← Back

Rumil

Personal AI Memory for macOS (Creator)

github.com/airferg/rumil

Overview

Rumil is a Mac app that gives every AI assistant I use the same memory. It keeps my context as linked Markdown files on my own computer and serves it to Claude, ChatGPT, and Perplexity over MCP (Model Context Protocol).

I designed it and built it with Claude Code in November 2026: I made the product and architecture decisions, and AI wrote most of the code.

Problem

Every assistant starts from zero. Claude remembers what I tell Claude, ChatGPT remembers what I tell ChatGPT, and none of them see my email, texts, calendar, notes, or the meetings where most of my context actually lives. I was re-explaining myself in every chat.

What It Does

• Pulls context in automatically: Gmail and Calendar, iMessage, Apple Notes, Obsidian, calendar links (Canvas, Outlook), news feeds, iPhone locations, voice memos, and meeting recordings transcribed on-device
• Copies my full ChatGPT, Claude, and Perplexity history through a Chrome extension, including chats from my phone

• Adds relevant context to prompts in the browser, only when the prompt is about me

• Shows everything as a rotatable globe of linked notes, people, and projects

• Runs a local model (Ollama) for anything too private to send to a cloud AI

Key Decisions

• Token cost over completeness: the first version made assistants search several times per question and spent about 15K tokens on one answer. I redesigned it around a router that resolves names and returns only the relevant lines, cutting per-session overhead from about 1,900 to 600 tokens
• Security after a real finding: a test showed API keys from one of my notes coming back in a search. I added secret redaction across search and reads, a key-rotation flow, and locked the internet tunnel to the memory tools only

• Local first: notes stay as files on my Mac, private folders can be hidden from every assistant, and a local model handles private summaries

Stack

Electron, Node.js, MCP SDK, MiniSearch, SQLite (iMessage), whisper.cpp, Ollama, Chrome extension (Manifest V3), ngrok. Nearly 300 automated tests.

What I'm Learning

• Memory is a retrieval problem, not a storage problem: having everything is worthless if the assistant has to dig
• Every tool an AI can call costs tokens before it is ever used, so fewer, smarter tools win

• The assistant decides when to look things up, so the rules for when to check memory matter as much as the memory itself