Secrets never leave the laptop

Cost, safety and controlAny team with .env files

We worried that a memory layer would ship our keys to someone's server.

Does an AI agent memory tool upload my .env file or API keys to its server?

Keep secrets in .env files, not in prompts: ArcRoom sends only metadata, never file contents, tool output or diffs. A secret file's path, or a command that names one, leaves only as [secret file]. Prompts are cut at 1,000 characters and secrets in them are redacted by pattern, but unusual formats could slip through.

Also asked as: I'm worried an agent memory layer will ship our keys to someone's server. Is that a real risk? · How do I use AI agent memory without leaking secrets from .env files? · Can an agent hook send my file contents to an outside server?

The work

Only metadata is sent: prompts (first 1,000 characters, secrets redacted), paths, commands, hashes and token counts.

ArcRoom keeps
  • Redaction
The next agent gets

sent: { tool: 'Read', path: '[secret file]' }

Step by step

  1. Only metadata is sent: prompts (first 1,000 characters, secrets redacted), paths, commands, hashes and token counts.
  2. File contents, tool output and diffs are never sent.
  3. A secret file's path, or a command that names one, leaves only as [secret file].

What the agent sees

sent: { tool: 'Read', path: '[secret file]' }
sent: { cmd: 'cat [secret file]' }

Compared with today's tools

Mem0 open source, self-hosted

The team runs Mem0's open-source server on its own infrastructure, so memory never leaves the network.

Setup
1–3 days
Ongoing
run, patch and back up the server

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init)
Ongoing
None

Self-hosting keeps everything in-house, which ArcRoom can't offer; ArcRoom instead never sends file contents, redacts secret paths, and needs no server to run.

Where it stops

A secret typed into a prompt is redacted by pattern. Unusual formats could slip through, so don't paste secrets into prompts.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in cost, safety and control