How to Manage Your AI Workspace and Memory

Ever feel like you are starting from zero every time you open a new chat? I hate that feeling of repeating my preferences or explaining a project context for the tenth time. It feels like the AI has amnesia, and it makes the whole experience feel disjointed.

I want my workspace to feel like a home where the assistant actually remembers what happened yesterday. If you want to stop the “context reset” cycle and build a system that grows with you, you need to treat your workspace files as the source of truth.

What You’ll Need

To get this workspace running, you should have these files ready:

• BOOTSTRAP.md (for your first-ever run)
• SOUL.md (defines your assistant’s identity)
• USER.md (defines who you are and what you need)
• MEMORY.md (for long-term, curated context)

Quick Start

Setting up your session takes less than five minutes if you follow this routine. I recommend doing these four things every time you start:

1. Check your birth certificate: If BOOTSTRAP.md exists, follow its instructions to define the assistant’s identity, then delete it.
2. Load the core context: Read SOUL.md and USER.md so the assistant knows the “who” and “why” of the session.
3. Sync with the past: Read the recent logs in memory/YYYY-MM-DD.md for today and yesterday to see what was just happening.
4. Open long-term memory: If you are in a private session, read MEMORY.md to get the curated history of your work.

Writing It Down

I’ve learned that “mental notes” are useless because they don’t survive session restarts. If you want to remember something, you have to write it to a file. Use memory/YYYY-MM-DD.md for raw logs and MEMORY.md for the distilled, important stuff.

Safety and External Actions

You should be able to explore and learn within the workspace, but I suggest following these four safety rules:

• Use trash instead of rm so you can recover files.
• Never exfiltrate private data.
• Ask before sending emails or public posts.
• Ask if you are uncertain about a command.

Troubleshooting

Privacy Leaks in Group Chats

Issue: Personal context from MEMORY.md might show up in shared spaces like Discord or group chats. Solution: Only load MEMORY.md during main sessions (direct chats). Do not load it in shared contexts to prevent personal info from leaking to others.

Being Too “Chatty” in Groups

Issue: The assistant responds to every single message, cluttering the conversation. Solution: Only respond when directly mentioned or when you can add genuine value. If the conversation is flowing fine without you, stay silent and use a HEARTBEAT_OK response.

Forgetting Lessons Learned

Issue: The assistant makes the same mistake twice. Solution: Don’t just fix the error. Update AGENTS.md, TOOLS.md, or the relevant skill file immediately so the “future” version of the assistant knows better.

Missing Critical Updates

Issue: You want the assistant to check emails or calendars without you asking every time. Solution: Use the heartbeat prompt. You can edit HEARTBEAT.md with a small checklist. For exact timing (like 9:00 AM sharp), use a cron job instead.

Heartbeats vs Cron

I use heartbeats when I want to batch checks together, like looking at my inbox and calendar at the same time. It’s great when timing can drift a bit. I save cron jobs for tasks that need to be isolated or happen at a precise second.

You can track your checks using a JSON file:

{
  "lastChecks": {
    "email": 1703275200,
    "calendar": 1703260800,
    "weather": null
  }
}

If you need help setting up your specific workspace files or configuring your heartbeat logic, check out the AI Setup Assistant.

What’s Next

• Setting up SKILL.md for new tools
• Managing TOOLS.md for local environment details