An open protocol, still pre-1.0, for moving and sharing AI-agent memory between agent systems. It defines a memory object, an exchange format and a runtime representation, so accumulated context, along with its provenance and permissions, can move between harnesses and memory providers. It is hosted at the AI Disclosures Project and co-convened with IBM and Mozilla.
Open Memory Protocol
The Open Memory Protocol is an attempt to make agent memory portable. Every coding-agent harness
keeps memory now, and each does it its own way: AGENTS.md, MEMORY.md, USER.md, HUMAN.md. The
initiative’s argument is simple. You cannot move a stateful agent from one harness to another, and an
enterprise cannot switch memory providers without re-ingesting everything. Provenance and permissions
get lost on every copy. OMP aims to be the smallest shared layer that fixes that, while each
implementer keeps its own storage and architecture.
- Memory object - The shared unit: a user-provided fact, a derived summary, an instruction, a prior decision, a tool output, or a reference to a transcript. It is scoped to a user, project, organization or group, and it is distinct from a raw chat export.
- Exchange format - How memory moves between systems, with provenance intact.
- Runtime representation - How a receiving system holds and uses imported memory.
- Security and threat model - Adversarial content, source trust, and stopping forgotten memories from coming back through import. This is open for review as PR #17.
Status, as of November 2026. It is pre-1.0, and no reference implementation exists. The current text is the exchange-and-runtime editors’ draft, v0.5, merged on November 6, 2026. It draws on earlier proposals from Cognee, IBM, Sapience, ContextNest and the AI Disclosures Project, which are kept in the repository as early drafts, along with a prior-art survey and Python prototypes. Code is Apache-2.0 and documentation is CC BY 4.0. Working-group discussion happens in a Discord where membership is by application.
It sits next to Agent Skills and MCP in the agent stack. Those describe what an agent can do. OMP is about what an agent remembers, and whether that memory can leave the system it was formed in.