Skip to content

Local & cluster mode

The same engine and the same data model run from one laptop binary to a multi-node cluster. You don't pick an architecture up front — you grow along it, and memory stays coherent across the spectrum via sync.

The three ways to run

Local modeControl planeCluster
Commandkyma (CLI / mcp / serve)kyma server (Docker)kyma server, N nodes
Catalogembedded SQLitePostgresPostgres (shared)
Storagelocal files (~/.kyma)object store (S3 / MinIO)object store (shared)
InfranonePostgres + object store+ a load balancer
Best forone developer, offline, instanta team, always-onread scale-out, HA
Extrason-demand ingest + query, local web UIscheduled data sources, dreaming, full web appeverything, horizontally

Local mode — kyma

One binary per developer. No Postgres, no Docker, no sudo — the catalog is embedded SQLite and data lives under ~/.kyma. It gives your agent the full toolset over stdio MCP (kyma mcp), an optional local web UI (kyma serve), and on-demand ingest + query. This is where Give your agent memory starts.

Control plane — kyma server

The team tier: Postgres for the catalog, an object store (S3 or MinIO) for extents. On top of local mode it adds the things that need to be always-on and shared — scheduled data sources, background dreaming, and the full web app (Graph Explorer · Memory · Discover · Agent). Stand it up with docker compose up -d (Five-minute start) or deploy it to production.

Cluster — many nodes, one source of truth

The control plane scales out without a rewrite, because of the five invariants: object storage is the only source of truth and compute is stateless. Run N kyma server nodes against the same Postgres catalog and the same object store, behind a load balancer. Any node can answer any query; ingest commits serialize through catalog CAS. Read throughput scales with node count.

How to choose

  • Just want your agent to remember? Local mode. Nothing to provision.
  • A team sharing memory + live data sources? Control plane.
  • High query volume or HA? A cluster of control-plane nodes.

You're never locked in: start local, sync to a control plane when your team needs it, add nodes when load demands it. Same binary, same data.

An open-source project · MIT licensed.