Adaptive Mutual–Human Shared Memory: A Universal Framework for Multi-Model Memory Sharing, Human-Assisted Memory Management, and Adaptive Mechanism Switching
The increasing deployment of large language models (LLMs), autonomous agents, and collaborative artificial intelligence systems has created a fundamental requirement for persistent, transferable, and continuously evolving memory. Although individual AI systems can maintain external memories and multiple agents can exchange information through shared memory repositories, existing approaches do not necessarily establish a unified mechanism for adaptively determining whether memory acquisition, retention, retrieval, consolidation, correction, updating, and reconstruction should be performed through interactions among multiple AI models, through human assistance, or through their coordinated combination. This paper introduces Adaptive Mutual–Human Shared Memory (AMHSM), a universal conceptual and architectural framework that integrates two independent but combinable dimensions of memory operation: Mutual Shared Memory and Human-Assisted Memory. Mutual Shared Memory enables multiple distinct AI models to access, contribute to, exchange, evaluate, synchronize, and collectively maintain shared memory. Human-Assisted Memory enables human individuals or groups to contribute knowledge, correct inaccurate memories, verify information, resolve conflicts, authorize memory operations, and guide memory organization. The central innovation is that neither mechanism is permanently designated as the exclusive memory pathway. Instead, AMHSM introduces Adaptive Memory Mechanism Selection, Adaptive Memory Mechanism Switching, Adaptive Memory Combination, and Adaptive Memory Reconfiguration, allowing a system to dynamically determine which memory mechanism is most appropriate for a particular information item, operation, task, and contextual condition. The framework distinguishes the identity of a memory operation from the mechanism through which that operation is performed and extends the Universal Mutual-Human Model (UMHM) into a specialized architecture for general-purpose intelligent memory systems. A formal two-dimensional representation, memory lifecycle model, adaptive decision formulation, consistency and provenance mechanisms, privacy-aware access architecture, theoretical propositions, application scenarios, and empirical evaluation methodology are developed. To the best of our knowledge, this paper is the first to formally articulate a universal, function-general shared-memory architecture specifically integrating multi-model mutual memory operations and human-assisted memory operations through an explicit adaptive mechanism-selection, switching, combination, and reconfiguration framework. This priority claim concerns the particular theoretical integration and architecture proposed here, rather than the broader established fields of multi-agent memory, shared databases, human-in-the-loop systems, or adaptive memory management. The framework provides a theoretical foundation for intelligent systems that can remember collectively, receive human assistance when appropriate, and adaptively reorganize their memory processes without depending on a single fixed memory pathway.
Authors
- Jincheng Zhang (ORCID: https://orcid.org/0009-0005-1860-0009)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24362547
- Primary Topic
- AI-based Problem Solving and Planning
- Type
- preprint