How to Make Codex Understand You Better: Master These Memory Techniques for a 90% Productivity Boost

After using the Codex App for a while, you’ve likely run into some familiar frustrations: having to repeat basic preferences every time you start a new conversation, or accidentally triggering system commands you didn’t mean to.

The root cause of these common headaches is that your Codex lacks stable long-term memory. To truly eliminate this inefficient back-and-forth, you need to build a complete memory system for it.

This article will show you how to configure AGENTS.md and the Memories feature to give Codex persistent memory — enabling it to truly understand your intent and boost your efficiency.

1. AGENTS.md

1.1 What It Is

AGENTS.md is a rules file you write in advance. Codex reads it automatically every time it starts up.

Without AGENTS.md, Codex starts with a blank slate every time — it doesn’t know your work habits, your project context, or the boundaries of what it should and shouldn’t do.

Once you configure those rules, the context becomes permanent. You’ll never have to repeat the same instructions again.

1.2 How to Configure and Use It

1️⃣ Global level – stored in ~/.codex/AGENTS.md on your local machine. It applies to all your projects, perfect for general personal preferences and universal rules.

In the Codex App, this corresponds to the “Custom Instructions” section in the Personalization settings.

Here’s a basic template you can copy and paste into that location, then adjust as needed:

(insert image placeholder)

2️⃣ Project level – stored in the project directory. Customize rules based on that project’s theme and characteristics.

A good project-level AGENTS.md should typically cover:

• Project background and structure – what the project is, where important directories are
• Execution methods – how to start, run, or operate the project
• Norms and constraints – standards to follow, and clear things not to do
• Completion criteria – how to determine that a task is done

1.3 AGENTS.md Best Practices

Setting up a basic AGENTS.md is just the first step. The following principles determine the ceiling of your memory system:

1️⃣ Layer your rules with clear responsibilities

Codex’s AGENTS.md follows a proximity principle: subdirectory > root > global. In other words, rules in the current working directory take the highest priority.

This means project-level rules can directly override similar global settings. If you lump all rules together, high-priority local rules can easily pollute or break the global logic.

Therefore, a good rule memory system should follow the separation of concerns principle:

• Global layer: only general rules like language preference, safety boundaries, and basic behavior habits
• Project layer: norms specific to that project, such as build commands, code conventions, and module descriptions

Each has its own job without interfering with the other, so no conflicts arise when they merge.

2️⃣ Start simple, iterate as needed

You don’t need to write a complete AGENTS.md all at once. Start with the most essential rules. When Codex repeatedly makes the same mistake, have it review the issue and update the rules, for example:

“You just made the same mistake as last time. Please summarize the reason and add the relevant experience to AGENTS.md.”

This way, rules always come from real friction points, and AGENTS.md evolves naturally with usage.

3️⃣ Keep it concise; split complex content into separate files

As you use Codex more, AGENTS.md can grow quite long. Instead of stuffing everything into one file, split topic-specific content into standalone files and reference them from the main file.

For example, put your project architecture description into a separate ARCHITECTURE.md, and just add a single reference line in AGENTS.md:

“For the core architecture and directory conventions of this project, see the ARCHITECTURE.md file in the current directory.”

2. The Memories Feature

Beyond manually configuring rules, Codex also supports an active memory feature. Enable it in Settings → Personalization → Memories as shown below:

(insert image placeholder)

Once enabled, Codex will automatically summarize useful information and key preferences from your conversations into memory. If you want to see what Codex has remembered, open the Memory.md file located in ~/.codex/memory. Here’s what mine looks like:

(insert image placeholder)

Final Thoughts

Building a memory system is an evolving process. Trust me — if you keep refining your action rules and accumulating memories, Codex will soon become the productivity tool that truly understands you and feels effortless to use.