the 15 highest-ranked candidates, with at most two passages per domain.If top similarity remains below 0.40, the system returns NEUTRAL without invoking the tribunal.
A sequential three-agent tribunal evaluates sufficient evidence through support, contradiction, and judging roles.The Judge returns a structured TRUE, FALSE, or NEUTRAL verdict with explanation, confidence, primary source, and correction when FALSE.
A single-pass web-verification routine is reserved for technical tribunal failures and is not triggered by a valid NEUTRAL verdict.

F. Knowledge Persistence

After successful web verification, the framework converts supported claims into normalized, provenance-preserving graph triples and uses verdict-dependent persistence to expand reusable knowledge.

Ontology extraction transforms verified claims and supporting evidence into Subject–Predicate–Object triples while distinguishing original assertions from supported factual statements.Vague entities, malformed predicates, placeholders, and discourse-level expressions are rejected.
Entities and predicates are normalized and resolved against Neo4j using exact matching, aliases, and fuzzy lexical matching.Verified facts reuse the retrieval encoder, enabling future MICRO-level semantic search.
TRUE results store the verified claim and supporting facts, FALSE results may link misinformation to a corrected claim through CORRECTED_BY, and NEUTRAL claims are not persisted.Stored elements retain source URL, source type, and verification timestamp provenance.
The persistence process closes the verification loop by converting externally verified claims into reusable graph knowledge.

IV. RUNNING SCENARIOS

The representative execution scenarios illustrate direct KG resolution, guarded web fallback for a logically unrelated match, and additional verification paths for other claims.

Scenario 1 resolves an asymptomatic-infectiousness claim directly from the KG in ∼2s after SSBERT = 0.95 and NLI ENTAILMENT against a TRUE node.
Scenario 2 routes a high-similarity but logically unrelated symptomatic-versus-asymptomatic claim to web investigation after NLI NEUTRAL with p = 0.958.The high-neutral guard prevents an unsupported KG verdict.

V. EVALUATION

The evaluation uses two COVID-19 claim collections and compares the hybrid framework with standalone LLM baselines, while also examining resolution paths and limitations. On both datasets, results show strong resolved-claim performance, but path analyses remain descriptive rather than controlled component tests.

Datasets: The evaluation excludes direct KG-test overlap by reserving 1,397 claims for evaluation and using 5,023 to construct the initial graph.The second dataset was also excluded from initial KG construction.
Datasets: Two COVID-19 claim collections were used, including a Patwa et al. dataset and a curated Check-COVID subset.Dataset 2 contains 333 self-contained claims; evaluation claims were excluded from initial KG construction.
Evaluation setup: Standalone baselines receive only the input claim and a fixed verdict instruction, without KG retrieval, web evidence, deliberation, or persistent memory.Coverage measures definitive TRUE or FALSE assignments; accuracy and macro-F1 are computed over resolved claims, with NEUTRAL treated as abstention.
Results: 97.4% accuracy at 90.7% coverage on Dataset 2 exceeds the Llama 3.3 70B baseline’s 87.7% accuracy at 97.3% coverage.The system leaves 31 claims unresolved as NEUTRAL when confidence thresholds are not met.
Results: 59.5% of Dataset 2 claims are resolved by the tribunal at 99.5% accuracy, compared with 31.2% resolved by the KG path at 93.3% accuracy.These resolution-path results describe different claim subsets and do not isolate component contributions.

VI. CONCLUSION

The paper concludes with a hybrid fact-checking system that combines semantic memory, evidence retrieval, multi-agent deliberation, and persistent knowledge enrichment. It reports 97.4% accuracy on resolved claims and identifies future work in temporal filtering, negation-sensitive NLI, ablations, and cross-domain evaluation.

Conclusion: The system integrates Knowledge Graph memory, NLI-augmented reasoning, trusted-source retrieval, adversarial deliberation, and persistent knowledge enrichment.These components support evidence-grounded explanations and principled abstention on ambiguous claims.
Conclusion: 97.4% accuracy on resolved claims outperforms standalone LLM baselines while retaining evidence-grounded explanations and principled abstentions.Verified web evidence is converted into reusable structured knowledge for incremental memory expansion.
Future work: Future work targets temporal evidence filtering, negation-sensitive NLI, component-isolating ablation studies, and generalization beyond COVID-19 misinformation.