User-Controlled Adaptive Memory Weighting and Configuration in Generative Artificial Intelligence: A Unified Architecture for Memory-Level Prioritization, Factor-Level Control, Contextual Adaptation, and Cross-Conversation Policy Management

Long-term memory has become an important component of generative artificial intelligence (AI), enabling large language models (LLMs) and AI agents to retain user preferences, contextual information, prior decisions, and task-related knowledge across interactions. However, existing memory systems commonly rely on automatically generated importance scores, relevance-based retrieval, recency decay, or system-defined weighting mechanisms. Although these methods can improve memory selection, they provide users with limited direct control over how individual memories are prioritized, how different weighting factors contribute to retrieval decisions, and how memory priorities should change across different tasks and conversational contexts. Furthermore, memory configurations are frequently treated as either globally persistent or locally applicable, leaving the relationship between temporary conversational preferences, persistent defaults, and temporary overrides insufficiently formalized as a unified user-controlled policy. This paper proposes a User-Controlled Adaptive Memory Weighting and Configuration Architecture (UCAM-WC) for generative AI systems. The architecture integrates three complementary weighting mechanisms: (1) memory-level weighting, through which users directly assign priority values to individual memory items or memory groups; (2) factor-level weighting, through which users configure the relative contributions of criteria such as semantic relevance, importance, recency, reliability, frequency of use, and task relevance; and (3) context-level dynamic weighting, through which users define how memory priorities and weighting factors should adapt to different tasks, topics, projects, or conversational situations. These mechanisms are governed by three configuration modes: Mode A, Conversation-Only Configuration, in which settings apply exclusively to the current conversation; Mode B, Persistent Default Configuration, in which user-defined settings are inherited by subsequent conversations; and Mode C, Hybrid Configuration, in which persistent defaults can be temporarily overridden within individual conversations without modifying the global configuration. The architecture introduces a hierarchical policy model, explicit configuration precedence, memory-selection functions, contextual adaptation rules, and cross-conversation isolation constraints. It also establishes research propositions and a controlled evaluation methodology for assessing user autonomy, retrieval quality, contextual appropriateness, configuration consistency, and memory utility. Rather than claiming that memory weighting or session-level configuration is individually unprecedented, this paper identifies the proposed integration of user-programmable weighting mechanisms and multi-scope configuration governance as its principal architectural contribution. The framework is conceptual and methodological; no empirical performance results are claimed.