Walrus Memory

Product Introduction

1. Definition: Walrus Memory is a portable, persistent memory layer and SDK designed specifically for AI agents. It is a cloud-based service built on the Walrus Verifiable Data Platform, providing a dedicated infrastructure for agents to store, retrieve, and manage contextual information across disparate sessions, applications, and runtimes.
2. Core Value Proposition: It solves the fundamental problem of AI agent memory loss, enabling reliable, multi-session workflows by providing portable, verifiable, and user-controlled context. It exists to be the foundational memory infrastructure for the next generation of autonomous and collaborative AI systems.

Main Features

1. Portable Context & Cross-Session Persistence: Memories stored via the Walrus Memory SDK are decoupled from any single session or application. An AI agent can remember a fact in one app and recall it weeks later in a completely different environment (e.g., from a chat interface to a backend workflow) without rebuilding context. This is achieved by storing all memories as encrypted objects on the Walrus decentralized network, indexed for fast semantic search.
2. Verifiable Data & Cryptographic Control: Every memory operation (remember, recall) requires a signed request from an Ed25519 key pair controlled by the developer. Ownership and delegate permissions are programmable and cryptographically enforced. This ensures immutable audit trails, proof of data integrity, and explicit, granular access control without relying on a central authority.
3. Native Agent Integration & MCP Support: The memory layer is designed for seamless integration with the modern AI stack. It provides native Model Context Protocol (MCP) support without requiring adapters, first-party plugins for platforms like OpenClaw and NemoClaw, and official SDKs for Python, TypeScript, and JavaScript. This allows developers to wire persistent memory into agents with just two core calls: remember and recall.
4. Secure, Encrypted Storage with Semantic Retrieval: All data is encrypted before being stored on the Walrus network. Upon a valid recall request, the memory layer processes the query using embedding models, retrieves the most relevant memories via semantic search, decrypts them, and injects them directly into the model context for the agent's next step.

Problems Solved

1. Pain Point: Context loss between sessions and applications, leading to fragmented workflows, repeated processing, conflicting decisions between agents, and unreliable production outputs. Agents cannot learn from or coordinate on shared history.
2. Target Audience: AI/ML Engineers, Backend Developers, and Technical Founders building autonomous agents, multi-agent systems, customer support bots, research assistants, or any AI-driven application that requires reliable statefulness. It targets teams using LLMs from providers like OpenAI, Anthropic, and Gemini.
3. Use Cases:
   • Multi-Agent Workflows: Enabling multiple agents in a complex pipeline to share context, coordinate state, and avoid contradictory actions.
   • Customer Support Systems: Allowing support agents to pick up exactly where a previous interaction left off, with full knowledge of past inquiries, solutions, and customer sentiment.
   • Persistent Personal Assistants: Creating assistants that remember user preferences, routines, relationships, and tone over long periods, becoming more personalized over time.
   • Research & Knowledge Management: Serving as a "second brain" for AI agents that continuously captures ideas and retrieves them naturally in future sessions.

Unique Advantages

1. Differentiation: Unlike general-purpose databases (Redis, Postgres) or raw vector databases, Walrus Memory is a fully managed, opinionated memory layer. It abstracts away the complex infrastructure of embeddings, access control, encryption, and decentralized storage. It provides a turnkey solution focused on the unique requirements of AI agents (portability, verifiability), which developers would otherwise need to build and maintain themselves.
2. Key Innovation: The core innovation is combining verifiable data (via cryptographic signatures and the Walrus platform) with a portable memory API. This creates a memory system where context is not siloed, is provably unaltered, and is controlled entirely by the user via programmable keys—shifting memory from an application-centric to an agent-centric model.

Frequently Asked Questions (FAQ)

1. How does Walrus Memory help AI agents remember across different apps and sessions? Walrus Memory decouples an agent's memory from its runtime environment. Memories are stored as encrypted, indexed objects on the Walrus network. An agent uses a cryptographic key pair to access the same memory from any application (e.g., a chat app, a Python backend, a TypeScript workflow) with the correct SDK installation and permissions, enabling seamless context handoff.
2. How is Walrus Memory different from using a standard vector database like Pinecone or a database like Postgres? A standard vector database or Postgres provides storage primitives; developers must separately handle encryption, access control, identity management, and data portability. Walrus Memory is a dedicated memory layer that delivers these features out-of-the-box, with built-in semantic search, cryptographic ownership verification, and a focus on cross-platform portability specifically for AI agent use cases.
3. Who owns and controls the data stored in Walrus Memory? The developer who holds the signing private key has sole ownership and control. Each memory is tied to that key, and permissions are programmable. No third party, including the platform, can access the data without the key's signature. This provides explicit, verifiable ownership and data sovereignty.
4. What security measures protect the memories stored? All memories are encrypted before they are stored on the Walrus decentralized network. They can only be decrypted by an authorized SDK using a signed request. The integrity of the data is independently verifiable through cryptographic means, ensuring memories cannot be tampered with.