Contexo: give your Cursor agent a shared, versioned memory of your context (open-source MCP)

Showcase Built for Cursor
,
1

Hey all — I built an open-source MCP server that gives Cursor a persistent, shared memory for your projects, and I’d love some feedback from people using Cursor day-to-day.

The problem that kept bothering me: every new agent session starts cold.

You end up re-explaining the architecture, the gotchas, and all the “we do it this way because X” decisions. Whatever the agent learns stays in that session, on that machine. Your teammate’s agent doesn’t get it, and often your own next session doesn’t either.

.cursorrules, docs, and memory features help, but they’re usually per-person, per-tool, and not really versioned.

So I built Contexo.

The idea is simple: treat AI context like code.

Contexo stores project knowledge in a dedicated git-backed repository that agents read and write through MCP. An agent can pull what’s already known before starting work and push back anything it learns afterward.

Because it’s git underneath, the context is versioned, diffable, and shareable. Because it’s MCP, it isn’t tied to Cursor—Claude Code, Codex, and Cursor can all share the same knowledge base.

A few things I think are interesting:

  • Every context update becomes a commit

  • Section-aware diffs instead of raw markdown line noise

  • Warnings when context has changed since the last pull

  • Agent-assisted 3-way merges when two agents update the same page

  • Captures the why, not just the what (decision, rationale, rejected alternatives)

Setup is basically:

curl -fsSL https://raw.githubusercontent.com/sugihAF/Contexo/main/scripts/install.sh | sh
ctx init
ctx login

Then add:

{
  "mcpServers": {
    "contexo": {
      "command": "ctx",
      "args": ["mcp"]
    }
  }
}

to Cursor’s MCP configuration.

Repo: GitHub - sugihAF/Contexo: Per-project, git-backed knowledge base for AI agents. Share research, decisions, and reasoning across your team's Claude / Cursor / Codex sessions via MCP. · GitHub
Site: https://contexo.live

Curious how people here are handling this today.

Are you relying on .cursorrules, Cursor memory, a wiki, docs, or just re-explaining things every session? And do you actually want context shared across teammates, or do you prefer keeping agent memory local to each developer?