Dual-Interface User-Controlled Memory Management for Generative Artificial Intelligence: A Unified Architecture for GUI-Based Manual Configuration, Programmatic Control, Memory Composition, and Adaptive Weighting

Memory mechanisms are becoming increasingly important in generative artificial intelligence (AI), enabling large language models (LLMs) to retain user information, retrieve previous experiences, and personalize subsequent interactions. However, existing memory systems frequently emphasize automatic memory extraction and system-managed retrieval, while providing limited unified control over memory composition, retrieval priorities, configuration persistence, and interaction methods. Furthermore, configuration mechanisms designed for developers may require programming knowledge, whereas simplified graphical interfaces may not expose the flexibility required by advanced users. This paper proposes a dual-interface, user-controlled memory architecture that enables both GUI-based manual configuration and programmatic configuration through APIs, code, structured configuration files, and automated workflows. Both interfaces operate on a shared memory-policy engine, allowing users with different technical backgrounds to manage the same underlying memory environment without maintaining separate configurations. The architecture supports user-directed memory addition from multiple sources, independent memory collections, arbitrary collection composition, individual memory weighting, user-defined retrieval-factor weighting, context-sensitive adaptive weighting, and three configuration modes: session-specific, persistent-default, and hybrid. In the hybrid mode, temporary session-level modifications are isolated from persistent defaults unless users explicitly save the changes. The proposed system also introduces cross-interface synchronization, policy versioning, configuration provenance, and explicit precedence rules to maintain consistency between manual and programmatic operations. A formal policy model, operational workflow, research propositions, and a proposed evaluation methodology are presented. This paper does not claim experimental validation or performance superiority. Its central contribution is a conceptual architecture in which memory management becomes simultaneously accessible to ordinary users and extensible for developers, while preserving a unified and verifiable configuration model.