Using SemVec AI for Detecting Agentic Drifts

Session track: Modern Applications

Session time: 03:00 pm - 03:30 pm UTC November 12

Session description:

Semantic Drift Detection in Agentic Systems Using SemVec AI and Neo4j As AI agents operate across extended multi-topic conversations, semantic drift — the gradual divergence of an agent's active context from its intended domain — represents one of the most critical and least visible failure modes in production agentic systems. This paper presents a hybrid architecture combining SemVec AI's Persistent Semantic State with Neo4j's graph-based knowledge representation to detect and recover from semantic drift autonomously. SemVec continuously compresses conversational history into a fixed-size semantic state vector, tracking phase transitions and computing turn-by-turn similarity scores. When the adaptive memory weight β drops sharply or semantic similarity falls below threshold, a drift event is flagged — without human intervention. Neo4j provides structural grounding: where SemVec detects that something has shifted semantically, Neo4j identifies what has shifted in the knowledge graph — which entities left scope, which relationships were severed, which new clusters are forming. Together they produce a drift signal that is both semantically aware and graph-anchored — interpretable, actionable, and recoverable through three strategies: soft β realignment, hard context reset, or cross-agent consensus correction via the SemVec Cortex cluster protocol. Validated across four maximally distant domains, the system achieved 100% autonomous drift detection accuracy across 50 conversation turns with no manual intervention.