Agent memory tools each use their own formats and cannot read one another, creating the same fragmentation observability once faced. The piece argues agent memory needs an OpenTelemetry-style shared data model, surveying existing formats and proposals. It also outlines what such a common spec could look like.