User-Controlled Adaptive Memory Configuration for Generative AI: Fine-Grained Memory Selection, Persistent Memory Profiles, and Conversation-Level Permission Management
Year: 2026
Type: preprint
Abstract: Persistent memory is increasingly becoming a foundational component of generative artificial intelligence (AI), enabling large language models (LLMs) to maintain continuity across conversations, retain user preferences, and adapt their responses to individual users. However, the growing integration of persistent memory introduces a fundamental interaction and governance challenge: users may have limited control over which specific memories are made available to an individual conversation, how memory selections persist across subsequent conversations, and whether temporary changes can override previously established memory configurations. Existing approaches have primarily emphasized memory extraction, storage, organization, retrieval efficiency, and automatic personalization. Although emerging systems provide memory-management interfaces, global memory controls, project-level separation, or conversation-level memory activation controls, these capabilities do not necessarily constitute a unified mechanism for user-defined, fine-grained memory composition across individual and future conversations. This paper proposes User-Controlled Adaptive Memory Configuration (UCAMC), a conceptual framework that enables users to determine not only whether an AI system can access persistent memory but also which specific memory units it is authorized to use, how selected memory configurations are organized into reusable profiles, and how configuration policies apply across different temporal and conversational scopes. The framework introduces five interrelated mechanisms: Fine-Grained Memory Selection, Memory Profile Management, Persistent Memory Policies, Conversation-Level Override, and Memory Permission Isolation. It distinguishes three configuration scopes: current-conversation configuration, persistent default configuration for subsequent conversations, and reusable named memory profiles. A formal model is developed to represent memory units, authorization states, configuration profiles, policy inheritance, and conversation-specific overrides. The framework further establishes a separation between memory availability, memory retrieval authorization, and actual context injection, requiring all memory-processing operations to respect the user's active permissions. The proposed architecture is intended to support research on user autonomy, contextual relevance, privacy, personalization, and transparent memory governance in multimodal and text-based generative AI systems. This paper presents a conceptual and testable research framework rather than an experimentally validated implementation; no empirical performance or usability results are claimed.
Author: Zhang, Jincheng
Language: English
Topic: Access Control and Trust
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225657167
Cites:
Cited by:
Related to: