AuthorsHyeonjeong Ha, Jeonghwan Kim, Cheng Qian et al.
TL;DR
MEMGUARD uses type-aware memory reorganization and dynamic routing to prevent heterogeneous memory contamination, reaching 89.53% anti-hallucination accuracy on HaluMem (+28.27% over MemOS).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Conversational memory mixes semantic facts, episodic events, and procedural rules, causing heterogeneous memory contamination when functional boundaries are weak.
On LoCoMo, 97.7% of unverifiability errors stem from write-time contamination, while 63.8% of factuality errors come from retrieval-time contamination.
HOW IT WORKS
Type-Aware Knowledge Decomposition, Self-Verified Extraction, and the Relational Knowledge Graph let MEMGUARD break conversations into atomic, typed memories and link them safely.
Think of MEMGUARD as a card catalog in a library: facts, episodes, and procedures live in separate drawers, with cross-references instead of merged cards.
This structure plus Query-Adaptive Type Routing lets MEMGUARD retrieve only compatible memory types and compose evidence across types, something a plain context window cannot control.
DIAGRAM
This diagram shows how MEMGUARD reorganizes a conversation into typed memory atoms, verifies coverage, builds a relational graph, and writes into type-isolated stores.
DIAGRAM
This diagram shows how MEMGUARD routes queries to memory types, retrieves primary contexts, and expands them via relational composition.
PROCESS
Write-Time Memory Reorganization
MEMGUARD converts each conversation into atomic memories using Type-Aware Knowledge Decomposition and Self-Verified Extraction, enforcing single-type constraints for every memory atom.
Relational Knowledge Graph Construction
MEMGUARD builds a directed, typed Relational Knowledge Graph over memory atoms, encoding cross-type dependencies without merging their contents.
Type-Isolated Memory Writing
MEMGUARD writes each atom into its corresponding store, using type-local comparison and ADD, UPDATE, or SKIP operations to avoid cross-type overwrites.
Retrieval-Time Dynamic Memory Routing
MEMGUARD applies Query-Adaptive Type Routing and Relational Knowledge Composition to retrieve type-compatible memories and expand them over the graph for answering.
KEY CONTRIBUTIONS
Identification of Heterogeneous Memory Contamination
MEMGUARD formalizes heterogeneous memory contamination and shows that 97.7% of unverifiability errors on LoCoMo arise from write-time contamination, with 63.8% factuality errors from retrieval-time contamination.
Type-Aware Memory Framework MEMGUARD
MEMGUARD introduces Type-Aware Knowledge Decomposition, Self-Verified Extraction, and a Relational Knowledge Graph to preserve functional boundaries during memory writing and retrieval.
Improved Hallucination Robustness with Fewer Tokens
MEMGUARD achieves 89.53% anti-hallucination accuracy and 71.49% update correctness on HaluMem while retrieving up to 5.8× fewer memory tokens than A-Mem on LoCoMo.
RESULTS
Acc. (%)
89.53
+28.27 over MemOS
F1
94.15
vs MemOS on HaluMem extraction
C (Update)
71.49
memory update correctness on HaluMem
# Avg. Token
1605
4.5× fewer tokens than A-Mem on LoCoMo (7244)
On HaluMem-Medium, which tests hallucination across memory extraction, updating, and question answering, MEMGUARD reaches 89.53% anti-hallucination accuracy and 94.15 F1. On LoCoMo, MEMGUARD attains 79.10% average accuracy with 2-hop retrieval using GPT-4.1-mini, showing that structured, type-aware memory improves reliability without scaling retrieval indiscriminately.
BENCHMARK
On HaluMem-Medium, which tests hallucination across memory extraction, updating, and question answering, MEMGUARD reaches 89.53% anti-hallucination accuracy and 94.15 F1. On LoCoMo, MEMGUARD attains 79.10% average accuracy with 2-hop retrieval using GPT-4.1-mini, showing that structured, type-aware memory improves reliability without scaling retrieval indiscriminately.
BENCHMARK
Acc. (%) for memory extraction anti-hallucination accuracy on HaluMem-Medium.
BENCHMARK
Avg. accuracy (%) across LoCoMo tasks under GPT-4.1-mini base LLM and GPT-4.1 judge.
KEY INSIGHT
MEMGUARD retrieves up to 5.8× fewer memory tokens than A-Mem on LoCoMo while still achieving higher average accuracy and stronger adversarial performance.
This is counterintuitive because many memory-augmented systems assume more retrieved context always helps, but MEMGUARD shows that selective, type-aware retrieval can be both leaner and more reliable.
WHY IT MATTERS
MEMGUARD unlocks long-term conversational agents that treat episodic, semantic, and procedural memories as distinct evidence sources with controllable interactions.
Builders can now design memory-augmented LLMs that maintain personalization and long-horizon reasoning while explicitly governing which memory types are allowed to influence each answer.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…