How to Transfer Your Claude Memory to OpenClaw

Two different memory systems

When you use Claude on claude.ai, it builds up a picture of you over time: your name, your work, how you like responses, what you're working on. That lives in Anthropic's cloud and gets pulled into your conversations automatically.

When you run OpenClaw with Claude as the model, that cloud memory doesn't carry over. OpenClaw is a local agent talking to the Claude API directly. It has no idea what the claude.ai version of Claude knows about you.

This guide bridges that gap.

How OpenClaw handles memory

OpenClaw stores memory as plain text files on your computer inside your workspace folder. The main one is MEMORY.md. Whatever you put in there gets loaded into every conversation with your agent.

It's transparent and easy to edit. The tradeoff is that recall isn't as reliable as cloud-based systems. OpenClaw reads from MEMORY.md on every message, but whether it actively uses that information depends on the model's judgment in the moment.

Why trimming matters

Claude's memory is thorough. By the time you export everything it knows about you, you might be looking at 60-100 entries. Your first instinct might be to paste all of that into MEMORY.md and move on. That technically works, but it gets expensive fast.

OpenClaw reads that file on every single message you send. Every word in there costs you tokens on the Claude API, even if 90% of it isn't relevant to what you're asking right now. People who've tried the "dump everything" approach usually land on the same conclusion: treat MEMORY.md like a business card, not a biography. Name, location, a few key projects, core preferences. The stuff that's almost always relevant no matter what you're talking about.

The sweet spot is around 20-30 entries. Enough for your agent to know who you are without burning extra tokens on every conversation.

Step 1: Export your Claude memory

Claude stores memory in two places, so check both.

First, paste this prompt into any Claude conversation:

Second, go to Settings > Capabilities > Memories from your chats. This shows Claude's synthesized memory separately from what it surfaces in conversation. The two don't always match. Grab both.

Step 2: Trim it in the ImportMemory editor

Paste everything into the ImportMemory editor. Cut anything outdated or irrelevant:

• Old roles, companies, or finished projects
• One-off preferences rather than ongoing habits
• Anything too personal for an API request
• Duplicate entries

Keep: your current role, active projects, response preferences, key tools, and standing instructions. Aim for 20-30 entries.

Step 3: Paste it into MEMORY.md

Open ~/.openclaw/workspace/MEMORY.md in any text editor. Structure it clearly:

## Identity
- Name: [your name]
- Location: [city, country]
- Role: [job title at company]

## Active Projects
- [Project name]: [brief description]

## How I like responses
- Keep things concise by default
- Use code examples for technical questions
- Don't add unnecessary caveats

Save and restart with openclaw gateway.

Tips for a better result

• Lean is better than complete. 20-30 entries that are always relevant beats 100 that are sometimes relevant.
• Keep it current. Claude's export is a snapshot. Update MEMORY.md every month or two.
• Use headers. Clear sections help the model find what it needs faster.
• Watch what you include. Your memory content goes through the API on every message.
• Recall isn't perfect. If the agent misses something important, state it directly in conversation.

Frequently asked questions