Claude Code Persistent Memory

By default, Claude Code forgets everything the moment a session ends. Rohrpost gives it persistent, project-scoped memory: every session's decisions and discoveries are captured, and relevant context is recalled automatically at the start of the next one — even weeks later, even in a brand new terminal.

Install the plugin

The Claude Code plugin is the fastest path: it registers the MCP server, installs a skill file, and wires lifecycle hooks for automatic capture — all in two commands.

curl -fsSL https://rohrpost.io/install.sh | sh

Then point it at your gateway (managed or self-hosted) with your API token:

rohrpost-plugin install --target code \
  --gateway-url https://app.rohrpost.io \
  --token YOUR_API_TOKEN

This writes the MCP server registration into your project's .claude/settings.json, installs a skill describing when Claude should recall and ingest, and adds hooks that fire automatic capture at natural session boundaries — without you having to remember to ask. See the full plugin download page for platform binaries and checksums.

Backfill existing history

A brand new tenant starts with an empty knowledge base — the classic cold-start problem. If you've already been using Claude Code on this project, you have months of transcripts sitting in ~/.claude/projects/. Import them once and Rohrpost starts warm instead of empty:

rohrpost-plugin import-history --dry-run

This reads only local transcripts already on your machine, previews what it found, and sends nothing until you confirm. Looks right? Drop --dry-run to actually import:

rohrpost-plugin import-history

Default limit is 500 messages per run — pass --limit 2000 to go further back, or--project /path/to/repo to backfill a different project than the current directory.

Verify it's working

Run the doctor command to confirm the MCP server, hooks, and token are all wired correctly:

rohrpost-plugin doctor

Then, in a Claude Code session, ask something you know is in your history — e.g. "what did we decide about auth?" — and Claude should call rohrpost_recall and answer from real project context instead of guessing.

Manual MCP config

If you'd rather not use the plugin installer, add the MCP server directly to.claude/settings.json (project-scoped) or ~/.claude.json (user-scoped):

{
  "mcpServers": {
    "rohrpost": {
      "command": "/path/to/rohrpost-mcp",
      "env": {
        "ROHRPOST_GATEWAY_URL": "https://app.rohrpost.io",
        "ROHRPOST_API_KEY": "rp_your_api_key"
      }
    }
  }
}

Download rohrpost-mcp for your platform from your admin panel'sDownloads tab, or build it from source. Full tool reference (all MCP tools, parameters, and example system prompt instructions) is in theMCP Server docs.

Troubleshooting

Claude never calls the Rohrpost tools

Prompt it explicitly at first — "call rohrpost_recall before answering" — or check that the skill file installed correctly with rohrpost-plugin doctor. The installed skill instructs Claude to recall at session start and ingest at session end automatically once it's been used a few times.

"command not found: rohrpost-mcp"

The plugin installer writes an absolute path to the binary. If you moved it after installing, re-run rohrpost-plugin install, or check the path in.claude/settings.json matches where the binary actually lives.

Recall returns nothing useful

New tenants start empty — run the backfill above, or just keep using Claude Code normally for a few sessions; capture happens automatically once hooks are wired.

Switching between local and managed gateways

rohrpost-plugin profile add local --gateway-url http://localhost:7878
rohrpost-plugin profile use local