Parallel-World Adaptive Self-Memory and Shared-Memory Switching for Large Language Models

Abstract

Large language model (LLM) agents increasingly rely on external memory to support long-term interaction, task continuity, experience reuse, and collaborative problem-solving. However, existing memory architectures often emphasize individual memory management, shared memory repositories, or predefined collaboration strategies. These approaches leave an important design question: how can multiple memory perspectives generated by the same model coexist with memories maintained by heterogeneous models, while the system adaptively determines when to operate independently and when to exchange information? This paper proposes Parallel-World Adaptive Self-Memory and Shared-Memory Switching (PW-ASMS), a unified framework that combines parallel memory worlds, adaptive self-memory, shared-memory collaboration, and cross-model memory integration. Parallel memory worlds are created by assigning different memory orientations to instances of the same underlying model. Each world independently selects, organizes, updates, and retrieves experiences according to its designated orientation. A dynamic controller subsequently determines whether an agent should rely on its own memory, exchange information with other parallel worlds, consult a shared memory repository, or combine these operations. The framework also supports heterogeneous LLMs whose memories can contribute complementary information to a common collaborative process. The proposed architecture consists of parallel memory worlds, individual memory stores, a governed shared memory layer, a memory evaluation module, and an adaptive switching controller. A utility-based decision mechanism considers task relevance, information novelty, confidence, memory consistency, communication cost, privacy constraints, and expected downstream benefit. The framework is designed to reduce unnecessary information exchange while preserving opportunities for collaborative knowledge acquisition. We formulate research questions, operational procedures, evaluation metrics, and experimental comparisons to assess its effectiveness. The paper provides a conceptual and methodological foundation for investigating adaptive memory organization in multi-agent LLM systems; its effectiveness remains to be established empirically.