How To: Add Deja Memory
Task: let each loop iteration recall what earlier runs learned.
Done when: the worker context has a recall field with learnings from Deja.
What memory does
Memory is optional. It does not change pass or fail. Only the gate decides that.
With a memory runtime, Plan.runLoop:
- Before each iteration, asks memory for learnings about the failing story. The worker gets them as
context.recall. - After each iteration, writes what happened: the failing story, the worker summary, the changes, and the commit.
- Keeps a working state for the proof run, and resolves it when the plan passes.
Use Deja
import { Effect } from "effect";
import { Plan, createDejaMemoryRuntime } from "gateproof";
await Effect.runPromise(
Plan.runLoop(plan, {
worker,
memory: createDejaMemoryRuntime({
endpoint: "https://your-deja-instance.example",
scope: "gateproof:my-project",
apiKey: process.env.DEJA_API_KEY,
recallLimit: 5,
}),
}),
);Options:
| Option | Required | Meaning |
|---|---|---|
endpoint |
yes | Base URL of the Deja instance. |
scope |
yes | Memory scope for this project. |
apiKey |
no | Sent as a bearer token. |
recallLimit |
no | Maximum learnings to recall. Default 4. |
updatedBy |
no | Name written with each state update. Default "gateproof". |
Gateproof calls /inject, /learn, /state/:proofRunId, and /state/:proofRunId/resolve on the endpoint.
Failure behavior
Recall and the working-state calls need a proofRunId. Without one, recall returns nothing and the state calls are skipped.
A memory error never fails the run. If Deja is down, recall returns nothing and writes are skipped. The proof loop continues.
Write your own memory
memory accepts any MemoryRuntime. Only writeIteration is required. recallIteration, updateWorkingState, and resolveWorkingState are optional.