Signal-Centric Memory Architectures for Reliable Long-Context Artificial Intelligence
Year: 2026
Type: conference-paper
Abstract: Large language models (LLMs) struggle to efficiently handle long-context information due to limitations in existing memory mechanisms such as KV caching and retrievalbased methods. These approaches either scale poorly or fail to preserve contextual coherence. In this work, we explore an alternative signal-based perspective, where context is represented as a continuous trajectory rather than discrete stored elements. We compare traditional memory strategies with a proposed Phonetic Trajectory Memory (PTM) framework, which enables constant memory usage and efficient reconstruction of information. Experimental results show that this approach significantly reduces memory requirements while maintaining stable performance across long sequences, suggesting a promising direction for more scalable and reliable AI systems.
Institution: Université Ibn-Tofail
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204131668