wtf is global-agent-memory?

Global Agent Memory gives AI coding assistants like Claude Code and Codex a shared, long-term memory that a human can review and control. Instead of each tool keeping its own scattered notes, this project stores knowledge such as project conventions, past solutions, and preferences in plain Markdown files that live in one central place, called a Vault. When an AI agent learns something worth remembering, it does not save it directly. It proposes the memory as a candidate. A person then reviews the dashboard and decides whether to approve, edit, reject, or mark it as protected before it becomes part of the permanent record. This human-in-the-loop step is meant to stop AI agents from quietly filling your memory with wrong or unwanted information. The system is built around three connected pieces. A local background service, called a daemon, speaks the Model Context Protocol so any compatible AI agent can search and propose memories. The Vault itself is readable in Obsidian, complete with linked notes and templates, for people who want to browse and edit their knowledge base by hand. And a local dashboard offers a simpler review screen for approving candidates, resolving conflicts, and checking system health. All of it runs on your own machine, with the daemon listening only on localhost and requiring an access token, so no data leaves your computer unless you choose to sync it yourself. Search combines simple keyword lookup with an optional local AI model for smarter, meaning-based results, and the whole system keeps working even if that optional piece is offline. Getting started is a single install command followed by a guided setup wizard that detects Claude Code or Codex on your machine and configures everything automatically. It requires Python 3.12 or newer and works on macOS and Linux.