One of the key architectural challenges is the inability of current transformer-based models to retain information from previous conversations or tasks beyond their maximum 'context window' of tokens, which prevents them from leveraging the information that has been generated. A major architectural challenge is the need for existing transformer-based models to retain information from previous conversations or tasks beyond a maximum 'context window' of tokens, which prevents them from utilizing the information that has already been generated. This restriction is extremely restrictive on the ability of either multi-session agents, or longitudinal user interactions, or complex reasoning chains of autonomous agents that need coherent access to events that are separated by thousands of tokens or many conversational turns. In this paper we propose a new architecture called Adaptive Memory-Augmented Agentic System (AMAS) that combines an agentic design with short-term memory and long-term memory modules and allows for arbitrarily long interaction horizon with the preservation of the contextual fidelity by using a dedicated memory retrieval agent operating on the adaptive relevance scores. The architecture of AMAS breaks down memory into four stages: acquisition, encoding, retrieval, and adaptive update, which is controlled by a mathematically principled relevance-weighted memory consolidation scheme. AMAS also outperforms standard LLM baselines by 20 and 25 percentage points on context accuracy and memory retention respectively, on the LongBench benchmark, multi-session conversational datasets and Natural Questions, while maintaining competitive latency at 1,500ms end-to-end.

Keywords : Agentic AI, LLM , Retrieval-Augmented Generation, Vector Databases, Spiking Neural Networks.

Authors : Kuppireddy Krishna Reddy

Title : Adaptive Memory-Augmented Agentic Systems for Long-Term Context Preservation in Large Language Model Environments

Volume/Issue : 2026;3(2 ( April - June ))

Page No : 30 - 36

New problems have been created in ‘aero-elastic stability' in modern Formula 1 ground-effect aerodynamics. Of particular interest is high-frequency oscillations, "porpoising" and structural wing flutter. This paper develops an idea of Multi-Modal Sensor Fusion framework which is applicable to real-time aerodynamic stability monitoring and mapping. This design is distinctly different from existing telemetry designs based on data transfers from individual sources operating with high speed optical deflection monitoring and high frequency chassis vibration monitoring (Inertial Measurement Units). The system uses a deep-learning network detection pipeline to detect the visual deformation of the structure and crosses this with a Z-axis impulse algorithm to calculate the quantity of vertical load fluctuations. These inputs are then added together to create a dynamic Instability Index that may be in one of three states: Critical, Transitional, Stable. Experimental simulations have demonstrated that this two modal concept has been possible with a very high level of accuracy (and a 94% success rate) to predict aerodynamic stalls as well as to reduce the “signal noise” generated by track irregularities or kerb strikes. This output is automatically visualized with an engineer ready prioritized visualization in a Spatial-Temporal Track Dashboard describing the location where an aero-elastic instability is impacting the performance of the vehicle. This can be implemented on a wide scale and becomes an edge computing solution to enhance lap time and vehicle safety in very high downforce.

Keywords : Aero-Elasticity, Formula 1, Sensor Fusion, DL, Ground-effect, RTSM.

Authors : Ramya Thudumu

Title : Multi - Sensor Fusion for Real-Time Aero-Elastic Stability Mapping in Ground-Effect Formula 1 Vehicles

Volume/Issue : 2026;3(2 ( April - June ))

Page No : 23 - 29

Reliable seizure detection in wearable electroencephalography (EEG) systems requires both high classification accuracy and strict computational efficiency. Conventional hospital-grade seizure monitors depend on dense 32–128 channel electrode arrays that are impractical for long-term ambulatory monitoring due to elevated power consumption, patient discomfort, and high hardware cost. This paper presents a low-channel wearable EEG seizure detection framework that achieves near-dense-array accuracy using only 4–8 scalp electrodes by exploiting Riemannian geometric representations. Multichannel EEG epochs are represented as symmetric positive definite (SPD) covariance matrices residing on a non-Euclidean Riemannian manifold. Geodesic distances between neural states are computed using the affine-invariant Riemannian metric, and tangent space projection yields Euclidean feature vectors that preserve the intrinsic curved geometry of the SPD manifold. A lightweight SVM classifier trained on geodesic features achieves 98.1% classification accuracy, an F1-score of 0.98, 29 ms inference latency, and 1.9 W power consumption on embedded hardware outperforming dense-channel CNN baselines while consuming 50% less power. Experiments conducted on a 21-subject benchmark EEG epilepsy dataset validate the robustness and efficiency of the proposed framework for real-time, battery-operated wearable deployment.

Keywords : Wearable Healthcare, Geodesic Features, Riemannian Geometry, Edge AI, Manifold Learning.

Authors : D Aruna Kumari

Title : Low-Channel Wearable EEG Seizure Detection Using Geodesic Features and Riemannian Manifold Learning

Volume/Issue : 2026;3(2 ( April - June ))

Page No : 15 - 22

The study of emotion recognition via EEG is important due to the role emotion plays in decision making, behaviour, mental health, human computer interaction and understanding of emotion from an EEG perspective. Traditional machine learning methods use hand-crafted features and previously have not been able to cope with non-linear brain cell activity patterns. While the deep learning architectures excel at building representations, the data and compute power needed are often large. In this paper, a quantum kernel learning (QKL) approach for emotions recognition using EEG data is suggested. The framework combines both the spectral preprocessing and the covariance-based representations on the symmetric positive definite (SPD) Riemannian manifold and tangent space embedding and quantum support vector learning. The signals from EEG are mapped into a structured covariance representation that is able to maintain inter-channel statistical structures. The filters in quantum feature maps are parameterized quantum circuits that lift classical EEG representations to exponentially larger Hilbert spaces, and thus offer a more powerfully discriminative kernel where classical EEG cannot compete. The QSVM dual optimization problem is tackled in a classical manner with the quantum kernel Gram matrix. A 32 channel 40 subject benchmark is used for the experimental analysis with 97.5% classification accuracy and higher robustness against the degradation of the signals when compared to the state of the art SVM, CNN-SVM, and CNN-QSVM baseline methods on all of the evaluation metrics.

Keywords : QML, Quantum Kernel, Quantum SVM , Riemannian Geometry, SPD Manifold, Brain Signal Analysis

Authors : Surya Pavan Kumar Gudla

Title : Quantum Kernel Learning for Emotion Recognition Using EEG

Volume/Issue : 2026;3(2 ( April - June ))

Page No : 7 - 14

Epilepsy is a common neural disease among the children and in such cases, early and proper diagnosis is of paramount importance to en-sure a successful treatment. The signals of EEG in paediatrics are very unstable and noisy which complicates the process of automatic seizure detection in traditional systems. The paper presents a quantum-classical hybrid frame-work of the detection of epilepsy by EEG topographic map. EEG is processed in-to topographic representations on the scalp and de-composed in-to standard frequency bands. A light-weight convolutional neural network derives spatial features out of these maps and they are then classified by a Quantum Support Vector Machine with amplitude embedding. The framework can be used to evaluate paediatric EEGs, especially when compared to classical and quantum classifiers, which proves the potential of the framework.

Keywords : Pediatric Epilepsy, EEG Topographic Maps, CNN, Quantum SVM, QML, Biomedical signal pro-processing.

Authors : M Anjan Kumar , D Ramana Reddy , Rakhee

Title : Intelligent Paediatric Epilepsy Detection using Quantum Computing

Volume/Issue : 2026;3(2 ( April - June ))

Page No : 1 - 6