A Knowledge Graph Agents Can Write To, Safely.

Abstract

The 2026 consensus in enterprise AI is that retrieval is not enough: agents need structured, relational, persistent knowledge, and the knowledge graph has become the beyond-RAG layer of choice, with graph retrieval reported at 86% multi-hop accuracy against 32% for vector search alone. Yet across the field one assumption is nearly universal: the ontology is read-only context the agent consults, and keeping it correct is a separate human discipline.

We present MNEME, the FlowX.AI Ontology Layer, which makes the opposite choice: the ontology is active, governed memory the agent helps build. Agents read, validate against, and write back to the same project-scoped knowledge graph, and every contribution is safe by construction.

Each Concept and Relation carries provenance (pdm_import, agent_write, or manual) and confidence as first-class fields; agent writes are staged below a confidence bar, screened by conflict detection of three kinds, and quarantined out of canonical retrieval until promoted by reinforcement across runs or by human review. The institutional source of truth, imported one-way from the platform data model, is authoritative and read-only to agents, so a hallucination can never become a regulatory taxonomy entry, while genuinely new structure is captured rather than lost. MNEME runs on two pluggable backends behind one contract, Postgres with pgvector for semantic-heavy ontologies and Neo4j for graph-heavy multi-hop, retrieves by a hybrid of vector similarity and graph expansion in a single call, and exposes a nine-node toolkit on the Agent Builder canvas.

We describe the architecture, the safe write-back loop, the toolkit and its agent patterns, an illustrative regulated taxonomy, and how active memory differs from the read-only knowledge graphs the rest of the market ships.