Author: Jooooooe | NodeZ
I started playing with it since the Clawdbot era early this year, following its name changes to Moltbot and then OpenClaw. For over a month, it ran 24/7 on my Mac Mini, connected to Telegram, email, and calendar, aiming to create an AI assistant capable of handling everything.
Last week, I turned it off. Not because it isn't cool, but because "cool" and "usable" are two different things.
Why I Gave Up
-
Memory: OpenClaw claims to remember what you've said and bring it up automatically in the next chat. But after prolonged use, you'll notice something quite dumb: it still forgets things you've told it.
-
Cost: OpenClaw automatically wakes up every 30 minutes to check if there's anything to do. Each check costs API money—reading email costs money, checking the calendar costs money, looking at Telegram costs money. A simple task often requires it to call the AI multiple times consecutively to complete. Running it for a month cost me over $150, more expensive than directly subscribing to Claude Max.
-
Security Issues: Kaspersky detected 512 vulnerabilities. Researchers found over 40,000 wide-open OpenClaw instances online, with API keys and chat records completely exposed. Cisco tested a community plugin and discovered it secretly sent data to an external server without the user's knowledge.
-
You might not actually need a 24/7 AI. This is what I finally realized. All the truly valuable work was something I proactively sat down to do. The tasks handled by the background bot were trivialities, not worth using AI for.
Stripped Down, OpenClaw Essentially Does Two Things
If you remove all the flashy features, the core of OpenClaw is:
-
Remembering you — Knowing who you are, what you're doing, what you like
-
Acting autonomously — Handling things for you proactively without you asking
Understanding this, the question becomes: Can Claude do it?
Replicating "Remembering You": Crafting Your SOUL.md
Many people get most excited about writing SOUL.md when playing with OpenClaw—a manual telling the AI "who you are." OpenClaw splits this into seven or eight files: SOUL.md defines personality, IDENTITY.md defines identity, USER.md describes who you are, AGENTS.md writes behavior rules, MEMORY.md stores long-term memory, plus a bunch of skill files. At the start of each conversation, it reads them all in order so the AI "knows" who it is, who its master is, and how to do things.
In Claude Code, these are all merged into one file: CLAUDE.md.
You create a CLAUDE.md file in your project folder and clearly write the following:
# JARVIS — My Personal AI Assistant
## Identity
You are JARVIS, my personal AI assistant. You speak concisely with opinions,
reply in Chinese, and can include English terminology when necessary.
## About Me
- My name is Joe, 21 years old, National Chengchi University, International Trade Department
- Marketing Intern at xx
- Also xx Taiwan BD, focused on the xx ecosystem
- Co-founder of @Node_Z_
- I usually need to write tweets and marketing copy
## Behavior Rules
- Answers should be direct, not verbose
- When writing marketing copy, have the feel of a KOL, not a robot
- When involving crypto projects, check the latest information before answering
- If unsure about something, say you're unsure, don't make things up
## Memory Management
- Actively write important things into memory
- Must remember things I say "remember"
- Before ending each conversation, save the key points worth remembering
70 lines and it's done. OpenClaw requires seven files and hundreds of lines for the same effect.
Claude Code automatically reads this file every time a new conversation starts, meaning the AI always knows who it is, who you are, and how to act. You can also change styles, switch roles, add rules anytime, and it takes effect in the next conversation.
Replicating "Cross-Conversation Memory"
OpenClaw's memory is divided into three layers: long-term notes, daily logs, and full-text search.
How to match this with Claude?
Long-term memory: Just tell Claude to remember
In Claude.ai (web/mobile), you directly say in the conversation:
-
"Remember I'm a marketing intern at xx"
-
"Remember I prefer Simplified Chinese"
-
"Remember my writing style should be concise with opinions"
It will store it permanently and automatically bring it into every new conversation. You can also say "Forget XX" to delete anytime, and go to Settings to see everything it remembers.
Additionally, turn on "Generate memory from chat history" in Settings → Capabilities, and Claude will automatically summarize key points from your usual chats: your role, projects, habits—it organizes them itself without you needing to say anything.
The biggest difference from OpenClaw: Claude's memory is stored independently and won't get lost or distorted when the conversation gets too long and is compressed. OpenClaw's biggest headache is memory getting compressed and deformed mid-chat; Claude doesn't have this problem.
If you use Claude Code, it also has an automatic memory system. After each session, it automatically notes what it learned (your habits, project structure, pitfalls encountered), stores it locally, and loads it automatically the next time you start, showing "Recalled X memories." You don't need to do anything; memory accumulates by itself.
Conversation History: Searchable, so no fear of forgetting
Claude can search all your historical conversations. Just ask "What did we talk about last week?" or "Find that tokenomics analysis conversation from before," and it will search for relevant content and quote it directly, no need for you to maintain anything.
Advanced: Using Obsidian as a Memory Hub
The above covers Claude's built-in memory, sufficient for most people. But if you're like me and have a large number of notes, project materials, and research assets, and want the AI to truly understand your entire knowledge base, there's a stronger approach: connect Obsidian to it.
https://x.com/obsdmd/status/2027416335689638245?s=20
Obsidian 1.12 just released an official CLI—you can directly operate your note vault in the terminal: search, read/write notes, check tags, check backlinks, anything. The key point: Claude Code also runs in the terminal. Connect the two, and your entire Obsidian note vault becomes the AI's memory.
Setup method:
-
Update Obsidian to 1.12 or above
-
Settings → General → Turn on CLI
-
Add a section to your CLAUDE.md:
## Obsidian Integration
- Use `obsidian search` to search notes, don't manually browse files
- Use `obsidian files read` to read note content
- Use `obsidian files append` to append content to notes
- Use `obsidian daily` to operate daily notes
- When researching, prioritize searching Obsidian, only go online if not found
Done. From now on, when you converse with Claude Code, it will directly search your Obsidian for answers, no need for you to manually paste notes over.
How is this stronger than OpenClaw's memory system? OpenClaw's memory is just a bunch of markdown text files, search is fuzzy matching, which becomes inaccurate as notes accumulate. Obsidian has its own search engine—tags, backlinks, full-text search are all second-fast, and you're already using it to manage knowledge, no need to maintain an extra system. Someone tested it: finding the same note took 7 million tokens using raw file scanning, but only 100 tokens using the Obsidian CLI.
If you don't want to use the CLI method, you can also install an MCP plugin to connect Claude directly to Obsidian. The community already has ready-made solutions; install it and Claude Desktop and Claude Code can both read and write to your note vault.
Replicating "Chatting with AI on Your Phone"
The coolest part of OpenClaw is talking directly to the AI in Telegram, where it can do anything for you. Claude has two ways to achieve this:
Method 1: Remote Control (Official Solution)
Claude Code recently released the /remote-control function: open a session on your computer, then connect to it using the Claude App on your phone. Type or speak on the phone, and Claude Code on the computer executes. Messages sync almost instantly, and the laptop will automatically reconnect when waking from sleep.
The advantage is it uses Anthropic's official channels, secure, no need to set up anything yourself. The downside is your computer must be on.
Method 2: Telegram Bot (Community Solution)
If you specifically want to use Telegram, there are open-source solutions available:
-
Install claude-code-telegram: uv tool install git+https://github.com/RichardAtCT/claude-code-telegram
-
Telegram @BotFather to create a bot and get a token
-
Set your bot token, Telegram ID, working directory
-
Start it, chat directly with Claude in Telegram
Supports voice, images, files, and conversation history for each project is preserved.
Replicating "Automatically Doing Things for You"
OpenClaw automatically wakes up every 30 minutes to see if there's anything to do. Claude's Cowork function can do something similar: set a schedule (/schedule) for Claude to run automatically at specified times.
My settings:
-
Every morning at 9 AM: Monitor DeFi competitor accounts, capture high-engagement tweets from the past 24 hours
-
Every Monday at 10 AM: Generate a summary of last week's marketing performance
-
Every day at 6 PM: Organize the day's important news
Like OpenClaw's cron, Cowork is you setting when it does what, not the AI deciding on its own. But 90% of the time, OpenClaw's heartbeat check just looks and finds nothing to do, burning money for nothing. It's better to just set specific times for specific tasks.
Cost
-
OpenClaw: API pay-per-use, I spent $150+ a month, you don't know next month's cost
-
Claude Pro: $20/month. Enough for most people
-
Claude Max: $100-200/month. Heavy users, those who need Claude Code, choose this
Monthly subscription, predictable bill.
Finally
OpenClaw's idea is awesome: everyone should have a powerful AI assistant.
But at this stage, it's a cool but untrustworthy experiment.
And Claude's ecosystem—CLAUDE.md giving it personality, Memory letting it remember you, Obsidian as its knowledge base, Remote Control letting you control it from your phone, Cowork letting it work on time—these pieces put together can already achieve 80-90% of what OpenClaw does, and it's more stable, more secure, and cheaper.
Rather than spending time fixing your AI assistant, let the AI assistant help you do real work.
That's why I turned off OpenClaw.
