Agent memory needs its OpenTelemetry moment — LoreKit blog

Why it matters

The article surveys the fragmented state of agent memory: instruction files like AGENTS.md or CLAUDE.md, Anthropic's memory tool writing to a memories directory, Copilot storing repo facts on GitHub's side, and MCP servers keeping JSONL knowledge graphs. None of these formats can read each other, so a lesson one agent records stays trapped in that tool. The author, who builds the open-source LoreKit, argues the obstacle is not a missing tool but a missing shared agreement, and points to observability as precedent: before OpenTelemetry unified data models, semantic conventions, and a protocol, every vendor shipped incompatible instrumentation. Agent memory is at that same early stage, and the timing is favorable because the formats are young and MCP and AGENTS.md already sit under a neutral standards home. Existing proposals, including Portable Memory, the Open Memory Protocol, and COGX, mostly solve export between vendors rather than live addressing among tools working in the same repository. The author sketches a record format covering where a memory is true, its identity, its content, and when it stops being true.