TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A LoreKit report argues that AI coding tools store memories in incompatible formats, leaving knowledge isolated between editors, CI agents and other systems. It proposes community agreement on shared ways to describe and address memories, while acknowledging existing portability proposals and that no standard has been adopted.
LoreKit has published a report arguing that AI agents need a shared memory format, as coding assistants and memory services store information in systems that often cannot exchange it. The proposal matters to teams using several agents on the same codebase: knowledge recorded by one tool may not be available to an editor, CI agent or review bot using another.
The report’s author describes a typical coding setup with instructions in AGENTS.md or CLAUDE.md, Cursor rule files, agent-managed memory files, repository facts held by GitHub Copilot, and a memory MCP server with a knowledge graph. These are examples from the report, not an independent survey of all users. The author’s core point is that each format has its own way to represent and locate information, so a lesson saved in one system may remain inaccessible to another.
LoreKit’s author says the problem is not simply a shortage of memory products. Separate tools keep emerging with their own storage, schemas and MCP interfaces, while users still lack an agreement underneath them. The report presents LoreKit’s own format as one proposal among several, and explicitly says it has gaps; it does not claim LoreKit has solved interoperability.
The proposed area for agreement includes how a memory is scoped, identified, represented and retired when it is no longer valid. The report distinguishes export—moving someone’s data between services—from live addressing, which would let several agents read and write shared information at the same time. It says a standard would need to account for both problems, particularly in a repository where an editor, CI job and review agent may operate concurrently.
If agents can use a common memory model, teams may be able to reuse project knowledge across tools rather than rebuilding it in each product. That could reduce duplicated setup and help an agent working in CI benefit from a finding recorded by an editor agent. These are potential benefits of interoperability, not outcomes measured by the report.
A common format would also shift where products compete. The report draws on OpenTelemetry as an analogy: observability tools agreed on shared data conventions and a transport protocol, while continuing to compete on storage and analysis. For agent memory, a comparable agreement could let vendors differentiate on how they search, manage or apply memories instead of making each system’s records inaccessible elsewhere.
There are practical questions as well as technical ones. Shared records need clear identity rules so an update does not accidentally create a duplicate, and lifecycle rules so outdated information is not treated as current. A format used by multiple tools would also need to handle concurrent changes and preserve enough context for users to understand where a memory came from.
Products Worth Considering
Logitech MX Master 4 for Mac Wireless Ergonomic Mouse, White Silver
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Twilight Blue
Jackery Explorer 300 Portable Power Station with SolarSaga 100W Solar Panel
As an affiliate, we earn on qualifying purchases.
From Observability to Agent Memory
The report compares the current state of agent memory with observability before 2019. At that time, it says, vendors commonly had their own agents, instrumentation libraries and data formats. OpenTracing and OpenCensus attempted to address fragmentation before merging into OpenTelemetry in 2019. The report identifies a shared data model, semantic conventions, the OTLP protocol and common software implementations as parts of that effort. The comparison is the author’s argument for a similar process; it does not establish that agent memory will follow the same path.
Several memory-related formats and projects already exist. The report describes MacPaw Research’s Portable Memory as a .mem bundle with reproducible serialization, deletion tombstones and optional signing, while saying its authors report limited adoption. It describes the Open Memory Protocol as an April 2026 draft focused on transferring conversations and derived memories in a ZIP bundle, and Cognee’s COGX format as supporting imports from Mem0, Letta and Zep. These examples address aspects of portability, but the report says shared, simultaneous access across tools remains a separate challenge.
The report also points to the Linux Foundation’s Agentic AI Foundation as a possible neutral venue, noting that MCP and AGENTS.md sit under its umbrella. That offers a potential setting for discussion; it is not evidence that the foundation has adopted or is developing the specific memory standard the author advocates.
“The problem isn’t a missing tool. It’s a missing agreement.”
— LoreKit report author
No Common Format Yet
The report advocates a standard but does not announce an agreement, governing body, implementation schedule or adoption commitment from the projects it discusses. It is not clear which group would lead the work, how competing proposals would be reconciled, or whether vendors would adopt a shared format.
The report’s format survey is presented as a snapshot from late 2026, and the cited tools may change. It also does not provide independent measurements of how often memories fail to transfer, how much work teams duplicate, or whether a shared format would improve agent performance. The author’s description of community posts as evidence of demand is an observation, not a quantified market study.
Important design questions remain open, including how a shared system would handle permissions, conflicting updates, provenance, expiration and removal. Exporting a record between services is not the same as enabling several agents to safely share and update it in real time.
Standards Discussion and Adoption
The next step proposed by the report is community discussion around a shared model for memory records, addressing and lifecycle—not immediate adoption of LoreKit’s implementation. The author points to the Agentic AI Foundation as a potential neutral home for standards work, but reports no formal process or timetable there.
Any proposal would need to clarify both portable export and live shared access, then define how records are identified, updated, sourced and expired. Evidence of progress would include a published draft, participation from multiple tool makers, reference implementations and examples showing that different agents can read and update the same records. Until those appear, interoperability remains a stated need rather than an established capability.
Key Questions
What is the news in the LoreKit report?
LoreKit has called for a shared agent-memory format so coding agents using different tools can exchange project knowledge. The report is a proposal, not an announcement that a standard has been adopted.
The report says tools store memories in different forms, such as instruction files, product-specific repository facts, files and knowledge graphs. Without shared conventions for records and addresses, information saved by one tool may not be readable by another.
How is this related to OpenTelemetry?
The report uses OpenTelemetry as an analogy for how shared data models and conventions can let different vendors work with the same information. It argues that agent memory could benefit from a similar agreement, but does not claim the two fields will develop in the same way.
Has LoreKit proposed the standard?
No. The author describes LoreKit’s format as one proposal among several and says it has gaps. The report calls for community agreement rather than adoption of LoreKit alone.
What remains unresolved?
No lead organization, shared specification, adoption timeline or common implementation is confirmed. The report also leaves open how a standard would manage permissions, simultaneous updates, memory sources and expiration.
Source: lorekit.io
