Mysten Labs launches Walrus Memory for portable AI agent memory

Mysten Labs has launched Walrus Memory, an encrypted, portable memory layer for AI agents that preserves context across apps, sessions and cloud providers. The platform integrates with Claude, ChatGPT and Gemini and offers plugins plus Python and TypeScript SDKs to add shared memory to agent workflows.

Walrus Memory is built to let multiple agents, applications and workflows share memory without tying them to a single runtime, session or provider. The product combines encrypted storage with retrieval and ranking tools and implements policy-based access controls so teams can set who can access data and how long it remains available. Shared memory spaces are intended to support coordination across long-running tasks and multiple sessions.

Kostas Chalkias, co-founder and chief cryptographer at Mysten Labs, described agentic memory as a limiting factor for advanced AI workflows: “The major misconception in AI is that compute is the only bottleneck. The major issue is we’re using a lot of memory as humans, and we want our LLMs to actually learn about us.” He said developers have often had to stitch together databases, vector stores and runtime state, producing systems that lose context during complex processes.

The platform incorporates cryptographic tools, including zero-knowledge proofs, to let agents verify contextual claims without exposing underlying data. Developers can program access rules on encrypted memory to control retention and sharing policies and apply filtering and ranking before returning context to a model.

Mysten offers plugins for OpenClaw and NemoClaw and SDKs to simplify integration. Early customers and partners building on Walrus Memory include Allium, Conso Labs, Inflectiv, OpenGradient, Talus Labs and Tatum. The company cited use cases such as portable agent identity systems and AI assistants that retain customer interactions across sessions.

Mysten reported improvements in memory quality after combining ranking and filtering techniques with encryption, with some internal metrics showing up to a 60% increase in relevant context returned to models. Chalkias attributed the gains to improved classification, filtered retrieval and encrypted processing.

Walrus Memory is part of the broader Walrus storage effort from the Sui developer community and the Walrus Foundation. The foundation marked its one-year anniversary on March 27 and reported about 450TB of stored data. The Walrus Foundation previously raised $140 million in private funding led by Standard Crypto with participation from a16z, Electric Capital and Franklin Templeton Digital Assets. Mysten positions Walrus Memory within that ecosystem while focusing on the needs of agent workflows rather than general-purpose decentralized storage.

The product supports multiple LLM providers to limit model lock-in and gives teams programmable control over how long data remains accessible. Chalkias emphasized transparency and user control, adding, “You don’t want your data to be there forever, you don’t want your data to be misused.”