Autosave, checkpoints & recovering lost Jupyter/Colab notebooks

Many threads start with “I lost a day of work in Jupyter/Colab, can I get it back?”. Let’s clarify what autosave and checkpoints actually do, when you can recover work, and how to reduce the risk of losing hours of experiments.

1. How autosave works in classic Jupyter

Classic Jupyter Notebook saves the notebook file (the .ipynb) periodically. Checkpoints create backup copies you can revert to. If the browser crashes before a save, you can lose work since the last autosave.

2. How Google Colab saves notebooks

• Colab notebooks stored in Drive are synced frequently to your Drive file.
• If you close a tab without saving or lose connection, you can sometimes recover older versions via Drive’s version history.
• Execution state (variables, outputs) is separate and lost when the runtime resets.

3. Can you recover a lost notebook?

3.1 Classic Jupyter

• Look for .ipynb_checkpoints/ next to your notebook.
• Sometimes an older, autosaved version is there.
• If the notebook file itself was overwritten, recovery may be difficult without backups.

3.2 Google Colab

• Open the notebook in Drive → File → Version history → See version history.
• Check if an earlier, more complete version exists.
• Execution state (variables, trained models in RAM) cannot be recovered.

4. A safer workflow: notebook + Git + Capsules

4.1 Use version control for the notebook

Put your project in Git and commit regularly:

git init
git add notebook.ipynb
git commit -m "Baseline analysis"

4.2 Take NoteCapsule snapshots at milestones

For important points (baseline model, best model, final report), create a Capsule that includes the notebook, environment snapshot, data manifest and metadata:

from notebookcapsule import create_capsule

create_capsule(
    name="after-baseline-model",
    notebook_path="notebook.ipynb",
    data_dirs=["./data", "./models"],
    base_dir=".",   # project root
)

If autosave, checkpoints or Colab versions fail you, a Capsule gives you a structured copy of the project that can be rerun on another machine.