Every AI coding tool remembers information differently. Cursor stores rules in .mdc files. Anthropic’s memory tool writes to /memories directories. GitHub Copilot keeps facts on its servers. A memory MCP server might maintain a JSONL knowledge graph. None of these formats can read the others.
This fragmentation mirrors a problem observability solved. Before 2019, every monitoring vendor shipped its own agent, instrumentation library and data format. Switching vendors meant rewriting everything. Two projects—OpenTracing and OpenCensus—competed to fix it, then merged into OpenTelemetry in 2019.
OpenTelemetry succeeded by establishing four things the vendors agreed to stop competing on: a data model defining what a span or log record is; semantic conventions with shared names like service.name; a protocol (OTLP) for sending data; and reference implementations. Vendors kept competing on storage, querying and analysis. Users instrumented once. One ecosystem of tools could read each other’s data.
According to the source, agent memory is now where observability was before 2019. The timing is favorable: the formats are young enough to change, and both MCP and AGENTS.md already sit under the Linux Foundation’s Agentic AI Foundation.
Today’s memory formats answer the same four questions differently: where a memory is true, what its identity is, what it contains, and when it stops being valid. A shared spec wouldn’t need to pick one answer. It needs to record which approach each system uses, so readers know how much to trust what they read.
The source proposes that a shared specification would include conventions for address levels (global, project, repo, branch), kinds of memory, and attribute namespaces; a contract with fixed operations like write, read, list and search over MCP; and conformance tests to verify records round-trip without data loss.
Several efforts already exist. MacPaw Research’s Portable Memory uses a .mem bundle with byte-reproducible serialization and optional signing. The Open Memory Protocol, an April 2026 draft, moves memories between platforms as a ZIP file. Cognee’s COGX format imports from existing systems.
These mostly solve export—moving memory between vendors. But coding tools face a second problem: live addressing. When a CI job, editor agent and review bot work in the same repository simultaneously, how do they read and write the same place and agree on meaning? A complete spec needs to address both.
Key facts
- AI tools currently store agent memory in four or more incompatible formats with no way to share data between them
- OpenTelemetry solved a similar problem in observability by establishing a shared data model, semantic conventions, protocol and reference implementations in 2019
- Agent memory formats differ on where memories are true, their identity, content and expiration—but there’s no standard way to record these differences
- A shared spec would need to cover conventions, a contract with fixed operations over MCP, and conformance tests, while leaving storage and ranking to individual tools
- Several proposals already exist including MacPaw’s Portable Memory and the Open Memory Protocol, but they mostly focus on export rather than live addressing across tools