Memory MCP

What it is

A memory server for Claude, backed by nothing more exotic than a GitHub repo full of markdown. It speaks the Model Context Protocol, so any MCP-capable client can read and write the vault: preferences, household details, travel context, notes on long-running projects. The same memory follows me from the phone to the desktop app to the web to Claude Code, because it lives in the repo and not in any one client’s storage.

Why I built it

Every assistant product has its own memory feature, and every one of them is a silo. Mine remembered things on my laptop that it had never heard of on my phone. Worse, none of them would tell me what they’d stored or let me correct it. Memory was a black box that occasionally produced a confidently wrong fact about my own life.

Markdown in git fixes all of that at once. I can read it, diff it, correct it in an editor, and roll it back when something gets written badly. It’s the least clever possible answer available. The durable thing should be plain text I own, and the assistant should be one more client of it.

How it works

The server exposes a small set of tools over MCP: read a file, read several at once, write, patch, list, tree, full-text search. All of them are implemented against the GitHub API. Writes commit straight to the repo, so every change to memory arrives as a commit with a diff and a timestamp, and it syncs down to my machine without any extra moving parts.

The decision that did most of the work was read_multiple. Loading context is almost always “give me these five files”, and collapsing that into one round trip noticeably changes how willing an assistant is to actually go and load its context.

Where it is now

Running as my own memory across every Claude client I use, and being tidied up for a public release. It’s built to be self-hosted: point it at a repo of your own, run it, and that’s the whole install.