💬 NLP

Know It, Act on It: Investigating Memory Utilization in LLM Personalization

This paper introduces a decoupled evaluation paradigm to reveal a significant gap between Large Language Model agents' ability to recall user preferences and their ability to act on them, finding that while memory architectures help, utilization remains critically weak for high-stakes domains like health and therapy.

Original authors: Zhaoxin Feng, Jianfei Ma, Emmanuele Chersoni

Published 2026-08-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhaoxin Feng, Jianfei Ma, Emmanuele Chersoni

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 are talking to a super-smart robot friend who has been chatting with you for months. You've told it your favorite movies, your fear of spiders, and that you're allergic to peanuts. You expect this robot to remember these things and use them to help you. This is the world of Large Language Model (LLM) agents. Think of them as digital companions that are evolving from simple tools (like a calculator) into long-term friends that can hold a conversation over weeks or even years. To do this, they need memory: a way to store what you've told them, organize it, and pull it up when needed.

But here is the tricky part: just because a robot has the information in its brain doesn't mean it will use it. This is called the knowledge utilization problem. It's like having a library full of books but never actually reading the one you need when you're in a crisis. Scientists have known for a while that these AI models sometimes forget things, but they haven't been sure if the problem is that the robot forgot the fact, or if it remembered it perfectly but just decided to ignore it. This distinction matters a lot. If a robot forgets you're allergic to peanuts, that's a memory failure. If it remembers you're allergic but still suggests a peanut-butter sandwich because it's "thinking creatively," that's a much more dangerous failure of judgment.

This paper, titled "Know It, Act on It," dives right into that messy middle ground. The researchers, Zhaoxin Feng, Jianfei Ma, and Emmanuele Chersoni from The Hong Kong Polytechnic University, wanted to figure out exactly where the robot breaks down. They built a clever testing game called KnowAct. Imagine they give the robot a secret note about a user's preference (like "I hate spicy food"). Then, they run two tests on that same secret. First, the Know Test: they ask the robot directly, "What food does this user hate?" If the robot says "Spicy food," it passes. Second, the Act Test: they give the robot a new scenario, like "The user wants to order dinner," but they don't mention the allergy. They see if the robot naturally suggests non-spicy food.

The results were a bit of a shock. The team tested 16 different memory systems (ranging from simple "long-context" models that just read everything at once, to complex "agentic" systems that try to manage their own notes). They found a huge gap between knowing and acting. In many cases, the robots passed the Know Test with flying colors—they could recite the user's preferences perfectly when asked. But when it came to the Act Test, they often failed miserably. For example, a robot might correctly tell you, "Yes, this user is allergic to peanuts," but then, in a separate conversation, enthusiastically recommend a Thai dish topped with crushed peanuts.

The researchers discovered that this gap isn't just a small glitch; it's a massive wall. Even the best-performing systems only managed to turn about two-thirds of their remembered facts into helpful actions. It gets worse depending on how the information was given. If a user explicitly said, "I am allergic to peanuts," the robot was good at remembering it. But if the user hinted at it indirectly (like complaining about a sneezy day while asking to edit an email), the robot often failed to even store the memory in the first place.

Perhaps the most worrying finding was about what kind of information was hardest to use. The robots were surprisingly bad at acting on health and therapy-related preferences. Even when they remembered that a user had hay fever or was feeling stressed, they frequently failed to adjust their advice. They might suggest a sunny picnic to someone with severe pollen allergies or recommend a high-energy workout to someone feeling emotionally drained. The paper suggests that while memory systems help the robot store more facts, they don't necessarily teach the robot how to care about those facts when it's time to make a decision.

In short, the paper shows that giving AI a better memory doesn't automatically make it a better friend. We are currently building robots that are great at memorizing your secrets but terrible at using them to keep you safe or happy. The authors suggest that future AI needs to focus less on just "storing" data and more on learning how to actually apply that data in real-world situations, especially when it comes to health and safety. Until then, you might want to double-check your robot friend's advice before eating that peanut-butter sandwich.

Technical Summary: Know It, Act on It: Investigating Memory Utilization in LLM Personalization

Problem Statement

As Large Language Model (LLM) agents evolve from stateless tools into personalized long-term companions, memory has become a core capability for storing, organizing, and retrieving user information. However, a critical gap exists between remembering information and acting on it. Current systems often suffer from a "knowledge utilization problem": an agent may successfully recall a user preference when explicitly queried but fail to incorporate that same preference into its behavioral responses in relevant scenarios.

Existing research has largely evaluated either recall (can the model retrieve the fact?) or behavior (does the model generate a tailored response?), but rarely both on the same piece of stored information. Furthermore, most studies operate in static long-context settings rather than simulating the incremental, multi-session pipelines used by real-world agentic memory systems. This lack of paired evaluation makes it unclear whether failures in personalization stem from memory retrieval errors or from a failure to utilize retrieved information.

Methodology

The authors introduce KnowAct, a decoupled evaluation paradigm designed to isolate the breakdown between memory retrieval and behavioral application.

1. Experimental Design

Paired Evaluation: For every user preference, the system undergoes two distinct tests:
- Know Test: A direct, first-person question (e.g., "Do I have seasonal allergies?") to verify if the agent can recall the specific preference.
- Act Test: A natural, task-oriented scenario (e.g., "Suggest a spring picnic location") where the preference should influence the response, but the preference itself is not mentioned in the prompt.
Expression Strength Gradient: To test robustness, preferences are embedded at three levels of directness:
- Explicit: The user directly states the preference as the central topic.
- Incidental: The preference is embedded as a side detail within a task-oriented request (e.g., polishing an email that mentions sneezing).
- Inferential: The preference must be deduced from behavioral patterns and contextual cues without being explicitly stated.
Dataset Construction: The study utilizes 50 synthetic personas from PersonaMem-v2, sampling 1,000 target preferences across five categories (stereotypical, anti-stereotypical, neutral, health/medical, therapy/emotional). These are injected into conversation histories alongside non-target "noise" chunks to simulate realistic retrieval environments.

2. Systems Evaluated

The framework evaluates 16 memory systems across five architectural categories:

Long-Context Baselines: Models like GPT-4o-mini, GPT-4o, Gemini 3.1 Flash, and Claude 4.6 Sonnet that rely on full context windows.
Simple RAG: Systems using sparse lexical matching (BM25).
Embedding-based RAG: Systems using dense vector retrieval (e.g., text-embedding-3-small/large).
Structure-Augmented RAG: Systems organizing memory into knowledge graphs or structured stores (e.g., Mem0, Zep, HippoRAG-v2).
Agentic Memory: Systems where the agent controls memory operations (e.g., Letta/MemGPT, Self-RAG, MemoryOS).

3. Metrics

Know Accuracy: The pass rate of the recall test.
Act Accuracy: The pass rate of the behavioral test (judged by comparing memory-augmented responses against zero-memory baselines).
Utilization Rate: The fraction of correctly remembered preferences that are also acted upon ( $N_{KP,AP} / (N_{KP,AP} + N_{KP,AF})$ ).

Key Results

1. The Know-Act Dissociation

A pervasive gap exists between knowing and acting. Even the best-performing systems convert only about two-thirds of successfully remembered preferences into appropriate behavioral responses.

GPT-4o-mini (Long-Context): Achieves high Know Accuracy on explicit preferences but a Utilization Rate of only 16.3%.
Claude 4.6 Sonnet: The top performer, yet still fails to act on roughly one-third of remembered preferences (Utilization Rate: 65.1%).
Memory Architectures: Dedicated memory systems (e.g., Mem0) significantly narrow this gap. Mem0 raised GPT-4o-mini's Utilization Rate from 16.3% to 54.6%, suggesting that effective architectures do not just retrieve more information but present it in a form the model can act upon.

2. Expression Strength Effects

Incidental is Hardest to Recall: Counter-intuitively, the Incidental condition (preferences hidden in task details) yields the lowest Know Accuracy across all architectures, even lower than the Inferential condition. This suggests that side details are difficult for both retrieval systems and long-context models to flag as noteworthy.
Explicit Reveals Utilization Failure: Under Explicit conditions, Know Accuracy is high, but the Know-Act gap is widest. This indicates that when information is clearly available, the bottleneck shifts from memory retrieval to utilization.

3. Domain-Specific Risks

Utilization rates diverge sharply by preference type, even when recall is comparable:

Health & Medical: These preferences consistently show the lowest Utilization Rates across all systems.
Therapy & Emotional: These show high recall but low utilization.
Significance: Failures in these domains carry the highest real-world stakes (e.g., recommending outdoor activities to a user with pollen allergies), highlighting a critical safety risk.

4. Failure Attribution

An oracle study on Mem0 cases (Know Pass / Act Fail) revealed the primary causes of failure:

Comprehension Failure (53%): The most common error. The relevant information is present in the context (retrieved), but the model fails to extract the constraint or preference from the surrounding text.
Retrieval Failure (30.7%): The system fails to surface the relevant dialogue chunk.
Application Failure (16.2%): The preference is explicitly stated in the context, yet the model still fails to incorporate it into the response, indicating an intrinsic limitation in the underlying LLM's ability to utilize provided constraints.

Significance and Claims

The paper claims that remembering is not sufficient for personalization. The primary contribution is the demonstration that current memory-augmented LLM agents suffer from a systematic "utilization bottleneck" rather than just a retrieval bottleneck.

Beyond Storage: The value of memory architectures lies not merely in increasing recall, but in structuring information to make it actionable.
Safety Implications: The inability to act on health and emotional preferences poses tangible risks in real-world deployments.
Evaluation Standard: The authors argue that future benchmarks must evaluate recall and behavior simultaneously on the same information to accurately diagnose personalization failures.

The paper concludes that while memory systems have improved retrieval, the field must now focus on ensuring that personal memories are selectively retrieved, correctly interpreted, and reliably translated into context-appropriate behavior.