One command installs a hook that captures every conversation turn as it happens, no asking Claude to remember anything. Setup takes about a minute.
On claude.ai, Claude Desktop, or Cowork instead? Those don't have a local hook to install, use the MCP connector and the remember tool instead.
In your terminal, run:
npx contexara initYou'll be asked which tool to install into. Pick one, or select several if you use more than one, the installer handles each in turn and tells you if any one of them fails without blocking the rest.
The installer prompts for your Contexara API key. Generate one from the Keys page if you don't already have it, it looks like ctx_live_....
Start a normal conversation in the tool you just connected. Say something worth remembering, a decision, a preference, what you're working on. Nothing to trigger manually, the hook sends it in the background as you talk.
Then open a brand new conversation, in the same tool or a different connected one, and ask about it. If it comes back, capture is working end to end.
The installer drops a small hook script into your tool's own config (Claude Code's settings, Codex's config.toml, or Cursor's hooks file) and wires it to fire on each turn. Every message and response gets sent to Contexara's capture endpoint, processed through the same extraction pipeline as anything saved manually through the remember MCP tool, so it lands with the same structure and quality either way.
Nothing leaves your machine except the conversation content itself, the hook doesn't read or touch anything else in your project.
Remove capture from a specific tool at any time:
npx contexara uninstall --platform claude-codeSwap claude-code for codex or cursor. This only removes that tool's hook, memory already captured stays in your account.
Check the key was copied in full and hasn't been revoked from the Keys page. Keys always start with ctx_live_.
Restart the tool after installing, some hosts only pick up config changes on the next launch. If it still doesn't appear, re-run npx contexara init for that platform, it's safe to run again.
The installer installs each platform independently and reports which ones failed without touching the ones that succeeded. Choose to retry when prompted, or re-run the command later targeting just that platform with --platform.