Codemix just open-sourced @codemix/graph, a TypeScript graph database built on CRDTs that does something most graph databases don't: type-check your traversals end to end. You define schemas with Zod, Valibot, or ArkType, and property types flow through every query and mutation with no casts. It's alpha software, but the team is running it in production already.
There's a Gremlin-style traversal API where every label, property key, and hop gets checked against your schema. And a Cypher-like query language. That's where the LLM angle comes in. You expose your graph to agents via an MCP server, and they run structured queries against your data. If you're building agent memory systems, this is an integration point you can actually ship with.
Swap InMemoryGraphStorage for YGraph and your entire graph lives in a CRDT document. Queries, traversals, indexes, they all work the same. You get conflict-free sync across clients and live queries that re-fire when remote peers mutate data.
Christian Bromann leads the project. He also maintains WebdriverIO, and that developer tooling background shows in the API design. It's built for people who want graph database features without leaving their TypeScript workflow.
Graph operations are expensive and JavaScript runtimes have limits. The Hacker News crowd flagged that, along with method chaining making unit testing harder. Fair points. But the agent knowledge graph potential got genuine interest, and the alpha label is honest. Worth watching if you need collaborative graph data without spinning up Neo4j.