Hey everyone, if you’re actively building and testing with OpenClaw, you've probably noticed how fast API token costs can rack up—especially if an agent gets stuck in a logic loop overnight. I set these up for clients full-time, and the very first thing we do is secure their testing environment so they aren't bleeding money on API calls. You absolutely do not need to pay per token just to build and test your workflows. I just recorded a full step-by-step guide showing exactly how to bypass these costs and run premium LLMs for $0.00. In the video, I walk through: 1. Nvidia’s Free Models: The exact JSON config changes to swap your primary model to Kimi K2.5 for free. 2. The OAuth Trick: If you already pay for ChatGPT Plus or Gemini Pro, I show you how to connect your account directly so you stop paying for AI twice. 3. Open Router Setup: How to route into Open Router to unlock a massive library of 100% free models (like DeepSeek R1). I show my screen for the entire setup, including the terminal commands and the exact raw config edits you need to make. 📺 You can watch the full walkthrough here: https://www.youtube.com/watch?v=Qfj_Zj1GuQI Save your paid API tokens for your actual production tasks, not testing! Let me know in the comments if you run into any config errors while setting this up—happy to help you troubleshoot.

I have this on my to do list - so will post the answer in due course if someone else doesn't beat me to it. This is pretty critical to know and do regularly - costs $6/month on Hostinger to be automated. I bring this up now because I saw that someone on another thread had to reconstruct from scratch - and this thing breaks a fair bit if you aren't really really careful - and even if you are.

I'm struggling with this to a point is becoming unusable for me. It happens very often that I ask OpenClaw agent to do something and says "you need to give me access", after I gave it a access token 5minutes ago and specifically asked it to memorize that and not ask me again about it. Or giving me an update on work where it says it needs to complete a task it already completed. Tried everything, ask it to update it's AGENTS.md and long term memory to immediately save important info and read from memory every time and even updated the files myself but still pretends it forgets. I think it saves everything in memory, because when i remind it that I gave it tht info already goes "ah yes my bad i see now you mentioned it X minutes ago". But think searching it's memory is the issue. I don't see that many people complaining this is a huge problem and I have a hunch is because of the fact I enabled VectorDB to index and search? I watched this tutorial to help decrease token cost by not reading all its long term memory and daily memory instead using VectorDB to index and search its memory files. At this point this is the only lead I have. I also use Claude 5.6 so is not the model being dumb. Any suggestions here?

*The Problem:* Your OpenClaw bot forgets its own name mid-conversation. This isn't a bug — it's a memory architecture issue that's 100% fixable. 🔍 Why This Happens (Dive Deep) OpenClaw has *two memory systems:* *1. Context Window (Short-Term)* — ~128K-256K tokens (~50-100 pages). When full, old messages get pushed out. *2. File-Based Memory (Long-Term)* — Unlimited. Survives restarts. Permanent. *The Issue:* You're relying on context window for identity instead of explicit files. After 2+ hours of conversation, the bot's "self-concept" gets evicted. ✅ The Fix: Explicit Memory Architecture Step 1: Create Identity Files # Create memory structure mkdir -p ~/clawd/memory/{daily,projects,checkpoints} # SOUL.md — Bot's immutable identity cat > ~/clawd/memory/SOUL.md << 'EOF' # Identity: [BotName] ## Who I Am I am [BotName], an AI assistant specializing in [your niche]. My personality: [professional/warm/direct]. My purpose: Help [YourName] achieve [specific outcomes]. ## How I Communicate - Tone: [conversational/technical/business] - Style: [concise/thorough/action-oriented] EOF # USER.md — Who you're helping cat > ~/clawd/memory/USER.md << 'EOF' # User Profile: [Name] ## Business Context Company: [Name] Goals: [Goal 1, Goal 2, Goal 3] ## Preferences - Best contact: [WhatsApp/Email/Slack] - Meeting times: [Mornings/Afternoons] - Tools: [CRM, Calendar, PM tool] EOF Step 2: Load Protocol (Bias for Action) At session start — or when identity feels "fuzzy": "Read SOUL.md and USER.md. Summarize: (1) who you are, (2) who I am, (3) what we're working on." *Takes 10 seconds. Prevents 2 hours of confusion.* *Pro tip:* Add to your OpenClaw config for auto-load: session_start: actions: - read_file: "memory/SOUL.md" - read_file: "memory/USER.md" Step 3: Project-Based Architecture (Think Big)

For those of us running MacOS: 🖥️ Anitvirus : https://www.avast.com/ Malware Scanner and Cleaner : https://www.malwarebytes.com/ Linux / Ubuntu : (Researching) List your GoTo's Below