The AI agent ecosystem is currently drowning in a sea of incompatible memory formats, a fragmentation that mirrors the chaos of pre-2019 observability tools. A recent analysis by LoreKit argues that agent memory has reached a critical inflection point where a shared standard, akin to OpenTelemetry, is no longer optional but essential for interoperability. Without a unified data model, lessons learned by one agent remain siloed, inaccessible to others, effectively resetting the intelligence of the system every time a new tool is introduced.

The Fragmentation Problem

Current setups often involve four or more distinct memory shapes, ranging from static instruction files like AGENTS.md to dynamic JSONL knowledge graphs. None of these formats can read the others, creating isolated silos where a CI agent’s learned lesson is invisible to an editor agent or a review bot. This lack of interoperability forces developers to rebuild the same basic concept—a note stored at an address with a lifespan—repeatedly across different vendors, wasting engineering cycles on reinventing the wheel rather than advancing agent capability.

Lessons from Observability

The industry has seen this movie before. Before 2019, every observability vendor shipped proprietary agents and data formats, making vendor switching a painful re-instrumentation process. The merger of OpenTracing and OpenCensus into OpenTelemetry solved this by standardizing the data model, semantic conventions, and protocol, while allowing vendors to compete on storage and analysis. Agent memory currently lacks this foundational agreement; we have plenty of implementations but no agreed-upon schema for what a memory record actually is, where it lives, or when it expires.

A Proposal for Standardization

LoreKit suggests a four-layer approach mirroring OpenTelemetry: a core data model defining memory records, semantic conventions for attributes, a protocol for transmission, and reference implementations. Key elements include standardized address levels (global, project, repo) and lifecycle rules, such as whether forgetting is default or opt-in. The proposal emphasizes that the spec should remain neutral, likely under the Linux Foundation’s Agentic AI Foundation, to avoid vendor lock-in while enabling tools to compete on how they process and rank memory.

Key Takeaways

  • Agent memory formats are currently fragmented across at least four major types, including instruction files, hosted facts, and knowledge graphs.
  • OpenTelemetry succeeded by standardizing the data model and protocol, leaving storage and analysis to vendor competition.
  • A proposed agent memory standard would define address levels, lifecycle rules, and provenance to enable cross-tool interoperability.
  • Existing proposals like Portable Memory and the Open Memory Protocol focus on export, but live addressing remains an unsolved problem.

The Bottom Line

Stop building proprietary memory silos; the ecosystem needs a boring, shared schema for agent memory now. Without it, we are just creating more walled gardens that prevent true agent collaboration.