Open Source Agents and Local LLMs: Where Does the Storage Go?

Open-source agent frameworks have made it easy to run powerful AI workflows entirely on your own machine. No API keys. No cloud. Everything local. That feels great — until your SSD is at 10% capacity and you are trying to figure out why.

The storage is not just the model

When most people think about local LLM storage, they think about the model weights. Those are large — a 7B model is around 4 to 5 GB, a 70B model is 40 GB or more. But the model file is only one part of what a local agent workflow leaves behind.

Depending on your setup, local agents running on open-source LLMs may also store:

• Agent workspace and session history
• Tool call logs and execution traces
• Local memory and context files
• Downloaded tool dependencies and executables
• Temporary files from multi-step workflows
• Cached model outputs and intermediate results

None of these are automatically cleaned up. They just accumulate.

Multiple models for multiple agents

It is common to run different models for different tasks — a large model for reasoning, a smaller faster model for quick responses, a vision model for image tasks, a code model for programming. Each model is a separate download. They do not share files. They do not know about each other.

Across Ollama, Hugging Face, and any other backends you might use, the same conceptual workflow can involve three or four separate model downloads in three or four separate hidden directories.

Why you cannot just check folder sizes

The storage for a local agent workflow is scattered. The model weights are in one place. The agent memory might be in another. The caches might be in a third. Some tools use system directories. Some use application support folders. Some use custom paths set by environment variables.

There is no single folder you can open to see the full picture. You would have to check a dozen locations, interpret cryptographic filenames, and manually cross-reference what belongs to what.