TuskPoint — langgraph-checkpoint-walrus

Verifiable LangGraph agent state on Walrus, with semantic recall over checkpoint history via MemWal.

Hackathon project — Walrus track. Built in small, testable increments.

What it is

Two layers:

1. WalrusSaver — a drop-in LangGraph BaseCheckpointSaver. Every checkpoint is serialized with LangGraph's own serde, gzipped, and stored as an immutable Walrus blob (the exact layer). A per-thread JSON manifest maps checkpoint_id -> {blob_id, parent, timestamp, summary}. On every save it also writes a one-sentence natural-language summary to MemWal (the semantic layer), so an agent can later search its own past.

2. mcp_server — an MCP server exposing six checkpoint tools (save / load / list / resume / diff / search) over stdio. It complements, and does not duplicate, MemWal's own MCP.

Related MCP server: memory-mcp

Architecture

                      ┌─────────────────────────────────────────┐
   LangGraph agent ──▶│             WalrusSaver                  │
   (put / get_tuple)  │  BaseCheckpointSaver[str]                │
                      │                                          │
                      │  EXACT layer            SEMANTIC layer   │
                      │  ───────────            ──────────────   │
                      │  serde→gzip→blob        build_summary()  │
                      │       │                       │          │
                      └───────┼───────────────────────┼──────────┘
                              ▼                        ▼
                    ┌──────────────────┐     ┌──────────────────┐
                    │  Walrus testnet  │     │      MemWal       │
                    │ publisher PUT /  │     │ remember / recall │
                    │ aggregator GET   │     │ (vector search)   │
                    └──────────────────┘     └──────────────────┘
                              ▲
                              │ latest manifest blob id cached in
                              │ .walrus_threads.json  (only local state)
                              ▼
                      ┌──────────────────┐
                      │  mcp_server      │  6 tools over stdio:
                      │  (FastMCP)       │  save load list resume diff search
                      └──────────────────┘

Exact vs. semantic — why both?

• Exact lookups are by ID, never fuzzy. checkpoint_load(thread, id) resolves the manifest entry → blob ID → Walrus GET → de-gzip → de-serialize. This is deterministic and content-addressed: the blob you read is byte-for-byte the blob you wrote. This is the part you rewind to.

• Semantic search is for discovery. checkpoint_search("when did the writer start?") asks MemWal for the nearest summaries. It returns pointers (summaries with thread/checkpoint IDs), which you then load exactly. Vector recall is never the source of truth — it's an index into the exact store.

How this differs from MemWal's own MCP

MemWal ships an MCP for free-form memories (remember / recall / analyze / restore / login / logout). TuskPoint manages durable, exactly-addressable checkpoints — agent state you can resume a graph from. The only overlap, checkpoint_search, is deliberately scoped to our checkpoint summaries, not general memories.

Quick start

python -m pip install -e ".[all]"
cp .env.example .env   # then fill in your keys

All secrets come from environment variables (loaded from .env). See .env.example for the full list. Never commit your real .env — it is git-ignored.

Proofs and demos

Each build step has a runnable proof.

1. Walrus blob round-trip

python scripts/check_walrus.py

Writes a random blob to a testnet publisher, reads it back from an aggregator, and asserts the bytes are identical — printing the blob ID.

2. MemWal remember / recall

python scripts/check_memwal.py

Remembers a sentence, then recalls it semantically and prints the distance.

3–4. Crash / resume demo (the headline)

# In-memory fake backend (single process, interrupt then resume):
python demo/run_demo.py

# REAL Walrus testnet, surviving a genuine process kill:
python demo/run_demo.py --real --part1   # run to interrupt, persist, EXIT
python demo/run_demo.py --real --part2   # FRESH process rehydrates from Walrus

A researcher→writer agent is interrupted before the writer runs. --part2 starts a brand-new process that reads only the manifest blob ID from .walrus_threads.json, pulls the checkpoint back from Walrus, and resumes the writer to completion. That is the "survive a process kill" proof.

5. Semantic self-search

python demo/run_demo.py --semantic

Runs the agent on real Walrus + MemWal, then asks "when did the writer start?" and prints the nearest checkpoint summaries — the agent searching its own history.

MCP server

Six tools over stdio: checkpoint_save, checkpoint_load, checkpoint_list, checkpoint_resume, checkpoint_diff, checkpoint_search.

Run it directly:

python mcp_server/server.py

Register with an MCP client

A ready-to-use .mcp.json is included. For Claude Desktop, add the equivalent to claude_desktop_config.json:

{
  "mcpServers": {
    "tuskpoint-checkpoints": {
      "command": "python",
      "args": ["mcp_server/server.py"],
      "cwd": "C:/Users/User/Documents/tuskpoint",
      "env": {
        "WALRUS_PUBLISHER_URL": "https://publisher.walrus-testnet.walrus.space",
        "WALRUS_AGGREGATOR_URL": "https://aggregator.walrus-testnet.walrus.space",
        "WALRUS_THREADS_CACHE": ".walrus_threads.json"
      }
    }
  }
}

checkpoint_search returns an explanatory message instead of failing if no MemWal credentials are present, so the server runs fine without them.

Tests

python -m pytest -m "not integration"   # 16 fast unit tests, no network
python -m pytest -m integration         # live Walrus round-trip + resume

Project layout

src/langgraph_checkpoint_walrus/
  walrus_client.py   BlobStore protocol, InMemoryWalrusClient, real WalrusClient
  manifest.py        ThreadManifest / CheckpointEntry (id -> blob_id, lineage)
  saver.py           WalrusSaver (BaseCheckpointSaver): gzip envelope per checkpoint
  memwal_layer.py    MemWalLayer: build_summary + summarize_and_remember + search
mcp_server/server.py 6 checkpoint tools over stdio (FastMCP)
demo/                researcher→writer agent + crash/resume/semantic demos
scripts/             check_walrus.py, check_memwal.py (standalone proofs)
tests/               unit (no network) + integration (live Walrus) suites

90-second video demo script

1. (0:00) Show check_walrus.py — "agent state lands on Walrus, byte-identical round-trip."

2. (0:15) Run demo/run_demo.py --real --part1 — agent interrupts before the writer, exits.

3. (0:35) Point at .walrus_threads.json — "only a blob pointer survives locally; the state is on the network."

4. (0:45) Run --real --part2 in a fresh shell — "new process, resumes from Walrus, writer finishes."

5. (1:05) Run --semantic — ask "when did the writer start?", show ranked summaries.

6. (1:20) Show the MCP server tools list — "any MCP agent can save/load/diff/search checkpoints."