DEV Community

Cover image for Walrus Session 8 Challenge: Build Chatbots That Remember & Win from a $2,500 Prize Pool
Vampire
Vampire

Posted on

Walrus Session 8 Challenge: Build Chatbots That Remember & Win from a $2,500 Prize Pool

#ai #agents #softwaredevelopment

We spend hours architecting full-stack web applications, managing state, and polishing UI flows. Yet, the moment we integrate an AI agent into these systems, we often accept a massive downgrade in user experience: the AI forgets everything the second the session closes.

Most conversational AI experiences break down precisely because they treat state as transient. A customer support bot repeatedly asks for account identifiers. An AI tutor or companion has no idea what concepts you discussed yesterday.

The industry’s current band-aid is simply expanding the context window. But stuffing hundreds of thousands of tokens of raw chat history into a prompt on every turn isn't true memory—it’s computationally expensive brute force. It destroys response latency and doesn't solve cross-session, cross-device persistence.

To build agents that genuinely adapt to users, we need a dedicated, durable state store.

Decentralizing Agentic Memory

The architecture needs to shift from ephemeral caches to persistent, addressable memory. When an agent needs context, it shouldn't just parse raw logs; it needs structured state tied to a user's identity that persists across sessions and interfaces.

This is the exact infrastructure challenge being addressed with decentralized state layers like MemWal. By using a decentralized network as the ground truth, you decouple the "brain" (stateless LLMs) from the "memory" (persistent state). The context follows the user, not the application instance.

Walrus Session 8: Chatbots That Remember

If you are building AI integrations, indie web apps, or experimenting with local LLMs, the Walrus Session 8: Chatbots That Remember developer challenge is specifically designed to tackle this problem.

Running from September 18 to October 9, 2026, this hackathon invites builders to create a new conversational agent or retrofit an existing chatbot using persistent memory.

Core Submission Requirements

  • Memory Integration: Integrate Walrus Memory via MystenLabs/MemWal to store and recall context between conversations.
  • Demonstrated Usage: Showcase at least 3 distinct users interacting with the agent, with at least 10 memories stored per user.
  • Live Deployment: Deploy to a real environment (a web widget, Discord bot, Telegram, WhatsApp, Slack, or public CLI).
  • Open Source & Setup: Provide a public GitHub repository with clear setup and local reproduction steps.
  • Technical Article (500–800 words): Publish a breakdown detailing your architecture, what broke during development, and the exact moment persistent memory changed the conversation.

Prize Tracks ($2,500 Total in Stablecoins)

  • Main Prizes — Best Chatbot ($900 total):
    • 1st Place: $500
    • 2nd Place: $250
    • 3rd Place: $150
  • Beyond the Big Two ($300 total - 2 × $150): Dedicated to builders using open-weight local runtimes (Ollama, vLLM, llama.cpp with Llama, DeepSeek, Mistral, Qwen) or alternative hosted engines (Google Gemini, Groq, Fireworks) instead of OpenAI or Anthropic.
  • Best Written Article ($300 total - 3 × $100): Awarded for the most practical, transparent technical write-ups.
  • Bug Bounty ($500 total - 5 × $100): Report reproducible edge cases, integration friction, or SDK bugs on the MemWal Issues Page.
  • Promo Bounty ($500 total - 5 × $100): Share the hackathon or your article in external developer communities outside the Sui ecosystem.

Getting Started

Top comments (0)