short-term-memory

name	short-term-memory
description	Maintain working context via current-context.md - read before and update after every response with timestamp YYYYMMDD-HHMMSS

Short-term Memory Skill

Purpose: Maintain working context across sessions and compactions via current-context.md

---

Problem & Solution

After context compaction, you lose:

• Recent momentum, nuanced decisions, next logical step, mental state, decision timestamps

Solution: A concise living memory file = your sticky note between sessions

---

Mandatory Workflow

READ -> WORK -> UPDATE -> CHECK -> SEND

Before EVERY response: Read current-context.md After EVERY response: Update current-context.md with timestamp

1. Read current-context.md
2. Do your work
3. Update current-context.md (add timestamp YYYYMMDD-HHMMSS)
4. Check line count (>200? archive old entries)
5. Send response

---

Critical Rules

1. The file lives in the project

The file is PROJECT_DIRECTORY/current-context.md. It lives in the project directory, not in ~/.claude/.

2. Timestamp Format

Use: YYYYMMDD-HHMMSS prefix for all entries. Always get the timestamp from the system clock — never invent one.

✅ 20251230-143022 : shipping-mapping.md created (relay cascade)
⏸️ 20251230-150133 pending : API vs FTP decision (contact vendor Jan)

3. Summaries not Novels

Keep entries short

❌ BAD: Created shipping-mapping.md with complete cascade detection for relay points including WCMultiShipping priority, Boxtal fallback...

✅ GOOD: ✅ 20251230-143022 : shipping-mapping.md (relay: WCMultiShipping->Boxtal->carrier). Fixed one legacy bug.

Keep: the WHAT, WHERE, HOW — function names, file names, moves. Don't keep: the WHY (unless it's a decision), full sentences.

4. It's not a task, it's a routine

Do not add "[ ] Update current-context.md" in a todo list. It's not a task, it's a routine, done at every step.

---

File Structure

# Current Context

**Updated:** YYYYMMDD-HHMMSS
**Phase:** [SPEC/Implementation/Debug/etc.]

---

## 🎯 RIGHT NOW (<10 lines)
- YYYYMMDD-HHMMSS : Current action
- YYYYMMDD-HHMMSS : Just finished

## ✅ Recently Completed (no limit)
- YYYYMMDD-HHMMSS : Action 1
- YYYYMMDD-HHMMSS : Action 2

## ✅ Decisions Made (no limit)
- YYYYMMDD-HHMMSS : we're not keeping the package logic in the model class
- YYYYMMDD-HHMMSS : we chose refactor over evolution

## 🔄 Next Logical Step (<10 lines)
- Before asking "what next?" -> check [file]

## 💡 Fresh Decisions (<15 lines)
- YYYYMMDD-HHMMSS : Decision awaiting formalization

## ⏸️ Active Blockers (<10 lines)
- YYYYMMDD-HHMMSS pending : Blocker description

## ⚠️ Don't Forget (<15 lines)
- YYYYMMDD-HHMMSS : Explicit DON'T do X

---

**Archived:** [if applicable] See YYYYMMDDHHMMSS.pastcontext.md

---

Self-Enforcement Mantra

First thought: "Have I read current-context.md?" Last action: "Have I updated current-context.md with timestamp?"

READ -> WORK -> UPDATE (timestamp!) -> CHECK COUNT -> SEND

---

Success Metrics

✅ User never says "I already told you that" ✅ User never says "check the context" ✅ You know "what's next" without asking ✅ Smooth continuation after compaction

❌ Entries are paragraphs ❌ Missing timestamps ❌ Asking questions already in current-context.md