LangGraph Research Example

apps/memforks-research is a multi-agent research pipeline built with LangGraph and @memfork/langgraph.

It demonstrates parallel agent branches, checkpointed graph state, compounding memory, and optional merge governance.

What It Shows

Feature	What to test
Checkpoint persistence	Kill and restart a run; graph state resumes from MemForks checkpoints.
Branch per worker	Each worker has its own thread branch.
Compounding memory	Run the same question twice; the second run recalls prior findings.
Supervisor synthesis	A supervisor branch combines worker findings.
Merge proposals	Worker findings can be proposed into the supervisor branch.

Run It

cd apps/memforks-research
npm install
cp .env.example .env
npm run research -- "What are the tradeoffs of microservices vs monoliths?"

Required environment:

OPENAI_API_KEY=sk-...
MEMFORK_TREE_ID=0x...
MEMFORK_PRIVATE_KEY=suiprivkey1...
MEMFORK_NETWORK=mainnet
MEMFORK_MEMWAL_ACCOUNT=0x...
MEMFORK_MEMWAL_KEY=...
MEMFORK_RELAYER_URL=https://relayer.memory.walrus.xyz

Optional:

TAVILY_API_KEY=tvly-...
MEMFORK_RESOLVER_ID=0x...
OPENAI_MODEL=gpt-4o

Graph Shape

Thread To Branch Mapping

LangGraph thread ID	MemForks branch
research-a	thread/research-a
research-b	thread/research-b
supervisor	thread/supervisor

With a fresh run, timestamped thread IDs can create brand-new branches. With stable IDs, repeated runs build on prior memory.

Worker Loop

Each worker:

1. Recalls prior research on its topic from its thread branch.
2. Searches the web if TAVILY_API_KEY is available.
3. Falls back to model knowledge if search is unavailable.
4. Synthesizes new findings with prior context.
5. Commits the synthesis to its branch.

Supervisor Loop

The supervisor:

1. Splits the question into complementary topics.
2. Starts workers in parallel.
3. Recalls prior supervisor-level synthesis.
4. Combines worker findings into a report.
5. Commits the report.
6. Optionally proposes worker-to-supervisor merges.

Why MemForks Instead Of A Local Checkpointer?

Local checkpointer	MemForks checkpointer
Process-local or file-local state	Walrus-backed persistent state
Hard to share across machines	Any configured machine can resume
No branch lineage	Thread branches inherit and diverge
No governed merge	proposeMerge() creates on-chain merge proposals
No durable audit trail	Checkpoints are content-addressed and branch-scoped

Minimal Checkpointer Setup

import { createMemForksCheckpointer } from "@memfork/langgraph";

const checkpointer = await createMemForksCheckpointer({
  treeId: process.env.MEMFORK_TREE_ID!,
  signer: process.env.MEMFORK_PRIVATE_KEY!,
  memwal: {
    accountId: process.env.MEMFORK_MEMWAL_ACCOUNT!,
    delegateKey: process.env.MEMFORK_MEMWAL_KEY!,
    serverUrl: process.env.MEMFORK_RELAYER_URL,
  },
  threadToBranch: (threadId) => `thread/${threadId}`,
});

Continue with LangGraph Adapter for SDK details.