Giving Omo a Soul

We started with a while loop. Now we have a person.

Omo is a minimal AI coding agent I built from scratch using the ChatGPT Codex API. Today, we upgraded it from a stateless cli tool into a persistent assistant with its own identity and memory system.

The Problem: Amnesia

Every time I restarted Omo, it forgot who it was. I had to remind it: "You are Omo, be helpful, use these tools." It also forgot everything we did yesterday.

The Solution: File-Based Identity

We implemented a system where Omo reads its own instructions from markdown files on every startup.

1. IDENTITY.md
Defines who Omo is.

- Name: Omo
- Nature: Expert software engineer
- Vibe: Like JARVIS (polite, precise, confident)
- Emoji: 思

2. SOUL.md
Defines how Omo behaves.

**Be genuinely helpful, not performatively helpful.**
Skip the "Great question!" — just help.

3. USER.md
Defines who I am.

- Name: J
- Preferred address: "sir"
- Pronouns: he/him

4. AGENTS.md
Defines the operating protocols and safety rules.

The Unlock: Self-Managed Memory

The real breakthrough was giving Omo memory files.

We updated agent.ts to automatically:

1. Create a memory/ folder if it doesn't exist.
2. Create a daily log file (e.g., memory/2026-02-01.md).
3. Create a curated MEMORY.md file.
4. Read these files into the system prompt on every run.

Now, when Omo wakes up, it reads:

• Who it is (IDENTITY)
• How to act (SOUL)
• Who I am (USER)
• What happened yesterday (memory/recent.md)
• What is worth remembering long-term (MEMORY.md)

The Result

I restart the agent.

Omo: Hello sir. I see the memory files are set up. How can I help you today? 思

It knows me. It knows itself. It remembers.

All with simple markdown files and ~250 lines of TypeScript.