Mysten Labs has launched Walrus Memory, a new product for AI agents. The company describes it as a portable memory layer designed to enable AI agents to retain context across different applications, sessions, and model providers, while keeping data control in the hands of users.
Maintain context across multiple devices
The company disclosed that many current AI agents still rely on developers to manually stitch together databases, vector stores, and runtime states, causing systems to lose context during complex tasks. Kostas Chalkias, co-founder of Mysten Labs, stated that AI’s primary bottleneck is not only computational power—memory capacity also limits agents’ ability to work continuously.
The design focus of Walrus Memory is to enable agents, applications, and workflows to share a unified memory, independent of any single runtime environment, session, or model provider. According to the company, this also allows multiple agents to collaborate effectively on long-term tasks.
Integrated with leading large models
The company states that Walrus Memory is now compatible with leading AI platforms such as Claude, ChatGPT, and Gemini, aiming to reduce users' dependence on a single model provider. Developers can also integrate this memory layer into existing agent workflows using the OpenClaw and NemoClaw plugins, as well as Python and TypeScript SDKs.
Teams such as Allium, Conso Labs, Inflectiv, OpenGradient, Talus Labs, and Tatum are currently building applications on Walrus Memory, including portable agent identity systems and AI assistants capable of remembering customer interactions across sessions.
Add encryption and access control
The company disclosed that this product incorporates programmable access control, allowing users to determine which models or agents can access the relevant data. Mysten Labs also noted that the system employs cryptographic tools, including zero-knowledge proofs, for context verification and managing access to encrypted memory.
In terms of performance, Chalkias stated that Walrus Memory currently focuses on optimizing the quality of memory provided to large models through four key stages: storage, retrieval, sorting, and encryption. According to him, on certain metrics, results can improve by approximately 60% after better sorting, filtering, and context processing.
Additional information: The original text includes the label “Brought to you by Walrus” and is part of a branded partnership. Statements regarding performance improvements and product advantages are primarily provided by the company.