Trustgraph
Every other framework launched in the last 18 months made a different architectural bet: property graphs.
Mem0, Zep/Graphiti, Cognee, Letta…different storage, different temporal models, same underlying commitment. Neo4j, Kuzu, Memgraph and more. Mem0 even went the other direction recently: its v3 pipeline ripped out graph memory entirely and replaced it with spaCy entity linking, a hub-and-spoke structure pointing entities back at memories, no typed edges at all. Is it an upgrade? That’s rather an admission that the graph layer wasn’t pulling its weight. TrustGraph is the one project building agent memory on the semantic web stack. And the reasons this matters aren’t esthetic.
A property graph (LPG) encodes relationships as edges with string labels your application code interprets. There’s no class hierarchy, no formal entailment, no shared vocabulary across deployments. Every “ownerOf” edge means whatever the application that wrote it intended it to mean, and nothing enforces that meaning when another agent reads it back. An OWL ontology is a contract. Classes carry semantics, properties carry domain/range constraints, and SHACL validates the data against those constraints before it ever reaches the agent. That’s the difference between an agent retrieving “facts that look relevant” and an agent retrieving facts it’s structurally guaranteed to be able to reason about correctly.
Provenance is the other gap. Most frameworks bolt a validity window onto a fact and call it temporal reasoning. That’s a timestamp, not provenance. I spoke already a few times about this, so I won’t repeat here. Try doing that cleanly in a property graph without inventing a parallel schema nobody else’s tooling understands.
And then there’s the part everyone skips: SPARQL federation. A LPG memory store is an island. An RDF store can be queried alongside every other RDF store on the planet using one query language, no translation layer, no vendor-specific Cypher dialect. If “agent memory” is ever going to interoperate across organizations the way the rest of enterprise data does, it has to speak a standard, not a product.
I am not against LPG efforts, only trying to point at the limitations and consequences. I know well why people like LPGs. Graphiti’s validity windows are a genuinely good idea. But a good idea bolted onto a graph model with no formal semantics is still a graph model with no formal semantics. The agent memory space will consolidate. When it does, the survivors won’t be the ones with the biggest GH star count. They’ll be the ones with a memory you can validate, audit, and federate because that’s what “trust” in production systems has always meant, long before anyone called it agent memory. There is a lot of innovation these days, which is great. Much of it is however geared towards single users, not enterprises or trust.
▶ https://trustgraph.ai/