Walrus Foundation has unveiled MemWal, a beta SDK and relayer architecture intended to give AI agents persistent, portable memory using Walrus storage and Sui-based ownership and access controls.
Walrus Foundation has introduced MemWal in beta, positioning it as a developer SDK and relayer architecture for AI agents that need to retain context beyond a single session or application. In its announcement, the foundation describes the system as a way to create agent memory that is durable, portable, shareable and verifiable. Walrus Foundation says the design stores memory data on Walrus while using Sui for ownership and access management.
The proposal addresses a familiar limitation in agent development: relevant state is often temporary or isolated within a particular model conversation, provider or application. Developers may want an agent to carry task history, user preferences or other working context when moving between tools and sessions. MemWal is intended to provide a storage and permissions layer for that use case rather than keeping that information solely inside an active model context window.
According to Walrus Foundation, MemWal separates the persistence of memory data from the rules that govern access to it. Walrus is used as the storage layer, while Sui-based logic is intended to determine who owns memory and who can read, update or share it.
That split could allow application developers to treat agent memory as a resource with programmable permissions. For example, a developer could preserve an agent's context over time while setting different access rules for users, applications or collaborating agents. Walrus frames the resulting memory as verifiable because it can be associated with the system's ownership and authorization model.
Tatum, which describes itself as a Walrus partner, similarly says the integration is designed to help agents carry context across sessions, providers and applications. Its overview highlights decentralized storage and Sui-based ownership logic as components for maintaining persistent context outside an individual AI product. Tatum's description focuses on the potential for agents to retain continuity as they operate across different environments.
A separate PR Newswire release refers to a product called Walrus Memory, describing it as a portable memory layer for agents with programmable permissions and shared-memory workflows. The supplied announcements use different product names—MemWal in the beta SDK announcement and Walrus Memory in the PR Newswire release—while describing substantially similar goals around portable and permissioned agent context.
The available materials do not establish the precise product relationship or rollout sequence between those names. Developers evaluating the offering should therefore consult Walrus's current technical documentation and release materials for supported frameworks, deployment details and availability.
MemWal is presented as infrastructure for builders rather than as a consumer-facing assistant. Its beta status also means that broad adoption, interoperability with agent frameworks and real-world operational performance remain open questions.
Still, the launch reflects growing work on systems that let AI agents maintain controlled, reusable context across executions. Walrus's approach is to make that memory persist independently of a single model session, while attaching ownership and sharing rules to the data developers choose to retain.
In its announcement, the foundation describes the system as a way to create agent memory that is durable, portable, shareable and verifiable.
Walrus Foundation says the design stores memory data on Walrus while using Sui for ownership and access management.
The proposal addresses a familiar limitation in agent development: relevant state is often temporary or isolated within a particular model conversation, provider or application.
Continue reading
