Memory Pod Fabric (MPF) is an open protocol specification that defines standardized methods for storing, retrieving, and securing semantic memories across AI applications, agents, and organizations.
As AI systems become more sophisticated and persistent, the need for standardized, verifiable, and securely shareable memory becomes critical. MPF addresses this challenge by providing a standardized protocol layer that can be implemented across diverse environments while maintaining consistent capabilities.
MPF serves as the essential third pillar in the open agent protocol stack:
- A2A lets agents talk to each other
- MCP lets models use tools
- MPF lets everyone remember
- Key Features
- Protocol Specification
- Quick Start
- Core Concepts
- Implementation Guidance
- Reference Implementations
- Contributing
- Governance
- License
- Security
- Community
- Standardized API for memory storage and retrieval
- Capability-based security with fine-grained permissions
- Cryptographic verification for regulatory compliance
- Multi-model vector support for future-proof embeddings
- Designed for interoperability with existing AI protocols
The complete protocol specification is available in the SPECIFICATION.md file. This specification defines:
- Core memory storage and retrieval operations
- Memory discovery and capability advertising
- Capability-based authorization model
- Cryptographic verification mechanisms
- Integration patterns with existing AI protocols
- Reference data structures and formats
To implement MPF in your project:
- Review the SPECIFICATION.md to understand the protocol
- Explore the examples directory for implementation samples (coming soon)
- Choose a reference implementation or create your own following the specification
- Implement the core endpoints required for MPF compliance
A Memory Pod is a logical boundary of memory with its own governance, access controls, and audit trail. A pod represents a collection of memory objects that share common access patterns and security requirements.
A Memory Object is the fundamental unit of storage in MPF, containing vector embeddings, metadata, content, and access control information.
MPF uses signed JWT tokens that implement the object-capability security model for fine-grained access control with specific, time-limited permissions.
MPF uses cryptographic audit trails through Merkle trees, providing tamper-evident verification for regulatory compliance and trust.
MPF can be implemented on various storage backends:
- Postgres + pgVector: Recommended for self-hosted deployments
- Snowflake VECTOR: Ideal for enterprise deployments
- Specialized vector DBs: Pinecone, Weaviate, Milvus, etc.
- Multi-tier storage: Hot/warm/cold storage separation
For detailed implementation guidance, see the implementation guide.
Official reference implementations will be available in multiple languages:
We welcome contributions to the MPF specification and reference implementations. Please see our CONTRIBUTING.md for guidelines on how to participate.
The MPF protocol is governed by a community working group that includes representatives from various organizations. For more information on governance and the decision-making process, see GOVERNANCE.md.
This specification is licensed under the Apache License 2.0.
The Apache 2.0 license provides patent protections that make it ideal for an open protocol:
- Explicit patent grant from contributors
- Defensive termination provisions
- Compatible with commercial use
- Industry standard for open protocols
For information about the security model and best practices, see SECURITY.md. For reporting security vulnerabilities, please follow the process outlined in the security document.
- Twitter/X: Follow @amotivv for updates
- Meetings: Regular community working group meetings (see GOVERNANCE.md)
Memory Pod Fabric: The Memory Protocol for AI Agents
amotivv, inc