Skip to main content

Project Context Files

Repository-root files that teach coding agents how to work inside a project.

How To Evaluate Entries​

Each entry in this family should make the following points clear:

  • What the convention is.
  • Where the file, URL, or protocol surface normally lives.
  • When a team should use it.
  • Adoption evidence from a public spec, canonical docs, or active ecosystem use.
  • Which example illustrates the convention, if one exists.
  • Related conventions that solve adjacent problems.

Registry Entries​

These files live in a repository root and tell AI coding agents how to behave in that specific project. They are the "README for AI" — not for humans, but for agents.

AGENTS.md​

The cross-tool standard for agent context. Maintained by the Agentic AI Foundation under the Linux Foundation (donated by OpenAI in December 2025 alongside Anthropic's MCP and Block's Goose). The pitch is one file readable by any agent: Claude Code, Cursor, Copilot, Gemini CLI, and others.

A typical AGENTS.md documents build commands, coding conventions, PR rules, and what the agent must not touch. In monorepos, each subdirectory can have its own AGENTS.md that inherits from the root. Agents read the nearest file in the directory tree.

CLAUDE.md​

Claude Code's native memory file. Loaded automatically before every session. Anything in it enters the context window without prompting. Supports @imports for referencing other files (up to 5 levels deep), which keeps the main file lean while allowing each team or subdirectory to own its own rules.

CLAUDE.md is Claude-specific. For multi-tool teams, the common pattern is to make it a symlink to AGENTS.md: CLAUDE.md → AGENTS.md.

MEMORY.md​

Claude Code's auto-memory file. Written by the agent, not the human. When Claude learns something about your codebase mid-session, it can persist that knowledge here. On next session, the index loads automatically; topic files load on demand. Complements CLAUDE.md: you write the instructions, the agent writes the learnings.

Cursor, Cline, and GitHub Copilot instruction files​

Tool-specific equivalents. Cursor uses YAML-frontmatter scoped rules by glob pattern. GitHub Copilot uses .github/copilot-instructions.md for repo-wide defaults and path-specific .instructions.md files. Open-source agents like Cline and Roo Code use .clinerules or .roorules in the project root to ingest behavior on initialization. For teams using multiple tools, the symlink pattern prevents content from drifting apart across files.

.aiignore​

Tells AI agents which files and folders to skip — analogous to .gitignore. JetBrains Junie adopted this pattern. Prevents agents from reading sensitive config files, large binaries, or generated code you don't want touched.

Memory Bank (cline_docs/ or .roo/)​

A project state architecture popularized by open-source agents like Cline and Roo Code. Instead of a single static file, the agent maintains a directory of Markdown files (e.g., activeContext.md, productContext.md, systemArchitecture.md). This allows the agent to document and update its own contextual state as the project evolves across long-living sessions.