SKILL: session-memory-writer
Use when: Writing memory/YYYY-MM-DD.md at the end of a session, or logging any subagent task outcome.
Why this exists: Memory files are the single source of truth for Langfuse backfill, Quill journal generation, Liv's CCP updates, and QMD search. Format drift = invisible tasks = broken observability. The regex parsers don't forgive variation.
1. Task Outcome Block (mandatory for every subagent completion)
Every agent task gets exactly this structure. No exceptions.
### [AgentName] Task — YYYY-MM-DD HH:MM
- **Task:** one-line description of what was asked
- **Outcome:** ✅ complete / ⚠️ partial / ❌ failed
- **Metric:** pass/fail, score, or N/A
- **Kept:** yes / no / pending-review
- **Time:** ~X minutes
- **Notes:** what worked, what didn't, key decisions, surprises
Real examples from 2026-03-24:
### Kit Task — 2026-03-24 06:45
- **Task:** Rebuild /setup page — full agent configuration wizard (5 tabs)
- **Outcome:** ✅ complete
- **Metric:** pass — clean build, 200 on live URL
- **Kept:** yes
- **Time:** ~15 min
- **Notes:** Built all 5 tabs with localStorage persistence. No features cut. Clean TypeScript build first pass.
Why the format is non-negotiable
langfuse-backfill-historical.py uses this regex pattern to find task blocks:
### [Name] Task — YYYY-MM-DD
The em-dash (—) and space on both sides are required. These all break the parser:
| ❌ Wrong | Why it fails |
|---|---|
### Kit Task 2026-03-24 |
Missing — separator — regex miss |
### Kit Task: 2026-03-24 |
Colon instead of em-dash — regex miss |
### Kit Task —\n2026-03-24 |
Date on separate line — regex miss |
- **Outcome:** done |
Not ✅ complete — field extraction returns empty |
- **Outcome:** complete ✅ |
Emoji at end, not start — field extraction may miss |
2. Daily File Structure
Naming: memory/YYYY-MM-DD.md
One file per day. Append throughout the day — never overwrite.
Dream cycle archives files >7 days old to memory/archive/.
Typical file structure (top to bottom):
# Daily Log — YYYY-MM-DD (Day)
_AEST timezone. Written end-of-day._
---
## Summary
[Optional 2-3 sentence overview — only for full sessions]
---
## Cross-Repo Activity — YYYY-MM-DD [Dream Cycle]
[Automatically captured — don't write this manually]
---
### [Agent] Task — YYYY-MM-DD HH:MM
[Task block — see Section 1]
### [Agent] Task — YYYY-MM-DD HH:MM
[Next task block]
---
### Loki Session — YYYY-MM-DD HH:MM–HH:MM (MAIN SESSION)
[Session summary block — see Section 3]
3. Loki Session Summary Block
For the main session (Loki's own summary at end of day):
### Loki Session — YYYY-MM-DD HH:MM–HH:MM (MAIN SESSION)
- **Projects touched:** [list]
- **Outcome:** ✅/⚠️/❌ + brief summary
#### Deliverables this session
1. [numbered list — one line per deliverable]
#### Key decisions
- [append-only, dated — architecture choices, trade-offs made]
#### Blockers / Next
- [what's pending, who owns it, expected resolution]
4. OUTBOX Entry Format
Write OUTBOX entries in OUTBOX.md (agents append here; Quill reads it to write journals).
## YYYY-MM-DD — [Short description of work]
**Tasks completed:**
1. [Brief description of each major deliverable]
**Key technical decisions:**
- [Non-obvious choices made]
**Repos touched:** [repo names + commit hashes if known]
Write an OUTBOX entry when:
- Agent completed >1 significant task in a session
- Work touched a shared codebase
- A lesson was learned that other agents should know
- A Notion journal entry would be worth writing
Skip OUTBOX when:
- Single minor task (e.g. fixed a typo, updated a README)
- No code shipped, no decisions made
- Routine heartbeat check with nothing found
5. What to Capture vs Skip
Always capture
- Every subagent task completion (task block, even short ones)
- Key decisions that affected architecture or priorities
- Non-obvious fixes — these become skill content
- Blockers with owner and expected resolution
- Tool commands that worked after debugging
Skip
- Internal tool calls that changed nothing
- Routine tasks with no learning ("updated the README")
- Heartbeat checks unless something was found
- Chat filler and clarification exchanges
- Dead ends that led nowhere useful
6. Downstream Dependencies (why format matters)
| Consumer | What it depends on | What breaks if format drifts |
|---|---|---|
langfuse-backfill-historical.py |
### [Name] Task — YYYY-MM-DD header pattern |
Task not extracted → gap in Langfuse observability |
| Quill journal pipeline | OUTBOX ## YYYY-MM-DD — headers |
Quill misses entries or writes empty journals |
| Liv CCP session update | Task outcome blocks to build Notion daily log body | Notion log is empty or incomplete |
| QMD memory search | Consistent structure across files | Retrieval quality degrades; semantic search misses context |
7. Common Mistakes
| Mistake | Fix |
|---|---|
### Kit Task: (colon) |
Use ### Kit Task — (em-dash, spaces) |
| Date on separate line from header | Keep date on same line: ### Kit Task — 2026-03-24 14:00 |
Outcome: done |
Use exactly ✅ complete / ⚠️ partial / ❌ failed |
| Skipping the Notes field | Most important field — this is what gets turned into skills and journals |
| One giant session block | One block per task, not per session |
| Writing prose narrative instead of task blocks | Prose is unreadable to parsers. 2026-03-23.md is the cautionary example. |
| Forgetting the time | Approximate is fine (~10 min) — needed for Langfuse duration metadata |
8. Quick Reference — Copy-Paste Templates
Task block:
### [AgentName] Task — YYYY-MM-DD HH:MM
- **Task:**
- **Outcome:** ✅ complete
- **Metric:** pass
- **Kept:** yes
- **Time:** ~X min
- **Notes:**
Session summary:
### Loki Session — YYYY-MM-DD HH:MM–HH:MM (MAIN SESSION)
- **Projects touched:**
- **Outcome:** ✅
#### Deliverables this session
1.
#### Key decisions
-
#### Blockers / Next
-
OUTBOX entry:
## YYYY-MM-DD — [Short description]
**Tasks completed:**
1.
**Key technical decisions:**
-
**Repos touched:**