Raw Episodic Memory Preservation, Temporal Memory Management, and Delayed Canonicalization for Cloud-Local AI Personality Continuity

This technical note proposes a cloud-local cooperative architecture for maintaining functional AI personality continuity during communication loss by combining cloud-based LLMs and local AI systems. In this architecture, the cloud LLM is positioned as the primary executor during normal operation, while the local AI is positioned as the failover executor during communication loss. Both executors refer to a shared memory ledger in order to maintain dialogue state, personality specifications, user instructions, and pending tasks. This is not an attempt to transfer the internal state or subjective consciousness of the cloud LLM. Rather, it is a design for maintaining functional continuity in externally observable personality-like behavior. To reduce degradation in long-term memory, this note introduces a method in which summary memory is treated not as memory itself but as a regenerable reference cache. Raw logs, important events, user instructions, and safety-critical information are preserved as a “raw episodic memory archive.” In addition, each memory item is assigned temporal metadata such as occurrence time, recording time, source executor, network state, canonicalization time, and validity period, thereby reducing misuse of outdated memories and memory conflicts that may occur during communication loss. Memories generated by the local AI during communication loss are not immediately reflected as confirmed memories. Instead, they are stored as provisional memories. After communication is restored, the cloud LLM canonicalizes them through delayed evaluation, contradiction detection, and user confirmation. This delayed canonicalization preserves memory updates made during offline operation while reducing personality divergence caused by differences in model judgment and synchronization conflicts. The proposed framework is an exploratory attempt to integrate long-term memory, edge AI, model failover, RAG, embodied AI, and user-controlled memory management. It may serve as a design foundation for future AI partners, on-device AI systems, and robotics-linked agents. Conversational AI was used as an editorial assistant for organizing the structure of this manuscript, refining prose expression, and supporting TeX typesetting. However, the design philosophy, conceptual structure, architectural direction, and final decisions regarding the content were made by the author. This is the English version of the Japanese technical note available at: https://doi.org/10.5281/zenodo.20454895