Give Your OpenClaw Agent a Personality and Memory

The Problem

Your agent works. You message it on Telegram, it responds. But it's generic. It doesn't know your name, doesn't remember what you asked yesterday, and speaks like every other AI chatbot. You didn't set up your own agent to get the same experience you'd get from any free chat app.

What You're Building

An agent that has a name, a communication style you defined, knowledge of who you are before you say a word, and memory that survives session resets. After this, it'll feel less like a tool and more like an assistant that actually knows you.

Verify before starting: Run openclaw gateway status — your agent should show "connected."

Milestone 1: Give Your Agent an Identity

Right now your agent is anonymous. Let's give it a name and a personality.

Update my ~/.openclaw/openclaw.json to give my main agent an identity. Set the name to "Atlas", the theme to "sharp, direct assistant who values clarity over politeness", and the emoji to "compass". Keep everything else in my config unchanged.

What Claude Code does: It adds the identity block to your agent configuration. The identity field controls how your agent presents itself — the name appears in responses, the theme shapes behavior, and the emoji shows in the dashboard. You can change "Atlas" to whatever you want — this is YOUR agent.

Try it: Message your bot on Telegram: "What's your name?" It should respond as Atlas (or whatever name you chose).

Milestone 2: Write a Custom System Prompt

Your agent has a name but no real personality. The system prompt is where you define exactly how it thinks and communicates.

Add a system prompt to my OpenClaw agent config. The prompt should tell the agent to: be direct and concise, skip pleasantries unless I initiate them, use concrete examples instead of abstract explanations, address me by my first name, and format responses for mobile (short paragraphs, no walls of text). Write it as a natural instruction, not a robotic ruleset.

What Claude Code does: It writes a system prompt that shapes every response your agent gives. The system prompt is the DNA of your agent — it's loaded before every conversation and defines the default behavior. Writing it as natural instructions (not rigid rules) produces better results because the AI interprets intent, not just keywords.

Try it: Ask your bot to explain something: "Explain what a cron job is." The answer should be direct, use an example, and feel different from a generic ChatGPT response.

Milestone 3: Add Long-Term Memory

Your agent has personality but amnesia. It doesn't know who you are, what you do, or what matters to you. Time to fix that.

Create a MEMORY.md file in my OpenClaw workspace at ~/.openclaw/workspace/MEMORY.md. Include key facts about me — my first name, my role, the projects I'm currently working on, my communication preferences, and what tools I use daily. Replace the example details with my real info. Format it as clean markdown sections.

What Claude Code does: It creates the long-term memory file that your agent loads at the start of every session. MEMORY.md is curated, stable knowledge — things that are always true about you. Unlike daily memory logs (which capture conversations), MEMORY.md is hand-crafted context you control. Your agent reads it before every conversation, so it always knows who you are.

Try it: Message your bot: "What do you know about me?" It should reference your name, role, and projects from MEMORY.md.

Milestone 4: Verify Memory Persists

The real test: does your agent remember you after a reset?

Reset my current OpenClaw session by sending the /reset command, then check the agent logs to confirm a new session started and MEMORY.md was loaded. Show me what the agent knows at the start of the fresh session.

What Claude Code does: It triggers a session reset and shows you the agent's initialization — including MEMORY.md being loaded into context. This proves persistence: your agent never starts from zero. Daily conversation logs (memory/YYYY-MM-DD.md) capture what you discuss, while MEMORY.md provides the stable foundation.

Try it: After the reset, message your bot: "What's my name and what do I do?" It should answer correctly — even though the conversation history was cleared. That's the power of persistent memory.

What You Built

Remember that generic chatbot that forgot you existed? It now has a name, a personality, knowledge of who you are, and memory that never resets. You built a personal assistant that:

• Introduces itself by name with a defined personality
• Communicates in the style you specified
• Knows your role, projects, and preferences before you say a word
• Remembers across session resets

The next time you open Telegram, your agent knows exactly who it's talking to.

Take It Further

• Add daily memory triggers — have the agent summarize key decisions at the end of each day
• Create topic-specific memory — separate MEMORY.md sections for different projects
• Build a multi-model brain — use different AI models for different types of questions (that's Lesson 3)