When your agents disagree, who decides? Mirsad — shared memory for AI agents
Hey everyone — I'm Koussay, an 18-year-old high-school student from Tunisia. I build open-source tools for AI agents, and I joined because this looks like the room where people shipping MCP servers actually compare notes.
What I shipped: Mirsad, a shared memory server for AI agents (MCP, 13 tools, on PyPI as mirsad-mcp).
The problem: when Claude, ChatGPT and other agents work on the same project, each keeps its own notes. When they disagree, there's no record of who decided what — or why.
Mirsad stores facts as bi-temporal records (what was believed, and when it changed), treats decisions as first-class objects with their reasons attached, and keeps an episode log as provenance for everything. When two agents write conflicting facts, the conflict goes into a queue for the owner to resolve — the system never silently overwrites, and never votes. Deliberately no embeddings: recall is keyword search plus graph structure, because I want every answer to be explainable.
It's local-first (SQLite), also runs on Cloudflare Workers, it's listed in the official MCP Registry, and I'm hardening it toward a 1.0 release in the open.
Question for the room: what are you using today when several agents need the same memory — and where does it break first?
GitHub: github.com/amkkoussay/Mersad







Comments
0 comments