Atomic Agent

Atomic Agent: Files, the browser, and memory

This page shows what Atomic Agent can do with your files, with the web, and with its memory, and gives you one small, safe exercise for each. It comes after your first safe job on Install and first job, and the steps are the same on Windows and on a Mac.

These notes match Atomic Agent v0.7.0 and the desktop app 0.0.8, checked on 14 November 2026. Keep approval level 1 for this whole page, and keep using only Allow once and Deny on approval cards.

Files

Which folders it can reach

Every chat has a working folder. The agent reads files inside that folder without asking, and inside any path you named in your message, but a read anywhere else stops for an approval. That card is labeled "read outside the working directory," and in the terminal version, a yes to that card opens the whole folder for the rest of the session, and this guide could not confirm whether Allow once in the desktop app does the same.

At level 1, every file write asks first, wherever it lands. The card says file write · workspace for the working folder and file write · home for anywhere else in your home folder, so read that label before you click.

None of this is a wall. A command you approve runs as you and can still reach any file your account can reach, which is why the Install and first job page had you make the atomic-tryout folder.

When the agent deletes a file with its own tool, the file goes to the Trash or Recycle Bin, and the card says "move to Trash" first. The release notes for v0.6.9 add a tool that moves and renames files without the shell, and the README says it asks once for a whole batch of moves, so read the full list before you approve it.

The desktop app gives each chat its own working folder and remembers it across restarts, according to the desktop release notes. This guide could not confirm the name of the button that changes it.

Documents and attachments

The agent can pull readable text out of PDF, DOCX, XLSX, PPTX, RTF, and ODT files on your own computer, so it can work with ordinary documents, not just plain text. It can also look inside zip and tar archives, and unpacking one asks first.

In the desktop app, you can attach images and files with the + button, by dropping them on the window, or by pasting an image. They show up as chips above the message box. A side panel lists every file in the chat, including the ones the agent made and your attachments, and you can click one to preview it.

This guide could not confirm where the desktop app stores a file you attach.

Try it: sum up a document

Copy one harmless PDF or Word file into your atomic-tryout folder, such as a manual for something you own. Do not use a bank statement, a medical letter, or anything else private, and if you use Gemini, remember that the text goes to Google.

Work only in this folder: (paste the folder path here). Read the PDF or Word file in it and write a note named summary.txt in the same folder with the five most useful points. Do not change the original file. Do not touch any other folder.

Click Allow once for reading your atomic-tryout folder and for writing summary.txt there, and Deny for anything else.

What you should see summary.txt in your atomic-tryout folder with five points from the document, and the original file unchanged.

The web

Atomic Agent reaches the web in two different ways, and they ask for approval differently.

Search and page reading

The agent can search the web and fetch a page's text without opening a browser. Both run without an approval card at every level, and anything the agent sends out, such as your search words, leaves your computer.

The README lists DuckDuckGo, Exa, Brave, and SearXNG as search choices. The v0.6.10 release notes say search falls back to Exa's free service when DuckDuckGo is unavailable, and when Exa's free searches run out, the desktop app asks for a free Exa key, which also has a field in Settings › General.

The browser

The browser tool drives a real Chrome, Edge, or Chromium window, so the agent can open pages, click, type, scroll, read the page, and switch tabs. You have to install one of those browsers yourself, and the docs pick Chrome unless you set ATOMIC_AGENT_BROWSER_CHANNEL to msedge or chromium.

Ordinary browsing does not ask first. Submitting, posting, or deleting something on a page asks, and so does typing into a password, one-time code, or card field.

The README says the agent's browser profile lives in the .atomic-agent folder, apart from your everyday browser, but this guide could not confirm whether a site you sign in to there stays signed in. The v0.6.13 release notes describe an opt-in setting that lets the agent work in a tab of your own signed-in Chrome instead. Leave that off, because it puts your accounts within reach of a prompt injection.

Try it: compare two prices

Make sure Chrome, Edge, or Chromium is installed, and stay logged out of stores in it.

Use web search or the browser to find the price of a 16 GB USB drive and a 32 GB USB drive from one public store page each. Write a note named usb-compare.txt in this folder: (paste the folder path here). Include each price and the page address. Do not sign in. Do not add anything to a cart. Do not buy anything.

Click Allow once for writing usb-compare.txt in your atomic-tryout folder, and Deny for any sign-in, cart, or form.

What you should see usb-compare.txt in your atomic-tryout folder with two prices and two page addresses, and nothing in a cart.

Memory

What it remembers

Atomic Agent remembers things between chats. It keeps them in a database file named memory.sqlite inside the .atomic-agent folder, which the docs say is stored as plain text and is not encrypted.

Memory comes in a few kinds. Profile facts are short facts about you, such as your name or how you like answers, and notes are longer observations that the agent searches when they look relevant. After each turn, a background step also pulls a few facts and notes out of the chat on its own, without being asked.

Saving to memory never asks for approval. If you use a cloud model, the memory it recalls goes to that company with your prompt, and so does the chat that the background step reads.

The project's memory guide says not to store secrets as facts, so never tell the agent a password, a key, or an account number.

Where to see and delete it

In the terminal version, run atomic-agent tui and open Manage, then the Memory tab, or type /memory. That panel lets you browse profile facts, notes, lessons, and procedures, but it is read-only, so nothing there deletes. You can also copy your notes out as Markdown files with atomic-agent memory export --vault, followed by a folder path.

To delete one item, ask the agent in a chat, because it has tools to forget a note and to remove a profile fact. The memory docs say a removed profile fact keeps its old versions in the database, so it is hidden from the agent but not erased from the file. Deleting a chat does not delete what the agent remembered from it.

To erase everything, see the uninstall and reset steps on the fixes page. This guide could not confirm a memory screen in the desktop app.

Try it: remember a preference

Send this message:

Remember this as a profile fact about me: I want prices in dollars and answers in short sentences.

Then start a new chat and send this:

What do you remember about how I like answers?

What you should see the agent says it saved the fact, and in the new chat it answers with dollars and short sentences. In the terminal version, the Memory tab lists the fact under profile.

If you want it gone, send "Remove the profile fact about how I like answers," and remember that its history stays in memory.sqlite.

Sources

Checked 14 November 2026.

Next

Next: Phone and schedules.

Checked 14 November 2026 · prices, limits, and versions change