Agent Memory Needs a Shared Standard

Every agent tool stores memory differently, and none can read the others. A typical developer machine today might hold an AGENTS.md, Cursor .mdc rules, Anthropic's memory tool writing to a /memories directory, and a JSONL knowledge graph via an MCP server—four formats that cannot exchange data. Thines, who builds LoreKit (an open-source memory tool), draws a direct parallel to the pre-2019 observability mess, when every vendor had its own agent, instrumentation library, and format. That fragmentation ended when OpenTracing and OpenCensus merged into OpenTelemetry, agreeing on a shared data model, semantic conventions, a wire protocol (OTLP), and reference SDKs. Vendors kept competing on storage and analysis; users stopped re-instrumenting their code for every switch. Thines argues agent memory needs the same four-layer agreement: a shared data model, common naming conventions, one transport protocol, and reference implementations. He points out that Reddit communities regularly surface "I built a memory layer" posts—each with its own schema—as evidence the problem is real but the solution keeps fragmenting. The timing may be right: MCP and AGENTS.md both sit under the Linux Foundation's Agentic AI Foundation, offering a neutral governance home. Thines is explicit that LoreKit's own format is "one proposal among several, with gaps of its own."