Dual-Memory Architecture with Trust Gradients for Cognitive AI Agents: Integrating Curated Knowledge and Working Memory via Model Context Protocol

Technical specification for a dual-memory architecture for cognitive AI agents, separating working memory from curated knowledge through explicit trust gradients and fact promotion pipelines. ARCHITECTURE: The system implements two complementary memory layers: (1) Working Memory (ARIADNE / Neo4j) — high-volume, low-trust, automatic ingestion with bi-temporal tracking. Design principle: "Remember everything, trust nothing." (2) Curated Knowledge (ATHENA / MCP Server) — human-validated, high-trust, formally verified with tri-color trust model (Blue: human-authored, Green: AI-suggested pending validation, Red: constraint violations). Design principle: "Trust everything here." (3) Fact Promotion Pipeline — confidence-based thresholds bridging the layers with human-in-the-loop validation. ATHENA MCP INTEGRATION: Nine MCP tools (JSON-RPC 2.0) organized as Read (3), Write/Propose (3), and Validate (3) operations implementing the propose-validate workflow. Complete tool schemas included. DESIGN PRINCIPLES: Anti-Taming Machine — preserves critical friction in human-AI decision-making. AI can only propose; it cannot auto-commit to curated knowledge. Grounded in episodic/semantic memory distinction (Tulving, 1972), metacognitive monitoring (Nelson & Narens, 1990), extended mind thesis (Clark & Chalmers, 1998). GAP ANALYSIS: 0/12 agent memory frameworks implement trust-level differentiation between verified facts and AI suggestions. This architecture addresses that gap. Parent project: DAEDALUS (Cognitive AI Agent Architecture). Institution: University of Oradea, Romania.