💬 NLP

MemGuard: Preventing Memory Contamination in Long-Term Memory-Augmented Large Language Models

MemGuard is a type-aware memory framework that prevents heterogeneous memory contamination in long-term memory-augmented LLMs by assigning explicit functional roles to memories and selectively composing evidence, thereby significantly improving reliability and retrieval efficiency.

Original authors: Hyeonjeong Ha, Jeonghwan Kim, Cheng Qian, Jiayu Liu, William M. Campbell, Yue Wu, Yuji Zhang, Kathleen McKeown, Dilek Hakkani-Tur, Heng Ji

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hyeonjeong Ha, Jeonghwan Kim, Cheng Qian, Jiayu Liu, William M. Campbell, Yue Wu, Yuji Zhang, Kathleen McKeown, Dilek Hakkani-Tur, Heng Ji

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, helpful assistant who remembers everything you've ever told them. They can recall your favorite coffee order, the name of your childhood pet, and the fact that you're allergic to peanuts. This is the promise of Memory-Augmented Large Language Models (LLMs).

However, the paper argues that current assistants are like a messy, chaotic filing cabinet. They throw all your memories into one big drawer without labels. If you say, "My friend took ibuprofen and felt better," the assistant might accidentally mix that up with the medical fact, "Ibuprofen is dangerous for people with asthma."

When the assistant tries to answer a question later, it grabs these mixed-up notes and creates a dangerous, incorrect answer: "You should take ibuprofen!" even though you have asthma. The paper calls this "Heterogeneous Memory Contamination." It's like trying to cook a meal by mixing a recipe, a grocery list, and a story about your last vacation into a single bowl of soup.

The Solution: MEMGUARD

The authors introduce MEMGUARD, a new system that acts like a strict, organized librarian for your assistant's brain. Instead of one messy drawer, MEMGUARD builds three separate, locked rooms:

The Semantic Room (Facts): This holds timeless truths, like "I am allergic to peanuts" or "The sky is blue."
The Episodic Room (Stories): This holds specific events, like "My friend took ibuprofen yesterday and felt better."
The Procedural Room (Rules): This holds instructions, like "If someone has a headache, suggest over-the-counter pain relievers."

How It Works (The Analogy)

1. Writing Memories (The Sorting Phase)
When you talk to the assistant, MEMGUARD doesn't just dump the conversation into the database. It acts like a smart sorter.

If you say, "I have asthma," it puts that in the Semantic Room.
If you say, "I took ibuprofen once and it helped," it puts that in the Episodic Room.
Crucially, it draws a map (a "Relational Graph") connecting these rooms. It knows the story in the Episodic Room is related to the fact in the Semantic Room, but it keeps them in separate buildings so they don't accidentally merge.

2. Retrieving Memories (The Search Phase)
When you ask, "I have a headache; can I take ibuprofen?", the assistant doesn't just search the whole messy pile.

Smart Routing: MEMGUARD asks, "What kind of answer do I need?" It realizes it needs a medical rule (Procedural) and a safety constraint (Semantic). It ignores the story about your friend for a moment because that's just an anecdote, not a rule.
The Map: It then uses the map to find the specific safety rule in the Semantic Room that says, "No ibuprofen for asthma."
The Result: It combines the rule and the safety warning to give you a safe answer: "No, you should avoid ibuprofen because of your asthma."

Why This Matters

The paper tested this system against other memory systems and found two major wins:

Fewer Hallucinations: By keeping the "stories," "facts," and "rules" separate, the assistant stops making up dangerous facts. It reduced "memory contamination" (mixing things up) by up to 28%.
Efficiency: Because the assistant knows exactly which "room" to look in, it doesn't have to read through thousands of irrelevant notes. It retrieved 5.8 times fewer memory tokens (chunks of text) than previous methods, making it faster and cheaper to run.

The Bottom Line

The paper claims that for an AI to be truly reliable over the long term, it can't just be a "brain" that remembers everything equally. It needs a structured memory system that respects the difference between a one-time story, a permanent fact, and a general rule. MEMGUARD provides that structure, preventing the AI from getting confused and giving you bad advice based on mixed-up memories.

Technical Summary: MEMGUARD

Problem: Heterogeneous Memory Contamination

Memory-augmented Large Language Models (LLMs) aim to extend reasoning beyond fixed context windows by maintaining long-term memory across interactions. However, existing systems often collapse distinct types of knowledge—stable user facts (semantic), episodic events, and behavioral rules (procedural)—into a shared retrieval space. This lack of functional separation leads to heterogeneous memory contamination, a failure mode where context-specific events are overgeneralized into unsupported facts, or semantically relevant but functionally incompatible memories mislead generation.

The authors identify that this contamination occurs across the memory lifecycle:

Write-time contamination: Unsupported or overgeneralized knowledge is written to memory, persisting as noise.
Retrieval-time contamination: Semantically related but functionally unsuitable memories are retrieved together, obscuring necessary evidence.
Composition failure: The model integrates contaminated evidence, leading to hallucinations.

Analysis of the LoCoMo benchmark reveals that unverifiability errors (answering when evidence is insufficient) are predominantly caused by write-time contamination (97.7%), while factuality errors (answering incorrectly despite available evidence) are frequently linked to retrieval-time contamination (63.8%). Existing methods, whether flat semantic stores, graph-based models, or cognitive-inspired hierarchies, largely fail to preserve functional boundaries, treating memory as a monolithic or weakly structured resource.

Methodology: MEMGUARD

To address this, the authors propose MEMGUARD, a type-aware memory framework that enforces functional boundaries during memory construction, retrieval, and evidence composition. The system is motivated by cognitive theories of long-term memory, which posit functionally specialized systems (episodic, semantic, procedural) that are selectively recruited based on reasoning goals.

1. Write-Time Memory Reorganization

MEMGUARD reorganizes conversational data before storage to ensure functional separation:

Type-Aware Knowledge Decomposition: The LLM decomposes raw dialogue into non-overlapping "memory atoms," assigning each exactly one functional type: semantic (timeless facts), episodic (time-anchored events), or procedural (recurring behaviors).
Self-Verified Extraction: A verification step ensures no information is missed during decomposition, recovering missing atoms while maintaining type constraints.
Relational Knowledge Graph: While atom contents are isolated by type, a directed graph $G=(V, E)$ encodes typed dependencies (e.g., supports, leads_to, contradicts) between atoms. This preserves cross-atom relationships without merging heterogeneous content.
Type-Isolated Writing: Atoms are routed to type-specific stores ( $M_{semantic}, M_{episodic}, M_{procedural}$ ). Deduplication and updates occur only within the same type store, preventing functionally distinct memories from overwriting one another.

2. Retrieval-Time Dynamic Memory Routing

At retrieval, MEMGUARD employs a two-stage process to select and compose evidence:

Query-Adaptive Type Routing: A soft router estimates the relevance of each memory type for a given query, allocating a retrieval budget $K$ proportionally across type-isolated stores. This prevents irrelevant memory types from dominating the context.
Relational Knowledge Composition: Primary retrieved atoms are expanded via Breadth-First Search (BFS) over the relational graph (up to $h_{max}$ hops). This recovers relevant cross-memory dependencies that may be weakly similar to the query but essential for reasoning. The expanded context is re-ranked with a hop-decay factor to balance relevance and structural depth.

Key Contributions

Identification of Contamination: The paper identifies heterogeneous memory contamination as a primary source of persistent hallucinations, caused by weak functional boundaries in existing memory systems.
MEMGUARD Framework: A novel architecture that structures memory writing, routes retrieval, and guides evidence composition based on explicit memory types (semantic, episodic, procedural).
Empirical Validation: Demonstration that preserving functional boundaries improves hallucination robustness and utility while reducing the volume of retrieved tokens.

Experimental Results

The authors evaluated MEMGUARD on hallucination benchmarks (HaluMem) and long-horizon conversational tasks (LoCoMo, PerLTQA, LongMemEval).

HaluMem (Hallucination Evaluation):
- MEMGUARD achieved 89.53% anti-hallucination accuracy, a 28.27% improvement over the best baseline (Mem0).
- It attained 71.49% memory update correctness, significantly reducing omission and hallucination rates during memory updates.
- Retrieval accuracy reached 89.87% (1-hop) and 90.24% (2-hop), comparable to extraction recall, indicating that type-aware writing improves evidence accessibility.
Utility Evaluation (LoCoMo, PerLTQA, LongMemEval):
- On LoCoMo, MEMGUARD achieved 75.53% average accuracy (GPT-4o-mini setting), comparable to the state-of-the-art MemOS, but retrieved 21% fewer memory tokens.
- Under a consistent GPT-4.1-mini setting, MEMGUARD surpassed A-Mem by 10.76% in accuracy while using 4.5× fewer retrieved tokens.
- The method showed particular strength on adversarial queries, where it more reliably abstained from answering when evidence was insufficient, demonstrating improved answerability calibration.

Significance and Claims

The paper claims that reliable long-term reasoning in LLMs depends not merely on scaling retrieval capacity, but on the principled organization and selective use of heterogeneous memory. By treating memory as a structured substrate with enforced functional boundaries, MEMGUARD mitigates cross-type interference and persistent hallucinations.

The authors emphasize that their approach targets contamination before generation (at write and retrieval times) rather than enforcing grounding during generation. While the method significantly reduces hallucinations and improves memory reliability, the paper modestly acknowledges limitations: it does not directly control generation-time behavior (composition errors may still occur if the base LLM misinterprets correct context) and incurs additional inference costs due to its agentic, LLM-based memory construction and retrieval pipeline. Future work is suggested to extend boundary-preserving management to embodied agents and decision-making tasks.