Building a Local AI Agent System using Strands SDK, Ollama, Mem0, and MCP

I built this project while trying to understand how modern AI assistants actually work behind the scenes. At the beginning, it felt complicated because there were many parts working together — models, tools, memory, and communication systems. But once I started building step by step, everything started to make sense.

The main idea of my project is to create a local AI assistant that doesn’t just answer questions, but can also use tools and remember information like a real personal assistant.

I used Ollama (llama3.2:3b) as the brain of the system. It runs completely on my local machine, so I didn’t need any external API. This made the project feel more private and powerful.

Then I added different tools to make the assistant useful in real situations. For example, I built a calculator tool that can safely solve mathematical expressions. Instead of guessing answers, the AI actually calls the tool and gets the correct result.

I also added a weather tool that gives simple weather information for different cities. Even though it is simulated, it helped me understand how external APIs or services can be connected to an AI system.

One of the most interesting parts was the age calculator tool. It takes a birth date and calculates the exact age along with the number of days left for the next birthday. This showed me how AI tools can handle real-world logic tasks.

But the most powerful feature in my project is memory.

I integrated Mem0 with FAISS vector storage, which allows the assistant to remember user details permanently. When I tell the system something like my name or any personal detail, it stores that information in a vector database. Later, even in a new session, the AI can recall it by searching through memory using meaning-based similarity, not just keywords.

This made the assistant feel more human because it doesn’t forget what I told it earlier.

Along with this, I built another system using MCP (Model Context Protocol). This part helped me understand how real AI systems separate tools into a different server. Instead of keeping everything inside one program, MCP runs tools like weather, calculator, and system info in a separate process, and the AI communicates with them when needed.

This architecture showed me how scalable AI systems are designed in real companies.

Overall, this project helped me connect many important concepts like AI models, tool usage, memory systems, embeddings, and communication protocols. It was not just about writing code, but about understanding how an intelligent assistant is built from the ground up.

In the end, I didn’t just build a chatbot — I built a small AI ecosystem that runs on my own machine and behaves like a real assistant.