Walrus Sessions 4How WalCup 26 Works
A World Cup 2026 prediction arena where every pick is stored permanently on Walrus, and an AI agent remembers everything across sessions.
What is WalCup 26?
WalCup 26 is a World Cup prediction game built on the Sui blockchain and powered by Walrus Memory (MemWal) — a decentralized blob storage protocol with semantic search, giving the AI agent genuine cross-session memory.
Built for the Walrus Sessions 4 hackathon, it demonstrates a real use case for persistent on-chain memory: an agent that knows your history, detects your prediction biases, and roasts you for them — across every session, forever.
48
Teams
∞
Walrus Blobs
1
Agent that never forgets
Making Predictions
Connect Wallet
Connect your Sui wallet. Your address becomes your unique on-chain identity — no sign-up needed.
Pick & Sign
Choose a match, pick the winner, predict the score, and sign with your wallet. Your wallet proves ownership of every prediction.
Stored on Walrus
The prediction is stored as a memory blob on the Walrus network. You get a blob ID you can verify live on WalrusScan.
🦭 Walrus Memory — How It Works
MemWal SDK wraps the Walrus blob storage protocol with a namespace-based memory layer. Each wallet gets a dedicated namespace: wc2026-sui-[wallet]. Predictions are written as natural-language memory texts, stored as encrypted blobs, and indexed with vector embeddings for semantic search.
// Prediction storage flow
1. User signs prediction with Sui wallet
2. API formats it as natural language memory
3. memwal.rememberAndWait(text) → blob stored on Walrus
4. Returns blob_id — verifiable on WalrusScan
5. blob_id saved to user record in Convex DB
Wallet-scoped
Each namespace is tied to a Sui wallet address. Only predictions from your wallet live in your namespace.
Decentralized blobs
Data lives on the Walrus network of storage nodes — no single server controls your memories.
Cross-session
Memories persist across browser sessions, devices, and app redeployments. The agent remembers Day 1.
Semantic recall
Vector-indexed blobs. The agent queries "Brazil predictions" and finds all relevant memory across all sessions.
🎙️ The Commentator — AI Agent
Powered by Groq Llama 3.3-70b via the Vercel AI SDK. Before every response, the agent queries your Walrus namespace for semantically relevant memories — past predictions, stated opinions, detected biases — and injects them into the system context.
// Agent memory retrieval flow
1. User sends message to /api/agent
2. memwal.recall(query, limit=15) → top blob matches
3. Blob texts injected into Groq system prompt
4. Llama 3.3-70b generates response with full context
5. Agent detects biases, tracks accuracy, roasts freely
Without a Sui wallet, the agent uses in-app predictions only — no persistent Walrus memory. Connect a wallet to unlock full cross-session recall.
🏆 Points System
⚙️ Tech Stack
⚠️ About Match Data
Fixtures and results are based on the official World Cup 2026 schedule but are not auto-synced with live scores. Results are updated manually. For official standings visit fifa.com ↗.
Built by Ed for Walrus Sessions 4