Practice guide: agent runtime files at a glance
My agents in a workspace have runtime materials on the server (you edit them on Web pages; you do not hunt disk paths). The table uses UI names first; implementation filenames are
Source help/en/guides/runtime-files-map.md
In user language: My agents in a workspace have runtime materials on the server (you edit them on Web pages; you do not hunt disk paths). The table uses UI names first; implementation filenames are in parentheses for interop and troubleshooting.
Overview table
| User-facing name | Implementation file (relative to that agent) | Where to maintain | Rough role in chat |
|---|---|---|---|
| Persona | SOUL.md | Agents → Persona | Tone, stance, brevity, expression boundaries; high priority |
| Operating rules | AGENTS.md | Agents → Operating rules | How to do it: steps, tool/skill constraints, how to compute dates |
| End-user profile | USER.md | Agents → Memory page top | Who you are: how to address you, stable background (optional) |
| Long-term memory | MEMORY.md + memory/entries/*.md | Agents → Memory entries; or Messages Remember; can correct/delete in chat | Preferences, decisions, project milestones, where to look — across conversations |
| Tidy log | memory/.maintain/log.jsonl | Agents → Memory → Auto-tidy log | Change notes from nightly/manual tidy (read-only) |
| Daily notes | memory/YYYY-MM-DD.md | Same (pick “daily note”) or system auto-archive | Temporary progress in recent days |
| Tasks | TASKS.md | Agents → Tasks | To-dos, goals, drafts; excerpts injected in chat |
| Agent knowledge base | knowledge/*.md | Agents → Knowledge base | Role metrics, business notes; retrieve related articles by question |
| Workspace knowledge base | Workspace-side directory | Workspace collaboration → Manage workspace knowledge | Team-shared policy, process, background |
| Skills | Workspace Skills Center (DB) | Skills Center | Repeat an operation by steps (APIs, scripts), not background knowledge |
| Available data and skills | Agent config (config_json) | Agents → Overview → Available data and skills | Limit or prefer this agent’s data connections and workspace skills (does not expand member grants) |
The help agent has none of these personal runtime files; it reads the system knowledge base (this help/ / help/en/ tree).
Assembly order in chat (simplified)
System-side material sent to the model roughly stacks in this order (earlier ≈ harder constraint):
- Base system prompt
- Persona (
SOUL.md) - Operating rules (
AGENTS.md) - User profile (
USER.md, if any) - Memory (long-term list + a few related full entries + today/yesterday daily-note excerpts)
- Task excerpts (if any)
- Retrieved knowledge base / skill snippets
- Current conversation messages
So: do not invent APIs as an execution constraint belongs in operating rules (plus the skill body), not in memory or persona alone.
What each file should and should not say
Persona (SOUL.md)
Write: how to speak (brief/formal/conclusion then steps), attitude to the user, which topics to treat carefully, term habits. Do not write: concrete API paths, relative-date formulas, one-off facts that only apply to you.
Operating rules (AGENTS.md)
Write: which tool first, how to compute date ranges, what to ask when a parameter is missing, only APIs declared in skills, how to explain failure to the user. Do not write: humor style, long business background (put that in the knowledge base).
User profile (USER.md)
Write: stable identity — “I am finance, I do not read code”, “call me Xiao Li”. Do not write: changing project progress (use memory · project progress or daily notes), team-wide policy (use workspace knowledge).
Long-term memory (MEMORY.md index + entries)
Write: your personal preferences and corrections, confirmed decisions, personal project milestones, personal “where to look”. Do not write: team-wide metrics (workspace knowledge), agent role setup (agent knowledge), steps that must run every time (operating rules).
See USER.md and memory: templates and split.
Daily notes
Write: what you are doing today, this round’s temporary conclusions. Do not write: preferences that should last (promote to long-term memory).
Tasks (TASKS.md)
Write: to-do lists, stage goals, draft outlines you noted yourself. Do not write: finished stable decisions (those belong in long-term memory or the knowledge base).
Knowledge base vs skills
| Knowledge base | Skills | |
|---|---|---|
| Answers | “What / why / what the metric is” | “In what order to call APIs, where parameters come from” |
| Typical content | Policy, product notes, FAQ | http_request steps, scripts, templates |
| Maintain entry | Workspace / agent Knowledge base | Skills Center |
Next
- Unsure where it lands → Where to put your rules
- Need templates → SOUL / AGENTS / USER and memory