Ghostwriter: liquidation cover on Sui, priced by an AI whose memory lives on Walrus.
Borrow on Sui and one bad hour for SUI can liquidate you. Most on-chain cover is an escrow and a price feed. We built an underwriter instead: it prices from experience, explains itself, says no, and learns when it's wrong.
Built for Total Recall: Tatum × Walrus Memory.
How it works
Ask for cover: "Pay me 10 gUSD if SUI drops below $0.80 in 7 days."
Ghostwriter recalls similar past cases from Walrus Memory and adjusts the price. Here, SUI touched the trigger in 81 of 180 similar windows, so memory made the cover cheaper (×0.90). On a riskier, near-the-money policy it went the other way (×1.35). Every factor links to the Walrus blob it came from.
No claim form. When the Pyth price crosses your trigger, the Move contract checks the price itself and the vault pays your wallet automatically.
What makes it different
It says no. Ask for 880 gUSD of cover and it declines, because one policy is capped at 25% of free capital. It counter-offers 495.36 gUSD, one click to accept.
It learns from losses. After one payout cost it 2.7 gUSD, it wrote itself a new pricing rule and saved it to Walrus:
SUI, mid volatility, 0–5% below spot → charge ×1.25
Nobody told it to.
It only knows what it remembers. Ask "What's the capital of France?" and it replies: "I have no memory of the capital of France." It answers from Walrus only, with blob IDs as receipts.
Agents buy cover too. Type one sentence, like "Cover my SUI if it drops 15% over 3 days, max 4 gUSD." An AI agent gets a quote over x402, checks it against the budget, and pays from its own wallet.
The stack
- Walrus Memory (Sui Mainnet): 105 memories of cases, decisions, policies and self-written rules. Borrowers own their own memory account.
- Tatum: every Sui read and write, on both networks.
- Sui + Move: policies, vault and automatic payouts.
- Pyth: price checks inside the contract.
- x402: agent-to-agent payments.
- Nautilus: an enclave that verifies Suilend liquidations on Mainnet.
The takeaway
When an agent's memory is something you can open and audit, you can trust its judgement. When it can write to that memory, it can learn.
An underwriter that remembers, shows its evidence, says no when it should, and learns when it's wrong.
🎥 Demo: [https://www.youtube.com/watch?si=oLevczEVpA8fNajQ&v=YLs4Wi_S6h4&feature=youtu.be]
💻 Code: [https://github.com/austinjeremiah/Ghost-Writer]
Top comments (0)