Current Large Language Models (LLMs) lack the critical ability to learn continuously, relying instead on ephemeral context windows that scale poorly, suffer from attention decay, and cannot retain knowledge between sessions. We propose the Memory Augmented Generative System (MAGS), a neuroscience-inspired architecture introducing dynamic memory modules to endow LLMs with human-like learning capabilities. MAGS integrates dual memory blocks: an Episodic Memory Block (eMB) for short-term, session-specific recall, and a Semantic Memory Block (sMB) for long-term, continuously updated knowledge. Memory updates occur via a reinforcement learning framework using Group Relative Policy Optimization (GRPO) and Low Rank Adaptation (LoRA), enabling models to autonomously modify and refine their knowledge graphs. Drawing inspiration from Hebbian plasticity and neurogenesis, MAGS strengthens memory paths that contribute positively to responses while pruning low-importance connections through k-core decomposition and a Game-of-Life-inspired algorithm. A novel recall and engram mechanism guides memory retrieval and updates, supported by explainable, inspectable graph structures. We validate MAGS through a custom-designed game environment that tests adaptability, recall accuracy, and memory update fidelity under evolving conditions. This architecture offers a scalable path toward lifelong learning in LLMs, enabling efficient, explainable, and robust memory systems without the need for costly retraining.

Despite being the fastest growing Neurodegenerative disease in the world, with over ten million cases worldwide, there is no definitive diagnosis method for Parkinson’s Disease currently. Current diagnosis technologies misdiagnose one in four patients, relying on tests and technologies that are subjective to the administrating clinicians. Furthermore, they are inaccessible to billions due to immobility, geographic barriers, or associated costs. With early and accurate diagnosis being vital to effective treatment, a tremendous issue emerges. ParkinSense addresses these issues, serving as an automated web application that contactlessly analyzes three cardinal symptoms of Parkinson’s Disease over a standard webcam and microphone: Hypomimia, Dysarthria, and Bradykinesia. With Ensemble Learning coupled with various classifiers, ParkinSense was able to achieve an accuracy rate of 99.72% when tested on synthetic patients. By examining and combining multiple modalities, ParkinSense boasts accuracy rates higher than many advanced unimodal technologies, as patients often do not substantially display all the symptoms of Parkinson’s Disease, leading to misdiagnosis when relying on one symptom for diagnosis. Furthermore, ParkinSense is able to accurately estimate disease severity for positively diagnosed patients, with a Mean Absolute Error of 9.7 on the Unified Parkinson’s Disease Rating Scale. ParkinSense strongly suggests that an entirely contactless diagnosis of Parkinson’s Disease through digital biomarkers can be effective, and web applications can be utilized for rapid, automated, and remote diagnosis and severity analysis, which can be crucial for those affected by barriers that prevent access to healthcare.

Despite being the fastest growing Neurodegenerative disease in the world, Parkinson’s Disease (PD) currently has no definitive diagnosis. Current methods of diagnosis are inaccessible to billions due to geographic, economic, and age-related mobility barriers, disproportionally so in rural areas or regions without ready access to healthcare; PD diagnosis can cost upwards of $5K, require a lengthy visit to a nearby neurologist, and can involve intensive scans. Notwithstanding, currently, almost 1 in 3 patients are misdiagnosed, and an estimated 40% of patients with PD are never successfully diagnosed with PD. Since treatment is often more effective in the earlier stages of PD, these issues are vital to remedy. ParkinSense, a web and mobile application was developed to combat these issues. ParkinSense allows one to diagnose themselves for PD from anywhere in the world in under 12 minutes, requiring only a device and internet connection. ParkinSense analyzes 3 cardinal symptoms of PD entirely contactlessly and automatically: Dysarthria, Hypomimia, and Bradykinesia. It involves over 16 machine learning models that utilize ensemble learning and multimodal data fusion to improve accuracy. ParkinSense boasts an overall accuracy rate of 99.72% when validated on public databases points, synthetic patients, and 106 authentic patients. Even after proper diagnosis, there is a critical need to monitor PD severity, to analyze disorder progression and the effectiveness of certain treatments. ParkinSense is able to estimate the disease severity on the two common PD severity classification scales: the UPDRS, and the H&Y Scale, with a relative mean absolute error of less than 4.5%. Furthermore, trends of severity can be analyzed over time via a GUI. Not only is ParkinSense contactless, rapid, accurate, and cheap, it is the first public method of remote PD diagnosis and severity analysis, potentially providing billions with a crucial healthcare tool.

Despite being the fastest-growing Neurodegenerative disease in the world, with cases expected to surge to 30 million by 2030, Parkinson’s Disease (PD) currently has no definitive diagnosis. Current methods of diagnosis are inaccessible to billions due to geographic, economic, and age-related mobility barriers, disproportionately so in rural areas or regions without ready access to healthcare, such as third-world countries; A single PD diagnosis can cost upwards of $5K, require a lengthy visit to a nearby neurologist, and can involve intensive scans. Notwithstanding, currently, almost 1 in 3 patients are misdiagnosed, and an estimated 40% of patients with PD are never successfully diagnosed with PD. Since treatment is often more effective in the earlier stages of PD, these issues are vital to remedy so that PD patients can be diagnosed as early as possible. ParkinSense, a web and mobile application, was developed to combat these issues, and understand whether Machine Learning can be exploited for accurate and rapid remote PD diagnosis and severity analysis. ParkinSense allows one to diagnose themselves for PD from anywhere in the world in under 12 minutes, requiring only a webcam, microphone, and internet connection. ParkinSense analyzes 4 cardinal symptoms of PD entirely contactlessly and automatically: Dysarthia, Hypomimia, Tremors, and Bradykinesia. ParkinSense is able to complete automated feature extraction. It was developed using recent innovations in Machine Learning, in conjunction with over 16 machine learning models that utilize ensemble learning, multimodal data fusion, and Bayesian Optimization to improve accuracy. ParkinSense boasts an overall accuracy rate of 99.72% when trained on public databases totaling 9 million data points, synthetic patients, and 124 authentic patients. Sensitivity, a valuable metric to maximize in medical Machine Learning solutions, was 0.999. Even after proper diagnosis, there is a critical need to monitor PD severity, to analyze disorder progression and the effectiveness of specific treatments. ParkinSense is able to estimate the disease severity on the two common PD severity classification scales with a normalized root-mean-square error of less than 0.0907. Together, ParkinSense outperforms average diagnosis methods by 34%, and average classification methods by 223%. Furthermore, trends of severity can be analyzed, for an assessment of treatment effectiveness, via a GUI. Results can be shared in real-time with a Neurologist or family member for monitoring. Neurologists can submit DaTScan, MRI scans, and other information to ParkinSense for further analysis and accuracy gain. Finally, ParkinSense enables treatment, by predicting possible drug treatments, and by providing tasks that can suppress and reduce the symptoms of Parkinson’s disease, entirely from home. This treatment saw a substantial decrease in disease severity, on average, in 52 patients of different backgrounds and cultures. During this study, all HIPAA Privacy, OSHA, local regulations, and IRB regulations were followed to ensure patient privacy and safety. Not only is ParkinSense contactless, rapid, accurate, and cheap, it is the first public method of remote PD diagnosis, telemonitoring, or treatment, potentially providing billions with a crucial healthcare tool, that is otherwise essentially inaccessible to them. Furthermore, ParkinSense was shown to be effective in both the early and late stages of PD, allowing for early and more effective treatment. These advances can be translated over to other neurodegenerative disorders, and diseases with digital biomarkers. Further research can be conducted on unutilized extracted features, such as eye saccades, which could offer additional success.

Abstract redacted due to paper being in review (Nature policy).