Atomic Agent

Atomic Agent: Persona and rules

Atomic Agent does not give you a persona file, a system prompt box, or a prompt folder. Its persona is built into the program, and personal preferences go in memory as pinned profile facts. On cloud-model chats, an AGENTS.md file in your folder can also add project rules once you trust that folder.

What Atomic Agent calls it

The architecture page calls the top of every prompt a stable prefix. That prefix holds a built-in persona and interaction rules alongside the tool list and the skill list. The project's prompt code names that persona DEFAULT_SYSTEM_PERSONA and keeps it fixed across a session. This guide could not confirm any screen, file, or command that replaces that built-in persona.

Preferences about you live in memory.sqlite as pinned profile facts, which Atomic Agent adds on later turns. Cloud-model chats, such as Google Gemini, OpenRouter, or a Claude or ChatGPT plan, also read project rules from the working folder. Once you trust that folder, AGENTS.md wins over AGENT.md, CLAUDE.md, .claude/CLAUDE.md, and rules written for Gemini CLI, Cursor, or Copilot. Local-model chats, including Atomic Agent's own local model, Ollama, and LM Studio, skip AGENTS.md and read no rules file.

Set a persona

There is no persona file to open. Keep the built-in persona intact and save your own preferences as a pinned profile fact.

  1. Leave the installed program files alone, and do not edit them to rewrite the persona. Updates replace those files, and the built-in persona sits next to the lines the model needs for tool calls.

What you should see the program folder is unchanged, and Atomic Agent still opens as before.

  1. Ask Atomic Agent in a chat to save a pinned profile fact with your name and a wish for short answers. The tips page says a pinned fact renders on every turn, and saving to memory never asks for approval. Never put a password, a key, or an account number in a fact.

What you should see the agent confirms it saved the fact, and Manage, then Memory (or /memory), lists it on a read-only panel.

Set rules

How you set standing rules depends on whether the chat uses a cloud model or a local model. Neither one replaces approval cards.

  1. On a cloud model, copy any AGENTS.md already in your work folder before you edit it. Create or edit AGENTS.md inside that folder with short rules phrased as things to do. Name the folder the job may use, and never tell the agent to skip approvals or ignore a safety setting.
# Rules

Work only inside this folder. Ask before you delete, move, or overwrite any file. Say when you are unsure.

Open a cloud chat in that folder in the desktop app or the terminal text screen, and Atomic Agent asks once whether to trust it. Approve that prompt only for a folder you made yourself, and decline it for any folder you downloaded from someone else. You can change that choice later under Settings › Skills in the desktop app, or with atomic-agent skill project on --workspace and off followed by the folder path. Headless runs (atomic-agent run, phone chats, and background tasks) never ask and skip untrusted files.

What you should see a one-time prompt asking whether to trust AGENTS.md in your folder, and the cloud chat following those rules after you accept.

  1. On a local model, Atomic Agent does not load AGENTS.md. Write the folder limit in each chat prompt instead, and keep approval level 1 on. Saved tool rules in approval-rules.json inside .atomic-agent (or inside a project's .atomic-agent folder, which can only add ask and deny rules) control approval prompts rather than voice. Leave those tool rules for Advanced features.

What you should see risky actions still show an approval card, and you did not add an allow rule while setting a preference.

This guide could not confirm a separate persona for phone chats, for sub-agents, or for one project folder. Profile facts in memory.sqlite are shared across every chat, except that Fusion worker turns do not read memory.

What not to edit

Do not edit the built-in persona, rules block, or tool instructions inside the installed program. Those lines control how the model formats tool calls, and an update overwrites them. Keep your preferences in a profile fact inside memory.sqlite. If a shared Git project already includes its own AGENTS.md, read that file carefully before you trust the folder.

How to undo

Copy the .atomic-agent folder first, as Fixes describes before an update, and keep a backup of any AGENTS.md file you edited. Ask the agent in a chat to remove the profile fact. The memory docs say a removed fact is hidden from the agent while older versions stay in memory.sqlite. To stop using a folder's AGENTS.md on cloud chats, rename AGENTS.md after keeping your copy, or turn trust off in Settings › Skills (or with atomic-agent skill project off --workspace and the folder path).

What you should see the Memory panel no longer lists the removed fact, and a new chat no longer uses the folder rules you removed or untrusted.

Rules are not a lock

Neither a profile fact nor an AGENTS.md file is a security wall. A model can ignore both, and a web page or file can try to override them through prompt injection. Approvals, a named folder, and a container are what actually limit an agent.

Profile facts and folder rules.

Sources

Checked 14 November 2026.

Next

Next: Persona and rules for every agent.

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