Ruva: Transparent On-Device Personal AI Graph Reasoning

• Storage Architecture: Hybrid engine using SQLite extended with sqlite-vec for vector operations and GraphRAG Edge operations, enabling both traditional database queries and graph traversals[1] • Graph Topology: 'Spiderweb' structure with a central User node as root; all entities (people, locations), Events, Photos, and Messages branch outward through typed edges representing relationships[1] • Retrieval Pipeline: (1) Anchor identification from query; (2) Configurable N-hop topological expansion to identify connected events; (3) Subgraph serialization with explicit timestamps and relationships[1] • Reasoning Component: Retrieved subgraph fed into a Small Vision Language Model (SVLM) for answer generation, leveraging explicit temporal and relational metadata rather than embedding similarity[1] • Multimodal Support: System ingests multimodal memories (text, images, messages) and represents them as explicit nodes in the knowledge graph[1] • Memory Deletion: Precise redaction of specific facts by removing nodes and edges from the graph, mathematically eliminating the concept rather than leaving probabilistic traces[2]

Ruva represents a paradigm shift in Personal AI architecture from opaque statistical matching to transparent, user-controllable reasoning systems. This addresses critical regulatory pressures around data privacy (GDPR's 'Right to be Forgotten'), AI accountability, and hallucination mitigation—issues that have plagued vector-based RAG systems. The on-device graph reasoning approach could influence how enterprises and consumers demand transparency from AI systems, potentially accelerating adoption of explainable AI architectures. However, scalability of graph-based reasoning compared to vector similarity search remains an open question for large-scale deployments. The work also signals growing recognition that Personal AI systems require fundamentally different architectural principles than general-purpose LLMs, emphasizing user agency and verifiable knowledge representation over statistical approximation.