ARIADNE: Neo4j Working Memory Specification for Cognitive AI Agents with Bi-Temporal Edges and Trust-Gradient Fact Promotion

ARIADNE is a Neo4j-based working memory specification for cognitive AI agents, designed as part of the DAEDALUS research project. The specification defines a three-tier graph schema for agent memory: Tier 1 — Episodic Memory: Episode nodes storing raw interactions (conversations, documents, API responses) with transaction-time immutability and multi-channel source tracking (WhatsApp, Telegram, Discord, CLI). Tier 2 — Semantic Memory: Entity nodes (canonical representations of people, projects, concepts) and Edge nodes (facts stored as first-class nodes rather than relationships, enabling rich metadata). Edges implement bi-temporal tracking with four temporal fields: created_at and expired_at (transaction time), valid_at and invalid_at (real-world validity time). This enables point-in-time queries such as "What did I believe about X in October?" Tier 3 — Community Memory: Leiden-clustered entity groups for global context retrieval, with hierarchical organization and LLM-generated summaries. Additional node types include Reflection nodes implementing the Reflexion pattern (Shinn et al., 2023) for failure-indexed learning with outcome tracking. Key architectural contributions: 1. Bi-temporal edge modeling distinguishing real-world fact validity from system recording time, validated by convergence analysis across three independent AI-assisted literature reviews. 2. Three-factor retrieval scoring (recency × importance × relevance) with configurable exponential decay, implemented as a complete scoring algorithm with formal mathematical specification. 3. Hybrid retrieval combining vector similarity search, BM25 keyword search, and graph traversal with Reciprocal Rank Fusion (RRF) for result merging. 4. Trust-gradient fact promotion pipeline connecting working memory to a curated knowledge layer (ATHENA) with confidence thresholds: >0.95 auto-promote, 0.85-0.95 batch review, <0.85 manual review. 5. Multi-stage entity resolution using Jaro-Winkler string similarity, embedding cosine similarity, and entity type matching, with negative link tracking to prevent repeated false comparisons. 6. Memory consolidation ("sleep") process with conflict detection, confidence decay on unreinforced facts, and community summarization. The specification includes: complete Cypher schemas with indexes and constraints; Python reference implementations of all retrieval algorithms; TypeScript interface definitions for gateway and ATHENA integration; a 12-week phased implementation roadmap; performance targets benchmarked against Graphiti (94.8% DMR accuracy baseline); and a testing strategy with unit, integration, and benchmark test specifications. Note: Reference algorithms are specified in Python for clarity; the target implementation stack is TypeScript/Bun. The Python serves as precise, executable specification rather than production code. ARIADNE forms the working memory component of a dual-memory cognitive architecture where it is paired with ATHENA (curated knowledge with tri-color trust validation). Design principle: ARIADNE "remembers everything, trusts nothing" while ATHENA "trusts everything here." Parent project: DAEDALUS (Cognitive AI Agent Architecture) Institution: University of Oradea, Romania