Technical paper · MNEMĒ
A Knowledge Graph Agents Can Write To, Safely.
MNEME, the FlowX.AI Ontology Layer: a provenance-tracked structured-memory graph that agents read, validate against, and contribute to, without corrupting the institutional source of truth.
- ontology
- knowledge graph
- active memory
- provenance
- GraphRAG
- entity resolution
- conflict detection
- structured memory
One new concept, earned
The ontology is memory the agent helps build, and every write is made safe on the way in. An agent’s contribution is deduplicated, screened for three kinds of conflict and staged where no production answer can see it, until reinforcement or a steward promotes it. The loan taxonomy below is the paper’s illustrative example.
A loan-product taxonomy assembling from the platform data model as read-only canonical concepts: Loan, Secured Loan, Unsecured Loan, Mortgage and Personal Loan. An underwriting agent grounds by classifying to Mortgage, expanding its IS_A chain and validating inherited attributes. A new regulation then leads the agent to propose a new concept, which is deduplicated, screened for label clashes, category mismatches and contradicting relations, and staged at low confidence outside canonical retrieval. A steward approves it and it joins the graph under Mortgage, keeping its agent_write provenance.
Core thesis
Everyone makes the ontology a read-only oracle the agent consults.The agent that just learned something the graph lacks is the agent best placed to extend it, and the one whose contribution we are most afraid to trust.Make the ontology active memory, and make the write safe.
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-based retrieval producing more comprehensive answers than vector retrieval on corpus-wide questions. Yet across production enterprise platforms one assumption is nearly universal: the ontology is read-only context the agent consults, and keeping it correct is a separate human discipline. Research systems have begun to evolve ontologies automatically, gated by a performance signal; that answers a different question from the one a regulated enterprise asks, which is who authorized each change.
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.
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 cannot become a regulatory taxonomy entry. MNEME runs on pluggable backends behind one contract-tested interface, retrieves by a hybrid of vector similarity and graph expansion, extends the same governance to access, history, and schema, and exposes a twelve-node Agent Builder toolkit and an MCP server.
We describe the architecture, the write-back loop, the toolkit, an illustrative regulated taxonomy, and a retrieval evaluation on HotpotQA.
Part of the FlowX.AI paper series.
Each paper names a framework and shows it running in production: governance, reliability, memory and measurement, engineered rather than hoped for.