Overview
- Mysten Labs has rolled out Walrus Memory, a portable memory layer tailored for AI agents that allows them to maintain context across various applications, sessions, and providers, while prioritizing user data control.
- Kostas Chalkias, the co-founder, claims that agentic memory represents the "true bottleneck" in AI development.
- The platform is compatible with major AI models such as Claude, ChatGPT, and Gemini, and includes plugins for OpenClaw and NemoClaw.
Kostas Chalkias, Co-Founder and Chief Cryptographer at Mysten Labs, who also contributed to Walrus, describes agentic memory as a reflection of human experience. Ideally, this memory should be transferable, enabling AI agents to manage and share context across different applications and sessions. However, current limitations in memory technology have hindered this capability.
AI developers often find themselves merging databases, vector repositories, and runtime states, resulting in inefficient systems that fail to handle intricate workflows, leading to agents that can easily forget information.
Chalkias pointed out that a significant misunderstanding in AI is the belief that computational power is the only limitation. He insists that humans utilize substantial memory, and it is essential for large language models (LLMs) to genuinely understand their users. Addressing the "real bottleneck" of agentic memory is crucial, he emphasized.
With the introduction of Walrus Memory, Mysten Labs aims to tackle these challenges by offering a memory layer specifically designed for AI agents that emphasizes portability, user authority, and coordination among agents.
Chalkias highlighted that Walrus Memory combines various essential features for AI agents: "Having rapid computation alone does not guarantee privacy; merely having encryption does not ensure that you can share your policies regarding the LLMs you choose," he explained. "Simply possessing large datasets is also insufficient."
Walrus Memory allows agents, applications, and workflows to share memory fluidly without being restricted to a specific runtime, session, or provider. Additionally, shared memory spaces facilitate coordination among multiple agents across extended workflows. Cryptographic mechanisms like zk-proofs enable agents to verify context and offer programmable access to encrypted memory.
Chalkias asserted, "I don't believe there is currently any other solution, particularly those focused on blockchain, that addresses all three aspects, which are significant barriers for most of them to function effectively."
Walrus Memory is designed to integrate smoothly with prominent AI platforms, including Claude, ChatGPT, and Gemini, ensuring users are not confined to a single model provider, thereby future-proofing their workflows.
Moreover, data stored within Walrus Memory features programmable access controls. Chalkias remarked, "It's not just about recall accuracy; transparency is key; you don’t want your data lingering indefinitely or being misused."
With plugins for OpenClaw and NemoClaw, alongside SDKs for Python and TypeScript, developers can effortlessly incorporate portable memory into their existing agent workflows. Teams such as Allium, Conso Labs, Inflectiv, OpenGradient, Talus Labs, and Tatum are already leveraging Walrus Memory to create applications like portable agent identity systems and AI assistants capable of remembering customer interactions across sessions.
Chalkias noted that memory management is continuously improving and that Walrus Memory focuses on four distinct services to enhance the quality of memory provided to LLMs, including storage, data retrieval, ranking, and encryption. "In some cases, we observed a 60% improvement due to enhanced ranking, filtering, and contextual analysis," he explained. "By classifying data differently and applying encryption, followed by filtering, we achieve significantly better outcomes," he added, noting, "We are no longer just a storage layer."
Explore Walrus Memory at walrus.xyz/memory.
Provided by Walrus
Discover More about collaborations with Decrypt.