Thread mind

Long coding sessions lose their context: the conversation fills the model's window, gets compacted or cleared, and the next session starts from scratch. You spend tokens and time re-explaining architecture decisions, constraints and where the work stands.

What it does

• Persistent memory: keeps a tree of thread summaries in a .threadmind/ directory inside your project.
• Inherited context: each thread inherits the decisions and constraints of its parent threads, so a sub-task starts with the rules it must follow.
• Structured summaries: the agent records Decisions, Constraints, State, Open questions and Next steps, one section at a time, without rewriting everything.
• Follows git: a thread created on a feature branch becomes the active thread whenever that branch is checked out, and summaries get flagged when the files they cover change.
• One thread per topic: create, switch, move, merge finished threads into their parent, or mark them done or abandoned with the reason.

How it works

1. Create a project with tm:project <title>. ThreadMind stores it in .threadmind/.
2. Load the context when a session starts: the agent calls context_get and gets the active thread's summary plus its ancestors' decisions and constraints, a few hundred tokens instead of the whole history.
3. Save as work progresses: tm:summary, or let the agent record each decision and update the state.
4. Branch into a new thread with tm:create <title> when a sub-task starts, and tm:merge it back when it's done.

Frameworks & tools

Requires the ThreadMind MCP server, the open-source thread-mind-mcp npm package (Node.js 18+), declared in your agent's MCP configuration. The skill includes the setup for Claude Code, Cursor, VS Code (GitHub Copilot), GitHub Copilot CLI, Codex, Gemini CLI, Windsurf and Claude Desktop. In Claude Code, the ThreadMind plugin also loads the context at session start and reminds the agent to save its summary.

Why this beats prompting it yourself

Notes in a prompt or a chat summary get lost at the next compaction, can't be reviewed, and grow until they no longer fit. ThreadMind keeps decisions in plain markdown files, organized by topic: the agent loads only the thread it works on plus the durable rules above it, and your team reviews the summaries in pull requests like code.

Use cases

• Resuming a complex refactor after a weekend, or after /clear, without re-explaining it.
• Sharing architecture decisions and constraints across a team, through summaries committed with the code.
• Running parallel sub-tasks in separate threads, each linked to its git branch.
• Keeping a record of why a library or pattern was chosen.
• Switching agents mid-project: the same .threadmind/ serves Claude Code, Cursor, Codex or Copilot.

Known limitations

• The ThreadMind MCP server must be set up once in each agent's configuration; the skill shows how, but never installs it itself.
• Outside Claude Code with the plugin, saving relies on the agent following the instructions: ask for tm:summary before closing a long conversation.
• Branch linking and outdated-summary warnings need git.
• Summaries are meant to be committed: keep secrets and personal data out of them.