Projects / agent-memory-workspace
A shared memory for AI coding agents
A personal project: a Git workspace that gives AI coding agents a lasting memory of projects, actions and decisions, split into private "brains" with guardrails that keep each piece of data where it belongs.
- Period
- 2026
- Role
- Sole author
- Context
- Personal project
[ In practice ]
How it works
- 01
Each "brain" is a private Git repository of Markdown pages: projects, actions, decisions, people and timelines.
- 02
The agents read a short map of the workspace and use small command-line tools to read and update the registers.
- 03
Skills turn meeting transcripts, messages and an inbox into dated notes and actions, once I have validated them.
- 04
Git hooks and checks keep secrets and private names out, and each brain inside its own boundary.
The need
AI coding agents forget everything between sessions. Working on several projects at once, I kept re-explaining the context, the decisions and the to-dos, and information from one project must never leak into another.
What I did
I built a Git workspace that acts as the agents’ long-term memory. Each “brain” is a private repository of plain Markdown pages, one per boundary, so work and personal notes never mix. Claude Code and Codex read the same map and the same rules, and small tools let them create actions, record decisions and log what changed.
The result
The agents start each session with the right context and keep the registers up to date as they work. I follow the actions in a small local web app or in Obsidian, and a single command sets the whole workspace up on a new machine.
My role
A personal project, designed and built alone: the memory format, the tools, the skills, the guardrails and the local app.
Technical approach
A Python and Bash toolkit around plain Markdown files and Git, with no database and no server.
- Brains. Private Git repositories, each declared with its scope and boundary in a small JSON file; a link from one brain to another is refused.
- Tools. Command-line scripts run with uv: a map of every project, action and decision registers, dated timelines, link checks and a read-only diagnosis of the machine.
- Agents. One
AGENTS.mdfor Claude Code and Codex; skills are installed into each agent’s folder from a shared configuration. - Guardrails. Git hooks refuse secrets and known private names and check the identity before a push; a Claude Code hook blocks risky commands.
- Interface. A local web app for the actions (today, board, timeline, search), also packaged as a macOS app; Obsidian opens the whole workspace as a vault.
- Code. The workspace kit is public: Work-Starter on GitHub.