The developer of open-source agent-memory tool LoreKit is urging the AI tooling community to agree on a common data format that would let different assistants read and exchange persistent memories. The proposal compares the fragmented state of agent memory with software observability before OpenTelemetry established a shared model.
In a technical essay published by LoreKit, the author argues that coding assistants, continuous-integration agents and review bots increasingly preserve instructions or learned facts in incompatible structures. A lesson recorded by one tool can therefore remain unavailable to another tool working on the same repository. The author presents this as an interoperability problem rather than a shortage of memory products.
The proposed direction is deliberately narrower than a complete memory platform. A shared record would identify where a memory applies, give it a stable identity, contain its substantive text, describe its lifecycle and state who or what wrote it. The essay says the last field matters because records can originate with a person, be written by a model or be extracted from a conversation by a separate service. That provenance could help a receiving tool decide how much confidence to place in the material.
LoreKit's author points to existing approaches rather than claiming the project already supplies a universal answer. The article cites Portable Memory, the Open Memory Protocol and Cognee's COGX format as efforts focused largely on exporting or transferring memory. It also highlights ideas from product-specific systems, including citations, automatic forgetting, compact key-value interfaces and explicit treatment of time. Details are described as a late-2026 snapshot because the field is changing quickly.
Under the suggested model, vendors would still compete on storage, retrieval, embeddings, ranking and interfaces. The common layer would cover the shape and meaning of the record, much as OpenTelemetry lets observability providers consume shared instrumentation data while differentiating their analysis products. The essay also distinguishes a record format from the Model Context Protocol, arguing that a standard way to call a tool does not by itself define the data exchanged.
Security remains unresolved. Shared memory can become a route for prompt injection if one tool records malicious or unreliable instructions that another later trusts. The proposal says records should expose their author, support signing and, where useful, point back to the code on which a claim depends. It acknowledges that a format cannot eliminate poisoning.
The post is an invitation to collaborate, not an adopted industry specification. LoreKit's own schema is offered as one input, and the author says any eventual standard should have a neutral home rather than belong to a single vendor. Whether developers of existing memory systems converge on that approach will determine if the proposal advances beyond a discussion document.
