| name | tm-import-memory |
| description | Use when the user wants to bulk-import external docs (docs 슬롯에 연결된 문서 서비스, docs slot connected document service) into team memory — "메모리 업로드", "문서 메모리로 옮겨", "메모리로 가져와/시드", "이 페이지(들) 메모리에 추가해" (including when the user refers to the connected document service by name, for example "노션 메모리 가져와"); English triggers include "upload docs to memory", "move docs into memory", "import/seed memory", "add this page/these pages to memory", and "import docs memory". Single-item CRUD from conversation belongs to tm-manage-memory. |
tm-import-memory — External Docs -> Team Memory Upload (Import)
Scrape pages from the document service connected to the docs slot, organize them by topic, and save them into team memory/.
This skill handles both cold-start seeding and later incremental reruns. This skill makes the judgment; the engine memory write performs the save.
Role vocabulary principle: This skill speaks only in terms of the "docs slot". The actual service is determined by services.docs.provider in team.config.json; read that provider name and use it in user-facing guidance and source attribution (do not hardcode a specific product name).
Boundary (With tm-manage-memory)
| Origin | Skill |
|---|
| External docs (docs slot), bulk, tree-shaped import | tm-import-memory (this skill) |
| Single-item knowledge/decisions from conversation | tm-manage-memory |
Reruns such as "add this page" also use this skill. Because filenames are chosen from the approved destination topic, the same topic converges into the same file and stays idempotent (see the partial rerun rule in section 4).
Procedure
0. Check the docs slot
Check services.docs in team.config.json and the corresponding MCP connected to the agent.
If it is not connected: respond in the user's language and say that the document service is not connected yet, and that tm-connect can connect it; then stop. Connecting the service belongs to tm-connect.
1. Scope discovery — do not read page bodies
From the link the user provided (or from workspace search), identify only the page list/tree. A hub-page fetch returns the child-page link list, so it can be enumerated without reading bodies.
Default cap: 20 pages and depth 2. For anything beyond that, show only the list and proceed only with the pages the user chooses. "All of it" is an unbounded run; do not proceed without list confirmation.
2. Preview confirmation gate (required — this one confirmation also approves the weight and route explanation in bulk)
Respond in the user's language. Before fan-out, show a plan table and get one confirmation:
| Source page | -> Destination (folder/filename) | Proposed weight | Rationale |
|---|
- Default weight is 📎(reference). Propose 📌 only for clear decisions/rules/operating principles. Never auto-propose 🔥 (the team can promote core items later). The point of the weight convention is "no silent confirmation"; once the user approves this table, confirmation has been received, and you do not ask separately per file.
- Prefer existing structure for destinations (
product/<product>/..., team/decisions, etc.). If a new top-level folder is needed (for example fundraise/), include the path plus a one-line routing map description (desc) in the table and approve it together. Engine route upsert must not guess --desc (required argument), so it must be settled here.
- State together: "Up to N files will be created/updated, with one commit/push attempt per file."
- Merge by topic. Do not create one file per page; aim for about 10 generated files or fewer.
3. Per-page fan-out (subagents)
Dispatch subagents in parallel, one per page. Each subagent:
- Reads only one page body, its own page. This is the reason for fan-out: large external docs should not pollute the main context.
- Returns only a structured result containing summary, topic tags, and source (page title and URL).
- Does not write files directly.
4. Aggregate and save (main)
If a new top-level folder was approved, register it first (all arguments are required; a bare call exits 2):
python3 infra/teammode.py memory route upsert --root <team-root> \
--path <folder>/ --desc "<one-line description approved in preview>" --author <member-name>
After merging and deduplicating subagent results by topic, save each file through the engine verb:
python3 infra/teammode.py memory write --root <team-root> \
--folder <folder> --filename <kebab-name>.md \
--content "<body>
## Sources
- [Page title](URL) (<docs provider name>, collected YYYY-MM-DD)" \
--author <member-name> --weight <approved weight>
- The body must include a
## Sources section at the bottom to mark external provenance. Use the value from services.docs.provider as the provider name.
- Partial rerun merge rule: if the target topic file already exists, read the existing file first, update only the rerun source/page portion, and build the full body while preserving content from other sources and the
## Sources list. Engine write is a full replace, so if you do not pass the entire merged result, content from other sources is lost. Do not create a new file from only the rerun page either (that violates the no 1:1 rule).
- ⛔ Do not directly Edit/Write
memory/; INDEX registration, frontmatter, backlinks, and commits are all handled by the engine.
5. Completion report
Respond in the user's language. Report the created/updated file list, INDEX reflection, and push success/failure separately (engine push failure is non-blocking; the local commit is preserved).
No need to collect re-onboarding/team members; they receive it through git pull.
What This Does Not Do
- Token issuance/connection (-> tm-connect) or single conversation-memory CRUD (-> tm-manage-memory)
- Source-document edits (read-only) or continuous automatic sync (v0.2+)
- Ignoring caps to collect "everything"; do not proceed without list confirmation
Common Mistakes
| Mistake | Correct approach |
|---|
| Main agent reads all page bodies directly | Only subagents read bodies; main agent handles list and aggregation |
| Asking about weight per file | The single preview table confirmation approves them in bulk |
| Page:file 1:1 dump | Merge by topic, about 10 files or fewer |
Directly editing memory/ | Always go through memory write |
| Writing directly to a new top-level folder | Register first with route upsert (all of --path, --desc, and --author) |
| Replacing a whole merged file during a partial rerun | Read the existing file and save the full body with only that source updated |
| Hardcoding guidance for a specific product name | Read services.docs.provider and speak using that name |
