Project memory
Project memory
For an agent to work well in your project, it needs to know your conventions: build commands, style, gotchas. Mini-claude reads that from a memory file.
How it works
At startup, Mini-claude looks for a memory file in the workspace, walking up to the root. First match wins, in this order:
MINICLAUDE.md: the native format.AGENTS.md: the cross-tool standard.CLAUDE.md: Claude Code.GEMINI.md: Gemini CLI..cursorrules: Cursor..windsurfrules: Windsurf..github/copilot-instructions.md: GitHub Copilot.
If it finds one, its contents are injected into the system prompt, the agent then knows your rules for the whole session.
Compatibility with other tools
This is intentional: if your repo is already set up for another agent (Claude Code, Codex, Gemini, Cursor…), Mini-claude picks up its instructions file as-is. No double maintenance. And it all stays local, these files are just read from your disk, nothing is sent to any service.
What to put in it
Keep it short, this file ships with every request.
# My project
- Build: `make build` · tests: `make test`
- Go 1.24, no comments in code
- Migrations live in `db/migrations/`, never hand-edited
- Commits short, imperative, no prefixProject memory is the right way to give the agent context without breaking the modes’ guardrails. Prefer it over
MINI_CLAUDE_SYSTEM, which replaces the mode’s system prompt. See Configure.