Multidimensional Memory for Persistent AI Assistants

This concept paper describes an architecture for persistent AI memory that grows over years, combines all interactions in a shared knowledge base, and nevertheless remains selectively, efficiently, and traceably retrievable. The approach replaces a rigid topic hierarchy with a mathematical, multidimensional memory space. Memories, preferences, corrections, research results, and further insights are stored as points with mutable coordinates and attributes. A query is not routed along a predefined path through a topic tree. Instead, the system directly determines the regions relevant to the current situation and combines multiple retrieval signals, such as topic position, relevance, informational depth, time, memory type, and epistemic confidence. Two storage layers separate higher-level insights developed with the user from general information. The immutable source archive preserves user inputs, agent responses, documents, and research in their original form. The memory layer above condenses this source information into dynamic points that can be moved, reweighted, corrected, and versioned. If a topic loses importance over the long term or content proves outdated, another entry is not merely appended. The active memory point itself is updated on the next retrieval. Over the long term, this structure enables three central capabilities. First, retrieval can depend on intent and therefore not merely search earlier chat histories, but selectively activate the knowledge areas relevant to a specific situation. Second, beyond direct matches, the memory system can reveal thematically related content, adjacent knowledge fields, and hidden cross-connections. Third, the distribution of points in the matrix can be shared with other systems or people without disclosing internal content: only the areas in which a user or their AI assistant has particular knowledge density and experience would be visible. On this basis, suitable contacts could be identified and selectively involved in an exchange.