OriginTrail's Decentralized Knowledge Graph lets organizations keep their data in separate systems while sharing selected, verifiable claims.
2
Tomaž Levak sees the DKG as collective memory for AI agents, with agents able to publish, exchange, and query knowledge with provenance information.
3
Graph RAG can federate questions across specialized knowledge bases and agents, then return a grounded response with the sources used.
Summary
Tomaž Levak explains how OriginTrail's Decentralized Knowledge Graph connects privately managed knowledge graphs into a shared protocol. Organizations can keep their data inside their own systems while exposing selected information to partners, with access controls and proofs that show when claims were published and whether they changed. He describes paranet neighborhoods, edge nodes, blockchain-based ownership and Merkle-root proofs, and the network's multi-chain design. Levak's main AI use case is collective memory for agents. Agents could store interactions, exchange data, and query public or private knowledge without relying on one central database. He also describes graph RAG, where a question can be federated across specialist agents and knowledge bases. The conversation covers enterprise supply-chain data, decentralized science, personal health data, data expiry, and hosted nodes. Levak is open about tradeoffs: global searches take time, while local reads can be fast when data is cached in an edge node.
The DKG is a shared layer for independently owned knowledge graphs
Tomaž Levak describes OriginTrail's Decentralized Knowledge Graph as middleware where organizations connect their systems and data without handing everything to a central company. It supports data ownership, transparency, and verifiability. He compares the network to a global knowledge graph made from smaller neighborhoods, called paranets. Each paranet can define its own rules, ontologies, and data structures. A group could create a paranet for the MLOps podcast and decide what contributions are acceptable before building an AI system on top of them.
Paranets can keep enterprise data private while exposing selected claims
Enterprise users can build paranets from private knowledge assets, which are atomic units in a graph. The assets stay on nodes controlled by the organization, while selected information can be published to make it discoverable. Levak gives a supply-chain example involving importers, overseas factories, third-party audits, Walmart, and the Department of Homeland Security. Different participants can see different parts of the same information. Access can be managed at a very granular level, including which entity can see particular lines of data.
Edge nodes connect existing systems without forcing one database or blockchain
An edge node is a lightweight gateway around an organization's knowledge graph. A logistics system can push events into its local graph, which then becomes accessible through the DKG. Existing knowledge graphs can also be connected directly. Levak stresses that the DKG is a protocol rather than a database. Users can choose among knowledge-graph databases, blockchains for storing proofs, and AI models. Multiple edge nodes can operate under the same paranet rules, and one node can potentially belong to more than one paranet.
On-chain ownership and proofs make published knowledge verifiable
When a knowledge asset is published, OriginTrail creates an NFT that proves ownership. The owner is the identity allowed to update that asset. The network also creates a Merkle tree root hash and publishes it on-chain. Readers can compare the fingerprint with the data they received to check whether the information was changed. Levak says this approach supports both public and private data and can work with different EVM chains. He also argues that AI systems can perform these fingerprint checks more easily than people clicking through records manually.
The DKG could give AI agents a shared memory with private and public modes
Levak sees AI agents as an important use case because agents can have social accounts, wallets, and assets, then transact with less friction than people. The DKG could become collective memory where agents store findings and interactions, keep some knowledge private, and publish other knowledge publicly. He gives the example of an agent negotiating on X to receive funding in TRAC so it could publish to a DKG edge node. Demetrios Brinkmann connects this idea to incident-response agents that map code, pull requests, and developer discussions into a graph to find root causes.
Decentralized science and personal data are practical experiments for agent systems
Levak describes a science-focused paranet that could collect open research and clinical-trial work in graph form. Agents could repeatedly examine that material while adding new findings to the public knowledge base. A more personal example would combine data from an Oura ring with public neuroscience and sleep research. A local agent could inspect the user's sleep patterns and return suggestions with sources. Levak says these examples are small today, but the value could increase when the system moves from a handful of papers to much larger and more contextual datasets.
Graph RAG adds structure and federation to retrieval
The DKG uses graph RAG rather than treating information only as text documents. Knowledge assets are connected in a graph, which gives the model more context. Levak says language models can generate SPARQL queries when given enough context or a template. A decentralized graph RAG system could send parts of a question to specialized agents and knowledge bases, receive graph-based responses, and combine them into a response for the original question. The returned answer would include provenance information for the knowledge assets used.
Speed depends on whether a query is global or local
Levak says Version 8 increased scale through random sampling and a sharding-like design. The network does not need to validate every knowledge asset, and an oversized network can be split into smaller parts. A global search still takes time because it traverses a large graph. Queries sent to a known edge node or paranet can run at production speeds. For the fastest reads, users can retrieve verified data once, store it in their local knowledge graph, and let the application read locally while adding new information when needed.
"Decentralized Knowledge Graph is a perfect middleware, like a common playing ground where folks can connect their systems, their data into this shared Knowledge Graph, Global Knowledge Graph, and build cool stuff on top of it."Tomaž Levak02:28
Who should watch
You are designing an agent system that needs persistent memory, source tracking, or access to information owned by several organizations.
Your team wants to share supply-chain or partner data while keeping the underlying records private.
You are comparing graph RAG with document-based RAG and want to understand how federation, provenance, and local edge nodes could work.
