AI memory proposal targets data model for coding agents
LoreKit’s author says a shared format should support memory export and live access while leaving storage and other tool features open to competition.
A LoreKit blog post is calling for AI coding tool builders to agree on a shared format for agent memory, arguing that incompatible systems make it difficult for one tool to use information recorded by another. The author compares the challenge with observability data before the formation of OpenTelemetry.
The proposed specification would cover both exporting memories between tools and allowing multiple agents to read and write shared memory. It could define where a memory applies, its identity and content, how long it remains valid, and who or what wrote it.
The author points to Copilot Memory, LangGraph, Graphiti and AGENTS.md as examples with features that could inform a shared approach. Existing proposals including Portable Memory, Open Memory Protocol and Cognee’s COGX are described as focused mainly on moving memories between platforms.
Storage, ranking, embeddings and user interfaces should remain outside the specification, the author argues, leaving those areas open to competition. The post also says a shared format should record authorship and support ways for readers to assess a memory’s reliability.
The author discloses building LoreKit and presents its format as one possible contribution, rather than a template. The post reports no agreed specification, standards body or development timetable; its descriptions of the fast-changing field are framed as a snapshot from late 2026.
