Kage — Onchain Agent Memory Protocol

影 (Kage) — Shadow. What an agent remembers is its shadow on the chain.

Kage is the missing memory layer for autonomous AI agents on Sui. Not just persistent storage — queryable, conflict-aware, event-driven, temporally-snapshotted, cross-graph agent cognition infrastructure.

Built for Sui Overflow 2026 — Walrus Track.

The Problem

Agents today forget. They rely on fragile local memory, can't share context across workflows, and produce zero verifiable audit trail. When a trade goes wrong, there is no record of what the agent knew, or why it decided.

LangGraph, Mastra, ElizaOS — none of them solve this. They're all offchain.

Kage Solves It With 5 Primitives

| # | Feature | What it does | |---|---------|-------------| | 1 | Associative Memory Index | Tag-based semantic query — find memories by asset, action, namespace, confidence | | 2 | Conflict Detection | Two agents disagree on BTC? Flagged automatically onchain | | 3 | Memory Inheritance | Child agent bootstraps from parent memory — linked provenance chain | | 4 | Temporal Snapshots | Freeze namespace state at any point — full time travel audit | | 5 | Cross-graph Sharing | Two separate agent graphs share a memory namespace via a permissioned bridge |

Plus event-driven triggers — subscriptions that fire a Sui event when a matching memory is written, waking the next agent automatically.

Architecture

Your Agent (TypeScript/Python)
        │
        ▼
   Kage Runtime SDK
   ┌──────────────────────────────┐
   │  write() / inherit()         │
   │  snapshot() / share()        │
   │  subscribe() / recall()      │
   └──────────┬───────────────────┘
              │
      ┌───────┴────────┐
      ▼                ▼
  Walrus            Sui Move
  (blob storage)    (truth layer)
  blob_id ──────────► memory index
  state JSON         conflict detection
  persistent         subscriptions
  verifiable         snapshots
                     cross-graph bridges

Quick Start

1. Deploy the contract

chmod +x deploy.sh
./deploy.sh          # mainnet
./deploy.sh --testnet

2. Configure

cp .env.example .env
# Fill in SUI_PRIVATE_KEY_B64
# KAGE_PACKAGE_ID and KAGE_PROTOCOL_ID are filled by deploy.sh

3. Install and run the demo

npm install
npm run demo

SDK Usage

import { Kage } from './src/kage.js';

const kage = new Kage({
  suiRpcUrl:           'https://fullnode.mainnet.sui.io:443',
  privateKeyB64:       process.env.SUI_PRIVATE_KEY_B64!,
  packageId:           process.env.KAGE_PACKAGE_ID!,
  protocolId:          process.env.KAGE_PROTOCOL_ID!,
  walrusAggregatorUrl: process.env.WALRUS_AGGREGATOR_URL!,
  walrusPublisherUrl:  process.env.WALRUS_PUBLISHER_URL!,
  agentName:           'MyAgent',
  namespace:           'myapp',
});

await kage.init();

// Write a memory
const result = await kage.write({
  state:      { signal: 'UP', price: 67420 },
  tags:       ['BTC', 'SCOUT', 'UP'],
  asset:      'BTC',
  action:     'BUY',
  confidence: 78,
});

// Inherit from parent
const child = await kage.inherit(result.memoryId, newState, 'EXECUTE', ['BTC', 'RISK'], 65);

// Snapshot the namespace
await kage.snapshot('Pre-trade state captured');

// Cross-graph share
await kage.createBridge(otherProtocolId, 'shared-signals');
await kage.shareMemory(bridgeId, otherProtocolId, result.memoryId);

// Subscribe to BUY signals
await kage.subscribe(['BUY'], 'BTC', 'BUY', 'myapp');

// Recall a blob from Walrus
const state = await kage.recall(result.blobId);

// Report outcome (updates reputation)
await kage.reportOutcome(result.memoryId, 'SUCCESS');

AlphaLoop Demo

AlphaLoop is a live 4-node trading agent that uses Kage as its memory layer. It pulls signals from Trend Pilot (X Layer) and runs Scout → Risk → Decide → Execute.

Every node writes to Walrus, inherits from the previous node, and the full decision trail is anchored onchain.

npm run demo

File Structure

kage/
├── move/
│   ├── Move.toml
│   └── sources/
│       └── kage.move          # The full protocol contract
├── src/
│   ├── kage.ts                # Main SDK
│   ├── walrus.ts              # Walrus client
│   ├── sui.ts                 # Sui Move client
│   ├── types/
│   │   └── index.ts           # All TypeScript types
│   └── demo/
│       └── alphaloop.ts       # Live trading agent demo
├── deploy.sh                  # One-command deploy
├── .env.example
├── package.json
└── tsconfig.json

Built by Semideus / @syke0x

Sui Overflow 2026 — Walrus Track — $70K prize pool