A notebook analysis where no step is redone

Data, ML and researchData scientist working in Jupyter with an agent

We were analysing churn in a 60-cell notebook over a week. Each session ended mid-way: features built, model not trained, or plots half done. Every new session re-read the whole notebook to work out where we were, about 40k tokens each time.

How do I stop my AI agent re-reading and redoing a Jupyter analysis every session?

With ArcRoom, save the findings that matter with remember and finished parts with complete_task, so a new session reuses them instead of rebuilding them. It also gets a last-run note: what was asked, what was edited and whether it stopped mid-way, so it opens only the cells it needs. The note lists files and steps, not cell outputs, so a restarted kernel still has to re-run its cells.

Also asked as: Each new agent session re-reads my entire notebook. How do I stop wasting tokens? · How do I continue a half-finished data analysis in a new AI session? · Can an AI agent remember where I stopped in a Jupyter notebook?

The work

ArcRoom records which notebooks and data files each session touched, never their contents.

ArcRoom keeps
  • Last-run note
  • Saved decision
The next agent gets

your last session in this folder (2 days ago) stopped mid-way.

Step by step

  1. ArcRoom records which notebooks and data files each session touched, never their contents.
  2. A new session in the folder gets a last-run note: what was asked, what was edited, whether it stopped mid-way.
  3. Findings that matter later ('tenure is the top feature, contract type second') are saved with remember.
  4. The agent opens only the cells it needs instead of the whole notebook.

What the agent sees

ArcRoom: your last session in this folder (2 days ago) stopped mid-way.
Asked: "train a gradient boosting model on the churn features"
Edited: churn.ipynb, features/churn_v2.parquet
Note: tenure top feature, contract type second

Compared with today's tools

Markdown cells at the top of the notebook

Write a 'where I am' cell at the end of each session and ask the agent to read it first.

Setup
None
Ongoing
write the status cell each session, ~5 min

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init)
Ongoing
None

A status cell works if you always write it; ArcRoom writes the note from what the session actually did, even when it stopped without warning.

Where it stops

The note lists files and steps, not cell outputs or kernel state. A restarted kernel still has to re-run its cells.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in data, ml and research