Push & pull
Clients send their change sets and read back everything new since their last sync, through a gap-free, resumable cursor.
The client keeps each person's memory local. The server is what turns many personal memories into one team memory over the wire — a self-hosted service that every teammate's TeamGraph binds to, so a problem solved once on any machine is answered everywhere.
A headless HTTP service that hosts a team's shared graph. Clients push what they learn and pull what the team learned; the server applies the same conflict-free merge the client does — only here it is the one convergent copy everyone syncs against.
Clients send their change sets and read back everything new since their last sync, through a gap-free, resumable cursor.
The CRDT rules run server-side: content-hash identity, last-writer-wins state by a hybrid logical clock, max edge weight, union authors and sources.
Semantic match, search and explore run over the shared graph, so recall draws on everything the team has pooled, not just one machine.
Every team is isolated at the database level with row-level security, and scoped bearer tokens gate read, write and admin access.
The server is a Go HTTP API backed by Postgres with pgvector for similarity search, and a local embedding sidecar so vectors are produced by a model you pin — never sent to a third party.
make compose-up and it's live on your network.On its own, the client is local-first — its memory lives in a SQLite file on one machine. The server is the piece that makes sharing real: one convergent graph the whole team binds to, with the same privacy guarantees intact.
gitlab.com/adam.allard/teamgraph-server
The Go service, its Postgres schema and migrations, the OpenAPI contract and the Docker Compose stack. Clone it, read the architecture, and run your own.
Self-host the shared graph, hand out scoped tokens, and let a fix on one laptop answer everywhere.