Build-A-Bot: Where Life-Long AI Companions Are Made

The First-Ever On-Chain Composable AI Agent Character Platform on Sui

Sui AI Agent Typhoon - ElizaOS and Sui Walrus Track

---

Getting started with Build-A-Bot:

1. Make copies of .env.example in both bab-inference-engin/ and backend/, fill in the parameters with API keys and MongoDB configs
2. cd bab-inference-engine/ && pnpm install
3. pnpm build && pnpm start to start the ElizaOS core
4. pnpm start:client to start the BAB front end
5. cd ../backend/ && npm install
6. ts-node src/main.ts to start the BAB backend

---

Project Summary

Build-A-Bot (BAB) is the first-ever on-chain composable AI Agent character platform on Sui. Inspired by the idea of customizing a teddy bear (choosing fabric, filling, and accessories), BAB allows users to:

• Cherry-pick and combine different on-chain tokenized AI Agent modules—Personality, Knowledge Bank, Speech, Tone, and Action—from an open marketplace.
• Compose these modules to create a truly unique AI Agent Companion with full on-chain ownership.

This is powered by BAB’s novel Module-as-a-Service system, enabling AI researchers and enthusiasts worldwide to:

1. Securely upload and tokenize their pre-trained AI modules (i.e., weights / configuration files).
2. Store these modules on decentralized storage via Walrus to ensure data availability and censorship-resistance.
3. Make them openly available in a composable and liquid marketplace so other users can buy their work and integrate it into their own AI Agents.

With BAB, you are no longer constrained to a handful of pre-made AI character templates (like those found on Character.ai). Instead, you can fully customize and own/trade each component of your AI character, akin to building Lego blocks for AI Agents.

---

Integration with ElizaOS

We enhanced the existing Eliza OS by modularizing its character file system, breaking it down into distinct parts and adding a memory module. Each agent’s behavior is defined by a character configuration file, which is now divided into separate modules: character, knowledge, voice, tone, and action. The knowledge modules store information in string arrays, while the voice modules, integrated with Eleven Labs, include settings like voice speed. These modules are tokenized into a Sui smart contract NFT, which can be listed on a marketplace. When purchased, an NFT is minted, allowing the use of the modules.

To initialize an agent, a memory module is minted, which stores the ownership to its chat history. If an existing memory module is used, the previous chat history is injected into the new session. The final agent is composed by integrating these modules into a character file and chat log, enabling a seamless and customizable user experience.

• The current state of ElizaOS is that characters are instantiated during runtime initiation
• We implemented an agent runtime reinstantiation system where whenever a user recomposes the AI agent, a new Eliza agent runtime will be initiated with the corresponding parameters used by composed agent
• This system is a lot flexible and introduces new opportunities

---

Core Features

1. Securely Tokenize AI Modules with Module-as-a-Service

   • AI researchers can list their ComposableModule (weights/configs) on-chain.
   • Storage is decentralized and censorship-resistant (powered by Walrus).
   • Ownership is tokenized, enabling you to buy or sell these modules as ComposableModuleInstance assets.

2. Composable AI Agent Companions

   • Users on BAB can own different tokenized AI modules and freely swap them at any time.
   • Don’t like your AI Companion’s current Tone? Purchase a different Tone module and replace it instantly.
   • Achieve near-infinite customization for personality, knowledge, voice, actions, and more.

3. Full Custody and Persistent Memory

   • The Memory Module is stateful and stores all past interactions or knowledge summaries.
   • It is tokenized, encrypted, and remains under the sole control of its owner.
   • As long as you have enough WAL tokens for storage, your AI Companion’s memory persists indefinitely on Walrus—ensuring your Agent lives on forever in the decentralized web.

4. Modular, Extendable, and Backward Compatible

   • Built with the BAB Sui Package for easy addition of future AI features.
   • As AI technology evolves, your AI Agent’s capabilities can grow with it.

---

What is the anatomy of a BAB Composable AI Agent?

A BAB Composable AI Agent composes of two categories of modules:

1. Stateless Functional Modules

   • Personality Module: Defines how the AI Agent speaks, behaves, and its general quirks or traits.
   • Knowledge Bank Module: Integrates verifiable external knowledge (e.g., live market prices, textbooks) to provide accurate, real-time responses.
   • Speech Module: Controls text-to-speech characteristics—pitch, speed, and voice style (masculine or feminine, formal or casual).
   • Tone Module: Applies a secondary filter or “character flair” to the speech—could mimic a Japanese anime character, celebrity, or historical figure.
   • Action Module: Lets the AI Agent perform external tasks—sending a tweet, executing a transaction on Sui, etc.

2. Stateful Memory Module

   • Memory Module: Holds the Agent’s “soul,” capturing conversation histories, personal details, and knowledge summaries. Swapping Memory Modules effectively replaces the Agent’s identity—because the memory is what retains continuity of personality and past experiences.

In the BAB WebApp, users can freely permute and compose new AI characters by layering these modules like building blocks. By choosing any combination of modules, you can instantly form and chat with a uniquely composed AI Agent.

---

How AI Researchers & Enthusiasts Can Sell Their Work

BAB provides a straightforward publishing flow:

1. Drag & Drop Pretrained Model Weights

   • Choose the module type (Personality, Knowledge Bank, etc.).
   • Upload your data to Walrus via a user-friendly UI.

2. Tokenize Your Module

   • BAB creates an on-chain ComposableModule object, binding your IP/content with a verifiable proof of data availability on Walrus.
   • Anyone who acquires your module’s token can seamlessly integrate it into their AI Agent’s composition.

This opens a liquid marketplace for modules, enabling module owners to profit from their AI research while preserving ownership over the underlying assets.

---

How It All Works (System Architecture)

The BAB system is built on top of the foundational work of ElizaOS, enhanced with modular loading capabilities for on-the-fly AI Agent module swapping. This architecture is powered by:

1. Walrus

   • A censorship-resistant and cost-efficient data storage layer.
   • Stores AI modules, ensuring high data availability and censorship resilience.

2. Sui

   • Used for on-chain module ownership coordination.
   • To do so, we implemented the ComposableModule object standards (defined in BAB::Core::ComposableModule).
   • Low gas fees and fast finality make Sui ideal for frequent on-chain transactions (buying, selling, or swapping modules).

High-Level Flow:

1. Upload & Tokenize AI Modules → Stored on Walrus, minted as Sui tokens.
2. Assemble AI Agent → Pick modules (Personality, Knowledge, Speech, Tone, Action) + attach a Memory Module.
3. Interact in Real-Time → Chat, swap modules, or perform actions, all while the Memory Module logs and persists state.