Agent Markdown Memory Bank Protocol

A lightweight protocol for AI coding agents (GitHub Copilot, Cursor, etc.) to maintain persistent project context across sessions using a structured .memory-bank/ directory.

For a real-world example have a look at the memory bank files for my openweather-controller project.

What is AGENTS.md?

AGENTS.md is a root-level instruction file that AI agents automatically read to understand how to behave in this repository. It enforces two core protocols:

1. Memory Bank Protocol

The agent maintains a .memory-bank/ directory as its single source of truth, containing:

File	Purpose
product-context.md	Project goal and purpose
active-context.md	Current focus and next steps
system-architecture.md	Architectural patterns and decisions
progress.md	Timestamped journal of completed work

Workflow:

1. Before any task — read product-context.md, active-context.md, system-architecture.md
2. During a task — update system-architecture.md immediately on architectural changes
3. After every task — append a timestamped entry to progress.md, update active-context.md with next steps

The agent runs date '+%Y-%m-%d %H:%M' to get an accurate timestamp rather than guessing.

2. Todo Workflow

When asked to pick up the next task, the agent reads the Next Steps (unsorted) section in active-context.md and recommends the top 3 items.

Mandatory Completion Checklist

Every session that involves code changes MUST end with:

• Append entry to .memory-bank/progress.md
• Update .memory-bank/active-context.md
• Confirm memory bank update to the user

Usage

Drop AGENTS.md into the root of any repository. AI agents that support root-level instruction files (Copilot with AGENTS.md, Cursor with .cursorrules, etc.) will automatically pick it up and follow the memory bank protocol.