Most conversational AI agents have a major flaw: the moment a user closes their browser tab or clears their session, all context is lost.
In Web3, this is especially frustrating. Users have to re-paste 66-character wallet addresses, re-state risk limits, and re-explain preferred settings every single time they start a chat.
Web2 chatbots paper over this by saving logs to centralized databases or proprietary cloud vector services, but that defeats the purpose of user data sovereignty.
To fix this, we retrofitted our on-chain agent, Barzakh AI, with decentralized, SEAL-encrypted memory on Walrus Mainnet as part of the Walrus Sessions 8 hackathon.
What changed when the agent could remember:
Zero-Parameter Recall: When a returning user says "gm", the agent greets them by name, remembers their tracked wallet, and surfaces active risk rules.
Active Tool Parameter Injection: Memory is not just decorative chat text. When the user asks "What are my recent transactions?" or "Check my portfolio", the agent auto-fills the remembered address into on-chain tools with zero user input.
Decentralized Storage: Memories are stored as SEAL-encrypted blobs on Walrus, giving users verifiable cryptographic control over their agent's memory, inspectable directly on Walruscan.
We wrote up the full technical breakdown, before/after comparison, architecture flow, and real conversation transcripts in our article: Read Here
Curious to hear how other developers here are handling cross-session state and user privacy in autonomous AI agents!
Top comments (0)