Evie™: A Bitemporal Memory Architecture for Longitudinal Health Personalization in Resource-Efficient AI Deployments — A Design Framework

This paper describes the conceptual architecture of Evie™'s memory system — a four-phase bitemporal memory design for a conversational AI health companion operating within the Fitodermonutrição™ framework. The AI amnesia problem — structural statelessness of LLM-based conversational agents — is characterized as a clinically significant limitation in longitudinal health support contexts. Evie™'s architecture addresses this through a bitemporal data model that records both valid time (when a health fact was true) and transaction time (when it was recorded by the system), enabling retrospective temporal reasoning and non-destructive conflict resolution without proprietary vector database infrastructure. The four memory phases are described: structured bitemporal schema (Phase 1), asynchronous narrative memory summary generation (Phase 2), vector similarity retrieval using a native database extension (Phase 3), and automated fact extraction with conflict resolution (Phase 4). The architecture is designed for resource-efficient deployment prioritizing open-source components and negligible marginal cost per user — a design constraint motivated by LATAM deployment economics. The competitive landscape is surveyed against Mem0, Zep, MemoryBank, Google Memory Bank, and MemOS. Integration with the IDIBS™ dimensional structure as a bounded semantic domain is described. A four-phase validation roadmap is proposed. No proprietary implementation details are disclosed.