As artificial intelligence (AI) agents become more prevalent, the challenge of agentic memory emerges as a critical issue within the field. Enterprises and individuals increasingly depend on these agents for complex and high-stakes tasks, yet current memory solutions present limitations affecting performance.
Walrus, in collaboration with its newly introduced SDK MemWal, aims to address these challenges by enhancing verifiability, availability, portability, and sharability of agentic memory, as explained by Mysten Labs Group Product Manager Abinhav Garg during an interview with Decrypt. “By utilizing Walrus along with MemWal, the memory exists on a transparent, verifiable data layer, freeing it from dependency on any particular model or vendor,” Garg clarified. This flexibility allows users to transition between different model providers such as OpenAI and Anthropic while ensuring that their data is stored securely with tamper-proof guarantees—crucial for agents operating in critical workflows where accuracy and accountability are paramount.
Walrus bestows inherent verifiability, portability, and availability on stored data, facilitating the seamless sharing of memory among agents across various teams and organizations, making it essential for agent collaboration. Moreover, MemWal integrates with leading agent orchestration frameworks like OpenClaw and NemoClaw through a newly released plugin. “Our goal was to simplify the adaptation of verifiable long-term memory in practical applications,” Garg stated, emphasizing that this integration allows builders to incorporate durable, verifiable memory into their systems using familiar tools.
Addressing privacy concerns is increasingly vital for AI systems handling sensitive and proprietary data, noted Garg. Agents are more frequently tasked with managing enterprise workflows, financial information, or personal contexts where confidentiality expectations rise significantly. MemWal and Walrus ensure privacy through an inherent encryption layer that guarantees content confidentiality and policy governance—even storage providers cannot access the data.
For users, it is imperative to move away from relying on opaque, centralized systems without clear assurances of privacy and control. Garg argued that private, controlled, and auditable storage for agentic memory will become a key requirement over time.
Enhancing agentic memory with these features opens up diverse applications, such as customer support agents retaining contextual user information or enabling inter-team agent collaboration using shared customer histories. Garg mentioned an exciting partnership exploring coordination between agents acting as publishers or consumers on marketplaces, allowing for messaging-based interactions that serve as a form of memory.
Other partners are investigating agentic memory for robots requiring context sharing to coordinate tasks in real-world environments, such as disaster response scenarios needing sustained shared memory over extended periods. Garg anticipates the standardization of agent technology stacks with distinct separations between compute, data, memory, and coordination. “Our vision is that memory and data should remain independent from any single model or platform—thus, Walrus serves as a robust data layer while MemWal provides an adaptable memory solution,” he concluded.
To integrate MemWal memory into your agents quickly, follow the provided quick start guide.
Brought to you by Walrus
Learn More about partnering with Decrypt.