Customize Gemini CLI: The guide to context & memory

Understanding Context in Gemini CLI for AI Agent Interaction

Key Concepts:

• Context: Information used by an AI agent (Gemini CLI) to understand user requests and respond effectively.
• Gemini.md: Files used to store context, available globally (user level) or locally (project/directory level).
• Memory: The AI agent’s stored context, accessible via the /memory commands.
• MCP (Model Context Protocol): A method for loading external context (discussed as a future topic).
• /init command: A built-in command to automatically generate a gemini.md file for a project.
• Ancestor Context Files: Gemini CLI searches for gemini.md files in the current directory and its parent directories to build a comprehensive context.

1. The Importance of Context for AI Agents

The video emphasizes that context is paramount when working with AI agents like Gemini CLI. Just as a human assistant needs to learn preferences (like coffee habits – 12 sugars in this example), Gemini CLI requires context to function efficiently. Without context, the agent lacks the necessary information to accurately interpret and respond to queries. Context allows the agent to understand how you work, not just what you ask. As stated in the video, “Context is a collection of information an AI agent uses to understand the user's request.”

2. Forms of Context & How to Provide It

Context can take several forms:

• General Instructions: Broad guidelines for the agent’s behavior.
• Project-Specific Instructions: Details about the technologies used (e.g., Python vs. PIP, React vs. Vit) and project conventions.
• Personas: Defining a role for the agent (e.g., “You are a QA tester. Make sure to test the changes end to end after every change.”). This influences the agent’s response style and focus.

The primary method for providing context is through gemini.md files. These files are read by Gemini CLI to build its understanding of the environment.

3. Managing Context with Gemini.md Files

• Global Context: A gemini.md file at the user level stores information applicable to all projects. This is ideal for personal preferences (name, preferred communication style, quirky rules like “I only like purple on Tuesdays”).
• Project Context: gemini.md files within a project directory provide context specific to that project.
• Context Hierarchy: Gemini CLI searches for gemini.md files in the current directory and then traverses up the directory structure (ancestor context files) to gather all relevant context. Subdirectories can also have their own gemini.md files for specialized context (e.g., a testing folder with testing-specific instructions).

4. Utilizing the /memory Commands

Gemini CLI provides commands to manage its memory (stored context):

• /memory show: Displays the current contents of the agent’s memory. Initially, the memory is empty.
• /memory add: Adds information to the global gemini.md context. The video demonstrates adding a name to the memory, which is then successfully retrieved when prompted. The addition requires confirmation and automatically refreshes the agent’s memory.

5. The /init Command for Project Context Creation

To streamline the process of setting up project context, Gemini CLI offers the /init command. This command:

• Reads key project files (e.g., package.json, index.js, data files).
• Analyzes dependencies and project structure.
• Automatically generates a gemini.md file containing a project overview, prerequisites, and key commands (e.g., npm install, development server start, production build).
• The output can be expanded (using 'S') to view the detailed context being stored.

The video highlights that using /init helps “steer [the agent] down the right path without it needing to explore your code or files every single time you start a new session.”

6. Case Study: Techstack Conference Website

The video presents a real-world scenario: onboarding to a new codebase for the Techstack conference website. The team received the code and used Gemini CLI to quickly understand the project.

• They used /init to generate a gemini.md file.
• The generated file included details about the technologies used (React and Vit), the project’s folder organization, and essential commands.
• This allowed them to quickly get up to speed on the project’s structure and how to work with it.

7. Demonstrating Context in Action

After adding a name to the global context via /memory add, the agent correctly responds to the question "What is my name?" demonstrating that the context has been successfully loaded and is being utilized. Similarly, after running /init for the Techstack website project, the contents of the generated gemini.md file are visible in the /memory show output, confirming that the project context is now available to the agent.

8. Future Directions: Model Context Protocol (MCP)

The video concludes by mentioning that the next lesson will cover loading external context using the Model Context Protocol (MCP). This suggests a more advanced method for providing context beyond the gemini.md file system.

Conclusion:

The video effectively demonstrates the critical role of context in maximizing the utility of AI agents like Gemini CLI. By strategically utilizing gemini.md files, the /memory commands, and the /init command, users can equip the agent with the knowledge it needs to understand their requests, work efficiently, and provide relevant responses. The emphasis on both global and project-specific context, along with the hierarchical file structure, provides a flexible and powerful system for managing the agent’s understanding of its environment. The key takeaway is that investing time in defining context upfront will significantly improve the overall experience and productivity when working with Gemini CLI.