@mubit-ai/langgraph

LangGraph.js BaseStore backed by MuBit.

MubitStore extends BaseStore (imported from @langchain/langgraph, which re-exports it from @langchain/langgraph-checkpoint) and calls only the public @mubit-ai/sdk Client methods — no private transport or operation tables. On top of the standard store surface it threads the MuBit attribution loop for long-running MAS and coding-agent runs.

Install

npm install @mubit-ai/langgraph @langchain/langgraph @mubit-ai/sdk

Basic usage

import { MubitStore } from "@mubit-ai/langgraph";

const store = new MubitStore({
  endpoint: "http://127.0.0.1:3000",
  apiKey: process.env.MUBIT_API_KEY,
});

await store.put(["memories", "user-1", "session-1"], "deploy-note", {
  text: "Canary before rollout.",
  intent: "lesson",
  source: "human",
});

const results = await store.search(["memories", "user-1"], {
  query: "What rollout lesson do we have?",
  limit: 5,
});

Namespace mapping

| Namespace | MuBit scope | | --- | --- | | ["memories"] | global | | ["memories", userId] | user-scoped | | ["memories", userId, sessionId] | user + session scoped |

Store behavior

• put() writes through client.remember with upsert_key and threads an idempotency_key (defaults to the key, or pass value.idempotency_key) so a re-run graph node does not double-write
• get() and search() use client.recall; each returned item exposes value.reference_id so callers can credit recalled evidence back through recordOutcome({ entryIds: [...] })
• evidence metadata is normalized from current MuBit response fields
• listNamespaces() remains local cache based
• delete() remains a no-op because MuBit deletion is explicit and scope-based

Attribution loop (v0.7.0)

Thread credit assignment through a run:

• recordOutcome(namespace, referenceId, outcome, options) — options.entryIds credits the recalled evidence reference_ids; options.verifiedInProduction gates reinforcement to production-confirmed outcomes
• recordStepOutcome(namespace, stepId, options) — per-step process reward with stepName, outcome, signal, rationale, and directiveHint

const results = await store.search(["memories", "user-1"], { query: "rollout" });
const entryIds = results.map((r) => r.value.reference_id).filter(Boolean);

await store.recordStepOutcome(["memories", "user-1", "session-1"], "canary", {
  outcome: "success",
  signal: 1,
  directiveHint: "Keep the canary soak before promoting.",
});

await store.recordOutcome(["memories", "user-1", "session-1"], "global", "success", {
  signal: 1,
  entryIds,
  verifiedInProduction: true,
});

A runnable end-to-end walkthrough lives in examples/attribution_loop.

MuBit extension methods

Existing MAS helpers:

• checkpoint(namespace, snapshot, options)
• surfaceStrategies(namespace, options)

Current observability and coordination helpers:

• getContext(namespace, options)
• memoryHealth(namespace, options)
• diagnose(namespace, options)
• archive(namespace, options)
• dereference(namespace, options)
• registerAgent(namespace, options)
• listAgents(namespace)
• handoff(namespace, options)
• feedback(namespace, options)

Config

| Option | Default | Purpose | | --- | --- | --- | | endpoint | http://127.0.0.1:3000 | MuBit HTTP endpoint | | apiKey / api_key | "" | MuBit API key |

For tests or advanced embedding, you can inject mubitClient directly.

Development

cd integrations/javascript/mubit-langgraph
npm install
npm test

License

Apache-2.0