Universal Memory Protocol launches as open standard for AI agent memory

The Universal Memory Protocol (UMP) has been introduced as an open standard designed to standardise how artificial intelligence agents store and retrieve memory across different sessions, platforms, and vendors. The initiative addresses a persistent fragmentation in the current AI landscape, where memory systems are typically reinvented privately within individual software harnesses. By providing a portable format for agent memory, UMP seeks to eliminate the need for developers to start from scratch when integrating new agents or storage systems.

The protocol builds upon two existing industry standards: the Model Context Protocol (MCP), which standardises tool access, and Agent2Agent (A2A), which governs inter-agent communication. UMP fills the gap by standardising how agents remember information. It operates as an application-level memory protocol that utilises signed JSON with bi-temporal tracking to manage data validity and provenance. The specification leverages W3C PROV and Decentralised Identifiers (DID) for identity management, requiring no new vocabulary beyond existing frameworks.

At its core, UMP defines a minimal set of operations for memory management, including recall, remember, revise, forget, and get. The system is designed to be lightweight, with a conforming client requiring approximately 100 lines of code. Implementations are available via an MCP server for immediate agent integration, a TypeScript SDK for building memory-aware applications, or HTTP for use in Python, Go, Swift, or browser-based environments. The protocol does not bundle native database drivers; instead, it provides wrappers for vector engines such as Qdrant, Pinecone, and Weaviate, as well as stores for SQL databases and Redis.

Interoperability is a central tenet of the specification, allowing for the import of existing memory formats from sources like Obsidian and Markdown files. Import bridges translate these formats into portable UMP drafts, preserving source paths in the provenance metadata. This approach ensures that teams can adopt the standard without vendor lock-in, keeping existing formats separate while establishing a common record and runtime interface. The system also includes a specific implementation called 'Recall', which offers a richer memory engine behind the same UMP interface without creating a dependency.

Security and data integrity are enforced through a verify, filter, and frame rehydration pipeline, ensuring that memory is never string-interpolated directly into prompts. The protocol separates structure, provenance, access, and trust from intelligence, leaving tasks such as extraction, ranking, decay, and consolidation to the underlying engine. This design allows implementations to compete on quality while maintaining interoperability, with four conformance levels enabling users to ship simple exports or wire full runtimes as their needs evolve.