Mysten Labs, the blockchain infrastructure company behind Sui, has released a new tool called Walrus Memory that lets AI agents hold onto context as they move between applications, sessions, and service providers. The portable memory layer gives users more control over how their AI agents remember and apply information across different digital environments.
Why Persistent Memory Matters for AI
Most AI agents today start each conversation or task with a blank slate. They lose track of user preferences, past instructions, and ongoing work the moment the session ends or the user switches to a different app. Walrus Memory aims to break that pattern by providing an external, user-managed memory store that agents can read from and write to no matter where they're running.
The technology works as an off-chain layer that attaches to AI models without requiring changes to the underlying model itself. An agent using Walrus Memory can remember a user's preferred tone, recent project details, or even a partially completed form across apps like a chatbot, a code editor, and a scheduling tool.
User Control Over Context
A key design choice in Walrus Memory is placing the user in charge of what gets remembered and shared. Instead of letting each service provider hold its own siloed memory — which users can't review or delete easily — Walrus Memory puts the data on a decentralized storage layer where the user holds the keys. That means a person can inspect exactly what their AI agent remembers, delete specific memories, or revoke access for a particular app.
The system uses cryptographic proofs to ensure that only authorized agents can read or write to a user's memory store. Users grant permissions per application, and those permissions can be revoked at any time.
Because Walrus Memory is not tied to any single AI provider or platform, it can work with agents from different vendors. A session started with an agent from one company can continue seamlessly when the user switches to another provider's assistant, as long as both agents have permission to access the same memory store.
Mysten Labs developed the tool partly to address fragmentation in the AI agent ecosystem, where users often have to reconfigure preferences and repeat instructions for each new tool they try. The portable memory layer could reduce that friction, though the company has not announced any integration partners yet.
Walrus Memory is available now as a developer preview. Mysten Labs has not set a timeline for a full production release, nor has it disclosed pricing or storage limits. Developers who want to test the layer can find documentation and code samples on the company's website. Whether mainstream AI providers will adopt a portable memory standard or continue building their own walled gardens remains an open question for the industry.
