Making Your OpenClaw Agent Actually Remember You: Memory & Soul Files Deep Dive

One of the coolest moments with OpenClaw is when your agent suddenly feels like it really knows you — not just repeating what you said five minutes ago, but actually remembering your style, your favorite Magic Hermes creations, and exactly how you want things built on the /ai/ sub-website.

That magic doesn't happen by accident. It lives in two special files: MEMORY.md and SOUL.md. I spent a lot of time figuring out how they really work on my Magic Hermes setup, and once I got them dialed in, my agents went from "helpful helper" to "my personal creative partner that never forgets how I like my games, book reviews, or blog posts built."

Here's exactly how persistent memory works in OpenClaw, how I trained mine with all my Magic Hermes preferences, and the exact commands I use every day to inspect and edit everything.

How Persistent Memory Actually Works

OpenClaw has two layers of memory:

Temporary context → The normal chat history that gets cleared with /new. This is what most agents use and why they forget you after a while.
Persistent memory → Two files that never get cleared unless you delete them:
- MEMORY.md — Factual knowledge, preferences, and things you explicitly want the agent to remember forever (your favorite game mechanics, writing style for book reviews, how you structure blog posts, etc.).
- SOUL.md — Personality, writing style, values, and "who you are" as the user. This is what makes the agent sound like it actually knows you and the Magic Hermes vibe.

Every time the agent thinks or replies, it quietly reads both files and injects the relevant parts into its reasoning. That's why a well-trained agent starts suggesting ideas that feel perfectly on-brand for the /ai/ section.

How I Trained My Agent with My Magic Hermes Preferences

I didn't dump a giant wall of text. I added things gradually so the agent could actually absorb the Magic Hermes style.

Here's what my files look like now (shortened for this post):

Excerpt from MEMORY.md:

# Ian's Magic Hermes Preferences - Updated April 2026

- Everything lives under the /ai/ sub-website at magic-ian-metal.com/ai/
- Core focus: Building fun, retro crypto-themed HTML5 games using OpenClaw
- Favorite games section: 2084 (crypto 2048 variant), Crypto Breakout, Magic Catcher, Infinite Runner, Crypto Runner
- Book reviews are synchronized slide presentations — clean, visual, and easy to follow
- Blog posts should be casual, helpful, and full of real "here's exactly what I did" examples
- All content must feel fun, slightly sarcastic, and made for crypto/gaming fans
- Always include working code examples, clear steps, and screenshots when possible
- Goal: Make the entire /ai/ section feel like a living AI playground

Excerpt from SOUL.md:

# Ian's Soul / Personality for Magic Hermes

- Super enthusiastic about AI agents building their own little world under /ai/
- Writes in a friendly, real-talk style — like chatting with a friend who's smart but new to OpenClaw
- Loves showing the exact process: "I tried this, it broke, here's the fix"
- Keeps everything fun and approachable — no corporate AI speak
- Obsessed with turning OpenClaw experiments into playable games, beautiful book reviews, and useful blog posts

I added these lines one piece at a time over a few days. After that, the agent started proactively suggesting new game features, better slide layouts for book reviews, and even reminded me of my own writing style when drafting blog posts.

Exact Commands to Inspect and Edit Memory & Soul

These are the commands I use daily on my Magic Hermes agents:

/memory → Shows the current MEMORY.md content
/soul → Shows the current SOUL.md content
/add-memory → Adds a new fact or preference (e.g. /add-memory When building games, always add a high-score table)
/add-soul → Adds or updates personality traits (e.g. /add-soul Keep all blog posts fun and slightly sarcastic)
/edit-memory → Opens the full file for editing in your preferred editor
/edit-soul → Same but for the personality file
/summarize-memory → Makes OpenClaw condense everything into a cleaner, tighter version

💡 Pro Tip: Refresh After Changes

After big changes, send /new in the chat so the agent immediately starts using the updated memory in its next replies.

What I Learned (and What I'd Tell My Past Self)

Don't try to write perfect memory files on day one. Start small, add one preference at a time, and let the agent use it for a day or two before adding more. The best memories are the ones that actually shape how it helps you create new stuff in the /ai/ section.

Once I got MEMORY.md and SOUL.md right, my OpenClaw agents started suggesting new game ideas, fixing slide layouts on their own, and writing blog posts that already sounded exactly like me. It finally felt like I had a real creative teammate instead of just another chatbot.

Want to make your agent remember you too?

Drop your biggest "I wish my agent remembered X" in the comments.

If this helped, check out my other OpenClaw posts in the /ai/blog/ section — I'm documenting the whole journey of building the Magic Hermes games, book reviews, and everything else with these agents.

Happy OpenClaw-ing! 🦞