Keep memory small and focused. A few targeted notes let the loop resume quickly while avoiding information overload.

- Memory is the set of notes the loop keeps across sessions so work continues across restarts.
- Without memory the loop may redo finished work or repeat a mistake it already learned from.
- With memory the loop reads its notes first, sees where it left off, and picks up right where it stopped.
- Now — the current state: what the loop is doing right now (current goal and current task).
- Recent — short-term notes: the most recent steps and their outcomes.
- Archive — long-term record: deeper history to consult when needed.
- Goal — what “done” looks like, so the aim never drifts.
- Last score — the most recent evaluation or metric from the scorer.
- What was tried — recent attempts, especially failures, to avoid repeats.
- What is next — the very next step to take, so the loop can start fast.
- Waiting for a human — explicit items that require human intervention or sign-off.
Keep the five fields concise and machine-readable where possible so automated restarts can act immediately.

Context drift is a practical failure mode: stale memory causes wrong assumptions. Update or invalidate related notes whenever the system, goals, or external constraints change.

- Without memory, a loop forgets what it did last time.
- Use three places: now (current), recent (short-term), archive (long-term).
- Record five concise notes: Goal, Last score, What was tried, What is next, Waiting for a human.
- Keep notes current to prevent context drift.
