Walrus launches MemWal SDK to give AI agents verifiable, portable memory

Walrus, the decentralized storage protocol built on the Sui blockchain, just launched something designed to fix that. The MemWal SDK is a developer toolkit that gives AI agents persistent, encrypted memory stored on Walrus’s decentralized infrastructure, complete with semantic search so agents can actually retrieve what they’ve learned.

What MemWal actually does

The SDK stores encrypted memories on the Walrus network and layers semantic search retrieval on top. That means agents don’t just dump information into a blob of text. They can query their own memory intelligently, pulling up relevant context based on meaning rather than exact keyword matches.

Abinhav Garg, Group Product Manager at Mysten Labs (the team behind both Sui and Walrus), framed the core value proposition around openness. The framework allows memory to reside on a data layer that is both open and verifiable, not dependent on any single AI provider’s infrastructure or whims.

The Sui blockchain handles ownership and access control. So users, not corporations, decide who gets to read, write, or share an agent’s memories.

The integration story and developer play

Walrus didn’t launch MemWal in a vacuum. The SDK ships with integrations for Vercel AI SDK, along with plugins for frameworks called OpenClaw and NemoClaw. Quick-start guides and comprehensive documentation are available, and the team is actively soliciting developer feedback through GitHub. The SDK is currently in beta.

Why decentralized memory matters for AI

MemWal’s pitch is that agent memory should be user-owned infrastructure. The four pillars the team highlights are verifiability, availability, portability, and shareability.

Verifiability means you can confirm the memory hasn’t been tampered with. Availability means it persists even if a specific provider goes offline. Portability means it moves between models and vendors. Shareability means multiple agents, or multiple users, can collaborate using shared memory pools.

This builds on work that Walrus and Mysten Labs have been developing since at least March 2025, when early foundations for this kind of decentralized storage infrastructure were being laid. The MemWal SDK represents the most explicit move yet to position that infrastructure as an AI-specific tool.