Bias
AI characters that remember you — long-term memory stored on Walrus, encrypted with Seal and recalled via MemWal (Sui Overflow 2026)
- Team
- 3 people
- My role
- Full-stack, Chat & memory
The challenge
AI character apps forget you when the chat ends. For a relationship to continue, memory has to persist, belong to the user, and never leak between rooms.
Key decisions
Memory scoped per room × character × user
Each turn recalls relevant memories from that scope into the prompt and writes the new turn back — so 1:1 memories stay out of group rooms and one character keeps a separate relationship with each user.
Store only what the LLM marks important
The reply comes with an importance score and low-importance turns are dropped — natural forgetting that also cuts storage cost.
Memory in the user's account, server via delegate keys
Onboarding creates a MemWal account per user, and the server reads and writes only through keys the user delegates — reusing an existing account and delegation when found.
Results
- 1:1 and group rooms (per-character speakers) with long-term memory
- Live demo on Vercel and a demo video
Tech
- Next.js — Chat rooms and API, onboarding with wallet connect
- AI Agent — Prompt from persona, memories and history; importance scoring
- Walrus — MemWal memory on Walrus with Seal encryption
Planning & collaboration
- Architecture & user flows — Onboarding → MemWal account and delegation → chat
Evidence
- repoGitHub
- demoDemo
- videoDemo video