# Write Memory

> Use at the end of non-trivial work in an opted-in .memobox project only when the result passes the durable-memory worthiness gate, or when the user explicitly asks to save a memory.

- Skill: `study8677/write-memory` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add study8677/write-memory`
- Raw SKILL.md: https://api.skillmd.com/api/skills/study8677/write-memory/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: study8677 (https://skillmd.com/u/study8677)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/study8677/write-memory

---


# Write MemoBox Memory

Write a Memory Mail only for a durable, verified outcome. Store decisions and reusable evidence, not a transcript dump. An existing `.memobox` directory is the opt-in signal; do not create one automatically.

## Worthiness Gate

Write only when at least one of these is true:

- A durable decision, constraint, or project convention was established.
- A non-obvious cause was found and the fix was verified.
- Exact files, commands, URLs, or other artifacts will save meaningful investigation later.
- An unresolved risk or next action must survive the current session.
- The result is a useful cross-session or cross-agent handoff.

Write nothing when the task was a simple lookup, formatting or mechanical edit, contains sensitive material that cannot be safely excluded, produced only speculation, duplicates an existing memory, or has no likely persistent value. Do not create a record merely because the task ended or merely to log that another memory was opened.

## Steps

1. Keep one coherent durable outcome per record; most tasks should write zero or one record.
2. Create a concise subject and a one-sentence summary useful from the index alone.
3. Include verified context in `--body`; put decisions, artifacts, risks, and next actions in their structured flags. When recency matters, set `--last-verified-at`; set `--valid-until` only when there is a real review horizon.
4. If this result replaces an older project memory, add one `--supersedes <id>` per replaced record. MemoBox will link the new record and mark those exact sources `stale`; review the old bodies before doing this.
5. Add one `--source-ref "memobox:<id>"` for each previously stored project memory that materially influenced this result, or `--source-ref "memobox-global:<id>"` for global memory. Do not cite ids that were merely opened.
6. Confirm `command -v memobox`. In a MemoBox source checkout only, use `PYTHONPATH=src python3 -m memobox.cli` as the fallback.
7. Run, omitting empty flags:
   ```bash
   memobox --store .memobox write \
     --subject "<short title>" \
     --summary "<index-level summary>" \
     --project "<project>" \
     --team "<team>" \
     --role "<role>" \
     --tags "<comma,separated,tags>" \
     --body "<verified context and result>" \
     --decision "<decision worth preserving>" \
     --last-verified-at "<ISO-8601 date or timezone-aware datetime>" \
     --valid-until "<ISO-8601 review horizon>" \
     --supersedes "<replaced-memory-id>" \
     --artifact "file:<exact path or evidence URI>" \
     --risk "<remaining caveat>" \
     --next-action "<follow-up if any>" \
     --source-ref "memobox:<reused-memory-id>" \
     --json
   ```
8. Confirm the returned id exists in `.memobox/mails/<id>.json`; if `--supersedes` was used, confirm the exact old records are now `stale` in the index.

Never store secrets, credentials, private tokens, personal data, confidential raw content, or unredacted logs. Use `--raw-trace-file` only when the user explicitly wants the evidence retained and it has been checked for sensitive data.

`PreCompact(auto)` is a safety net, not the primary write path. Its `needs_review` checkpoints may be partial and should not replace a deliberate, structured end-of-task Memory Mail.