How to Give Your OpenClaw a Memory That Actually Works

The Amnesia Problem

Every time you start a new conversation with your OpenClaw agent, it forgets everything.

It does not remember what you told it last week. It does not recall that you prefer short responses or that you are working on a specific project. Every session starts from scratch — like talking to someone with complete short-term memory loss.

This is the single biggest frustration people have with AI assistants. And it is solvable.

OpenClaw memory system is not broken. It is just not configured by default. Once you set it up, your agent remembers. Not in some fuzzy, unreliable way — but in a way that actually works.

How OpenClaw Memory Works

OpenClaw uses a file-based approach. Your agent memory lives in markdown files on disk, not in some opaque database. This means you can read it, edit it, and back it up like any other file.

The default system gives you two tiers:

Daily logs. Every conversation gets saved to memory/YYYY-MM-DD.md. Your agent loads today and yesterday logs when it starts. This gives it recent context without bloating the context window.

Curated long-term memory. MEMORY.md is the important stuff — your preferences, projects, facts about you that should persist across sessions. Your agent reads this on every start.

But there is a third tier that most people do not know about:

Searchable knowledge. OpenClaw indexes your memory files and lets you query them semantically. memory_search("API decisions") returns relevant snippets from months ago, even if you do not remember exactly where you wrote them.

This is the difference between an agent that remembers and one that does not.

Setting It Up

The default setup works out of the box, but you are not getting the full benefit until you configure it properly.

Enable memory search in your openclaw.json:

{
  "memorySearch": {
    "enabled": true,
    "provider": "openai",
    "remote": { "apiKey": "sk-..." }
  }
}

Without this, you are limited to keyword search. With embeddings enabled, semantic search works — you can ask "what did I decide about the API?" and get the right answer even if you never used those exact words.

Add memory instructions to AGENTS.md. This is the step most people skip. Your agent needs to know when to write to memory. Add something like:

On every session start:
- Read MEMORY.md
- Read memory/YYYY-MM-DD.md (today and yesterday)

When I say "remember this":
- Write the key points to MEMORY.md

Before making significant decisions:
- Search memory for relevant past context

Create a MEMORY.md with the basics. Start with:

# About Nat
- Timezone: US/Eastern
- Prefers: Direct answers, no fluff
- Working on: Newsletter, blog

# Current Projects
- OpenClaw setup
- Claw Mart content

What the Three-Tier System Actually Gives You

If you want the full system — the one that remembers everything reliably — you need the three-tier approach:

Tier 1: Short-term. The context window. Anything your agent is actively working with. This is volatile — gone when the session ends.

Tier 2: Session. Daily logs. Everything that happened today and yesterday. Automatically loaded. Searchable but uncurated.

Tier 3: Long-term. The curated stuff in MEMORY.md plus your searchable subdirectories (projects, people, decisions). This is what survives over months and years.

The Three-Tier Memory System skill from ClawMart sets all this up for you. It adds structure: separate directories for people, projects, decisions. Automatic promotion of important session notes to long-term memory. A workflow for reviewing and curating what gets remembered.

Without it, you are manually managing what goes into long-term memory. With it, the system handles the lifecycle — you just do the reviewing.

Common Mistakes

1. Missing embeddings. You configured everything but forgot the API key. Your searches work, but poorly. You get BM25 keyword matching and wonder why memory feels unreliable.

2. Read-only workspace. Your agent cannot write to disk. Memory files never update. Check your file permissions.

3. Group chat memory leaks. MEMORY.md only loads in DMs by default. If you are testing in a group chat and wondering why your agent has amnesia, that is why.

4. File bloat. Without compaction enabled, daily logs pile up. After a few months, search gets slow and noisy. Turn on compaction or set up an auto-archive routine.

5. No promotion workflow. Daily logs fill up with raw conversation. Nobody — human or agent — is moving the important bits to MEMORY.md. So your long-term memory stays empty while your session logs grow endlessly.

What to Do Right Now

1. Check if memorySearch is enabled in your openclaw.json. If not, add it.
2. Create a MEMORY.md with your basic info and current projects.
3. Add memory instructions to AGENTS.md so your agent actually reads and writes to it.
4. Verify your embedding provider is configured — run a memory_search() and see if semantic queries return relevant results.
5. If you want the full system without the manual wiring, grab the Three-Tier Memory System and get it running in a few minutes.

Your agent is only as good as what it remembers. Set up OpenClaw persistent memory once, and every future session starts with context instead of a cold start.