How the config rule works (and how to control it)

If you connect the Memory MCP server and then notice your agent’s rules file changed, this page explains exactly what happened, why, and how you stay in control. Nothing here is hidden, and everything is reversible.

What it does

On startup, the server writes one short priority rule into your AI agent’s always-loaded rules file - the file your agent reads on every session, such as CLAUDE.md or AGENTS.md. The rule simply tells the agent to reach for Ragionex Memory first when you ask it to remember or recall something.

That is the entire change: one small, clearly-marked block of guidance, in the one file the agent always reads.

Why a rules file, and not just the MCP protocol

MCP tools alone are not reliably visible to the agent at the right moment. Across popular clients, a few common behaviors work against an always-on memory rule:

• Tools are lazy-loaded. A tool’s full description often stays hidden until the model goes looking for it - so the agent may not realize memory is available unless something already pointed it there.
• Tool descriptions are capped. Clients cap how much text a single tool description can carry, so guidance packed into the tool itself gets clipped.
• Server instructions get truncated. Once a few MCP servers are installed, client-level server instructions are trimmed to fit a budget.

The net effect: you cannot count on the MCP protocol alone to keep an always-on memory rule in front of your agent. The always-loaded rules file is the one place the guidance is read every session - never hidden, clipped, or truncated.

How you stay in control

The change is designed to be observable, safe, and reversible. You are never locked in.

• It is clearly marked. The block is wrapped in stable begin/end markers so you can see exactly where it starts and ends, and tell it apart from your own content.
• Your file is backed up. Before the block is first written into an existing file, that file is backed up alongside the original - so the prior state is always recoverable.
• Your content is never touched. Only the bytes between the markers are managed. Everything else in the file - your own rules, notes, and formatting - is left exactly as it was.
• You can freeze it. If you edit either marker, the server detects that and stops touching the file - your opt-out is respected, and it warns rather than silently overwriting.
• You can remove it entirely. Run the command below and the block is stripped from every agent file (with a backup taken first).

npx @ragionex/memory-mcp init --remove