Mastra•3mo ago

Workflow short term memory across multiple runs

Do workflows have a short-term memory with threads?

Coming from LangGraph : you don't distinguish workflows and agents. So basically when you use short term memory it applies to both.

Say that you want to craft a travel recommender. There are 2 patterns :

A simple workflow with a big prompt, and you run the workflow multiple times with a persistent state to refine the response => I get how to implement it as an agent but not a workflow
A more complex workflow with human-in-the-loop interruption, that works in a single run => I get this part but this makes up convoluted workflow, not suited for a chatbot

My understanding is that I would be expected to show the workflow into an agent and benefit from the agent's state memorization, right?

Mastra Triager•10/22/25, 12:06 PM

Created GitHub issue: https://github.com/mastra-ai/mastra/issues/9117

GitHub

[DISCORD:1430277060302995618] Workflow short term memory across mul...

This issue was created from Discord post: https://discord.com/channels/1309558646228779139/1430277060302995618 Do workflows have a short-term memory with threads? Coming from LangGraph : you don&#3...

_roamin_•10/22/25, 9:48 PM

Hi @Eric Burel ! Workflows don't really have memory, there is a "state" you can use to persist data at the "workflow run" level, but it's just a plain json, nothing else:

const firstStep = createStep({
  id: "first-step",
  execute({ setState }) {
    // set the state withing a workflow step
    setState({
      myValue: "a value",
    });
  },
});

const secondStep = createStep({
  id: "second-step",
  execute({ state }) {
    // read the state from within another workflow step
    console.log(state.myValue);
  },
});

createWorkflow({
  id: "my-worfklow",
  // describe the state's schema
  stateSchema: z.object({
    myValue: z.string(),
  }),
}).then(myStep);

const firstStep = createStep({
  id: "first-step",
  execute({ setState }) {
    // set the state withing a workflow step
    setState({
      myValue: "a value",
    });
  },
});

const secondStep = createStep({
  id: "second-step",
  execute({ state }) {
    // read the state from within another workflow step
    console.log(state.myValue);
  },
});

createWorkflow({
  id: "my-worfklow",
  // describe the state's schema
  stateSchema: z.object({
    myValue: z.string(),
  }),
}).then(myStep);

However, you can run agents inside workflow steps, and then that's when you have access to all the memory features, like semantic recall and working memory

Eric BurelOP•10/24/25, 9:15 AM

Cool, I definitely need to dig state, because sometimes playing around with the input/output of the step is a tad clumsy
Eg for a RAG that optionnaly fetches, docs, I can shove results in state and consume them anywhere else only if needed/available for instance
I haven't explored how you actually run the workflow outside of the playground anyway, I see they can be attached to an agent (and the workflow itself can use the agent, recursively or not I don't know)

const firstStep = createStep({ id: "first-step", execute({ setState }) { // set the state withing a workflow step setState({ myValue: "a value", }); }, }); const secondStep = createStep({ id: "second-step", execute({ state }) { // read the state from within another workflow step console.log(state.myValue); }, }); createWorkflow({ id: "my-worfklow", // describe the state's schema stateSchema: z.object({ myValue: z.string(), }), }).then(myStep);

Workflow short term memory across multiple runs

Similar Threads

Similar Threads

Similar Threads