Overview
AI coding agents lose everything between sessions, so the same context has to be rebuilt over and over.
Work
I built it as a TypeScript monorepo: an MCP server that exposes ten tools by default (an admin tool reaches the other fifty by name, so the schema stays small), a CLI, and hooks that capture findings and tasks into plain markdown. No database, no vendor lock-in. Findings have a lifecycle (supersede, retract, contradict) and trust that decays over time, and a fragment graph links concepts across projects. The store has seven front ends on top of the same files: a full-screen terminal shell, a braille-canvas terminal graph that lights up as agents read and write, a 3D web viewer, a VS Code extension, a SwiftUI iOS app with widgets and Siri intents, and Herdr and Omarchy plugins.
Result
It works across Claude, Copilot, Cursor, and Codex, installs as a Claude Code plugin, syncs shared stores over git, and reloads cleanly on a new machine with a single init command.
Stack
- TypeScript
- MCP
- SwiftUI
- Node.js