Agent Knowledge
Keep shared notes and skills that the Judgment Agent reads every time it runs.
Agent Knowledge is a set of text files per project for context the Judgment Agent can’t get from traces, like what your team counts as a large order. The agent reads it in chats, in Scheduled Tasks and as the Slack Agent. Everyone in the project can read the files, and your coding agent can read and write them through MCP.
Open it
Open Settings, pick the project, then select Agent Knowledge. The page is titled Knowledge.
Kinds of files
| Where | What goes there | Who writes it |
|---|---|---|
| Any file or folder you create | Team context: a glossary, what “good” looks like, known incidents, escalation rules | You |
memory/ |
Notes the agent keeps for itself, for example what it learned during an investigation | The agent, mostly |
skills/ |
Repeatable instructions. Each file becomes a / command in the Judgment Agent. |
You |
memory/, skills/, and judges/ are reserved folders and can’t be deleted. Don’t put API keys or other secrets in any file, because everyone on the project can read them.
Add a file
- In the file tree, select New file (or New folder to group files).
- Name it, for example
lunch-order glossary. - Write the content, then select Save or press Cmd+S (Ctrl+S on Windows and Linux). Save stays disabled until the file has a name and unsaved changes.
A glossary for the lunch-order agent:
# Lunch-order glossary
- Large order: more than $500. Medium: $200 to $500. Small: under $200.
- A "dropped restaurant" means get_menu failed and the agent continued without it.
Cedar Grill's menu API has returned 504s on and off since March; treat those as
a known vendor issue; the agent should still tell the user.
- Orders for 25+ people must be checked with check_catering_capacity.
Turn a question into a skill
Put a file in skills/ and it becomes a slash command. A file named skills/weekly-lunch-review runs as /weekly-lunch-review:
Review the last 7 days of lunch-order-agent traffic in production.
1. Count requests and the share labeled Negative by the User sentiment judge.
2. List the top 3 failing tools with one example trace each.
3. Compare large orders with last week.
Answer in a short table, then three bullets on what changed.
Type / in the Judgment Agent’s composer to see your skills. Skills are also a good fit for Scheduled Tasks: the same instructions, run on a schedule.
Use a file in a chat
The agent picks Knowledge files by their names and paths, so give each file a name that says what’s in it. To make sure it uses a particular file, type @, select Knowledge, and select the file.
Organize and export
- To rename a file, edit its name in the editor and save. Folders have a Rename action.
- Delete a file or folder from its menu. Deleting a folder deletes everything in it.
- Export memory files downloads every file as one JSONL file. It contains Knowledge files, not chat history.
From your coding agent
With the Judgment MCP server connected, your coding agent can use the same files:
| Tool | Role | Use |
|---|---|---|
list_memory_entries |
Viewer | List file and folder paths |
search_memory_files |
Viewer | Search by path and content (MCP only; the Judgment Agent doesn’t have this tool) |
read_memory_files |
Viewer | Read files by ID or path |
write_memory_file |
Developer | Create or update a file by path |