> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentdesk.team/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Context

> Inspect, edit, and reset the files that shape an agent's behavior and memory.

Agent Context contains the agent's persistent instructions, identity, and memory. Use it for behavior that should follow the agent across work. Changes apply to its next runs; saving a context file does not rewrite an answer already produced.

## Open Manage contexts

1. Open **Agent Chat → Agents**.
2. Open the action menu beside the active agent, where **Edit profile** is also available.
3. Choose **Manage contexts**.
4. Check **Manage contexts · \[agent name]**, then select a file tab.

This is a separate menu action beside Edit profile. Access to these context files requires agent management permission; view-only profile access does not imply access to the context editor.

## Choose the right file

| File | What belongs here |
| - | - |
| **AGENTS.md** | Durable working instructions: how the agent approaches tasks, checks results, and follows team conventions. |
| **SOUL.md** | Personality, communication style, and behavioral principles. |
| **IDENTITY.md** | The agent's identity and role description. Change its display name or avatar through Edit profile. |
| **MEMORY.md** | Long-term knowledge and decisions the agent should retain. The agent also updates this file itself. |

The editor exposes these four files. **TEAM.md** and **AGENTDESK.md** are managed by AgentDesk and are not listed for editing here.

Context text guides behavior. It does not change the agent's fixed role template, grant permissions, connect a provider, or install a skill. Put shared project instructions in [PROJECT.md](/user-guide/projects), and reusable procedures in [skills](/user-guide/agent-skills).

## Edit and save

1. Read the current file before editing, particularly MEMORY.md.
2. Make a focused change in the text editor.
3. Choose **Save** for the selected file and wait for its saved confirmation.
4. Save each changed file separately before closing the dialog.

An unsaved-change indicator marks modified tabs. Switching tabs keeps those drafts in the dialog; saving one file does not save every tab. Closing with unsaved changes offers **Keep editing** or **Discard**.

Files must be text and no larger than **256 KiB**. If a save fails, review the displayed error before retrying.

For example, an AGENTS.md addition could say:

```markdown theme={null}
When summarizing project progress:
- Link the source updates.
- Separate confirmed results from pending work.
- End with blockers and the next action.
```

Keep instructions clear and consistent with the workspace's rules. Store credentials through the appropriate connection or secret workflow rather than in context text.

## Handle a changed file

A person or the agent may update a file while you have it open. Saving checks the version originally loaded and reports a conflict instead of silently overwriting a newer version.

1. Keep a copy of the changes you intended to make.
2. Choose **Reload file**.
3. Compare the current content, including any agent-written memory.
4. Apply only the change still needed and save again.

Review unsaved drafts in other tabs before reloading the files.

## Reset from a role template

**Reset** restores the selected file from the agent's role template when a template is available.

1. Select the file and choose **Reset**.
2. Read the confirmation: the current contents will be replaced.
3. Confirm and wait for the restored message.
4. Read the restored content before the agent's next work.

Reset is a template restore, rather than an undo history. If there is no template for the selected file, the control is unavailable.

MEMORY.md has an empty memory baseline when the role does not provide one; resetting it can clear accumulated knowledge. IDENTITY.md is restored with the agent's current name and role. Resetting these files does not change the assigned role template.

## Check the effect

After saving, use the agent's next run to check the intended behavior. Give it a small request that exercises the instruction, then inspect its response and evidence. Context changes do not guarantee a task will succeed when required tools, permissions, or external connections are unavailable.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.