Agentic Memory: Walrus Takes On AI’s Next Big Bottleneck

As AI agents become increasingly ubiquitous, agentic memory is becoming one of the most important issues in the artificial intelligence space.

Enterprises and individuals are coming to rely on agents for ever more complex, high-stakes tasks—but the memory layer that most agents run on today has limitations that impact on the quality of their work.

That’s something Walrus, coupled with a recently launched SDK called MemWal, is looking to solve—bringing verifiability, availability, portability and sharability to agentic memory, Mysten Labs Group Product Manager Abinhav Garg told Decrypt.

“With Walrus plus MemWal, memory lives on an open, verifiable data layer, so that means it’s not tied to any one model or vendor,” Garg explained. That means users can switch between model providers such as OpenAI and Anthropic, while data is stored with verifiable guarantees, so it’s tamper-proof—something that’s “especially important as agents start operating in more critical workflows where correctness and auditability matter,” he said.

Data stored on Walrus inherits its built-in guarantees around verifiability, portability and availability, enabling “easier sharing of memory between agents across teams and organizations,” he added, making it a “must for agent collaboration.”

MemWal also integrates with popular agent orchestration frameworks OpenClaw and NemoClaw, through a plugin released this week. “We wanted to make the verifiable long term memory easy to adapt in real systems,” Garg said, adding that it enables a “seamless” workflow for builders.

“Without this, developers would have to understand the integration of a decentralized storage layer like Walrus, which could add friction and complexity,” he explained. “With the integration, they can just equip their agents with durable, verifiable memory directly with the tools that they’re already using.”

MemWal and privacy

Privacy is becoming “a much bigger issue for AI systems in general,” Garg said, noting that agents are increasingly being called upon to handle sensitive and proprietary data. “Whether that’s enterprise workflows, financial information or personal context, the expectations around confidentiality increase significantly,” he added.

MemWal and Walrus have privacy and programmable access control through a native encryption layer, meaning that “even though the storage itself is decentralized, the contents remain confidential and governed by policy—even the storage providers cannot read it,” Garg explained.

For users, he argued, “It’s no longer acceptable for that data to sit in some opaque, centralized system without clear guarantees,” noting that private, controlled and auditable storage for agentic memory will become “a defining requirement over time.”

New use cases for agentic memory

Empowering agentic memory with verifiability, availability, portability and sharability opens up an array of applications, Garg said, ranging from customer support agents that retain contextual cues about users, to collaboration between agents in different teams “working off the same customer history.”

“There is an amazing partner who is trying to figure out how there can be coordination between agents as a publisher or a consumer on a marketplace,” he added. “So how would those agents interact with each other and engage in sort of a messaging over a period of time? And that messaging can be used as a sort of memory itself.”

Other partners have been exploring agentic memory for robots that need to share context with each other to coordinate tasks in the real world. “So, imagine if they’re doing that over hours or even weeks—let’s say during a disaster response scenario, they would need that shared memory,” Garg explained.

Ultimately, he anticipates a “standardization of the stack” for agents. “You’ll see clear separation between compute, data, memory and coordination,” he said. “Our view is that memory and data shouldn’t be tied to any single model or platform—so Walrus becomes that durable data layer and MemWal becomes a memory layer on top of it.”

Use the quick start guide to add MemWal memory to your agents now.