MemGraphRAG: Memory-based Multi-Agent System for Graph

Retrieval-Augmented Generation
Chuanjie Wu∗ Zhishang Xiang∗ Yunbo Tang
wuchuanjie@[Link] xiangzhishang@[Link] tangyunbo@[Link]
Xiamen University1, 2 Xiamen University2, 3 Xiamen University1
Xiamen, China Xiamen, China Xiamen, China

Zerui Chen Qinggang Zhang† Jinsong Su†

chenzerui1@[Link] qinggangzhang@[Link] jssu@[Link]
Xiamen University1 Jilin University Xiamen University1, 2, 3
arXiv:2606.00610v1 [[Link]] 30 May 2026

Xiamen, China Changchun, China Xiamen, China

Abstract CCS Concepts

Retrieval-Augmented Generation (RAG) has become an essential • Information systems → Retrieval models and ranking; •
method for mitigating hallucinations in Large Language Models Computing methodologies → Information extraction; Knowledge
(LLMs) by leveraging external knowledge. Although effective for representation and reasoning.
simple queries, traditional RAG struggles with large-scale, unstruc-
tured corpora where information is highly fragmented. Graph- Keywords
based RAG (GraphRAG) incorporates knowledge graphs to capture RAG, GraphRAG, Multi Agent, Agent Memory, Indexing Graph
structural relationships, enabling more comprehensive retrieval
ACM Reference Format:
for complex reasoning. However, existing GraphRAG methods rely Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang,
on isolated, fragment-level extraction for graph construction, lack- and Jinsong Su. 2026. MemGraphRAG: Memory-based Multi-Agent System
ing a global perspective on the whole corpus. As a result, these for Graph Retrieval-Augmented Generation. In Proceedings of the 32nd ACM
methods frequently lead to thematically inconsistent, logically SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD
conflicting, and structurally fragmented graphs that degrade re- 2026), August 9–13, 2026, Jeju Island, Republic of Korea. ACM, New York, NY,
trieval performance. In this paper, we propose MemGraphRAG, USA, 20 pages. [Link]
a novel framework that introduces a memory-based multi-agent
system to ensure high-quality graph construction. Specifically, 1 Introduction
MemGraphRAG employs a collaborative society of agents sup- Recently, Retrieval-Augmented Generation (RAG) effectively ex-
ported by shared memory, which provides a unified global context tends the capabilities of Large Language Models (LLMs) by leverag-
throughout the extraction process. This mechanism allows agents ing external knowledge [16, 31, 63]. However, existing RAG systems
to dynamically resolve logical conflicts and maintain structural suffer from critical challenges in real-world scenarios. This is due to
connectivity throughout the corpus. Furthermore, we propose a the unstructured and heterogeneous nature of large-scale corpora,
memory-aware hierarchical retrieval algorithm tailored for the con- where relevant information is often sparsely distributed. The con-
structed graph. Extensive experiments on multiple benchmarks texts retrieved by RAG systems are often noisy and lack structural
demonstrate that MemGraphRAG outperforms the state-of-the-art coherence. Although recent methods attempt to segment docu-
baseline models with comparable efficiency. Our code is available ments into smaller chunks for efficient indexing [2, 27, 29], this
at [Link] strategy disrupts long-range dependencies and loses critical contex-
tual details. As a result, the retrieved contexts are often incoherent
or insufficient for complex reasoning tasks [23, 62].
To address these limitations, Graph Retrieval-Augmented Gener-
∗ Contributed equally.
† Corresponding
ation (GraphRAG) [41, 52, 62] has emerged as a powerful paradigm,
author.
1 School of Informatics leveraging external structured graphs to model the hierarchical
2 Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural structure of background knowledge [23, 57]. Early efforts, such as
Heritage of Fujian and Taiwan,Ministry of Culture and Tourism RAPTOR [44] and Microsoft’s GraphRAG [12], organize knowledge
3 Institute of Artificial Intelligence
through recursive summarization and community-level abstrac-
tions to support coarse-to-fine retrieval, thereby facilitating compre-
hensive response generation. Subsequent methods, including GFM-
RAG [37], G-Retriever [24], and LightRAG [17], further incorporate
specialized retrieval mechanisms and learning objectives to improve
This work is licensed under a Creative Commons Attribution 4.0 International License. multi-hop generalization, scalability, and efficiency. Most recently,
KDD 2026, Jeju Island, Republic of Korea. HippoRAG [19]and its enhancement HippoRAG2 [20] have drawn
© 2026 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-2259-2/2026/08 inspiration from cognitive associative memory, utilizing algorithms
[Link] such as Personalized PageRank to simulate multi-hop reasoning
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

a) GraphRAG Local Information Only While some recent studies attempt to improve graph quality
before extraction by filtering triples using predefined schema [9,
Graph
g
ki n Construction 45], these approaches suffer from limited generalization and high
un
Ch Retrieval manual costs. Other efforts seek to improve graph quality through
Merging
Ch
un bottom-up clustering-based community summarization [12, 20, 50]
ki n Graph
Corpus g
Construction Knowledge Graph SubGraph or topic modeling [44]. Nevertheless these unsupervised approaches
remain susceptible to error propagation, because inaccuracies in
b) MemGraphRAG Inject Global Memory entity relations tend to be amplified at high-level summaries.
To address this, we revisit the pipeline of existing GraphRAG
Facts
Retrieval systems and propose a Memory-Based Multi-Agent Framework
Chunking for Graph Retrieval-Augmented Generation (MemGraphRAG).
Graph Construction
Hierarchical Graph Passages
Specifically, MemGraphRAG employs a collaborative society of
Corpus Global View + Local Details agents supported by a novel Three-Layer Global Memory. This
shared memory structure serves as a unified knowledge repository,
Figure 1: Comparison between existing GraphRAG and Mem- providing a global perspective that enables agents to dynamically
GraphRAG. Exiting GraphRAG performs isolated chunk- coordinate the extraction process, resolve conflicts upon detection,
level extraction without the global view, resulting in a noisy and integrate fragmented information across the entire corpus. To
and inconsistent indexing graph. While MemGraphRAG in- summarize, our contributions are listed as follows:
corporates global memory to ensure global consistency.
• We identify the root cause of performance degradation in
existing GraphRAG systems: the reliance on isolated local
extraction. We demonstrate how this lack of global context
inevitably leads to three critical deficiencies: thematic irrele-
vance, logical inconsistency, and structural fragmentation.
• We propose MemGraphRAG, which introduces a memory-
based multi-agent system into graph construction. The shared
memory not only maintains global thematic consistency to
prevent irrelevance and fragmentation, but also provides
grounded evidence to resolve local logical inconsistencies.
Figure 2: (Left) Evaluation of representative RAG and
Besides, we propose a memory-aware hierarchical retrieval
GraphRAG systems. The radius reflecting performance of
algorithm tailored for the constructed graph.
each systems. Relevance measuring context relevance to the
• We conduct extensive experiments on four benchmark datasets,
query. Recall measuring whether sufficient evidence is cov-
demonstrating that MemGraphRAG consistently outperforms
ered. (Right) Impact of removing irrelevant triples based on
state-of-the-art baselines in terms of graph quality, retrieval
schema frequency on the final performance (LLM-ACC).
quality and generation accuracy, validating its practicality
for real-world applications.
pathways. These strategies demonstrate the potential of graph-baed
retrieval in addressing the core limitations of traditional RAG.
However, contrary to their theoretical advantages, GraphRAG 2 Problem Statement
systems frequently underperform naive RAG systems in many To facilitate subsequent discussion, we first introduce key defi-
real-world applications [22, 52, 69, 70]. This performance decline nitions for the knowledge representation, and then present the
is primarily due to the low quality of automatically constructed complete problem formulation of GraphRAG.
knowledge graphs [52, 70]. Although graph-based retrieval en-
hances relevant knowledge recall, errors in graph construction
introduce substantial noise into the retrieved contexts simultane- 2.1 Key Definitions
ously. Fundamentally, these challenges persist because existing We first provide formal definitions for the core components of our
pipelines typically derive knowledge from isolated local segments, knowledge representation:
lacking a global perspective on the previously processed context. (i) type (𝑡) and entity (𝑒): a type 𝑡 (e.g., person) denotes an
This isolation leads to three critical deficiencies that undermine abstract category, while an entity 𝑒 (e.g., Einstein) is a concrete
graph quality: (i) thematic irrelevance: extracted triples are often instance. Formally, a typing function 𝜙 assigns each entity to its
irrelevant to the central theme, introducing meaningless facts. (ii) specific type, denoted as 𝜙 (𝑒) = 𝑡.
logical inconsistency: contradictory facts may emerge within a (ii) schema (𝑠) and fact (𝑓 ): a schema 𝑠 = (𝑡ℎ , 𝑟, 𝑡𝑡 ) (e.g., (person,
single subgraph, compromising semantic coherence. (iii) structural born_in, country)) specifies a logical constraint. 𝑡ℎ , 𝑡𝑡 represent the
fragmentation: the built graphs often suffer from fragmentation head and tail types, respectively, 𝑟 denotes a semantic relation.
issues, where the isolated nodes and disconnected components Based on this structure, a fact 𝑓 = (𝑒ℎ , 𝑟, 𝑒𝑡 ) (e.g., (Einstein, born_in,
weaken the core advantage of the knowledge graph in supporting Germany)) is a concrete instantiation of a schema, where 𝑒ℎ , 𝑒𝑡
global comprehension and multi-hop reasoning. represent the head and tail entity.
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

1) Mutually Exclusive Conflict 2) Temporal Conflict 3) Granularity Conflict

how automatic graph construction affects retrieval quality and
downstream generation.
ear 1879 nt Biden ace LA
t hy side h pl
Bir P re Birt
Einstein Conflict Be Included in
USA Conflict James
Birt P re
3.1 Performance Degradation
hye a side Birt
r 1880 nt Trump h pl
ace USA
We first compare Vanilla RAG with recent GraphRAG systems (MS-
GraphRAG, HippoRAG, and GFM-RAG) in the G-Medical dataset[52].
nt
side ty As shown in Figure 2(a), GraphRAG methods achieve higher re-
P re Biden h ci LA
Birt
th
Birthyear 46 trieval Recall (e.g., GFM-RAG: 84.3% vs. RAG: 71.8%), but suffer a
Einstein 1879
USA James
substantial drop in Relevance (38.5% vs. 62.9%), leading to noisier
45 th Birt
P re h co USA contexts and lower generation accuracy. These results indicate that
side Trump untr
nt y
existing GraphRAG pipelines often expand the retrieval coverage
at the cost of introducing excessive irrelevant information, which
Figure 3: Illustration of three conflict types in extracted ultimately harms the QA performance.
graphs: 1) Mutually Exclusive Conflict from logically incom-
patible facts, 2) Temporal Conflict caused by missing tem-
poral grounding for time-varying states, and 3) Granularity 3.2 Error Analysis
Conflict arising from inconsistent abstraction levels for the To further investigate why graph construction introduces noise
same entity or concept. Details are in Table 8 in Appendix C and conflicts, we hypothesize that the root cause lies in the isolated
(iii) ontology (O): the ontology O is defined as the collection local extraction paradigm adopted by most baselines. Without a per-
of all valid schemas, denoted as O = {𝑠 1, . . . , 𝑠 | O | }. It includes the sistent global memory, extraction LLMs process document chunks
theme and logical rules of the whole knowledge graph. independently, which leads to systematic issues in graph quality.
(iv) passage (𝑝): a passage 𝑝 denotes the specific text segment Specifically, we summarize the major failure modes as follows:
acting as the source of the extracted information. We define a Thematic Irrelevance. Without a global view of the corpus
function 𝜓 (𝑓 ) = 𝑝 to trace each fact 𝑓 back to its origin. theme, local extraction tends to introduce off-topic triples. To quan-
Detailed definitions are provided in Appendix ??. tify this effect, we conduct a filtering experiment (Figure 2(b)) that
removes triples based on schema frequency. Interestingly, filtering
2.2 Problem Formulation out 40% of low-frequency triples slightly improves accuracy (65.28%
We formally formulate the task of GraphRAG as a unified frame- vs. 64.85%), suggesting that a large fraction of extracted triples are
work composed of two distinct phases: thematically irrelevant noise.
(i) Offline Graph Structure Construction. Given a corpus of Logical Inconsistency. Independent extraction also introduces
unstructured documents D = {𝑑 1, 𝑑 2, ..., 𝑑 | D | }, the primary objec- semantic contradictions into the merged graph. As illustrated in
tive is to transform raw text into a structured graph G = (V, E). Figure 3, we observe mutually exclusive conflicts, temporal con-
In our framework, the vertex set V is heterogeneous, comprising flicts, and granularity conflicts, which create inconsistent reasoning
entities, types, and passages (V = V𝑒 ∪ V𝑡 ∪ V𝑝 ) , and the edge paths and confuse downstream retrieval. More conflict analyses are
set E encodes the semantic dependencies between them. Formally, provided in Appendix C.
this construction process is formalized as Structural Fragmentation. Due to missing global coreference
resolution and schema alignment, key entities are often duplicated
G = GraphConstructor( D ) (1) or scattered across disconnected subgraphs. This fragmentation
where GraphConstructor(*) maps the unstructured corpus to a prevents effective multi-hop traversal and reduces the usefulness
semantic graph topology, facilitating the efficient navigation from of the graph for global reasoning.
abstract concepts to concrete evidence.
(ii) Online Graph-Enhanced Retrieval and Reasoning. Based
on the constructed graph G, the system processes a user query 𝑞 to 3.3 Discussion
generate a final answer 𝑎. Unlike extracting isolated text segments, Current GraphRAG systems exhibit two fundamental limitations.
this phase involves identifying optimal reasoning paths within the First, existing GraphRAG systems exhibit a fundamental
graph to curate a structured context. The process is formulated as trade-off between recall and relevance. Although graph ex-
𝑎 = LLM(Retriever(𝑞, G) ) (2)
pansion improves coverage, it often retrieves irrelevant evidence
that overwhelms the LLM and degrades generation accuracy. Sec-
where Retriever(*) identifies the most relevant graph elements ond, current GraphRAG systems lack a global memory mech-
(subgraphs) to support grounded answer generation. anism during graph construction. Most systems rely on iso-
lated local extraction, processing document chunks independently
3 Preliminary Study without maintaining a persistent global state. As a result, the con-
Although knowledge graphs can model complex dependencies, re- structed graph fails to preserve thematic coherence and resolve
cent benchmarks show that advanced GraphRAG systems may cross-document conflicts, leading to thematic irrelevance, logical
underperform naive RAG in real-world QA tasks [52, 70]. To inves- inconsistency, and structural fragmentation in downstream retrieval
tigate this issue, we conduct two preliminary studies to analyze and reasoning processes.
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

Memory-based Indexing Graph Construction Memory-guided Online Retrieval

Multi-Agent Group Three-Layer Global Memory

Ontology Graph Hierarchical Graph Type Entity
Ontology Layer (𝑴𝒐𝒏𝒕 ) Question
Unstructured
Extraction Agent Pending Schema Stable Schema
Documents Schema
Ontology
Schema Fact
（Person, Rule, Country) Frequency (Country, Capital, City) Type: Person
Filter Relation: Job
(Person, Native, Location) (Person, Job, Profession) Type: Profession 𝑴𝒑𝒂𝒔 𝑴𝒇𝒂𝒄 𝑴𝒐𝒏𝒕 Passage
… … Weight: 5
(Company, Create, Product) (Person, Birthyear, Year) Answer
Fact Graph
Conflict Detector Conflict Fact Layer (𝑴𝒇𝒂𝒄 )
trigger
Found Fact
Ontology
Conflict Active Facts Inactive Instances
Propagation (Newton, Birthyear, 1643) (Louis XIV, Rule, France) Entity: Coach
Relation: Job Retrieved Retrieved
(Newton, Birthyear, 1645) (Drake, Native, Toronto) Entity: Simpson Sean Facts Passages
(𝑠𝑖𝑚 > 𝜏)
Conflict Handler … … Weight: 1
(Simpson, Job, Coach) (OpenAI, Create, ChatGPT)

Update Passage Layer (𝑴𝒑𝒂𝒔 ) Passage Graph

Passage
Ontology Personalized
Article 288
Original PageRank
Simpson Sean is a Canadian ice ho- Entity: Simpson
UK Sean
Document Conflict Triple ckey coach and former professional … The Unit-
Document: Simpson
(Newton, Birthyear, 1645) ice hockey player. He last served as ed Kingdom
Sean (UK), off-
is a Canadian ice
Audit head coach of Adler Mannheim in
the Deutsche Eishockey Liga …
icially the
hockey United
coach and……
(Newton, Birthyear, 1643)
Correct Triple

Figure 4: Overview of the MemGraphRAG framework with two phases: (i) Memory-Based Indexing Graph Construction, where
Global Memory (M) and the Knowledge Graph (G) co-evolve via unified schema filtering, global adjudication, and memory-
guided bridging; and (ii) Memory-Guided Online Retrieval, which leverages multi-layer memory filtering, structure-aware
node initialization, and Personalized PageRank to identify globally relevant contexts for generation.

4 Our Framework Graph G𝑓 𝑎𝑐 , constructed from M 𝑓 𝑎𝑐 , which represents instanti-

To overcome fragmented extraction and enable coherent graph evo- ated entity-relation triples for multi-hop reasoning; and (iii) Source
lution, we propose MemGraphRAG, a memory-based framework Evidence Graph G𝑝𝑎𝑠 , induced from M𝑝𝑎𝑠 , which grounds facts in
for constructing and maintaining high-quality knowledge graphs. G𝑓 𝑎𝑐 back to their supporting passages. This hierarchical design
Our key insight is that reliable graph construction requires not enables reasoning to traverse from abstract semantics to grounded
only structured storage, but also persistent coordination and cor- evidence. More details are provided in Appendix D.2.
rection across documents. As illustrated in Figure 4, it consists of Multi-Agent Group (A). We introduce a group of agents A =
two collaborative modules: Memory-based Graph Construction and {𝐴𝑒𝑥𝑡 , 𝐴𝑑𝑒𝑡 , 𝐴𝑟𝑒𝑠 }. Specifically: (i) the Extraction Agent 𝐴𝑒𝑥𝑡 extracts
Memory-guided Retrieval. We first introduce the foundational archi- schemas, facts, and passages into M with evidence grounding; (ii)
tecture, followed by the graph construction and retrieval pipelines. the Conflict Detection Agent 𝐴𝑑𝑒𝑡 monitors M 𝑓 𝑎𝑐 to detect redun-
dancy, structural anomalies, and logical inconsistencies; and (iii) the
Conflict Resolution Agent 𝐴𝑟𝑒𝑠 leverages schema constraints from
4.1 MemGraphRAG Architecture
M𝑜𝑛𝑡 and historical evidence from M𝑝𝑎𝑠 to resolve conflicts and
MemGraphRAG consists of three core components: a Global Mem- maintain global consistency in G. This design separates extraction,
ory that stores schemas, facts, and passages at different granularities diagnosis, and correction for reliable graph construction.
and supports the construction of the Hierarchical Indexing Graph,
and a Multi-Agent Group that interacts with memory to iteratively
extract, detect, and resolve conflicts. Specifically:
4.2 Memory-based Indexing Graph
Global Memory (M) organizes the extracted knowledge into Construction
a three-tier hierarchy, including an Ontology Layer (M𝑜𝑛𝑡 ) that Traditional graph construction often processes document chunks
stores schemas with extraction frequencies, a Fact Layer (M 𝑓 𝑎𝑐 ) that in isolation, resulting in index fragmentation and noise accumu-
maintains concrete facts, and a Passage Layer (M𝑝𝑎𝑠 ) that preserves lation. To address the critical limitations of Thematic Irrelevance,
original text passages for evidence grounding. To strengthen cross- Logical Inconsistency, and Structural Fragmentation identified in
layer associations, we introduce a dense indexing mechanism that our pilot study, we reformulate knowledge graph construction as a
enforces schema consistency and evidence traceability through two dynamic co-evolution process between the Global Memory M and
bidirectional interactions, where schema-instance alignment links the Knowledge Graph G. Distinct from static extraction pipelines,
schemas with facts and fact-evidence grounding connects facts with our approach adheres to three core principles designed to system-
their supporting passages. (See more details in Appendix D.2). atically resolve these issues: (i) Thematic Denoising via Unified
Hierarchical Indexing Graph (G). It provides a unified rep- Schema Filtering: Addressing Thematic Irrelevance, we employ a
resentation spanning abstract schemas, concrete facts, and textual unified schema to rigorously filter and manage extracted triples,
evidence. It consists of three interconnected graph views: (i) Se- ensuring that only thematically relevant knowledge is retained; (ii)
mantic Ontology Graph G𝑜𝑛𝑡 , derived from M𝑜𝑛𝑡 , which encodes Consistency Maintenance via Global Adjudication: To resolve
schema-level type relations and structural constraints; (ii) Fact Logical Inconsistency, we utilize the global memory to assist agents
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

in detecting and adjudicating semantic contradictions across dis- edges, enabling multi-hop reasoning over concrete facts. To improve
parate documents, thereby ensuring the logical unity of the graph; connectivity and reduce fragmentation, we further augment G𝑓 𝑎𝑐
iii) Structural Unification via Memory-Guided Bridging: To by introducing additional bridging edges, including type-based
overcome Structural Fragmentation, we leverage the global mem- connections derived from shared stable schema types in G𝑜𝑛𝑡 and
ory to identify and merge equivalent entities across disconnected similarity-based connections between entities with high embedding
subgraphs. By connecting isolated local extractions and aligning similarity. Finally, we induce the Source Evidence Graph G𝑝𝑎𝑠 from
them with the global ontology, we construct a cohesive and in- M𝑝𝑎𝑠 , which links facts and entities in G𝑓 𝑎𝑐 back to their originating
terconnected knowledge representation. Specifically, our graph passages, ensuring that every reasoning path remains traceable to
construction procedure is described as follows: grounded textual evidence.
4.2.1 Thematic Denoising via Unified Schema Filtering. Graph con-
struction begins with the Extraction Agent 𝐴𝑒𝑥𝑡 , which transforms
4.3 Memory-guided Online Retrieval
each document chunk 𝑐𝑖 into structured memory entries. Rather Building upon the Global Hierarchical Graph G and Global Memory
than producing triples alone, 𝐴𝑒𝑥𝑡 jointly constructs entries for all M, we perform memory-guided retrieval and reasoning in three
three layers of Global MemoryM by generating candidate schemas, stages: (i) Multi-Layer Memory Retrieval, which retrieves candi-
instantiated facts, and their supporting passages: date schemas, facts, and passages from M𝑜𝑛𝑡 , M 𝑓 𝑎𝑐 , and M𝑝𝑎𝑠 ; (ii)
Structure-Aware Node Initialization, which maps the retrieved
𝐴𝑒𝑥𝑡 (𝑐𝑖 ) → { S𝑐𝑎𝑛𝑑 ∈ M𝑜𝑛𝑡 , T𝑐𝑎𝑛𝑑 ∈ M 𝑓 𝑎𝑐 , P𝑠𝑟𝑐 ∈ M𝑝𝑎𝑠 }. (3)
evidence to initial node weights based on semantic relevance and
This design ensures that each extracted triple is strictly aligned structural signals; and (iii) Graph Propagation, which runs Per-
with a schema and grounded in source evidence. To mitigate halluci- sonalized PageRank (PPR) over the heterogeneous graph to rank
nation accumulation, newly generated schemas are initially treated globally important nodes and passages for LLM generation.
as candidates and are promoted to stable schemas only when their
empirical frequency exceeds a threshold: 4.3.1 Multi-Layer Memory Filtering. The retrieval phase initiates
𝑠𝑡𝑎𝑏𝑙𝑒
by querying the three distinct layers of the Global MemoryM in
M𝑜𝑛𝑡 = { 𝑠 ∈ M𝑜𝑛𝑡 | Freq(𝑠 ) ≥ 𝜏 }. (4) parallel. Given a user query q, we retrieve top-𝐾 candidates from
Accordingly, only facts aligned with stable schemas are activated for M in parallel, including schemas from M𝑜𝑛𝑡 , facts from M 𝑓 𝑎𝑐 , and
downstream graph construction and reasoning. Detailed extraction passages from M𝑝𝑎𝑠 . To reduce noise before graph reasoning, we
procedures are provided in Appendix D.3.1. retain only schemas and facts whose semantic similarity satisfies
Sim(q, x) > 𝜏. This filtering ensures that subsequent node initial-
4.2.2 Consistency Maintenance via Global Adjudication. During ization is seeded with high-confidence structural evidence. If no
evolutionary extraction, newly activated triples may introduce re- valid structural candidates remain (i.e., S𝑟𝑒𝑡 ∪ F𝑟𝑒𝑡 = ∅), we fall
dundancy or semantic conflicts with existing facts. To ensure the back to standard RAG retrieval by directly selecting passages from
long-term consistency of the Fact Layer M 𝑓 𝑎𝑐 , We deploy a decou- M𝑝𝑎𝑠 based on query similarity.
pled diagnosis and correction loop, where the Conflict Detection
Agent (𝐴𝑑𝑒𝑡 ) and the Conflict Resolution Agent (𝐴𝑟𝑒𝑠 ) collaborate 4.3.2 Structure-Aware Node Initialization. To seed graph propaga-
to continuously maintain memory integrity. Specifically, when a tion with query-specific context, we project the retrieved evidence
new triple 𝑡𝑛𝑒𝑤 ∈ M 𝑓 𝑎𝑐 becomes active, 𝐴𝑑𝑒𝑡 asynchronously scans onto the heterogeneous graph by defining an initial reset proba-
existing facts and identifies a conflict set F𝑐𝑜𝑛𝑓 based on semantic bility distribution 𝑃𝑖𝑛𝑖𝑡 (𝑣) for each node 𝑣 ∈ G. This distribution
similarity and ontology-level structural constraints: assigns the starting importance of nodes before propagation. We
F𝑐𝑜𝑛𝑓 = { 𝑡 ′ ∈ M 𝑓 𝑎𝑐 | Sim(𝑡𝑛𝑒𝑤 , 𝑡 ′ ) > 𝛿 ∨ Match(𝑡𝑛𝑒𝑤 , 𝑡 ′ ) }. (5) then initialize 𝑃𝑖𝑛𝑖𝑡 (𝑣) along three complementary dimensions, as
detailed below.
If F𝑐𝑜𝑛𝑓 is non-empty, 𝐴𝑟𝑒𝑠 is triggered to resolve the detected Entity Node Initialization via Facts: To ensure that graph propa-
inconsistencies. Rather than generating corrections heuristically, gation originates from grounded evidence, we initialize each entity
𝐴𝑟𝑒𝑠 leverages fact-evidence grounding to retrieve the provenance node 𝑒 based on the relevance of its associated retrieved facts. Specif-
passages from M𝑝𝑎𝑠 and adjudicates conflicts by comparing the ically, its initial weight is defined as the mean similarity over all
corresponding textual evidence. This evidence-driven resolution query-relevant facts containing 𝑒:
enables reliable corrective actions such as filtering invalid facts,
merging redundant triples, and resolving temporal or granular- 1 ∑︁
𝑃𝑖𝑛𝑖𝑡 (𝑒 ) = Sim(q, f ), (6)
ity inconsistencies, thereby ensuring that M 𝑓 𝑎𝑐 remains globally | F𝑒 |
𝑓 ∈F𝑒
coherent throughout the graph construction process.
where F𝑒 ⊆ F𝑟𝑒𝑡 denotes the subset of retrieved facts that contain
4.2.3 Structural Unification via Memory-Guided Bridging. In the entity 𝑒. If F𝑒 = ∅, we set 𝑃𝑖𝑛𝑖𝑡 (𝑒) = 0.
final phase, we project the refined Global Memory M into the Type Node Initialization via Schemas: We further initialize
Hierarchical Indexing Graph G by constructing three interconnected type nodes 𝑡 ∈ Gschema based on the retrieved schemas from Mont
graph views. Specifically, we build the Semantic Ontology Graph to avoid introducing irrelevant semantics. A critical challenge is
G𝑜𝑛𝑡 directly from M𝑜𝑛𝑡 , where nodes and edges encode schema- that type nodes often exhibit exceptionally large degrees (e.g., a
level types and their valid relations, serving as the logical backbone generic “Person” node connected to thousands of entities). Directly
of the overall structure. We then construct the Fact Graph G𝑓 𝑎𝑐 activating such high-degree nodes would spread importance across
from M 𝑓 𝑎𝑐 , where entities form nodes and instantiated triples form overly many nodes, introducing significant noise. To address this
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

Table 1: Generation performance of different GraphRAG methods. The best result is bold, and the second is underline. The
column Δ indicates the performance gain of our MemGraphRAG (59.25) compared to each baseline. Background colors in Δ
columns represent the magnitude of improvement (Darker green = larger gap).
HotpotQA 2WikiMultiHopQA MuSiQue G-Medical G-Novel Overall
Method
Str-Acc. LLM-Acc. Str-Acc. LLM-Acc. Str-Acc. LLM-Acc. LLM-Acc. LLM-Acc. Avg. Δ
Direct Zero-shot LLM Inference
Llama3-8B 30.80 28.20 34.00 16.00 6.70 7.40 26.43 15.20 20.59 +38.66
Llama3-13B 24.90 17.00 22.30 9.50 4.20 5.00 28.28 19.30 16.31 +42.94
GPT-3.5-Turbo 32.70 42.50 28.30 31.00 10.10 21.40 45.82 29.41 30.15 +29.10
GPT-4o-mini 38.10 39.70 36.00 31.30 14.10 15.20 42.13 31.42 30.99 +28.26
Vanilla Retrieval-Augmented-Generation
Retrieval (Top-1) 48.80 50.40 38.10 34.00 19.90 23.80 50.90 43.94 38.73 +20.52
Retrieval (Top-3) 55.40 58.90 46.50 41.80 26.60 28.50 55.16 46.06 44.87 +14.38
Retrieval (Top-5) 58.50 60.30 49.80 45.40 28.30 32.00 61.07 48.35 47.97 +11.28
Graph-based Retrieval-Augmented-Generation Methods
KGP [51] 62.70 62.10 33.10 32.70 28.40 32.50 56.29 49.01 44.60 +14.65
G-retriever [24] 44.00 41.80 47.80 29.70 16.20 17.60 52.40 45.90 36.93 +22.32
RAPTOR [44] 57.00 61.00 51.70 43.60 24.70 28.90 57.88 44.24 46.13 +13.12
MS-GraphRAG [12] 51.60 43.50 47.30 38.60 20.60 23.70 55.67 50.43 41.43 +17.82
LazyGraphRAG [8] 52.70 43.80 46.80 37.90 21.50 24.80 56.63 51.56 41.96 +17.29
LightRAG [17] 61.40 62.00 56.90 40.50 28.60 30.50 56.42 46.09 47.80 +11.45
HippoRAG [19] 58.40 61.40 67.50 61.30 30.40 26.00 57.06 45.77 50.98 +8.27
HippoRAG2 [20] 65.20 67.20 64.20 57.90 32.20 38.30 64.85 56.48 55.79 +3.46
E2 GraphRAG [64] 63.10 65.70 57.20 40.90 26.10 29.00 60.24 54.28 49.57 +9.68
GFM-RAG [37] 64.10 67.70 69.10 61.10 32.50 36.10 58.19 53.39 55.27 +3.98
LogicRAG [6] 55.80 65.60 64.80 63.40 30.10 34.60 56.75 49.84 52.61 +6.64
LinearRAG [70] 65.30 67.30 70.20 65.70 33.20 37.20 65.70 52.57 57.15 +2.10
MemGraphRAG (Ours) 67.20 71.60 70.30 69.80 34.40 37.90 68.40 57.41 59.25 –

issue, we introduce a structural regularization term that combines 4.3.3 Personalized PageRank. After initialization, we run Person-
semantic relevance with a log-degree penalty: alized PageRank (PPR) on the heterogeneous graph to propagate
query-specific importance. Starting from the normalized distribu-
tion v (0) , the iteration is defined as v (𝑘+1) = (1 − 𝜆)Wv (𝑘 ) + 𝜆v (0) ,
© 1 ∑︁ 1 where W denotes the transition matrix and 𝜆 is the damping factor.
𝑃𝑖𝑛𝑖𝑡 (𝑡 ) = ­ Sim(q, s) ® ×
ª
(7)
| S𝑡 | log(deg(𝑡 ) + 1) We set 𝜆 = 0.5 to limit propagation within a local neighborhood
« 𝑠 ∈S𝑡 ¬ | {z }
| {z }
Hub Suppression and reduce semantic drift. After convergence, we select the top-𝐾
Schema Relevance
passages and top-𝑀 entities ranked by v (∞) for LLM inference.
where S𝑡 denotes the retrieved schemas associated with 𝑡. This
design incorporates schema-level relevance while preventing overly 5 Experiments
generic types from dominating propagation. In this section, our aim is to answer the following questions: Q1
Passage Node Initialization via Information Density: Finally, (Generation Accuracy): How does MemGraphRAG perform com-
we initialize the Passage Nodes (𝑝 ∈ 𝐺 𝑝𝑎𝑠 ) by combining semantic pared to state-of-the-art GraphRAG methods in terms of genera-
relevance with an information density prior: tion performance? Q2 (Retrieval Analysis): How does our retrieval
Í ! method compare to other frameworks in terms of performance and
𝑒 ∈E𝑝 IDF(𝑒 )
𝑃𝑖𝑛𝑖𝑡 (𝑝 ) = Sim(q, d𝑝 ) × 𝛼 × 𝜎 (8) efficiency?Q3 (Graph Adaptability Analysis): Can the graph con-
log( | E𝑝 | + 1) structed by MemGraphRAG generalize to other GraphRAG meth-
| {z } ods? Q4 (Ablation Study): What contribution does each component
Information Density Term
of MemGraphRAG make to the overall performance? (Note that ad-
This scoring function combines semantic alignment Sim(q, d𝑝 ), a ditional experiments and case studies are provided in Appendix A.)
dampening factor 𝛼 (set to 0.05) to prevent passage nodes from dom-
inating propagation, and an Information Density Term that favors 5.1 Experimental Setting
passages containing rare and informative entities by aggregating Datasets. We first evaluate the effectiveness of MemGraphRAG on
their IDF scores with log-normalization. Detailed initialization pro- three widely-used multi-hop QA datasets, including HotpotQA [59],
cedures are provided in Appendix E.1. 2WikiMultiHopQA (2Wiki) [25], MuSiQue [47]. We follow the
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

Table 2: Retrieval performance of different GraphRAG methods on G-Bench(Medical).

Fact Retrieval Complex Reasoning Contextual Creative Gen

Method Retrieval Time
Recall Relevance Recall Relevance Recall Relevance Recall Relevance
RAPTOR [44] 85.40 69.38 89.70 53.20 88.86 58.73 72.70 52.71 0.171
Lazy-GraphRAG [8] 74.29 19.90 78.65 17.50 78.72 21.35 83.41 15.09 9.835
LightRAG [17] 80.32 41.27 82.91 42.79 85.71 43.11 81.34 45.17 11.052
HippoRAG [19] 87.25 52.44 83.80 42.19 83.46 49.13 81.66 45.03 1.586
HippoRAG2 [20] 78.70 87.96 77.00 80.94 77.40 86.85 61.12 78.64 2.157
GFM-RAG [37] 90.08 57.90 85.03 33.06 78.62 40.14 83.51 22.87 1.375
LinearRAG [70] 88.86 86.09 87.03 81.58 89.13 87.89 89.08 72.74 0.123
MemGraphRAG(ours) 89.56 88.53 90.42 82.64 89.57 86.91 89.86 79.12 0.061

settings used in [19, 20] for a fair comparison, choosing 1,000 MemGraphRAG across four benchmark datasets. The detailed ex-
questions from each validation set. We also test our approach perimental results are presented in Table 1. Based on our analysis,
on G-Bench(Medical) and G-bench(Novel) [52] to evaluate Mem- we derive the following key observations.
GraphRAG on complex reasoning across medical, novel knowledge. RAG system significantly enhances the LLM generation
More details about datasets can be found in Appendix F. performance. Direct inference (without retrieval) yields the lowest
Baselines. We categorize all baselines into three groups: (i) Zero- scores across all benchmarks. For instance, GPT-4o-mini achieves a
shot LLM Inference: We evaluate several foundational models in- mere 14.65% average accuracy on MuSiQue in a zero-retrieval set-
cluding LLaMA3 (8B) and LLaMA3 (13B) [11], as well as GPT-3.5- ting. Integrating retrieved contexts via Vanilla RAG (top-5) doubles
turbo and GPT-4o-mini [40]. (ii) We deploy Vanilla RAG across mul- this performance to 30.15%. This confirms that retrieval augmenta-
tiple retrieval configurations (retrieving 1, 3, or 5 top passages). (iii) tion is essential for knowledge-intensive tasks.
State-of-the-art GraphRAG Systems: We compare against leading Graph-based retrieval is more effective for multi-hop rea-
GraphRAG implementations, including KGP [51], G-retriever [24], soning. While increasing the retrieval count (𝑘) improves Vanilla
LightRAG [17], RAPTOR [44], MS-GraphRAG[12], HippoRAG [19, RAG, the performance gains quickly plateau. This limitation stems
20], GFM-RAG [37], LazyGRAG[8], E2 GraphRAG [64], LogicRAG[6] from Vanilla RAG’s reliance on surface-level keyword matching,
and LinearRAG[70]. which often overlooks the logical bridges required for multi-hop
Evaluation Metrics. We evaluate our method using four met- reasoning. In contrast, GraphRAG methods explicitly capture struc-
rics across two categories. For QA performance, following existing tural dependencies and consistently, and often deliver stronger
work[6, 70], we use: 1) String-based accuracy (Str-Acc.), which com- results. Notably, HippoRAG 2 emerges as a competitive baseline,
putes whether the gold answer is included in the generated answer achieving 38.30% and 56.48% LLM-based accuracy on MuSiQue and
after normalizing them to lowercase words, and 2) LLM-based accu- G-novel, respectively.
racy (LLM-Acc.), which lets an LLM decide whether the generated MemGraphRAG consistently surpasses existing GraphRAG
answer correctly matches the gold answer. For GraphRAG-bench, baselines. While exiting GraphRAGs attempt to align semantics
since golden answers consist of lengthy descriptive statements, we through graph structures, they are often sensitive to noise and
only evaluate using LLM-ACC. For retrieval quality assessment, we low-quality indexing introduced by solated chunk- level extraction.
adopt metrics from GraphRAG-Bench [52]: 1) Context Relevance, In contrast, MemGraphRAG mitigates these issues by providing
which measures semantic alignment between questions and re- more reliable indexing and retrieval, achieves the best results across
trieved passages, and 2) Evidence Recall, which evaluates whether all datasets. It reaches 59.25% average accuracy, yielding a 2.10%
the retrieved contents contain all the necessary information that absolute gain over the strongest baseline.
used for generating the correct answer.
Implementation Details. For consistency, all methods use the 5.3 Retrieval Analysis (Q2)
same embedding model (i.e., NV-Embed-v2 [39]). We set 𝑘=5 for To evaluate the retrieval performance of MemGraphRAG, we con-
top-𝑘 retrieval in all methods. For both offline indexing (graph ducted tests across four distinct task levels on the GraphRAG-
construction) and online generation, we adopt GPT-4o-mini as the Bench. We utilized Recall and Relevance as metrics to assess the
default LLM (additional open-source LLM results are reported in GraphRAG’s capacity for retrieving both comprehensive and pre-
Appendix A). For evaluation, we use GPT-4o-mini to compute the cise information. Additionally, to assess practical deployment feasi-
LLM-Acc metric. To ensure reproducibility, we set the inference bility, we recorded the average retrieval time (in seconds) across all
temperature to 0 for all LLM calls. queries. The experimental results are presented in Table 2.
MemGraphRAG achieves consistently strong retrieval per-
formance, balancing high recall with high relevance. Mem-
5.2 Generation Accuracy (Q1) GraphRAG consistently ranks at the top in Complex Reasoning tasks
To address Q1, we conduct a comprehensive evaluation of gener- (Recall: 90.42, Relevance: 82.64) and Fact Retrieval tasks, signifi-
ation performance by comparing various baseline methods with cantly outperforming baselines such as HippoRAG2 and LightRAG.
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

Table 3: Adaptability Analysis: MemGraphRAG as a universal graph constructor across different frameworks. The blue rows
indicate experiments using MemGraphRAG’s constructed graph, while the purple row represents our full framework. The
rightmost column shows the performance gain.
GraphConstructor Retriever HotpotQA 2Wiki MuSiQue G-Medical G-Novel Average Δ
HippoRAG [19] HippoRAG [19] 59.90 64.40 28.20 57.06 45.77 51.07 +8.61
MemGraphRAG HippoRAG [19] 60.65 65.25 29.00 57.75 46.24 51.78 +7.90
HippoRAG2 [20] HippoRAG2 [20] 66.20 61.05 35.25 64.85 56.48 56.77 +2.91
MemGraphRAG HippoRAG2 [20] 66.00 61.20 35.40 65.42 56.76 56.96 +2.72
MS-GraphRAG [12] MS-GraphRAG [12] 47.55 42.95 22.15 55.67 50.43 43.75 +15.93
MemGraphRAG MS-GraphRAG [12] 48.00 43.20 22.45 56.53 50.88 44.21 +15.47
LazyGraphRAG [8] LazyGraphRAG [8] 48.25 42.35 23.15 56.63 51.56 44.39 +15.29
MemGraphRAG LazyGraphRAG [8] 48.75 42.55 23.50 57.98 52.06 44.97 +14.71
MemGraphRAG MemGraphRAG 69.40 70.05 36.15 68.40 54.41 59.68 -

These results indicate that our approach effectively filters noise

and invalid entity relationships, enabling the system to precisely
pinpoint entities and relations relevant to the query. Unlike meth-
ods that sacrifice precision for coverage, MemGraphRAG maintains
superior relevance while capturing broad context, thereby validat-
ing the effectiveness of our Global Adjudication mechanism for
consistency maintenance in constructing high-quality graphs.
MemGraphRAG achieves the lowest retrieval latency, show-
ing superior online inference efficiency. MemGraphRAG re-
quires an average of only 0.061 seconds per retrieval, which is signif-
icantly faster than LightRAG (11.052s) and HippoRAG (1.586s). This
efficiency is attributed to our lightweight retrieval process, which Figure 5: Ablation study of MemGraphRAG on three datasets.
relies on efficient Personalized PageRank (PPR) rather than compu-
tationally expensive real-time LLM filtering or iterative reasoning
loops. Consequently, MemGraphRAG delivers high-precision com- mitigates structural fragmentation and logical inconsistencies. By
plex reasoning while maintaining low latency in practice. producing a more cohesive and thematically aligned knowledge
structure, MemGraphRAG substantially strengthens the effective-
5.4 Indexing Graph Adaptability Analysis (Q3) ness of existing retrievers, demonstrating its robustness as a foun-
dational indexing solution.
To evaluate whether our constructed index graph can seamlessly
adapt to different GraphRAG frameworks, we conducted a trans-
ferability experiment. Our pilot study previously identified that 5.5 Ablation Study (Q4)
existing graph construction methods suffer from critical deficien- To verify the contribution of each module in MemGraphRAG, we
cies, including Thematic Irrelevance, Logical Inconsistency, and conduct an ablation study on HotpotQA, 2WikiMultiHopQA, and
Structural Fragmentation. Consequently, we assess whether the G-Medical. We compare the full model with four variants that
structural unification enabled by MemGraphRAG through Memory- remove the Schema Filter, Conflict Resolution, Hub Suppression, and
Guided Bridging can mitigate these issues for other frameworks. the Information Density Term, respectively. As shown in Figure 5,
Specifically, we replaced the native graph construction modules MemGraphRAG consistently achieves the best performance across
of HippoRAG, HippoRAG2, MS-GraphRAG, and LazyGraphRAG all datasets (e.g., 69.40% on HotpotQA), which indicates that these
with the graph constructed by MemGraphRAG, while retaining memory-driven graph construction and initialization mechanisms
their original downstream retrieval and reasoning mechanisms. are jointly crucial for building a robust knowledge graph.
The comparative experimental results are presented in Table 3. w/o Schema Filter: Removing Unified Schema Filtering causes a
MemGraphRAG consistently improves retrieval perfor- clear degradation, especially on 2WikiMultiHopQA and G-Medical
mance across all evaluated GraphRAG frameworks, serving (68.10% and 65.92%). Without the frequency-based stability con-
as a universal high-quality graph constructor. As shown in straint (Freq(𝑠) ≥ 𝜏), low-frequency and off-topic schemas are
the results, replacing the original graphs with MemGraphRAG- retained, introducing noisy triples that weaken semantic focus.
constructed graphs leads to consistent improvements for all baseline w/o Conflict Resolution: Excluding Global Adjudication leads
retrievers across all datasets. For instance, the average performance to the largest drop on HotpotQA (66.95%). Without conflict detec-
of HippoRAG increases from 51.07 to 51.78, and MS-GraphRAG tion and resolution, the fact layer accumulates contradictory or
improves from 43.75 to 44.21. This consistent improvement shows redundant triples, which disrupts multi-hop reasoning chains and
that our memory-driven global construction mechanism effectively increases the chance of retrieving inconsistent evidence.
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

w/o Hub Suppression: Removing Hub Suppression reduces ac- [2] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Ruther-
curacy (67.22% on HotpotQA). Without degree-based regularization, ford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bog-
dan Damoc, Aidan Clark, et al. 2022. Improving language models by retrieving
generic high-degree nodes dominate propagation, causing semantic from trillions of tokens. In International Conference on Machine Learning (ICML).
drift toward irrelevant subgraphs. [3] Mingyang Chen, Linzhuang Sun, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng
Zhu, Haofen Wang, Jeff Z. Pan, Wen Zhang, Huajun Chen, Fan Yang, Zenan
w/o Information Density Term: Dropping the Information Zhou, and Weipeng Chen. 2025. ReSearch: Learning to Reason with Search for
Density Term yields a smaller but consistent decline (68.67% on LLMs via Reinforcement Learning. arXiv:2503.19470 [[Link]] [Link]
HotpotQA). Without IDF-style weighting, passage initialization abs/2503.19470
[4] Shengyuan Chen, Zheng Yuan, Qinggang Zhang, Wen Hua, Jiannong Cao, and
cannot prioritize discriminative evidence, weakening the model’s Xiao Huang. 2025. Neuro-Symbolic Entity Alignment via Variational Inference.
ability to anchor reasoning on informative documents. The Thirty-ninth Annual Conference on Neural Information Processing Systems
(2025).
[5] Shengyuan Chen, Qinggang Zhang, Junnan Dong, Wen Hua, Qing Li, and Xiao
6 Conclusion Huang. 2024. Entity alignment with noisy annotations from large language
In this paper, we propose MemGraphRAG, a novel GraphRAG models. The Thirty-Eighth Annual Conference on Neural Information Processing
Systems (2024).
framework that integrates a global memory mechanism into the [6] Shengyuan Chen, Chuang Zhou, Zheng Yuan, Qinggang Zhang, Zeyang Cui, Hao
knowledge graph construction process. By leveraging a shared hier- Chen, Yilin Xiao, Jiannong Cao, and Xiao Huang. 2025. You Don’t Need Pre-built
Graphs for RAG: Retrieval Augmented Generation with Adaptive Reasoning
archical memory structure, our multi-agent system collaboratively Structures. arXiv preprint arXiv:2508.06105 (2025).
maintains a global perspective throughout both the extraction and [7] CircleMind-AI. 2024. FastGraphRAG: High-speed graph-based retrieval-
retrieval phases. This paradigm effectively overcomes key limita- augmented generation. CircleMind-AI Blog (2024).
[8] Jonathan Larson Darren Edge, Ha Trinh. 2024. LazyGraphRAG: Setting a new
tions of traditional GraphRAG approaches that rely on isolated standard for quality and cost. Microsoft Blog (2024).
local extraction. It systematically mitigates thematic irrelevance, [9] Junnan Dong, Siyu An, Yifei Yu, Qian-Wen Zhang, Linhao Luo, Xiao Huang,
logical inconsistency, and structural fragmentation, thereby en- Yunsheng Wu, Di Yin, and Xing Sun. 2025. Youtu-GraphRAG: Vertically Unified
Agents for Graph Retrieval-Augmented Complex Reasoning. arXiv:2508.19855
abling a globally consistent indexing graph. Extensive experiments [Link]
demonstrate that MemGraphRAG consistently outperforms state- [10] Su Dong, Qinggang Zhang, Yilin Xiao, Shengyuan Chen, Chuang Zhou, and Xiao
Huang. 2026. Use Graph When It Needs: Efficiently and Adaptively Integrating
of-the-art baselines in terms of graph quality, retrieval precision, Retrieval-Augmented Generation with Graphs. arXiv preprint arXiv:2602.03578
and generation accuracy, providing a robust solution for deploying (2026).
reliable RAG systems in complex real-world scenarios. [11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad
Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan,
et al. 2024. The llama 3 herd of models. arXiv e-prints (2024), arXiv–2407.
Limitation [12] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva
Mody, Steven Truitt, and Jonathan Larson. 2024. From local to global: A graph
While MemGraphRAG demonstrates strong robustness in process- rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130
ing large-scale textual corpora and constructing globally consistent (2024).
knowledge graphs, its current design is limited to unimodal textual [13] Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang,
Xiangnan He, and Tat-Seng Chua. 2024. Alphaedit: Null-space constrained
inputs. However, real-world knowledge repositories are inherently knowledge editing for language models. arXiv preprint arXiv:2410.02355 (2024).
multimodal, containing heterogeneous formats such as statistical [14] Junfeng Fang, Yukai Wang, Ruipeng Wang, Zijun Yao, Kun Wang, An Zhang,
Xiang Wang, and Tat-Seng Chua. 2025. Safemlrm: Demystifying safety in multi-
charts, technical diagrams, document layouts, and embedded im- modal large reasoning models. arXiv preprint arXiv:2504.08813 (2025).
ages in academic papers or financial reports. Currently, our frame- [15] Linfeng Gao, Baolong Bi, Zheng Yuan, Le Wang, Zerui Chen, Zhimin Wei,
work requires non-textual elements to be transcribed or described Shenghua Liu, Qinggang Zhang, and Jinsong Su. 2025. Probing Latent Knowl-
edge Conflict for Faithful Retrieval-Augmented Generation. arXiv preprint
in text before processing, which may lead to the loss of critical vi- arXiv:2510.12460 (2025).
sual semantics and spatial relationships. For example, quantitative [16] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai,
trends in line charts or complex structures in scientific diagrams Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large
language models: A survey. arXiv preprint arXiv:2312.10997 (2023).
often contain dense information that textual descriptions cannot [17] Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. LightRAG:
fully capture, potentially causing information loss during indexing. Simple and Fast Retrieval-Augmented Generation. arXiv preprint arXiv:2410.05779
(2024).
Extending the Global Hierarchical Graph to incorporate multimodal [18] Anton Gusarov, Anastasia Volkova, Valentin Khrulkov, Andrey Kuznetsov, Ev-
nodes (e.g., embedding visual patches into the Fact Layer M 𝑓 𝑎𝑐 or genii Maslov, and Ivan Oseledets. 2025. Multi-Agent GraphRAG: A Text-to-
the Passage Layer M𝑝𝑎𝑠 ) is a promising direction for future work. Cypher Framework for Labeled Property Graphs. arXiv:2511.08274 [[Link]]
[Link]
Such an extension could enable cross-modal reasoning, allowing the [19] Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su.
multi-agent system to verify textual claims against visual evidence 2024. HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Lan-
and further improve the versatility of MemGraphRAG. guage Models. In Advances in Neural Information Processing Systems (NeurIPS).
[20] Bernal Jiménez Gutiérrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. 2025.
From rag to memory: Non-parametric continual learning for large language
Acknowledgments models. arXiv preprint arXiv:2502.14802 (2025).
[21] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei
The project was supported by Natural Science Foundation of Fujian Chang. 2020. REALM: Retrieval-Augmented Language Model Pre-Training.
Province of China (No. 2024J011001) and the Public Technology arXiv:2002.08909 [[Link]] [Link]
[22] Haoyu Han, Harry Shomer, Yu Wang, Yongjia Lei, Kai Guo, Zhigang Hua, Bo
Service Platform Project of Xiamen (No.3502Z20231043). We also Long, Hui Liu, and Jiliang Tang. 2025. Rag vs. graphrag: A systematic evaluation
thank the reviewers for their insightful comments. and key insights. arXiv preprint arXiv:2502.11371 (2025).
[23] Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, Ma-
References hantesh Halappanavar, Ryan A Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al.
2024. Retrieval-augmented generation with graphs (graphrag). arXiv preprint
[1] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. arXiv:2501.00309 (2024).
Self-rag: Learning to retrieve, generate, and critique through self-reflection. In [24] Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh V Chawla, Thomas Laurent, Yann
International Conference on Learning Representations (ICLR). LeCun, Xavier Bresson, and Bryan Hooi. 2024. G-retriever: Retrieval-augmented
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

generation for textual graph understanding and question answering. arXiv Transactions of the Association for Computational Linguistics 10 (2022), 539–554.
preprint arXiv:2402.07630 (2024). [48] Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal.
[25] Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. 2023. Interleaving retrieval with chain-of-thought reasoning for knowledge-
Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning intensive multi-step questions. In Proceedings of the 61st annual meeting of the
steps. arXiv preprint arXiv:2011.01060 (2020). association for computational linguistics (volume 1: long papers). 10014–10037.
[26] Zijin Hong, Zheng Yuan, Qinggang Zhang, Hao Chen, Junnan Dong, Feiran [49] Hong Ting Tsang, Jiaxin Bai, Haoyu Huang, Qiao Xiao, Tianshi Zheng, Baixuan
Huang, and Xiao Huang. 2024. Next-Generation Database Interfaces: A Survey Xu, Shujie Liu, and Yangqiu Song. 2025. AutoGraph-R1: End-to-End Reinforce-
of LLM-based Text-to-SQL. arXiv preprint arXiv:2406.08426 (2024). ment Learning for Knowledge Graph Construction. arXiv:2510.15339 [[Link]]
[27] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo [Link]
Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. [50] Shu Wang, Yixiang Fang, Yingli Zhou, Xilin Liu, and Yuchi Ma. 2025. ArchRAG:
2023. Atlas: Few-shot learning with retrieval augmented language models. The Attributed Community-based Hierarchical Retrieval-Augmented Generation.
Journal of Machine Learning Research (JMLR) (2023). arXiv preprint arXiv:2502.09891 (2025).
[28] Houcheng Jiang, Junfeng Fang, Ningyu Zhang, Guojun Ma, Mingyang Wan, Xiang [51] Yu Wang, Nedim Lipka, Ryan A Rossi, Alexa Siu, Ruiyi Zhang, and Tyler Derr.
Wang, Xiangnan He, and Tat-seng Chua. 2025. AnyEdit: Edit Any Knowledge 2024. Knowledge graph prompting for multi-document question answering. In
Encoded in Language Models. ICML (2025). Conference on Artificial Intelligence (AAAI).
[29] Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, [52] Zhishang Xiang, Chuanjie Wu, Qinggang Zhang, Shengyuan Chen, Zijin Hong,
Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Active retrieval augmented Xiao Huang, and Jinsong Su. 2025. When to use graphs in rag: A compre-
generation. In Empirical Methods in Natural Language Processing (EMNLP). hensive analysis for graph retrieval-augmented generation. arXiv preprint
[30] Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, arXiv:2506.05690 (2025).
Hamed Zamani, and Jiawei Han. 2025. Search-R1: Training LLMs to Reason and [53] Zhishang Xiang, Chengyi Yang, Zerui Chen, Zhimin Wei, Yunbo Tang, Zongpei
Leverage Search Engines with Reinforcement Learning. arXiv:2503.09516 [[Link]] Teng, Zexi Peng, Zongxia Li, Chengsong Huang, Yicheng He, et al. 2026. A
[Link] Systematic Survey of Self-Evolving Agents: From Model-Centric to Environment-
[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Driven Co-Evolution. (2026).
Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. [54] Yilin Xiao, Chuang Zhou, Qinggang Zhang, Su Dong, Shengyuan Chen, and Xiao
2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Huang. 2025. LAG: Logic-Augmented Generation from a Cartesian Perspective.
Advances in Neural Information Processing Systems (NeurIPS). arXiv preprint arXiv:2508.05509 (2025).
[32] Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian [55] Yilin Xiao, Chuang Zhou, Qinggang Zhang, Bo Li, Qing Li, and Xiao Huang. 2025.
Zhang, and Zhicheng Dou. 2025. Search-o1: Agentic Search-Enhanced Large Reliable Reasoning Path: Distilling Effective Guidance for LLM Reasoning with
Reasoning Models. arXiv:2501.05366 [[Link]] [Link] Knowledge Graphs. arXiv:2506.10508 [[Link]]
[33] Lei Liang, Mengshu Sun, Zhengke Gui, Zhongshu Zhu, Zhouyu Jiang, Ling Zhong, [56] Cehao Yang, Xiaojun Wu, Xueyuan Lin, Chengjin Xu, Xuhui Jiang, Yuanliang Sun,
Yuan Qu, Peilong Zhao, Zhongpu Bo, Jin Yang, et al. 2024. Kag: Boosting llms Jia Li, Hui Xiong, and Jian Guo. 2025. GraphSearch: An Agentic Deep Searching
in professional domains via knowledge augmented generation. arXiv preprint Workflow for Graph Retrieval-Augmented Generation. arXiv:2509.22009 [[Link]]
arXiv:2409.13731 (2024). [Link]
[34] Yujie Lin, Kunquan Li, YiXuan Liao, Xiaoxin Chen, and Jinsong Su. 2026. Bi- [57] Chang Yang, Chuang Zhou, Yilin Xiao, Su Dong, Luyao Zhuang, Yujing Zhang,
directional Bias Attribution: Debiasing Large Language Models without Modify- Zhu Wang, Zijin Hong, Zheng Yuan, Zhishang Xiang, et al. 2026. Graph-
ing Prompts. In The Fourteenth International Conference on Learning Representa- based Agent Memory: Taxonomy, Techniques, and Applications. arXiv preprint
tions. [Link] arXiv:2602.05665 (2026).
[35] Yujie Lin, Chengyi Yang, Zhishang Xiang, Yiping Song, and Jinsong Su. 2026. [58] Diji Yang, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Jie Yang, and
ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models. Yi Zhang. 2024. Im-rag: Multi-round retrieval-augmented generation through
arXiv:2605.18879 [[Link]] [Link] learning inner monologues. In Proceedings of the 47th International ACM SIGIR
[36] LINHAO LUO, Yuan-Fang Li, Reza Haf, and Shirui Pan. 2024. Reasoning on Conference on Research and Development in Information Retrieval. 730–740.
Graphs: Faithful and Interpretable Large Language Model Reasoning. In The [59] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan
Twelfth International Conference on Learning Representations. Salakhutdinov, and Christopher D Manning. 2018. HotpotQA: A dataset for di-
[37] Linhao Luo, Zicheng Zhao, Gholamreza Haffari, Dinh Phung, Chen Gong, and verse, explainable multi-hop question answering. In Empirical Methods in Natural
Shirui Pan. 2025. GFM-RAG: graph foundation model for retrieval augmented Language Processing (EMNLP).
generation. arXiv preprint arXiv:2502.01113 (2025). [60] Zheng Yuan, Hao Chen, Zijin Hong, Qinggang Zhang, Feiran Huang, Qing Li, and
[38] Renqiang Luo, Huafei Huang, Shuo Yu, Fengqi Yu, Feng Xia, Sajal K. Das, and Xiao Huang. 2025. Knapsack optimization-based schema linking for llm-based
Chengqi Zhang. 2026. Utility-Preserving Federated Graph Learning with Dual- Text-to-SQL generation. arXiv preprint arXiv:2502.12911 (2025).
Perspective Fairness. IEEE Transactions on Pattern Analysis and Machine Intelli- [61] Fangyuan Zhang, Zhengjun Huang, Yingli Zhou, Qintian Guo, Zhixun Li, Wen-
gence (2026). sheng Luo, Di Jiang, Yixiang Fang, and Xiaofang Zhou. 2025. EraRAG: Efficient
[39] Gabriel de Souza P Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt and Incremental Retrieval Augmented Generation for Growing Corpora. arXiv
Schifferer, and Even Oldridge. 2024. NV-Retriever: Improving text embedding preprint arXiv:2506.20963 (2025).
models with effective hard-negative mining. arXiv preprint arXiv:2407.15831 [62] Qinggang Zhang, Shengyuan Chen, Yuanchen Bei, Zheng Yuan, Huachi Zhou,
(2024). Zijin Hong, Junnan Dong, Hao Chen, Yi Chang, and Xiao Huang. 2025. A Sur-
[40] OpenAI. 2023. GPT-4 Technical Report. OpenAI Blog (2023). vey of Graph Retrieval-Augmented Generation for Customized Large Language
[41] Tyler Thomas Procko and Omar Ochoa. 2024. Graph retrieval-augmented genera- Models. arXiv preprint arXiv:2501.13958 (2025).
tion for large language models: A survey. In Conference on AI, Science, Engineering, [63] Qinggang Zhang, Zhishang Xiang, Yilin Xiao, Le Wang, Junhui Li, Xinrun Wang,
and Technology (AIxSET). and Jinsong Su. 2025. FaithfulRAG: Fact-Level Conflict Modeling for Context-
[42] Hongjin Qian, Zheng Liu, Peitian Zhang, Kelong Mao, Defu Lian, Zhicheng Dou, Faithful Retrieval-Augmented Generation. arXiv preprint arXiv:2506.08938 (2025).
and Tiejun Huang. 2025. MemoRAG: Boosting Long Context Processing with [64] Yibo Zhao, Jiapeng Zhu, Ye Guo, Kangkang He, and Xiang Li. 2025. Eˆ 2GraphRAG:
Global Memory-Enhanced Retrieval Augmentation. arXiv:2409.05591 [[Link]] Streamlining Graph-based RAG for High Efficiency and Effectiveness. arXiv
[Link] preprint arXiv:2505.24226 (2025).
[43] Meng Qu and Jian Tang. 2019. Probabilistic Logic Neural Networks for Reason- [65] Baolin Zheng, Guanlin Chen, Hongqiong Zhong, Qingyang Teng, Yingshui Tan,
ing. In Advances in Neural Information Processing Systems (NeurIPS). Vancouver, Zhendong Liu, Weixun Wang, Jiaheng Liu, Jian Yang, Huiyun Jing, et al. 2025.
Canada, 7710–7720. USB: A Comprehensive and Unified Safety Evaluation Benchmark for Multimodal
[44] Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Large Language Models. arXiv preprint arXiv:2505.23793 (2025).
Christopher D. Manning. 2024. RAPTOR: Recursive Abstractive Processing for [66] Qihuang Zhong, Haiyun Li, Luyao Zhuang, Juhua Liu, and Bo Du. 2024. Iterative
Tree-Organized Retrieval. In International Conference on Learning Representations data generation with large language models for aspect-based sentiment analysis.
(ICLR). arXiv preprint arXiv:2407.00341 (2024).
[45] Kartik Sharma, Peeyush Kumar, and Yunqing Li. 2024. OG-RAG: Ontology- [67] Chulun Zhou, Qiujing Wang, Mo Yu, Xiaoqian Yue, Rui Lu, Jiangnan Li, Yifan
Grounded Retrieval-Augmented Generation For Large Language Models. arXiv Zhou, Shunchi Zhang, Jie Zhou, and Wai Lam. 2025. The essence of contex-
preprint arXiv:2412.15235 (2024). tual understanding in theory of mind: A study on question answering with
[46] Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun story characters. In Proceedings of the 63rd Annual Meeting of the Association for
Gong, Lionel Ni, Heung-Yeung Shum, and Jian Guo. 2024. Think-on-Graph: Deep Computational Linguistics (Volume 1: Long Papers). 22612–22631.
and Responsible Reasoning of Large Language Model on Knowledge Graph. In [68] Chulun Zhou, Chunkang Zhang, Guoxin Yu, Fandong Meng, Jie Zhou, Wai Lam,
International Conference on Learning Representations (ICLR). and Mo Yu. 2025. Improving Multi-step RAG with Hypergraph-based Memory
[47] Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. for Long-Context Complex Relational Modeling. arXiv preprint arXiv:2512.23959
2022. MuSiQue: Multi-hop Questions via Single-hop Question Composition. (2025).
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

[69] Yingli Zhou, Yaodong Su, Youran Sun, Shu Wang, Taotao Wang, Runyuan He, This improvement indicates that our memory consistency mainte-
Yongwei Zhang, Sicong Liang, Xilin Liu, Yuchi Ma, et al. 2025. In-depth Analysis nance mechanism effectively links entities scattered across different
of Graph-based RAG in a Unified Framework. arXiv preprint arXiv:2503.04338
(2025). document chunks. As a result, it bridges fragmented subgraphs and
[70] Luyao Zhuang, Shengyuan Chen, Yilin Xiao, Huachi Zhou, Yujing Zhang, Hao enables more robust long-range reasoning paths.
Chen, Qinggang Zhang, and Xiao Huang. 2025. LinearRAG: Linear Graph
Retrieval Augmented Generation on Large-scale Corpora. arXiv preprint
MemGraphRAG demonstrates superior subgraph-level se-
arXiv:2510.10114 (2025). mantic clustering than existing GraphRAG methods. Mem-
GraphRAG also attains the highest Average Clustering Coefficients,
A Additional Experiments with 0.865 on the G-Novel and 0.527 on the G-Medical. These re-
sults indicate that nodes in MemGraphRAG tend to share common
A.1 Ablation on Backbone LLMs neighbors, leading to denser local connectivity and clearer seman-
To further evaluate the universality and robustness of MemGraphRAG, tic clusters. This further shows that MemGraphRAG integrates
we conducted experiments utilizing the stronger llama-3-70b- dispersed knowledge into a more unified and highly structured
instruct as the underlying backbone model. We compared our index graph, instead of yielding sparse graphs composed of loosely
method against a comprehensive suite of baselines, ranging from related facts.
non-structured methods (e.g., Vanilla RAG) to state-of-the-art graph-
based approaches (e.g., HippoRAG2, E2GraphRAG). The results are
detailed in Table 4.
MemGraphRAG consistently achieves state-of-the-art per-
formance across all evaluated datasets, highlighting its com-
patibility and robustness across different backbone models.
As shown in the table, MemGraphRAG achieves the highest average
performance of 58.41%, significantly outperforming the strongest
baseline, HippoRAG2 (55.41%), and surpassing standard graph-
based methods like LightRAG (47.81%) by a substantial margin.
First, compared to non-structured methods, our approach exhibits
a dominant advantage over Vanilla RAG (Top-5 average: 47.52%),
validating that our memory-driven graph structure effectively cap-
tures long-range dependencies that vector retrieval misses. Second,
in the realm of graph-based RAG, MemGraphRAG excels partic-
ularly in multi-hop reasoning tasks. On the 2WikiMultiHopQA
dataset, we achieve a Containment Accuracy of 69.40% and an LLM
Accuracy of 66.80%, notably higher than HippoRAG2 (61.90% and Figure 6: Multi-dimensional assessment of graph quality.
54.40%, respectively). This indicates that our method constructs
a more connected and logically coherent graph, enabling the re-
triever to accurately locate multi-hop evidence chains. Furthermore, A.3 Case Study
on domain-specific datasets like G-Medical, MemGraphRAG main-
We conduct a qualitative analysis in Table 6 and Table 7 to illustrate
tains its lead (67.13%), proving its robustness in handling specialized
how MemGraphRAG overcomes the limitations of isolated extrac-
knowledge. Collectively, these results confirm that MemGraphRAG
tion by ensuring logical consistency and thematic purity through
provides a high-quality, globally consistent indexing structure that
its global memory mechanism.
universally enhances the reasoning capabilities of LLMs.
1) Case Study on Conflict Resolution. Table 6 illustrates a
representative scenario of Mutually Exclusive Conflict, where dis-
A.2 Graph Analysis
parate documents claim conflicting birth years for the same entity
To more intuitively assess the quality of the index graphs produced ("1645" vs. "1643"). Traditional pipelines simply aggregate these con-
by our memory-based construction approach, we analyze their tradictions, leading to ambiguous reasoning paths. MemGraphRAG
topological properties and compare MemGraphRAG with existing addresses this through Global Adjudication. Upon detecting the
baselines in terms of connectivity, redundancy, and semantic ag- conflict, the Resolution Agent (𝐴𝑟𝑒𝑠 ) retrieves the original prove-
gregation. Following previous study [52], we assessed the Average nance from the Passage Layer (𝑀𝑝𝑎𝑠 ) and validates the correct fact
Degree and Average Clustering Coefficient of the index graphs ("1643") before indexing. This mechanism effectively eliminates
constructed by various GraphRAG frameworks on the G-Medical logical incoherence, enabling the retriever to provide an accurate
and G-Novel datasets. The comparative results are presented in context for the LLM.
Table 5 and Figure 6. 2) Case Study on Thematic Denoising. In domain-specific
MemGraphRAG demonstrates superior entity-level con- tasks (e.g., medical protocols), LLMs often extract irrelevant noise
nectivity compared to existing GraphRAG methods. Mem- alongside core facts. As shown in Table 6, the baseline graph is pol-
GraphRAG achieves the highest Average Degree on both datasets, luted by irrelevant triples (e.g., Patient prefers Tea), which distracts
reaching 14.37 on the Medical dataset (surpassing HippoRAG2’s the retrieval process. MemGraphRAG mitigates this via Unified
13.31) and 9.26 on the Novel dataset (surpassing HippoRAG2’s 8.75). Schema Filtering. By treating extracted schemas as candidate and
only stabilizing those that exceed a frequency threshold (𝜏), our
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

Table 4: Comparison of different methods. The column Δ shows the improvement of MemGraphRAG (58.41) over baselines.
Darker green in Δ indicates a larger performance gap.

HotpotQA 2WikiMultiHopQA MuSiQue G-Medical G-Novel Overall

Method
Contain-Acc. LLM-Acc. Contain-Acc LLM-Acc Contain-Acc LLM-Acc LLM-Acc LLM-Acc Avg. Δ
Non-structure Methods
llama-70B-instruct 38.20 40.50 33.80 29.30 14.40 16.50 39.31 29.73 30.22 +28.19
Vanilla RAG (Top-1) 48.40 51.70 40.20 34.40 22.20 23.10 52.73 45.13 39.73 +18.68
Vanilla RAG (Top-3) 54.50 55.10 46.60 39.20 29.50 31.70 55.38 48.90 45.11 +13.30
Vanilla RAG (Top-5) 56.00 58.10 53.20 46.70 28.20 32.30 59.81 45.83 47.52 +10.89
Graph-based RAG Methods
KGP 63.20 62.10 34.70 32.60 24.80 30.20 56.94 47.04 43.95 +14.46
G-retriever 44.70 43.70 50.50 30.80 18.90 19.80 50.77 43.09 37.78 +20.63
RAPTOR 57.10 59.70 54.70 45.00 26.50 32.10 57.63 42.83 46.95 +11.46
MS-GraphRAG 49.90 42.70 50.00 39.80 17.60 22.20 53.22 47.71 40.39 +18.02
LazyGraphRAG 50.80 41.40 46.60 36.80 23.10 27.70 58.79 48.63 41.73 +16.68
LightRAG 64.30 63.10 55.30 40.70 28.00 27.70 58.68 44.67 47.81 +10.60
HippoRAG 59.50 64.00 69.00 63.80 31.80 28.40 57.30 47.44 52.66 +5.75
HippoRAG2 64.90 67.30 61.90 54.40 33.90 37.60 67.11 56.16 55.41 +3.00
𝐸 2 GraphRAG 61.50 65.70 58.60 40.50 24.20 27.40 62.47 53.82 49.27 +9.14
MemGraphRAG(ours) 65.60 69.40 69.40 66.80 33.70 39.50 67.13 55.76 58.41 –

Table 5: Quality evaluation of indexing graph construction at inference time to enable adaptive retrieval planning. While ef-
in GraphRAG frameworks. fective, these methods typically operate within the constraints of
fixed resources or rely on the LLM’s inherent reasoning capabilities
G-Novel G-Medical HotpotQA rather than structured knowledge representation.
Method
Degree Clust. Coeff Degree Clust. Coeff Degree Clust. Coeff
MS-GraphRAG [12]
HippoRAG2 [20]
1.48
8.75
0.315
0.657
1.82
13.31
0.300
0.497
1.56
7.96
0.334
0.613
B.2 Graph Retrieval-Augmented Generation
LightRAG [17] 2.10 0.212 2.58 0.139 2.18 0.236
Fast-GraphRAG [7] 3.19 0.324 5.50 0.347 3.04 0.336 To overcome the limitations of unstructured text chunks, GraphRAG
HippoRAG [19] 1.73 0.100 2.06 0.087 1.86 0.140 focuses on explicit graph structure construction to capture global
MemGraphRAG(ours) 9.26 0.865 14.37 0.527 8.92 0.725
dependencies and structural patterns. Current approaches can be
categorized into two primary construction paradigms:
system successfully filters out irrelevant noise while retaining sta- Relation-extraction-based Construction. This line of work [6,
𝑡𝑟𝑒𝑎𝑡𝑠 17–19, 38, 49, 55, 56, 64, 69] structures text corpora into Knowledge
ble clinical patterns (e.g., 𝐷𝑟𝑢𝑔 −−−−−→ 𝐷𝑖𝑠𝑒𝑎𝑠𝑒). This results in a
Graphs (KGs) by extracting triples to form atomic knowledge units.
cleaner Fact Graph (𝐺 𝑓 𝑎𝑐 ) that strictly follows the domain ontology,
These units are subsequently unified via entity alignment [4, 5],
significantly improving retrieval precision.
enabling the application of sophisticated graph reasoning algo-
B Related Work rithms [36, 43, 46]. Some methods augment reasoning by integrating
these static KGs as navigational aids, such as Think-on-Graph [46]
B.1 Retrieval-Augmented Generation and RRP [55]. However, independent OpenIE extraction often leads
While Large Language Models (LLMs) have demonstrated impres- to inconsistency. Although schema-guided approaches [33, 45] at-
sive capabilities, they remain prone to hallucination [10, 13–15, 26, tempt to standardize this, they entail high manual costs. Addressing
28, 34, 35, 60, 65, 66]. Retrieval-Augmented Generation (RAG) miti- these inefficiencies, LinearRAG [70] proposes a relation-free “Tri-
gates this by grounding generation in external evidence [2, 21, 27, Graph” based on lightweight entity extraction, achieving linear
42, 61, 67, 68]. However, effectively organizing fragmented knowl- scalability without the noise associated with traditional triple ex-
edge from distributed documents to support complex reasoning traction.
remains a persistent challenge. Clustering-based Hierarchy Construction. Complementary
To address this, recent research has evolved from simple retrieval to triple-based methods, this category focuses on capturing global
to Reasoning-enhanced RAG [1, 3, 30, 32, 53]. Departing from static information by identifying dense structural patterns. Methods typi-
index construction, this paradigm focuses on interleaving the re- cally employ community detection algorithms, such as Louvain or
trieval process with the logical flow of the LLM. Several approaches Leiden, to recursively aggregate entities into clusters [12, 19, 44].
optimize the retrieval process through Chain-of-Thought prompt- These clusters serve as hierarchical summaries, abstracting raw
ing, recursive inner monologues, or logical decomposition, such as passages into topic-level communities to provide a macro-level per-
IRCoT [48], IM-RAG [58], and LAG [54]. LogicRAG [6] advances spective. Despite its utility in summarizing high-level themes, this
this direction by eliminating pre-built graphs entirely, instead con- unsupervised approach faces limitations regarding precision, as in-
structing a reasoning Directed Acyclic Graph (DAG) dynamically accuracies in low-level entity relationships can propagate upward,
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

Table 6: Case Study: Resolving Logic Conflicts via Global Adjudication. Comparing how MemGraphRAG handles contradictory
birth years across documents versus a Traditional GraphRAG baseline.

Pipeline Phase Traditional GraphRAG (Baseline) MemGraphRAG (Ours)

1. Input Corpus Doc A: “Newton was born in 1645.” Same Corpus: Contains mutually exclusive facts due to
Doc B: “Isaac Newton, born 1643...” source errors or extraction noise.
2. Graph Construction Isolated Extraction: Global Adjudication:
𝑇1 : (𝑁 𝑒𝑤𝑡𝑜𝑛, 𝑏𝑜𝑟𝑛_𝑖𝑛, 1645) 𝐴𝑑𝑒𝑡 detects Conflict: 𝑇1 ⊥ 𝑇2
𝑇2 : (𝑁 𝑒𝑤𝑡𝑜𝑛, 𝑏𝑜𝑟𝑛_𝑖𝑛, 1643) → 𝐴𝑟𝑒𝑠 checks Evidence (𝑀𝑝𝑎𝑠 )
→ Both edges added to Graph 𝐺. → Update: Keep 𝑇2 , Discard 𝑇1 .
3. Retrieval Query Q: “When was Isaac Newton born?”
4. Retrieval Process Noisy Activation: Consistent Path:
Query triggers both nodes: {1645, 1643} Query triggers verified node: {1643}
→ Retriever fetches conflicting context. → Trace back to 𝑀𝑝𝑎𝑠 evidence.
5. Final Answer “Newton was born in 1645 or 1643...” “Isaac Newton was born in 1643.”
(Ambiguous / Hallucinated) (Precise & Verified)

Table 7: Case Study: Thematic Denoising in Medical Protocols. Demonstrating how MemGraphRAG filters irrelevant extraction
noise using Unified Schema Filtering.

Pipeline Phase Traditional GraphRAG (Baseline) MemGraphRAG (Ours)

1. Input Corpus Chunk 1: “Osimertinib treats EGFR-mutant NSCLC.” Same Corpus: Mixture of clinical facts and irrelevant patient
Chunk 2: “Patient prefers tea over coffee.” anecdotes.
2. Graph Construction Full Extraction: Schema Filtering:
𝑇1 : (𝑂𝑠𝑖𝑚𝑒𝑟𝑡𝑖𝑛𝑖𝑏, 𝑡𝑟𝑒𝑎𝑡, 𝑁 𝑆𝐶𝐿𝐶 ) Schema 𝑆 1 (𝐷𝑟𝑢𝑔, 𝑡𝑟𝑒𝑎𝑡, 𝐷𝑖𝑠 ) freq ≥ 𝜏 → Stable
𝑇2 : (𝑃𝑎𝑡𝑖𝑒𝑛𝑡, 𝑝𝑟𝑒 𝑓 𝑒𝑟,𝑇 𝑒𝑎) Schema 𝑆 2 (𝑃𝑎𝑡, 𝑝𝑟𝑒 𝑓 , 𝐵𝑒𝑣) freq < 𝜏 → Pending
→ Noise 𝑇2 pollutes the graph. → Result: Only 𝑇1 activated in 𝐺 𝑓 𝑎𝑐 .
3. Retrieval Query Q: “What is the standard treatment for NSCLC?”
4. Retrieval Process Drifting Path: Focused Path:
Node NSCLC → Patient → Tea Node NSCLC → Osimertinib
→ Retrieves irrelevant dietary info. → Strictly follows clinical ontology.
5. Final Answer “Osimertinib is used. Patients may prefer tea.” “Osimertinib is the recommended treatment.”
(Unprofessional / Distracted) (Professional & Concise)

and the iterative clustering of large-scale graphs poses significant (AI, subclass, NLP) vs. (AI, subclass, Unsupervised Learning).
bottlenecks for real-time deployment. These inconsistencies create redundant paths that dilute the
reasoning focus.
C Details of Preliminary Study
Independent extraction across different chunks may introduce con- D Details of the Proposed Method
flicting information into the merged graph, resulting in semantic D.1 Key Definitions
contradictions. In our preliminary study, we identify three major
types of such conflicts, as summarized in Table 8. Specifically: To establish a rigorous foundation for the subsequent methodology,
we first provide formal definitions for the core components of our
• Mutually Exclusive Conflict: Facts that cannot coexist
hierarchical knowledge representation:
in reality. For example, Chunk A yields (Newton, Birth year,
(i) Type (𝑡) and Entity (𝑒): We distinguish between abstract
1643) while Chunk B yields (Newton, Birth year, 1645).
concepts and concrete instances. A type 𝑡 ∈ T denotes a high-level
• Temporal Conflict: Contradictions arising from time-variant
taxonomic category (e.g., Person) that serves as a semantic anchor.
facts. A corpus spanning different years may generate both
An entity 𝑒 ∈ E refers to a specific instance grounded in the text
(Biden, President, USA) and (Trump, President, USA) without
(e.g., Einstein), where each entity is associated with a type through
temporal qualifiers, confusing the retriever.
a mapping function 𝜙 (𝑒) = 𝑡.
• Granularity Conflict: Facts describing the same reality at
(ii) Schema (𝑠) and Fact (𝑓 ): We define knowledge triples at two
incompatible abstraction levels. For instance, connecting an
levels of abstraction. A schema 𝑠 = (𝑡ℎ , 𝑟, 𝑡𝑡 ) specifies a structural
entity to both specific and general concepts, such as (Xiao
constraint, where 𝑡ℎ , 𝑡𝑡 ∈ T represent the head and tail types, and
Ming, born_in, Shanghai) and (Xiao Ming, born_in, China), or
𝑟 denotes a semantic relation (e.g., (Person, born_in, Country)). A
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

Table 8: Taxonomy of Knowledge Conflicts in Graph Retrieval-Augmented Systems.

Conflict Type Definition Mechanism & Impact Illustrative Examples

Mutually Exclusive Logically incompatible facts Mechanism: Distinct sources at- Attribute Value Clash:
Conflict that cannot simultaneously tribute divergent values to a func- Source A: (Newton, born_in,
hold true within a single do- tional property (single-value at- 1643)
main of discourse. tribute). Source B: (Newton, born_in,
Impact: Introduces binary logical 1645)
contradictions that halt determinis-
tic reasoning.
Temporal Conflict Inconsistencies arising Mechanism: Facts valid in disjoint Role Evolution:
from state changes in time intervals (𝑇1 ≠ 𝑇2 ) are flattened 𝑇2020 : (Trump, President, USA)
time-variant facts when into a static knowledge base. 𝑇2021 : (Biden, President, USA)
temporal metadata is Impact: Confuses the retriever by (Both retrieved without times-
absent. presenting outdated or competing tamps)
truths as currently valid.
Granularity Conflict Discrepancies in the level of Mechanism: Simultaneous map- Geospatial:
abstraction or specificity re- ping of an entity to hierarchically (Xiao Ming, born_in, Shanghai)
garding the same entity or distinct nodes (e.g., specific vs. gen- vs. (Xiao Ming, born_in, China)
concept. eral) within an ontology. Taxonomical:
Impact: Creates redundant infer- (AI, subclass, NLP) vs. (AI, sub-
ence paths and dilutes reasoning class, Machine Learning)
precision.

fact 𝑓 = (𝑒ℎ , 𝑟, 𝑒𝑡 ) is a concrete instantiation of a schema, where Fact Layer (M 𝑓 𝑎𝑐 ) stores instantiated triples derived from these
𝑒ℎ , 𝑒𝑡 ∈ E (e.g., (Einstein, born_in, Germany)). schemas. The lowest Passage Layer (M𝑝𝑎𝑠 ) preserves the original
(iii) Ontology (O): The ontology is defined as the structured col- source passages, ensuring that extracted facts remain grounded in
lection of all valid schemas, denoted as O = {𝑠 1, . . . , 𝑠𝑛 }. It governs their linguistic context.
the structural rules of the knowledge graph by enforcing semantic To strengthen associations across layers, we introduce a dense
constraints, ensuring that all extracted facts conform to predefined indexing mechanism that enforces structural consistency through
schema specifications. bidirectional interactions. Specifically, Schema–Instance Alignment
(iv) Passage (𝑝): A passage 𝑝 ∈ P represents a granular segment is established not merely as a one-way classification, but as a mutual
of raw text from the corpus, serving as the evidence grounding binding between abstraction and instantiation. On the bottom-up
unit. Specifically, each extracted fact 𝑓 is explicitly linked to its direction, we define a mapping
supporting textual evidence through a mapping function 𝜓 (𝑓 ) → Φ : M 𝑓 𝑎𝑐 → M𝑜𝑛𝑡 , (9)
𝑝𝑖 .
which enforces strict typing by assigning each triple 𝑡 ∈ M 𝑓 𝑎𝑐 to
D.2 MemGraphRAG architecture a schema constraint 𝑠 ∈ M𝑜𝑛𝑡 . On the top-down direction, each
schema 𝑠 induces its instantiation set
To overcome fragmented extraction and support the coherent evo-
lution of knowledge graphs, we propose the MemGraphRAG archi- T (𝑠) = { 𝑡 ∈ M 𝑓 𝑎𝑐 | Φ(𝑡) = 𝑠 }, |T (𝑠)| ≥ 0, (10)
tecture. Our core premise is that high-quality graph construction capturing the duality that schemas constrain facts while facts sub-
requires not only structured storage, but also active management of stantiate schemas.
knowledge. The system is built upon two complementary compo- Simultaneously, Fact–Evidence Grounding is modeled via a bidi-
nents: (i) a Hierarchical Memory Architecture that organizes schemas, rectional relation
facts, and passages across different abstraction levels, and (ii) a Ψ ⊆ M 𝑓 𝑎𝑐 × M𝑝𝑎𝑠 , (11)
Multi-Agent System that serves as the dynamic execution engine,
which links each fact to its supporting passages (provenance) while
leveraging memory to drive the iterative “extract–verify–modify”
allowing passages to index the facts they yield (extraction). For any
process. In the following sections, we describe how these compo-
triple 𝑡, we define its evidence set as
nents work together to ensure global consistency.
Global Memory, which organizes knowledge in a three-tier E (𝑡) = { 𝑝 ∈ M𝑝𝑎𝑠 | (𝑡, 𝑝) ∈ Ψ }, |E (𝑡)| ≥ 1. (12)
structure that aligns abstract schemas, concrete facts, and sup- Together, these bidirectional mappings ensure that the graph is
porting evidence. The top-level Ontology Layer (M𝑜𝑛𝑡 ) maintains both logically governed by the ontology and rigorously grounded
schema patterns with their statistical frequencies, providing seman- in textual evidence.
tic structure and global theme for graph construction. The middle
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

Hierarchical Indexing Graph, which provides a unified rep- D.3.1 Thematic Denoising via Unified Schema Filtering. To mitigate
resentation spanning abstract schemas, concrete facts, and textual the stochastic hallucinations inherent in LLMs and ensure statistical
evidence. Concretely, we organize G into three interconnected consensus, we implement a “Probationary Extraction Protocol.”
graph views that enable hierarchical navigation from high-level This protocol enforces a strict separation between raw extractions
semantic concepts to fine-grained supporting passages. (i) Semantic and validated knowledge.
Ontology Graph (G𝑜𝑛𝑡 ): Derived from the ontology layer M𝑜𝑛𝑡 , G𝑜𝑛𝑡 First, Composite Extraction into Memory. The process ini-
forms a high-level network of domain types and schema relations. tiates by partitioning the document stream into uniform chunks
It serves as the logical backbone of the overall graph by encoding 𝑐𝑖 ∈ C. For each chunk, the Extraction Agent (𝐴𝑒𝑥𝑡 ) generates a
valid relational patterns and domain constraints. (ii) Fact Graph Composite Extraction Record that simultaneously populates all three
(G𝑓 𝑎𝑐 ): Constructed from the fact layer M 𝑓 𝑎𝑐 , G𝑓 𝑎𝑐 represents an memory layers:
entity-relation graph over instantiated triples, which acts as the pri- 𝐴𝑒𝑥𝑡 (𝑐𝑖 ) → {𝑂 cand,𝑇cand, 𝑃src } (13)
mary substrate for multi-hop reasoning. (iii) Source Evidence Graph
(G𝑝𝑎𝑠 ): Induced from the passage layer M𝑝𝑎𝑠 , G𝑝𝑎𝑠 grounds entities where 𝑂 cand and 𝑇cand represent candidate schemas and triples, and
and relations in G𝑓 𝑎𝑐 back to their originating text passages, provid- 𝑃src anchors them to the source text.
ing fine-grained evidence support for faithful answer generation. Second, The Ontology Filter Mechanism. Crucially, newly
Together, this multi-view architecture enables structured reasoning extracted schemas are initially assigned a logical “Candidate State”
that progressively traverses from G𝑜𝑛𝑡 to G𝑓 𝑎𝑐 , and finally to G𝑝𝑎𝑠 (Pending). While physically stored in memory for tracking, they
for evidence retrieval. remain invisible to the global graph structure G. This isolation
Multi-Agent System, which introduces the dynamic execution prevents low-frequency noise from polluting the index.
units that drive the system’s evolution, is formulated as a collabo- Finally, Confidence-Driven State Promotion. We formalize
rative ecosystem of specialized agents interacting with M through the evolution of knowledge using a frequency-based confidence
distinct cognitive roles. Specifically, the Multi-Agent System is de- function. A schema transitions from “Pending” to “Stable” only
fined as A = 𝐴𝑒𝑥𝑡 , 𝐴𝑑𝑒𝑡 , 𝐴𝑟𝑒𝑠 , where each agent focuses on a sepa- when its extraction frequency across the corpus exceeds a statistical
rate function. Our design philosophy emphasizes the decoupling of threshold 𝜏:
generation, diagnosis, and correction to ensure high-fidelity graph  Stable,


 if Freq(𝑜) ≥ 𝜏,
construction: (i) the Extraction Agent (𝐴𝑒𝑥𝑡 ), which initializes the State(𝑜) = (14)
 Pending, otherwise.

graph by processing input documents and populating all three lay- 
ers of M (Schema, Fact, and Passage) in parallel, ensuring that each This transition triggers a cascading activation: only triples governed
extracted fact is grounded in supporting evidence; (ii) the Conflict by a stable schema are flagged as “Active.” Only these active triples
Detection Agent (𝐴𝑑𝑒𝑡 ), which is triggered by updates in the Fact are permitted to enter the subsequent conflict detection phase,
Layer (M 𝑓 𝑎𝑐) and performs purely diagnostic checks to identify ensuring the graph is constructed solely from consensus-verified
structural anomalies, redundancy, and logical inconsistencies; and knowledge.
(iii) the Conflict Resolution Agent (𝐴𝑟𝑒𝑠), which resolves con-
flicts flagged by 𝐴𝑑𝑒𝑡 by leveraging the global context stored in M, D.3.2 Consistency Maintenance via Global Adjudication. Dynamic
including historical evidence in M𝑝𝑎𝑠 and schema constraints in graph updates inevitably introduce contradictions. To ensure trust-
M𝑜𝑛𝑡, thereby maintaining the global consistency of G. worthiness, we implement a collaborative mechanism where agents
utilize Global Memory as the “ground truth” for adjudication.
D.3 Memory-based Indexing Graph Step 1: Asynchronous Conflict Triggering. The Conflict De-
tection Agent (𝐴𝑑𝑒𝑡 ) is triggered strictly when a triple 𝑡 new tran-
Construction sitions to an “Active” state. 𝐴𝑑𝑒𝑡 performs a hybrid scan over the
Traditional graph construction often processes document chunks in existing Fact Memory (Mfac ), utilizing both vector similarity and
isolation, leading to redundant entities and fragmented subgraphs. symbolic matching to identify potential conflict candidates 𝑇conf :
To address this, we reframe graph construction not as a one-off
extraction task, but as a dynamic co-evolution process between 𝑇conf = {𝑡 ′ ∈ Mfac | Sim(𝑡 new, 𝑡 ′ ) > 𝛿 ∨ Match(𝑡 new, 𝑡 ′ )} . (15)
the Global Memory M and the Knowledge Graph G. Driven by the If 𝑇conf ≠ ∅, the resolution protocol is initiated.
memory system, we implement two strategic paradigms to ensure Step 2: Evidence Retrieval and Adjudication. Unlike black-
structural integrity: (i) Structure Optimization via Progressive box resolution, our approach is evidence-driven. The Conflict Reso-
Construction: Instead of trusting LLM outputs immediately, we lution Agent (𝐴𝑟𝑒𝑠 ) leverages the memory mapping Ψ to retrieve the
treat extractions as hypotheses. The memory acts as a “probation- original provenance for both the new assertion and the conflicting
ary sandbox,” allowing the graph to evolve via an iterative “ex- facts. It constructs a context window 𝐶 ctx containing the raw source
tract–verify–modify” cycle that filters noise before it pollutes the passages: Ø
graph structure. (ii) Conflict Resolution via Global Perspec- 𝐶 ctx = Ψ(𝑡 new ) ∪ Ψ(𝑡 ′ ). (16)
tive: By maintaining a persistent global state, our shared memory 𝑡 ′ ∈𝑇conf
enables the system to detect and resolve semantic contradictions Based on 𝐶 ctx , 𝐴𝑟𝑒𝑠 reasons to determine factual validity, effectively
(e.g., logical, temporal, or granular conflicts) that span across dis- acting as a judge reviewing case files.
parate documents, ensuring a unified and consistent knowledge
representation.
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

Algorithm 1 Memory-based Indexing Graph Construction

Require: Document stream D; Global memory M = {𝑀𝑜𝑛𝑡 , 𝑀 𝑓 𝑎𝑐 , 𝑀𝑝𝑎𝑠 }; schema threshold 𝜏; conflict threshold 𝛿; bridging threshold 𝛿𝑏
Ensure: Global hierarchical graph G
1: Stage I: Composite Extraction into Memory (Sandbox)
2: for each chunk 𝑐 𝑖 from D do
3: {𝑂𝑐𝑎𝑛𝑑 ,𝑇𝑐𝑎𝑛𝑑 , 𝑃𝑠𝑟𝑐 } ← 𝐴𝑒𝑥𝑡 (𝑐𝑖 ) ⊲ Extract candidate schemas, triples, and provenance
4: Store 𝑂𝑐𝑎𝑛𝑑 ,𝑇𝑐𝑎𝑛𝑑 , 𝑃𝑠𝑟𝑐 into (𝑀𝑜𝑛𝑡 , 𝑀 𝑓 𝑎𝑐 , 𝑀𝑝𝑎𝑠 ) ⊲ Probationary storage: extraction as hypotheses
5: end for
6: Stage II: Unified Schema Filtering and Triple Activation
7: for each schema 𝑜 ∈ 𝑀𝑜𝑛𝑡 do
8: if Freq(𝑜) ≥ 𝜏 then
9: State(𝑜) ← Stable ⊲ Promote only consensus schemas
10: end if
11: end for
12: for each triple 𝑡 ∈ 𝑀 𝑓 𝑎𝑐 do
13: if State(Schema(𝑡)) = Stable then
14: State(𝑡) ← Active ⊲ Activate only triples governed by stable schema
15: end if
16: end for
17: Stage III: Conflict Detection and Evidence-based Adjudication
18: for each newly active triple 𝑡𝑛𝑒𝑤 do
19: F𝑐𝑜𝑛𝑓 ← {𝑡 ′ ∈ 𝑀 𝑓 𝑎𝑐 | Sim(𝑡𝑛𝑒𝑤 , 𝑡 ′ ) > 𝛿 ∨ Match(𝑡𝑛𝑒𝑤 , 𝑡 ′ )} ⊲ Global scan for logical/temporal/granularity conflicts
20: if F𝑐𝑜𝑛𝑓 ≠ ∅ then
𝐶𝑐𝑡𝑥 ← Ψ(𝑡𝑛𝑒𝑤 ) ∪ 𝑡 ′ ∈ F𝑐𝑜𝑛𝑓 Ψ(𝑡 ′ )
Ð
21: ⊲ Retrieve provenance passages as evidence
22: 𝐴𝑟𝑒𝑠 updates 𝑡𝑛𝑒𝑤 and F𝑐𝑜𝑛𝑓 based on 𝐶𝑐𝑡𝑥 ⊲ Discard / refine / temporally augment conflicting facts
23: end if
24: end for
25: Stage IV: Multi-view Projection and Memory-guided Bridging
26: Construct G𝑜𝑛𝑡 from stable schemas in 𝑀𝑜𝑛𝑡
27: Construct G𝑓 𝑎𝑐 from active triples in 𝑀 𝑓 𝑎𝑐
28: Construct G𝑝𝑎𝑠 from provenance passages in 𝑀𝑝𝑎𝑠 ⊲ Project memory layers into graph views
29: Add type-based edges linking entities with shared schema types ⊲ Type-based bridging for disjoint subgraphs
30: Add similarity-based edges if Sim(𝑒𝑖 , 𝑒 𝑗 ) > 𝛿𝑏 ⊲ Embedding-based bridging for long-range connectivity
31: Merge all views into global hierarchical graph G
32: return G

Step 3: Taxonomy-Based Resolution Strategies. Based on corresponding graph views: G𝑜𝑛𝑡 (Schema View), G𝑓 𝑎𝑐 (Fact View),
the evidence, 𝐴𝑟𝑒𝑠 executes targeted updates to resolve specific and G𝑝𝑎𝑠 (Source View).
conflict types: To address the common issue of disjoint subgraphs in extracted
• Mutually Exclusive Conflict (Logical): For contradictory facts knowledge, we augment the primary reasoning substrate, G𝑓 𝑎𝑐 ,
(e.g., conflicting birthplaces), the agent compares evidence with two memory-enabled connectivity mechanisms:
reliability to discard the erroneous fact. (1) Type-Based Bridging: Leveraging M𝑜𝑛𝑡 , disjoint entities
• Temporal Conflict: For facts valid in different periods (e.g., are explicitly connected if they map to the same high-level
distinct presidential terms), the agent resolves ambiguity by schema type (e.g., connecting all Researchers regardless of
appending temporal attributes (e.g., adding “46th” vs. “47th”). their document origin).
• Granularity Conflict (Structural): For facts describing the (2) Similarity-Based Bridging: Leveraging embedding stor-
same reality at different abstraction levels (e.g., “Shanghai” age in M, we introduce implicit edges between entity pairs
vs. “China”), the agent refines predicates to allow logical whose vector similarity exceeds a threshold 𝛿.
coexistence (e.g., born_city vs. born_country). These mechanisms leverage the global nature of memory to connect
D.3.3 Structural Unification via Memory-Guided Bridging. The fi- long-distance entities, significantly enhancing the graph’s ability
nal phase transforms the validated contents of the memory system to support multi-hop reasoning across documents where explicit
into a navigable Global Hierarchical Graph G. We adopt a multi- textual links are missing.
view projection strategy that maps the three memory layers into
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

Algorithm 2 Memory-guided Online Retrieval

Require: Query embedding q; Graph G with transition matrix M; Memory M = {𝑀𝑜𝑛𝑡 , 𝑀 𝑓 𝑎𝑐 , 𝑀𝑝𝑎𝑠 }; top-𝐾; threshold 𝜏; damping 𝜆; balance
𝛼
Ensure: Evidence set C for downstream LLM generation
1: Stage I: Multi-layer Retrieval and Filtering
2: S𝑟𝑎𝑤 ← TopK(𝑀𝑜𝑛𝑡 , q, 𝐾); F𝑟𝑎𝑤 ← TopK(𝑀 𝑓 𝑎𝑐 , q, 𝐾); P𝑟𝑎𝑤 ← TopK(𝑀𝑝𝑎𝑠 , q, 𝐾) ⊲ Align query with ontology, facts, and passages
3: S𝑟𝑒𝑡 ← {𝑠 ∈ S𝑟𝑎𝑤 | Sim(q, s) > 𝜏 }; F𝑟𝑒𝑡 ← {𝑓 ∈ F𝑟𝑎𝑤 | Sim(q, f) > 𝜏 }
4: if S𝑟𝑒𝑡 ∪ F𝑟𝑒𝑡 = ∅ then
5: return P𝑟𝑎𝑤 ⊲ Fallback to standard RAG
6: end if
7: Stage II: Structure-aware Node Initialization
8: Define reset weights 𝑃𝑖𝑛𝑖𝑡 (𝑣) on nodes 𝑣 ∈ G:
1 Í
9: Entity nodes: 𝑃𝑖𝑛𝑖𝑡 (𝑒) = | F | Sim(q, f), F𝑒 = {𝑓 ∈ F𝑟𝑒𝑡 | 𝑒 ∈ 𝑓 } ⊲ Ground by query-relevant facts
 𝑒 Í 𝑓 ∈ F𝑒 
1 1
10: Type nodes: 𝑃𝑖𝑛𝑖𝑡 (𝑡) = | S | 𝑠 ∈ S𝑡 Sim(q, s) · log(deg(𝑡 )+1) , S𝑡 = {𝑠 ∈ S𝑟𝑒𝑡 | 𝑡 ∈ 𝑠} ⊲ Schema relevance + hub suppression
𝑡
Í 
𝑒 ∈E𝑝 IDF(𝑒 )
11: Passage nodes: 𝑃𝑖𝑛𝑖𝑡 (𝑝) = Sim(q, d𝑝 ) · 𝛼 · 𝜎 log( | E |+1) ⊲ Semantic alignment + information density
𝑝

12: Normalize 𝑃𝑖𝑛𝑖𝑡 into p (0) with 𝑣 p (0) (𝑣) = 1

Í

13: Stage III: PPR Propagation and Evidence Selection

14: repeat
15: p (𝑘+1) ← (1 − 𝜆)Mp (𝑘 ) + 𝜆p (0) ⊲ Personalized PageRank with restart
16: until convergence
17: Select top-ranked passages P ∗ and entities E ∗ by p (∞)
18: C ← P ∗ ∪ E ∗
19: return C

E Prompt Set parallel. Given a user query q, we parallelly retrieve top-𝐾 can-
To provide a more intuitive illustration of our graph construction didates from 𝑀𝑜𝑛𝑡 , 𝑀 𝑓 𝑎𝑐 , and 𝑀𝑝𝑎𝑠 respectively. To prevent low-
procedure and ensure reproducibility, we present the Conflict De- relevance noise from propagating into the graph reasoning stage,
tection and Conflict Resolution components used in MemGraphRAG we apply a strict relevance filter. For the retrieved schemas S𝑟𝑒𝑡
indexing, as shown in Figure 7 and 8. and facts F𝑟𝑒𝑡 , only candidates satisfying a semantic similarity
threshold Sim(q, x) > 𝜏 are retained. This filtering ensures that
the subsequent node initialization is seeded exclusively with high-
E.1 Memory-guided Online Retrieval confidence structural evidence. Crucially, to guarantee system ro-
Building upon the constructed Global Hierarchical Graph G and bustness, if the filtering process yields no valid structural evidence
the Global Memory M, this section details our memory-guided (i.e., 𝑆 ret ∪ 𝐹 ret = ∅), the framework adaptively falls back to a stan-
retrieval and reasoning mechanism. To bridge the gap between the dard RAG mode, relying solely on the direct similarity between the
user query and the complex graph topology, the inference workflow query and the content in 𝑀𝑝𝑎𝑠 for answer generation.
unfolds through three logically progressive stages: The workflow E.1.2 Structure-Aware Node Initialization. To seed the subsequent
consists of three key steps: i) Multi-Layer Memory Retrieval, graph propagation process with specific semantic context, we must
which retrieves initial initial candidate evidence, including schemas project the retrieved evidence onto the heterogeneous graph topol-
𝑠, facts 𝑓 , and passages 𝑝 from 𝑀𝑜𝑛𝑡 , 𝑀 𝑓 𝑎𝑐 , and 𝑀𝑝𝑎𝑠 , respectively. ogy. Formally, we define an initial reset probability distribution
It then applies a preliminary noise filtering process to ensure rele- 𝑃𝑖𝑛𝑖𝑡 (𝑣) for any node 𝑣 ∈ G. This distribution provides an initial im-
vance. ii) Structure-Aware Node Initialization, which projects portance score for the inference algorithm, quantifying the intrinsic
the retrieved evidence onto the graph structure by mapping them to significance of each node prior to information diffusion.
initial node weights. We apply distinct scoring strategies for Entity 1. Entity Node Initialization via Facts: To ensure that graph
nodes 𝑒, Type nodes 𝑡, and Passage nodes 𝑝, integrating seman- propagation originates from grounded evidence, we first initialize
tic relevance, topological constraints, and information density. iii) entity nodes based on the relevance of their associated facts re-
Graph Propagation, which executes the Personalized PageRank trieved from M 𝑓 𝑎𝑐 . Formally, we quantify the initial importance
(PPR) algorithm on the heterogeneous graph, initiating from the of an entity e as the mean semantic similarity of all filtered facts
weighted nodes. This propagation diffuses importance across the containing it:
graph to identify the most globally significant passages and nodes,
which are then selected for downstream LLM generation. 1 ∑︁
𝑃𝑖𝑛𝑖𝑡 (𝑒) = Sim(q, f) (17)
|F𝑒 |
E.1.1 Multi-Layer Memory Filtering. The retrieval phase initiates 𝑓 ∈ F𝑒
by querying the three distinct layers of the Global MemoryM in
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

Conflict Detection

Task Definitions: You are an expert fact checker. Given a target triple and a list of related triples. Your
task: Detect whether target triple conflicts with any triple in the list of related triples, and classify
conflicts into three types:
• mutual conflict (mutual exclusivity / one-to-one relations)
• Temporal conflict (time-dependent facts; conflicts arise when time scopes overlap or are missing)
• Granularity conflict (different levels of specificity; may be compatible via containment)
Definitions and rules:
1. mutual conflict (type = “mutual") A mutual conflict happens when:
• Same subject and predicate, but different objects, AND the predicate is one-to-one / mutually exclusive.
Example: (X, birthplace, Shanghai) vs (X, birthplace, Beijing)
• Or cyclic/contradictory relational structure that cannot both be true under common-sense constraints.
Example: (A, father, B) vs (B, father, A)
2. Temporal conflict (type = “temporal") A temporal conflict happens when:
• The predicate describes a role/state that can change over time and is typically unique at a given
moment (e.g., president/CEO/champion/current location).
• If both triples claim different objects for the same subject-predicate:
– If explicit time scopes exist and overlap → hard temporal conflict.
– If time scopes exist and do NOT overlap → not a conflict.
– If time scopes are missing but the predicate is time-variant and moment-unique → suspected temporal
conflict (ask for time ranges; do NOT assert a hard conflict without time info).
3. Granularity conflict (type = “granularity")
• Triples differ due to specificity/abstraction level.
Example: (X, birthplace, Shanghai) vs (X, birthplace, China)
• If one object is a parent/superset/contains the other (hypernym/meronym/administrative containment),
then it is usually compatible → classify as "granularity".
• If objects are incompatible (cannot contain each other and cannot both be true) → Logical conflict.
Output MUST be a valid JSON object following the required schema.

Figure 7: The prompt used for Conflict Detection Agent.

where F𝑒 ⊆ F𝑟𝑒𝑡 denotes the subset of query-relevant facts leverages ontology as a weak supervision signal while strictly con-
contain entity 𝑒. If F𝑒 = ∅, the weight defaults to 0. This aggregation straining the diffusion radius of overly generic concepts.
strategy ensures that entities are activated strictly by explicit, query- 3. Passage Initialization with Information Density: Finally,
relevant factual support. we need to initialize the Passage Nodes (𝑝 ∈ 𝐺 𝑝𝑎𝑠 ). We formulate the
2. Type Node Initialization via Schemas: To incorporate comprehensive scoring function to prioritize semantically relevant
macro-level domain knowledge and avoid introducing irrelevant sources with high-value information, while avoiding dominance
semantics, we further initialize type nodes 𝑡 ∈ Gschema based on over finer-grained entity nodes, as follows:
the retrieved schemas from Mont . A critical challenge is that type Í !
𝑒 ∈ E𝑝 IDF(𝑒)
nodes often exhibit disproportionately large degrees (e.g., a generic 𝑃𝑖𝑛𝑖𝑡 (𝑝) = Sim(q, d𝑝 ) × 𝛼 × 𝜎 (19)
“Person” node connected to thousands of entities). Activating such log(|E𝑝 | + 1)
high-degree nodes directly would cause importance to diffuse too | {z }
Information Density Term
broadly across the graph, thereby introducing substantial noise. To
address this issue, we introduce a structural regularization term This formula integrates three critical dimensions: (i) Semantic Align-
that combines semantic relevance with a log-degree penalty: ment (Sim), which measures the vector similarity between the query
! 𝑞 and the passage embedding 𝑑𝑝 ; (ii) Structural Balance (𝛼), a damp-
1 ∑︁ 1 ening coefficient empirically set to 0.05, which prevents dense pas-
𝑃𝑖𝑛𝑖𝑡 (𝑡) = Sim(q, s) × (18) sage nodes from overwhelming sparse entity nodes during the
|S𝑡 | log(deg(𝑡) + 1)
𝑠 ∈ S𝑡
| {z } | {z } initial propagation phase and ensures a balanced importance distri-
Schema Relevance
Hub Suppression bution; and (iii) Information Density Term, which quantifies content
where S𝑡 denotes the subset of retrieved schemas corresponding
to type 𝑡, deg(𝑡) is the node degree. This formulation effectively
MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea.

Conflict Resolution

Task Definitions: You are an expert knowledge graph curator. Given a set of conflicting triples and their
source passages, your task is to resolve the conflicts and produce corrected triples.
Conflict Resolution Strategies:
1. Mutual Conflict (type = “mutual"): These are contradictory claims about the same entity (e.g., same
subject-predicate but different objects)
• Resolution: Analyze the source passages to determine which triple is more accurate
• Keep only the CORRECT triple, discard the incorrect one(s)
• If both seem equally valid based on context, prefer the one with more specific/credible source
2. Temporal Conflict (type = “temporal"): These are time-dependent facts where time scopes overlap or are
missing
• Resolution: Add time information to the relation to distinguish the facts
• Modify the predicate to include time context (e.g., "was president of [2000-2005]" vs "was president
of [2005-2010]")
• If time info is not in sources, note it as “temporal_conflict_unresolved"
3. Granularity Conflict (type = “granularity"): These are facts at different levels of specificity (e.g.,
"born in Shanghai" vs "born in China")
• Resolution: Add granularity description to the relation to clarify the scope
• Modify the predicate to include granularity context (e.g., "was born in [city: Shanghai]" vs "was born
in [country: China]")
• Both can be kept if they are compatible (containment relationship)
Output MUST be a valid JSON object following the required schema.

Figure 8: The prompt used for Conflict Resolution Agent.

quality by summing the Inverse Document Frequency (IDF) of enti- multiple supporting documents to answer complex queries, thereby
ties 𝐸𝑝 within the passage and applying log-normalization, thereby testing the system’s ability to perform effective cross-document
rewarding passages that contain rare and discriminative facts rather evidence retrieval.
than generic, verbose content. (ii) 2WikiMultiHopQA (2Wiki) [25]: A benchmark derived
from Wikipedia knowledge graphs, specifically constructed to test
E.1.3 Personalized PageRank. Following the initialization phase,
structured reasoning. It consists of queries that necessitate aggre-
We execute the Personalized PageRank (PPR) algorithm on the
gating evidence chains from two to four specific articles, focusing
heterogeneous graph to diffuse the initial semantic energy. The
on the model’s capacity to handle complex entity relationships and
propagation uses the normalized vector p (0) as the starting distri-
maintain logical consistency.
bution and follows the iteration:
(iii) MuSiQue [47]: A challenging dataset designed to minimize
v (𝑘+1) = (1 − 𝜆)Wv (𝑘 ) + 𝜆v (0) (20) reasoning shortcuts often found in earlier benchmarks. It features
where W is the transition matrix of the graph. We specifically set connected reasoning chains of 2-4 hops, requiring systems to per-
the damping factor 𝜆 = 0.5 to restrict the random walk to a local form strictly sequential logical inference across multiple documents
neighborhood, thereby preventing semantic drift into irrelevant to derive the correct answer.
multi-hop connections. Upon convergence to v (∞) the top-K pas- (iv) G-bench (Novel) & G-bench (Medical) [52]: Two domain-
sages and top-M entities with the highest scores are selected as the specific benchmarks tailored to evaluate GraphRAG performance
context window for LLM inference. on hierarchical retrieval and deep contextual understanding. The
Medical subset utilizes NCCN guidelines to test the handling of
F Benchmark Dataset dense, rule-based clinical protocols, while the Novel subset em-
We first evaluate the effectiveness of MemGraphRAG on three ploys literary texts from Gutenberg to assess the comprehension of
widely-used multi-hop QA datasets, including HotpotQA [59], 2Wiki- implicit, non-linear narrative structures.
MultiHopQA (2Wiki) [25] and MuSiQue [47] and two GraphRAG
benchmarks: G-bench (Novel) and G-bench (Medical) [52]. We pro-
G Implementation Details of Baselines
vide a concise overview of each dataset’s key characteristics below. In our experiments, we compare our method against several widely
(i) HotpotQA [59]: A widely adopted dataset for evaluating used GraphRAG models.
multi-hop reasoning across disparate texts. It requires models to fil-
ter through distractor paragraphs and synthesize information from
KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, and Jinsong Su

KGP [51] facilitates multi-document question answering by con- PageRank. It acts as a dual-system memory model to enable deep
structing a graph where nodes represent passages or document knowledge integration, facilitating robust retrieval for scenarios
structures. It employs an LLM-driven traversal agent to navigate requiring the synthesis of information from multiple sources.
semantic and structural connections, progressively aggregating HippoRAG2 [20] extends the Personalized PageRank-based
supporting context for the final response. framework of its predecessor by optimizing passage contextualiza-
G-Retriever [24] targets real-world textual graphs by formulat- tion and the online interaction with LLMs. These enhancements
ing the subgraph retrieval task as a Prize-Collecting Steiner Tree enable the model to mimic human long-term memory more ef-
(PCST) optimization problem. This approach extracts the most rel- fectively, balancing robust factual recall with complex associative
evant subgraph to fit within the LLM context window, enabling reasoning.
effective conversational QA while mitigating hallucination and E2 GraphRAG [64] optimizes the GraphRAG paradigm by estab-
ensuring scalability. lishing bidirectional indexes between document chunks and entities.
RAPTOR [44] employs a recursive abstraction approach to con- It combines a summary tree with a lightweight entity graph to facil-
struct a hierarchical tree structure. By clustering and summarizing itate fast lookup, enabling an adaptive retrieval process that seam-
text chunks from the bottom up, it enables the retrieval of infor- lessly integrates local context and global understanding without
mation at varying levels of granularity, capturing both high-level manual query mode selection.
context and fine-grained details for holistic understanding. GFM-RAG [37] introduces a Graph Foundation Model (GFM)
MS-GraphRAG [12] enhances global corpus understanding by designed for zero-shot application on unseen datasets. It employs
building an entity-relation graph and pre-computing community- a pre-trained Graph Neural Network to reason over graph struc-
level summaries. It answers queries by synthesizing insights from tures, effectively capturing complex query-knowledge relationships
these communities, offering improved comprehensiveness for ques- while mitigating the impact of noise and incompleteness in the con-
tions that span the entire document collection. structed graphs.
LazyGraphRAG [8] introduces a cost-effective paradigm that LogicRAG [6] introduces a dynamic retrieval paradigm where
eliminates the need for expensive up-front summarization of source query-specific logic is modeled as a directed acyclic graph at infer-
data. By avoiding the pre-computation of community hierarchies, ence time. By linearizing this graph via topological sort, it guides
it reduces indexing costs to the level of standard vector RAG while the retrieval process through a logically consistent sequence of sub-
maintaining superior performance on local queries and competitive problems, significantly reducing token usage compared to static
quality on global queries compared to full-graph approaches. graph approaches.
LightRAG [17] introduces a two-tier retrieval strategy designed LinearRAG [70] challenges the reliance on costly and unstable
to capture both detailed entity relationships and broader thematic relation extraction in existing methods. It constructs a relation-free
contexts. It utilizes graph-enhanced indexing to facilitate rapid hierarchical structure termed “Tri-Graph” using lightweight entity
access to relevant information and allows for seamless integration extraction and semantic linking. This approach scales linearly with
of new data via an incremental update algorithm. corpus size and employs a two-stage retrieval strategy involving
HippoRAG [19] proposes a neurobiologically inspired frame- local entity activation and global importance aggregation.
work that orchestrates LLMs, knowledge graphs, and Personalized