Skip to main content

Command Palette

Search for a command to run...

Temperature, System Prompts, and Why AI Has No Memory: The "Personality" of LLMs

Updated
•2 min read
Temperature, System Prompts, and Why AI Has No Memory: The "Personality" of LLMs
G
I'm a Senior Software developer who loves solving real-world problems and building meaningful products 💡 I currently focus on crafting clean, user-friendly experiences using React Native ⚛️ I enjoy working on challenging projects and constantly learning new things — whether it’s exploring a new framework or diving deeper into existing ones. This space is where I share my journey, the issues I tackle, and the lessons I pick up along the way 🚀

Context

I used to think LLMs were like databases—you ask a question, you get the stored answer. But after playing with Temperature and System Prompts, I realized they are more like improv actors. They need a script and a "mood" setting to stay on track.

LLM is Stateless. It’s like a gold-fish; every time you hit "Enter," it’s seeing you for the first time. To make it "remember," we have to feed it its own history. we can control its "vibe" using Temperature and give it a permanent identity using System Prompts.

Temperature: The Creativity Dial

  • Low Temperature (0.1 - 0.3) : The AI is boring but reliable. It picks the most likely word every time.

  • High Temperature (0.7 - 1.0+) : The AI takes risks. It picks less likely words, leading to "creative" (and sometimes weird) outputs.

System Prompts: The Background Instructions

It’s a hidden instruction we send before the user even types a word.

  • Example: "You are a professional React Native expert. Keep answers short and use TypeScript."

  • The AI uses this as a North Star for every following response.

Stateless Nature: The "No Memory" Problem

LLMs don't have a database of your past conversations. If you say "Hi, I'm Govind" in Prompt 1, and "What's my name?" in Prompt 2, the LLM will say "I don't know."

The fix: We have to manually send the previous chat history back to the model with every new message.

In mobile apps, Statelessness is a challenge for local storage. Since we have to send the whole conversation history back every time to keep the AI "smart," our local state management becomes the AI's "memory." We have to be careful—sending too much history will hit that Context Window limit.

Building with AI is less like traditional programming and more like managing a very talented, very forgetful intern. You have to give them clear instructions (System Prompts) and decide how much caffeine they’ve had (Temperature).

AI for Mobile Developers: Learning Local LLMs

Part 3 of 7

AI for Mobile Developers: Learning Local LLMs is a public learning journey documenting how a React Native developer explores practical AI integration for mobile apps. This series focuses on understanding how Large Language Models work and how they can run directly on mobile devices using local inference. Instead of deep AI theory, the goal is to learn from a developer perspective — experimenting with tools, running models locally, and eventually integrating AI features inside mobile applications.

Up next

GGUF, Quantization, and Pruning: The Three Keys to "Shrinking" an AI Brain

I used to think that "smaller model" just meant "worse model." But today I learned that there are two separate ways to make an AI fit on a phone: you can make its memory less precise (Quantization), o