By the end of this guide, your Clawdbot will:

• Actually remember corrections you made weeks ago
• Find context from old conversations, not just the last 2 days
• Learn from feedback and save structured rules it follows automatically
• Get better at its job over time instead of resetting every session

Three layers. Each one takes 2-10 minutes. Let's fix your bot's brain!!

PS: just give this guide to Claude and ask it to hold your hand through setting up all the steps.

First, Why Your Bot Keeps Forgetting

Every time you start a conversation with your Clawdbot, here's what actually loads:

  You send a message
       |
       v
  ┌─────────────────────────────────────────────────┐
  |              YOUR AGENT WAKES UP                 |
  |                                                  |
  |   1. Loads MEMORY.md        <- pinned rules        |
  |      ~/.openclaw/workspace/MEMORY.md               |
  |   2. Loads last 2 days     <- recent memory files  |
  |      ~/.openclaw/workspace/memory/YYYY-MM-DD.md    |
  |   3. Loads session         <- current chat         |
  |                                                  |
  |   Everything else?                               |
  |   Invisible. Gone. Not loaded.                   |
  └─────────────────────────────────────────────────┘

Your 20-minute onboarding conversation from 3 weeks ago? Not loaded. The correction you made last Tuesday? Invisible, unless it was saved to a memory file within those 2 days.

And it gets worse. When a conversation gets long, OpenClaw compacts it — compresses the chat to free up space. If your agent learned something useful during that conversation but hadn't saved it yet, it's gone.

So you've got two gaps: old memories your agent can't find, and mid-conversation insights that get lost before they're saved. The three layers below fix both.