How to Implement Chat History in Gemini AI API

In this tutorial, I will show you how to implement chat history in the Gemini AI API using JavaScript and Node.js. This allows the API to remember past conversations, enabling context-aware interactions. We'll start by setting up the project and then explore various methods to manage chat history efficiently.

First, visit Google AI Studio at astudio.google.com to access the documentation. Under text generation, find the sample code for multi-turn conversations. Copy this code and set up a basic Node.js file. You'll need an API key from Google AI Studio—create one in your project and keep it secure, ideally in an environment variable. Install the required npm package with npm i @google/generative-ai. To avoid warnings, add "type": "module" to your package.json file.

Next, we'll convert the code into an API endpoint using Express. Install Express with npm i express. Import Express and set up a basic server. Create a POST endpoint at /chat that accepts a prompt in JSON format. Inside the endpoint, use the Gemini AI model to send the prompt and return the response. Start the server on port 3000. Test it with Postman by sending a POST request to http://localhost:3000/chat with a JSON body containing the prompt.

To implement chat history, we'll use a global list to store prompts and responses. Each time a user sends a prompt and receives a response, append both to the list in the correct format: role "user" for the prompt and role "model" for the response. Pass this list to the model's history parameter so it retains context. Note that the Gemini SDK automatically appends prompts and responses, but you can override this by spreading the list.

However, storing full prompts and responses can increase token usage and costs. To optimize, consider these methods: First, trim prompts and responses using slice to store only the first 50 or 100 characters. Second, limit output tokens by setting maxOutputTokens in the model configuration. Third, remove old chat history entries when the list exceeds a certain length, such as 50, by deleting the first 10 elements. This ensures the model isn't overloaded with outdated context.

To clear chat history, create another endpoint at /clear that resets the global list to empty. For multiple users, use a map instead of a list. Each user has a unique ID, and the map stores individual chat histories as key-value pairs. When a user sends a prompt, check if their ID exists in the map; if not, initialize an empty list. Then, append prompts and responses to that user's list. This prevents data mixing between users. Similarly, create a clear endpoint that accepts a user ID to reset only that user's history.

For testing, use Postman to send prompts with user IDs. Verify that each user's chat history is maintained separately and that clearing works per user. This method is suitable for up to 100 users; for larger scales, use a database for permanent storage. Remember to handle errors and validate inputs in production.

In summary, implementing chat history in Gemini AI involves setting up an Express server, managing storage efficiently, and optimizing for multiple users. These techniques help build scalable, context-aware AI applications.