Open Memory Protocol Initiative

A multi-stakeholder ecosystem for portable AI memory

OMPI incubates open, interoperable specifications for portable AI memory. It is housed at the AI Disclosures Project, a project of Code for Science & Society, and runs as a light-weight working group with in-person convenings and monthly digital meetings.

The goal is decentralized, competitive AI markets with specialized producers and modular technological pieces.

Specifications are model-, harness-, and application-agnostic. Multi-user memory, provenance, and privacy are first-order design considerations. In AI systems, memory is part of the context layer that enables personalization and agentic action — user-provided or model-learned, scoped to a project, user, or organization.

Artifacts and specs

Specs

Other

Memory is becoming a core agent layer without shared exchange semantics

Agent memory is persistent context that shapes future model or agent behavior — user facts, derived summaries, project instructions, prior decisions and actions.

Every major coding-agent harness (Claude Code, Codex, goose, OpenClaw, Letta Code) implements memory with a different convention. Developers cannot move a stateful agent across harnesses. Enterprises cannot switch memory providers without re-ingesting derived objects. Open-source memory projects duplicate one another for lack of a shared object model.

Fragmentation has engineering and security costs: bespoke adapters, lost provenance on copy, permissions that don't survive export, harder security review. A common open layer addresses this without dictating how each agent reasons, summarizes, indexes, or retrieves.

Partners and implementers

From ecosystem launch to sustainable stewardship