The ideas and papers that inform NOVA’s architecture — memory, multi-agent coordination, preference learning, and geometric structure.
Temporal-hierarchical memory consolidation (TiMem/HEMA-style): raw inputs → categorization → pattern recognition → synthesis. Progressive consolidation and complexity-aware recall inform how Living Memory grows with every interaction.
Debate vs. vote, complementary agents, and robust aggregation. Research on when voting suffices vs. when debate helps, and how to select complementary perspectives, guides the 12-Pillar council design. Governance (conflict resolution, policies, resource allocation) and incentive alignment (Shapley credit, reputation, social welfare) ensure agents cooperate even with competing priorities.
DPO and RLHF from behavioral telemetry; preference pairs from game and advisory interactions. Heterogeneous feedback and clustering for personalized, robust preference learning.
Six-layer geometry: The Core; Risk Intelligence, Governance, Living Memory; Cube (structural bridge); 12 Pillars with 100–1000 processing floors each and adaptive routing; Sphere (parallel processing, fair weighting); The Orchestrator. Stateful processing units with outcome-based lifecycle management, confidence self-calibration via per-entity SGD, and evolutionary pattern caching; consulted pillars run simultaneously (2–4× faster, no race conditions). Each entity has parameterized cognitive processing (e.g. The Analyst, The Innovator, Sentinel, Arbiter). Governance and incentive alignment ensure conflict resolution and collective outcomes.
Detailed research notes and implementation status live in the project docs. For partnerships or citations, get in touch.