How Do You Back Up a Private AI System?
Updated November 2, 2026
You back up a private AI system by copying the directory where its models live, saving its configuration, and keeping a list of the models installed, so the whole setup can be rebuilt on another machine or after a disk failure. A private AI system has no provider cloud copy to fall back on, so if the local files are lost, the models must be downloaded again.
The good news is that the parts worth backing up are ordinary files in known locations. The models are the large part, the configuration is the small part, and your own documents and prompts are a separate part that depends on the app you use to chat.
Where the models live
Ollama, the most common local runtime, stores its models in a single directory that depends on the operating system. On macOS the directory is the .ollama/models folder in the user's home directory, on Linux it is /usr/share/ollama/.ollama/models for the system service, and on Windows it is the .ollama/models folder inside the user's profile.
Inside that directory, the model weights sit in a blobs folder and the list of layers for each model lives in a manifests folder, so a backup must copy the whole directory, not just one file. If a different directory needs to be used, the OLLAMA_MODELS environment variable points the runtime at the chosen directory, which is also how you restore to a new location.
Check what you have before backing up by listing the installed models. That list is your inventory, and it tells you whether the restore worked, because the restored machine should show the same models.
What to copy, and what to skip
Copy the entire models directory, including both the blobs and manifests folders, and copy the runtime configuration, such as the server settings file that holds options like the cloud-features switch. Those two pieces are enough to rebuild the runtime and its models on another machine.
Skip nothing inside the models directory, because the manifests and blobs work as a pair, and a partial copy may leave a model unloadable. If disk space is tight, a reasonable alternative is to back up the configuration and the model list, and accept that the models themselves will be re-downloaded, since model files are publicly available and do not contain your personal data.
Your conversations are a separate question. A local runtime serves models but does not necessarily keep a chat archive, so whether your chat history is backed up depends on the app you chat with. Check where that app stores its data, and back that location up separately if the history matters to you.
A backup routine that works
Start by stopping the runtime or confirming no model is being downloaded, so the copy is not taken mid-write. Copy the models directory to an external drive or a network location, and copy the configuration file alongside it, with a dated note of which machine and runtime version the backup came from.
Verify the copy by comparing sizes or checksums, and by restoring it to a test location and pointing the runtime at it with the OLLAMA_MODELS variable. A backup that has never been restored is a guess, not a backup, and the time to discover a problem is during a calm test, not during a disk failure.
Restoring after a loss
To restore, install the runtime on the replacement machine, copy the models directory back to its default location or to a location of your choice, and set the OLLAMA_MODELS variable if you chose a non-default location. Then list the models and run one, to confirm the restore is complete and the model loads.
If the models directory was not backed up, the fallback is to re-download each model from the model list you saved. That path costs time and bandwidth but loses nothing personal, because the model files are identical for every user.
What cannot be re-downloaded is your own data: your configuration choices, your custom model files if you created any, and your chat history if your app stored it locally. Those are the items that make the backup routine worth doing, and they are the items to check first after a restore.
Frequently asked questions
Where are Ollama models stored for backup?
On macOS in ~/.ollama/models, on Linux in /usr/share/ollama/.ollama/models, and on Windows in C:\Users\%username%\.ollama\models. Copy the whole directory, including blobs and manifests.
Can I restore a backup to a different folder?
Yes. Set the OLLAMA_MODELS environment variable to the chosen directory, and the runtime will use that location for its models.
Do I need to back up my chat history separately?
Yes, if it matters to you. The runtime serves models but does not necessarily archive chats, so your chat app's own storage location must be backed up on its own.
Sources
Sources checked November 2, 2026. Where a detail could not be confirmed from an official source, this page says unknown rather than guessing.