Most chatbots forget everything the second a conversation ends. Ask it
something on Monday, come back Tuesday, and you're a stranger again.
Walrus Sessions 8: "Chatbots That Remember" is a live hackathon
session tackling exactly that — building (or retrofitting) a chatbot
that uses Walrus Memory to persist context across sessions, users,
and devices.
The challenge
Build or retrofit a chatbot that:
- Stores and recalls context with Walrus Memory (on mainnet)
- Gets deployed somewhere real and used for a few days
- Gets written up — what it does, how memory was integrated, before/after, real evidence it works
Any use case counts: customer support, onboarding, tutoring, sales,
community bots, even game NPCs.
Why I got interested
I spent a few days going deep on the underlying open-source SDK
(MemWal) — not just building with it, but actually reading through the
codebase: the auth layer, the on-chain contract, the sample apps. It's
a genuinely well-built system (constant-time auth checks, on-chain
delegate-key verification, careful quarantine/replay protection), and
digging into how persistent memory is implemented securely taught me
more than most tutorials do. Ended up finding and reporting a couple of
real bugs along the way, which the maintainers fixed fast.
The details
- Dates: Sept 18 – Oct 9, 2026
-
Prizes: $2,500 in WAL total
- Best Chatbot: $500 / $250 / $150
- Beyond the Big Two (non-Claude/GPT): 2 × $150
- Best Article: 3 × $100
- Bug Bounty (no chatbot required, just a reproducible bug on GitHub): 5 × $100
- Repo: https://github.com/MystenLabs/MemWal
- Full rules & submission: [https://walform.wal.app/f?formId=0x09b022796f9cb7ce24247e3097c5c8ae2b414317c90c8aeb6ce335e7caf31ff5]
If you've ever wanted an excuse to actually understand how persistent,
encrypted agent memory works under the hood instead of just calling an
API — this is a good one to dig into, whether you build the full
chatbot or just go bug-hunting in the SDK.
Top comments (0)