Project Potemkin
AI Agent & Memory Systems
Project Potemkin is an experimental AI agent system focused on persistent memory, autonomous interaction, and long-term user relationships.
Overview
This project was our submission for UK AI Agent Hackathon EP3 by ASI (Artificial Superintelligence Alliance), one of the largest Web3 x AI hackathons in Europe.
From first idea to final demo and promo video, we pushed it out in one week. For me, this project was a big turning point because it pulled me into Web3 while forcing me to learn a lot of web development in real time.
The core concept is autonomous AI behavior plus on-chain memory ownership, so users keep sovereignty over their memory agent and can still support decentralized third-party sharing. It is still evolving, but we genuinely believe the direction is worth building. This project also won an award in the Unibase track at this hackathon. You can watch the trailer below and try the demo directly.
Project Partners
Built with: Zhewen (Ryan) Zhen, Shian (Andy) Ye.
Key Ideas
- Persistent memory architecture
- Autonomous agent behavior
- Human-AI interaction over time
- Multi-session context evolution
Implemented Highlights
- Family-scoped identity with isolated memory storage per family.
- Automatic long-term memory + profile refresh every 16 short-term turns.
- Memory buckets: shared, important events, private, and user-approved public.
- Signup/login, bearer token auth, and multi-member family accounts.
- Family-level language setting with multilingual input support.
- Optional BNB Testnet on-chain space and permission purchase flow.
Memory & Profile Flow
- Each turn is written to STM (SQLite + vector retrieval store).
- Every 16 turns are summarized into LTM automatically.
- Family profile is refreshed after long-term summarization.
- Optional hub sync enables cross-device memory reuse.
Demo & Links
Technologies
Python · LLM APIs · Memory Systems · Agent Architecture
Project Trailer
Project Potemkin introduction video.