Agentic memory is “a mirror of ourselves,” says Kostas Chalkias, Co-Founder and Chief Cryptographer at Mysten Labs, original contributors to Walrus. In an ideal world, that memory should be portable, with AI agents coordinating and carrying context between apps and sessions—but to date they’ve been hamstrung by the limitations of memory.

Developers building AI agents have been left to stitch together databases, vector stores, and runtime state, resulting in unreliable systems that struggle to cope with complex workflows—and agents that forget.

“The major misconception in AI is that compute is the only bottleneck,” Chalkias said. “The major issue is we're using a lot of memory as humans, and we want our LLMs to actually learn about us.” That, he said, means solving the “real bottleneck” of agentic memory.

That’s what Mysten Labs is aiming to fix with its newest offering, Walrus Memory, a memory layer built specifically for AI agents and designed around portability, user control and agent coordination.

Chalkias explained that Walrus Memory brings together multiple features that are a “necessity” for AI agents. “Just having fast compute, you don't necessarily have privacy; just having an encryption layer, you don't necessarily have a way to share your policies on whatever LLMs you want,” he said. “If you just have large data, this is also not enough.”

Walrus Memory enables agents, apps and workflows to seamlessly share memory without being tied to a single runtime, session or provider, while shared memory spaces empower multiple agents to coordinate across long-running workflows. Cryptographic tools such as zk-proofs, meanwhile, are deployed to enable agents to perform contextual verification, and allow for programmable access to encrypted memory.

“I don't believe any other, especially blockchain-focused solution at the moment, is solving all of these three elements, which is pretty much the major bottleneck for most of them to work,” he added.

Walrus Memory integrates with leading AI platforms including Claude, ChatGPT and Gemini, Chalkias noted, ensuring that users aren’t locked into working with a single model provider—future-proofing user workflows.

Data stored on Walrus Memory also comes with programmable access control. “It's not only recall accuracy, it's also transparency; you don't want your data to be there forever, you don't want your data to be misused,” Chalkias said.

Plugins for OpenClaw and NemoClaw together with Python and TypeScript SDKs mean that developers can easily add portable memory to existing agent workflows. Already, teams including Allium, Conso Labs, Inflectiv, OpenGradient, Talus Labs, and Tatum are working with Walrus Memory to build applications including portable agent identity systems and AI assistants that remember customer interactions across sessions.

Memory handling is “getting better and better” by the day, Chalkias said, noting that Walrus Memory targets four different services to improve the quality of the memory supplied to LLMs, including storage, data retrieval, ranking and encryption. “In some metrics we had 60% improvements by having better ranking, better filtering and context,” he explained. “You're classifying the data differently, and by encrypting the data, and then doing some filtering over the data, this gives us by far better results,” he said, adding that, “We're not just a storage layer anymore.”

Get started with Walrus Memory at walrus.xyz/memory.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。