State that outlives the run
A workflow that has to remember something usually means provisioning a database first. Here the store is a component on the canvas — append-only streams, key/value records and agent memory, scoped to your workspace, with a retention window you choose.
Streams, key/value records and agent memory as canvas components — scoped to your workspace, with retention you set.
What the store gives you
- A component, not a cluster — Remembering something between runs normally means standing up a database, sizing it, patching it and paying for it whether or not a workflow ran. Here it is a node you place between two other nodes, with the same run history and the same error handling as every other step.
- Three shapes, one store — Append to a stream when order and history matter, set and read a record when only the latest value does, and give an agent a memory it can carry across turns. All three are the same store addressed three ways, so nothing has to be copied between them.
- Scoped to your workspace, not by convention — Every read and write is bounded by the workspace that made it, and the store itself refuses a request that names another workspace rather than quietly answering with something else. The boundary is enforced where the data lives, not by the caller asking politely.
- Retention you set, filtering you declare — Each namespace carries its own retention window, so short-lived working data expires on its own without touching the records you keep. Declare which fields should be filterable and entries are indexed by them as they are written, giving you exact-match filtering on the dimensions you named.
Platform capabilities
- None: Databases to provision before the first write
- Per namespace: How the retention window is set
- Step or tool: Two ways to reach the same store
Do I have to provision a database?
No. The store runs as part of the platform, so there is nothing to size, patch, connect or pay for separately. You place the component, pick a namespace and write to it — and usage is metered per workspace, so what a workflow costs to remember is a number you can see rather than a surprise.
How long is my data kept?
For as long as the retention window on that namespace says. Working data can expire in days while the records you care about are kept far longer, and the window is a setting rather than a cleanup job you have to write. There is deliberately no "keep forever" — anything you need indefinitely belongs somewhere you back up.
Can an AI agent read and write it?
Yes, if you grant it. The same store is available as an ordinary step you place in order and as a tool an agent may call, which is how an agent keeps memory across turns. Both routes are scoped to the same workspace and metered the same way, so granting the capability does not widen what can be reached.