A Framework for Dynamic Memory Management and Personalized Agent Decision-Making in Retrieval- Augmented Generation

In the rapidly evolving world of artificial intelligence, large language models have become remarkably skilled at generating human-like text, answering questions, and holding conversations. However, these systems face a fundamental limitation: they possess a static knowledge base that does not naturally grow or adapt through conversation. When a user asks a series of questions over days or weeks, the model often forgets earlier details, treats each interaction as a fresh start, and fails to remember what the user already knows or prefers. To solve this, researchers have developed systems that combine these powerful language models with external databases, a method known as retrieval-augmented generation. This allows the AI to look up facts before answering, but even these improved systems often struggle to maintain a coherent story over long conversations or to tailor their responses to the specific needs and learning styles of individual users.

A team of researchers from Hechi University in China and INTI International University in Malaysia has proposed a new framework designed to fix these specific gaps. They call their system DMAP-RAG, a structure that integrates dynamic memory management and personalized decision-making into the retrieval process. Rather than treating a conversation as a simple list of past messages, the researchers built a system that organizes information like a growing tree, separating recent interactions from long-term history. This allows the AI to remember the immediate context of a chat while also storing summarized events from days or weeks ago in a structured way. Furthermore, the system actively builds a profile of the user as they speak, noting their knowledge level, preferred learning style, and areas of interest, and then uses this profile to decide how to search for information and how to phrase the final answer.

The researchers tested this framework across several different scenarios, including academic research questions, complex multi-step reasoning tasks, and a custom-built dataset for teaching computer programming. In these tests, the new system consistently outperformed standard approaches. When asked to answer questions based on scientific papers, it achieved a score of 78.9 percent, a significant improvement over existing methods. In tests requiring the AI to connect information across multiple documents, it reached a knowledge accuracy score of 66.1 percent. Perhaps most notably, when human evaluators judged the quality of the conversations, they rated the system's ability to stay on topic and maintain logical consistency at 4.4 out of 5, and its ability to personalize responses to the user at 4.3 out of 5. These results suggest that by giving the AI a structured way to remember and a specific way to understand the person it is talking to, the system can provide guidance that feels more natural and helpful.

The core of this improvement lies in how the system handles memory. Instead of storing every past message in a flat, unorganized list, the researchers created a dual-storage system. One part acts as a short-term cache, holding the last ten exchanges of a conversation to ensure the AI understands the immediate context. When a conversation pauses or ends, the system summarizes these exchanges and moves them into a long-term memory bank. This long-term bank is not a simple folder of text; it is organized as a tree structure where related topics branch out from one another. If a user asks about a specific concept today and returns next week to ask a related question, the system can navigate this tree to find the connection, rather than getting lost in a sea of unrelated data. This structure helps the AI avoid the common problem of forgetting the thread of a conversation after a break.

Alongside this memory system, the researchers implemented a module that continuously updates a profile of the user. As the user speaks, the system looks for clues about their background. If a user asks simple questions about basic concepts, the system infers they are a beginner. If they use technical terms or ask about complex debugging issues, the system upgrades its assessment of their expertise. It also tracks their learning preferences; if a user responds well to analogies or visual descriptions, the system notes this as a "visual learner." This profile is not static; it grows and changes with every interaction. When the system is ready to answer a question, it consults this profile to decide how to phrase the response, ensuring that a beginner gets a simple explanation while an expert receives a detailed technical breakdown.

The final piece of the framework is an agent decision core that acts as the brain of the operation. Before generating an answer, this component decides whether the system needs to look up information or if it can answer from what it already knows. It weighs factors such as the complexity of the question, the user's knowledge level, and the current state of the conversation. If a search is needed, the system rewrites the user's question to make it more effective for searching, combining the original query with the user's profile and recent history. This ensures that the information retrieved is not just relevant to the words typed, but also tailored to the specific person asking. Once the relevant information is gathered, the system synthesizes it into a final response that is coherent, accurate, and personalized.

In a specific test case involving a student learning the C programming language, the difference between this new system and older methods became clear. When a student who identified as a visual learner asked why a function sometimes failed to change a value, a standard system gave a dry, technical definition of "pass-by-value." The new system, however, remembered that the student was a visual learner and had previously discussed pointers. It responded by using an analogy of a book and a photocopy, explaining that the function was working on a copy of the book rather than the original. This response not only answered the question correctly but also connected it to previous learning and adapted the explanation style to the student's needs. Human evaluators rated this response significantly higher in both coherence and personalization than responses from other models.

The researchers acknowledge that this approach requires more computing power than simpler systems, as it involves maintaining complex memory structures and running additional decision-making steps. However, they found that the increase in computational cost was manageable on modern hardware. The system also has limitations, such as the time it takes to build a reliable user profile from scratch and the fact that it currently only handles text. Despite these constraints, the study suggests that combining structured memory with personalized agent decision-making offers a promising path forward. By giving artificial intelligence the ability to remember contextually and adapt to the individual, these systems move closer to becoming truly helpful partners in education, technical support, and long-term dialogue.