Hermes · Decision-Grade Survey · 2026-07-05

OSS Knowledge-Graphs for a CRM

Adopt-path lens over the Pantheon crm_* spine · T-deep · every candidate live-verified via gh api

4 lanes · 30+ reposstars/license/last-commit verified source-coverage ledgerresearch-orchestrate (no banned Workflow)

The frame that decides everything

"Build a KG for a CRM" has two answers on opposite sides of our write-integrity gate (rule #20). Our edges (crm_edges, FKs) already exist and are trusted — so this is a projection problem, not an extraction problem.

→ Recommendation is a two-layer adopt path, not a single tool.

FAMILY 1 · WRITE-INTEGRITY-SAFE · ADOPT NOW

Layer 1 — Ontop

Virtual RDF knowledge graph over the live Postgres CRM via R2RML. Zero data movement, no freshness bug, invents nothing — every triple is an existing row/FK/edge.

Apache-2.0 · 873★ · commit 2026-06-19 · SPARQL

FAMILY 2 · GATED · PHASE 2

Layer 2 — Graphiti

Temporal property graph over crm_notes/meetings text → relationship suggestions. Inferred edges are un-provenanced → operator review before any crm_edges write. Never auto-writes.

Apache-2.0 · 28.4k★ · bi-temporal · MCP server

crm_* Postgres (SSOT, unchanged) contacts · organizations · deals · notes · documents · crm_edges · meetings │ │ │ Family 1 (deterministic, safe) │ Family 2 (LLM-inferred, gated) ▼ ▼ ┌──────────────────────────┐ ┌──────────────────────────────┐ │ LAYER 1 · ONTOP │ │ LAYER 2 · GRAPHITI │ │ virtual RDF KG (R2RML) │ │ derived temporal graph │ │ zero data move │ │ → edge SUGGESTIONS │ │ SPARQL · invents nothing │ │ → operator review (P0) │ │ ✓ rule #20 │ │ → then write crm_edges │ └──────────────────────────┘ └──────────────────────────────┘

Lane A — Relational → KG mappers (Family 1, sharpest match)

Tool	★	License	Commit	Model	Verdict
Ontop	873	Apache-2.0	2026-06-19	Virtual SPARQL→SQL, live	WINNER — zero data move
Morph-KGC	280	Apache-2.0	2026-06-28	Materialize (batch)	#2 — needs freshness
RMLMapper	201	MIT	2026-02-17	Materialize to file	no query surface
SDM-RDFizer	137	Apache-2.0	2026-05-18	Billion-triple scale	overkill
D2RQ	324	unclear	archived 2021	Virtual	dead (Ontop's ancestor)
Sparqlify / Squerall	134/—	none / gone	dormant / 404	—	disqualified

Lane B — LLM KG builders (Family 2, suggestion layer)

Tool	★	License	Release	CRM-fit	Verdict
Graphiti	28.4k	Apache-2.0	2026-06-08	Bi-temporal, custom Contact/Org/Deal types, dedup, MCP	WINNER
Cognee	27k	Apache-2.0	2026-06-26	Plausible; dedup NOT verified	#2 — verify first
Neo4j LLM Graph Builder	4.9k	Apache-2.0	2026-06-11	Docs only, no structured path	secondary
iText2KG · KG-Gen	951 · 1.2k	Apache/MIT	dormant	Text-only research impls	reject

Lane C — GraphRAG frameworks (RULED OUT — wrong shape)

Every tool infers a graph from unstructured text — the inverse of our problem. Feeding CRM rows = paying LLM cost to re-derive edges we already hold, with drift + rule-#20 collision.

Tool	★	Status	Note
Mem0	60.1k	graph store REMOVED (OSS v3)	Hosted-only, untyped ranking signal. Disqualified. Correct any prior "Mem0 graph memory" claim.
LightRAG	37.3k	only escape hatch #1	`insert_custom_kg()` writes pre-built edges direct. Pick if `crm_notes` RAG ever wanted.
R2R	7.9k	~8mo dormant	Postgres-native + REST CRUD + MCP — closest architecture, stalled momentum. Health-check first.
MS GraphRAG · nano · HippoRAG · fast-graphrag	34k…3.8k	doc-first / research-grade	Graph = RAG byproduct, not addressable. Not it.

Lane D — Substrate + Customer-360

Engine	★	License	Deploy	Note
TerminusDB	3.3k	Apache-2.0 (cleanest)	1 Docker	Versioned document-graph = free audit trail (dovetails rule #20). WOQL niche.
Neo4j	16.8k	GPL-3.0 (CE)	server (JVM)	Most mature, standard Cypher, heaviest ops.
Memgraph	4.2k	BSL 1.1 + MEL	server	Source-available, not OSI until change-date.
FalkorDB	4.7k	SSPL	server	Non-OSI; SDK layer doc-shaped, no Postgres ingest.
Kùzu	4.0k	MIT	ARCHIVED 2026	Embedded option is dead. Do not adopt.

No turnkey "customer-360 / CRM knowledge graph" OSS build exists — confirmed by direct search + full audit of Neo4j's own example orgs. Don't wait for one; adopt at the tool altitude. (Note: Layer-1 Ontop needs no separate store — it is virtual over Postgres.)

Adopt path & open questions

Adopt now (~1–2 days)

1. Ontop (Docker) → read-only role on CRM Postgres.
2. R2RML maps: one per table; FKs + crm_edges → triples; crm_edge_types → predicates.
3. Internal SPARQL endpoint (read-only; no public mutating surface).
4. Validate a traversal query against real rows.

Route to metis / operator

1. Who consumes the graph? (human/agent/analytics) → decides SPARQL vs property-graph.
2. Is read-only projection enough, or is the real ask relationship intelligence (→ Layer 2)?
3. Layer-2 write-gate workflow = metis-domain design.
4. Cognee dedup deep-read — only if Graphiti rejected.

Full report + verified lane sections + source-coverage ledger: research/oss-kg-for-crm-2026-07-05/ · Prior art (not re-surveyed): oss-crm-landscape (#49) · aion-graphify-adoption (#7). Load-bearing rows all [live-verified]/[deep-read] except Cognee-dedup (flagged; does not gate headline).