| name | team-memory-governor |
| description | Use when reading, updating, or governing long-lived national-team markdown memory files so agent runs keep reusable football context without mixing stale facts, speculative claims, and append-only observations. |
Use this skill when the task involves persistent team memory.
Purpose
The team memory layer exists so orchestrators and specialist agents can reuse durable context about a national team without rebuilding everything from scratch on every run.
These files are not free-form notes. They are governed knowledge artifacts.
File model
Each national team should have one markdown file in memory/teams/.
Every file should separate:
- canonical profile facts that can be overwritten when verified,
- current tournament snapshot facts that should be refreshed often,
- historical notes that should only change with source-backed updates,
- append-only update logs that preserve provenance and recency.
Update rules
- Overwrite stale canonical facts only when newer verified evidence exists.
- Refresh current-cycle stats whenever newer structured match or player data is available.
- Append to the memory update log instead of deleting prior update history.
- Never silently replace historical tournament claims without citing the new source and update date.
- Mark unknowns clearly instead of hallucinating.
Source discipline
- Prefer official competition or team data first.
- Preserve source URLs or source labels for every meaningful update.
- Keep timestamps in ISO format where possible.
- Distinguish observed facts from agent interpretations.
Agent behavior
When this skill is active:
- Load the relevant team memory files first.
- Compare them with the newly retrieved evidence.
- Decide which sections need overwrite vs append treatment.
- Produce a compact structured delta before any write.
- Write back only the affected team files.
Safe write policy
Never let one noisy run rewrite the whole file.
Prefer:
- targeted canonical updates,
- targeted snapshot refreshes,
- short append-only memory log entries.
Avoid:
- replacing the whole file with a model summary,
- removing provenance,
- mixing rumor with verified fact,
- storing one-off user chatter as team memory.
