You're making a mistake by focusing on AI tools; you need to focus on AI memory. Your agent may have 100+ skills and be MCP connected, ok cool, but ask him what you talked about yesterday, and you'll see: "I don't have access to previous conversations." Here's how to set it up: 1/ 𝗖𝗥𝗘𝗔𝗧𝗘 𝗧𝗛𝗘 𝗠𝗘𝗠𝗢𝗥𝗬 𝗦𝗧𝗔𝗖𝗞 Create a workspace folder. inside it, make these 5 files: 2/ 𝗪𝗥𝗜𝗧𝗘 𝗦𝗢𝗨𝗟𝗺𝗱 It's your agent's identity. Open the file and write who it is. Example: "you are a builder. you write concise code. you don't say 'great question'. you have opinions. when you're unsure, you say so. you prefer shipping." Without this, every session starts with a generic assistant. 3/ 𝗪𝗥𝗜𝗧𝗘 𝗨𝗦𝗘𝗥𝗺𝗱 Tell the agent about YOU. Things it would normally have to re-learn every session. "timezone: europe. language: english for code, french for chat. prefers short answers. hates corporate speak. works on crypto + AI projects." 4/ 𝗦𝗘𝗘𝗗 𝗠𝗘𝗠𝗢𝗥𝗬𝗺𝗱 This is the big one. Start with what your agent needs to know long-term: • project names, repo URLs, deploy addresses • key decisions you've made and why • mistakes to avoid ("never use X for Y because Z") • anything you'd hate to repeat This file grows over time. The agent adds to it. You can edit it too. Think of it as a shared brain between you and your agent. 5/ 𝗧𝗘𝗟𝗟 𝗬𝗢𝗨𝗥 𝗔𝗚𝗘𝗡𝗧 𝗧𝗢 𝗨𝗦𝗘 𝗜𝗧 This is the part most people skip. In your agent's system prompt or whatever config your framework uses, add these instructions: "Every session: read SOUL, USER, and MEMORY, and memory/today before doing anything. After completing tasks, write what happened to memory/YYYY-MM-DD.md. periodically review daily files and update MEMORY with key learnings. That's it. 6/ 𝗛𝗘𝗔𝗥𝗧𝗕𝗘𝗔𝗧.𝗺𝗱 (𝗢𝗣𝗧𝗜𝗢𝗡𝗔𝗟) If your agent runs on a timer (cron, heartbeat polling), add a checklist: "Check unread emails. Check the calendar for the next 24h. Check if any deploys failed." The agent reads this periodically and acts without you asking.

An agent that remembers your last 50 decisions starts anticipating the 51st which is a category difference.