Everfern

Everfern: Files, the browser, and memory

This is page 3 of 8 for Everfern. You finished the first safe job, so Everfern is installed, your personal apps are closed, and the everfern-tryout folder holds one harmless file. This page shows what Everfern can reach on your computer, what it can do on the web, and what it remembers, with one small exercise for each.

Everfern's README says little about these three areas, so much of this page comes from its release notes and its source code on GitHub, read on 14 November 2026. The labels here were checked in the code at the v1.3.0 tag, the version you download. This guide read the code but did not run the app to check these labels. Code shows what is built, not a promise of how it behaves.

Files: what Everfern can reach

The README says Everfern reads and writes "PDFs, Word, Excel, PowerPoint, and CSVs out of the box." Its work happens in a few places, and they are walled off very differently.

Computer use is your real desktop. Clicking and typing happen on your actual screen, so anything you can open, Everfern can open.

Shell commands go to the Linux VM. In the code, a settings section named Linux VM calls it "the local Linux virtual machine (WSL 2)" on Windows. On a Mac it says "Docker is required to run the Linux VM on macOS," so there the VM is really a Docker container named everfern-ubuntu. With no VM set up, the code blocks most commands. Only a few read-only ones, such as dir, ls, and pwd, may then run on your real computer.

The VM can still see your files. On a Mac, the code starts that container with your whole /Users folder shared into it as /host/Users. On Windows, Microsoft's documentation says WSL mounts your C: drive at /mnt/c by default, and Everfern's code turns Windows paths into that form. So the VM keeps commands in a separate place, but it does not keep them away from your documents.

File tools work on the host. The code names six file tools, read, write, edit, grep, find, and ls, and says they "run on the host" rather than in the VM. A file Everfern writes with them lands on your real disk.

Commands can leave the VM with your approval. The model can ask to run a command "natively on the host machine," and it must give a reason. A card then offers Deny, Always allow, and Allow once. Approval cards for other risky actions say Reject instead of Deny and add an Allow prefix button. Pick Deny or Reject unless you understand the action. Never pick Always allow or Allow prefix during this tryout, because both save a rule that stops Everfern from asking again.

Projects are a starting point, not a wall. The chat code offers Select folder as project..., which ties a chat to one folder. Everfern's built-in instructions tell the model that reading and writing "inside the project workspace" never needs permission. They also say deleting or moving files "outside the project workspace" always needs your confirmation. Those are instructions to the model, so treat them as habits it should follow, not a lock it cannot break.

Attachments. The chat code has an Add files or photos option with the shortcut Ctrl+U. It pulls the text out of Word files and scanned pages after you attach them. This guide could not confirm whether a Mac uses the same shortcut or Command instead.

Try it: a small spreadsheet in your folder

Click Select folder as project... and choose your everfern-tryout folder before you paste anything. Without a project, Everfern's code treats your whole home folder as the workspace, and its instructions tell the model that reading and writing there needs no permission. Then paste:

Work only in my everfern-tryout folder. Make a spreadsheet file named fruit.xlsx with three fruits in one column and a price for each in a second column. Do not use computer use. Do not read, change, move, or delete any other file. If you need to run a command on my real computer, ask me first and explain why.

If a card asks to run something on the host, read the command first. Click Deny unless it only makes that one file in that folder.

What you should see fruit.xlsx in your everfern-tryout folder next to first-note.txt, and nothing else new. If the file is missing, ask Everfern in the chat for the full path it used.

The web: Navis, search, and page reading

Everfern has three ways to use the web. Any page it reads could hold a prompt injection, so all three need the same care.

Navis, the browser agent. The README says Navis "fills forms, scrapes data, and drives web apps in plain English." Its demo opens a browser, searches, and reads several sources before writing up what it found. A June 2026 release note says Navis now runs as a browser extension. The extension's own README says to install it "when you want Navis to control your normal browser profile," which means your everyday browser. In the code, the Tool Settings section has Add to Chrome and Add to Firefox buttons, and setup offers Install Navis Extension.

The code's default settings also point Navis at your Chrome profile, and they reset that choice each time the settings load. This guide found no switch that gives Navis a separate, logged-out browser. So assume Navis works in your everyday browser, with whatever you are logged into there. That is why safety switch 1 on the previous page matters so much.

Web search. In the code, Tool Settings has a Web Search section with an Execution Mode setting. Local is described as "Playwright search," and API as "External search provider." The API choice takes a key from a search company such as Exa or Firecrawl, and the screen says that key is "Stored locally in ~/.everfern/" on your computer.

Web Fetch. A July 2026 release note added Web Fetch, which pulls "any page's readable content and metadata directly into a task." The note says it is labeled, so you can see when the agent is reading the web. The same note says you can watch Navis write its report as findings.md while research runs.

A step limit. The Navis settings have a Max Steps Per Task slider that runs from 10 to 200 in the code. A May 2026 release note set it to 12, but the current code's default is 200. Check yours and keep it low for early jobs, because a lower limit makes a confused run stop sooner.

Everfern's built-in instructions tell the model to treat web content as "untrusted data" and to reject commands hidden in it. That helps, but it is a request to the model, not a wall, so stay logged out.

Try it: a public-page question

Paste this, and change the city if you want:

Use the web to find the opening hours of two public libraries in [your city]. Use only public pages. Do not sign in anywhere, do not fill in any form, and do not save any files. Give me the answer in the chat with the web address for each fact.

Keep watching while it works. Quit with your shortcut if it opens a login page or a form.

What you should see two sets of opening hours in the chat, each with a web address you can open yourself, and no form filled in and no sign-in anywhere.

Memory: what it keeps, and how to delete it

The README says Everfern "remembers your preferences across sessions" and compresses old chats "with no information loss." In the code, you see what it saved in the Memory Graph section of the settings. That section says it lets you "Manage and visualize your long-term preferences, habits, and knowledge facts."

It sorts memories into Preferences, Habits, and Facts, and you can switch between a graph and a List view. It has buttons for Add Memory, Export ZIP, and Import & Merge. To delete one memory, the code uses a Forget Memory button, which first asks, "Are you sure you want EverFern to forget this memory?"

Everfern keeps all of this as files on your computer. The Privacy & Data section in the code lists where:

The memory-saving tool in the code also writes plain notes to ~/.everfern/memory.md. The screenshots folder matters most, because computer use takes pictures of your screen, and they stay on your disk until you delete them.

The same section says "Nothing is sent to EverFern servers," but that covers only what Everfern stores. A cloud model still receives your requests. A June 2026 release note also added Thumbs Up and Thumbs Down buttons, and their feedback window lets you choose how much of the chat to share "with our servers." Skip those buttons if you want nothing to leave.

Two August 2026 releases list changes named "memory fix," so memory has had bugs. Page 6 covers wiping everything at once.

Try it: one preference

Tell it one preference:

Remember that I want answers in short sentences and prices in dollars. Do not use any tools other than memory.

Then start a new chat with New chat in the sidebar, or the shortcut the code lists as Ctrl+N, and ask:

What do you remember about how I like my answers?

What you should see the new chat repeats your preference about short sentences and dollars. In Memory Graph, the preference shows up as an entry you can remove with Forget Memory, although this guide could not confirm which of the three groups it lands in.

When the spreadsheet, the library answer, and the preference are in place, you have tried all three areas.

Sources

Opened 2026-11-14. Source-code links point to the v1.3.0 tag.

Next

Continue to Phone and schedules.

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