Agent memory needs its OpenTelemetry moment

The cost to standardize agent memory formats falls to the AI tool makers, not users.

The proliferation of distinct agent memory formats presents an interoperability challenge for AI agents, mirroring an earlier issue in the observability sector. Currently, each AI agent stores its memory in a unique format, rendering memories created by one agent unreadable by others. This lack of a universal standard impedes the sharing of learned information across different tools and workflows. For example, a CI agent's learned lessons are inaccessible to an editor agent or any subsequent tool. This situation is analogous to the pre-2019 observability landscape, where vendors used proprietary formats, requiring re-instrumentation for vendor switches. The emergence of OpenTelemetry as a unified standard for observability data, encompassing a shared data model, semantic conventions, and protocols, offers a precedent for agent memory. Initiatives are underway, such as proposals for memory data models and export formats, but a comprehensive standard encompassing both data structure and live addressing for concurrent access is still needed.

The current fragmentation of agent memory formats is hindering the development of a cohesive AI agent ecosystem, similar to the challenges faced by the observability industry before the adoption of OpenTelemetry. Without a standardized data model and conventions, developers are forced to build custom integrations for each tool, slowing innovation and increasing complexity. The Linux Foundation's Agentic AI Foundation provides a neutral venue for establishing such standards, analogous to the role the foundation played in unifying observability efforts. The key will be to establish shared conventions for memory attributes, operations, and data serialization, allowing different agents to read and write to shared memory spaces without loss of fidelity.

The path forward for agent memory standardization involves creating a data model that captures essential elements like addressability, identity, content, and lifecycle, along with a protocol for data exchange and clear semantic conventions. This will enable tools to compete on value-added features such as storage, querying, and analysis, rather than on proprietary memory formats. By defining clear boundaries for what the standard covers—and what it does not—the community can foster interoperability and accelerate the adoption of more capable and interconnected AI agents, much like OpenTelemetry did for observability.