wire-memory CLAUDE.md

wire-memory

Claude Code plugin that gives AI agents persistent memory via Wire containers.

How it works

This is a native Claude Code plugin — no build step, no dependencies, no CLI.

1. User runs /wire-memory:connect — opens browser, authenticates via Wire's nonce flow, selects a container
2. Connect script saves MCP endpoint + API key to ~/.wire-memory/config.json (owner-only permissions)
3. Connect script writes real values directly into .mcp.json and marks it --skip-worktree to prevent accidental commits
4. After restart, agent has native access to wire_write, wire_search, wire_explore, wire_navigate, wire_delete
5. The skills/memory/SKILL.md teaches the agent when and how to use these tools for persistent memory
6. The hooks/hooks.json registers a UserPromptSubmit eval hook that prompts the agent to check memory on every prompt

Plugin structure

wire-memory/
├── .claude-plugin/plugin.json   # Plugin manifest
├── .mcp.json                    # MCP server config (placeholder until connected)
├── skills/memory/SKILL.md       # Behavioral instructions for memory usage
├── hooks/hooks.json             # UserPromptSubmit eval hook registration
├── commands/connect.md          # /wire-memory:connect
├── commands/status.md           # /wire-memory:status
├── commands/disconnect.md       # /wire-memory:disconnect
├── scripts/connect.mjs          # Nonce flow (Node.js, zero deps)
├── scripts/eval-hook.mjs        # Eval hook — prompts agent to use memory tools
├── scripts/status.mjs           # Show connection info
├── scripts/disconnect.mjs       # Clean up config
├── package.json                 # @usewire/memory
└── LICENSE                      # MIT, Superloops LLC

Development

Test locally:

claude --plugin-dir .

Then run /wire-memory:connect to test the flow.

No build step

Everything is plain .mjs and markdown. No TypeScript, no bundler, no dependencies. The scripts use only Node.js built-ins (crypto, fs, path, os, child_process).