Operating Memory in Production

Budgeting

Estimate before scheduling hygiene or heavy recall:

• Light scratch load: ~2k tokens
• Daily episodic append: ~1–5k
• Weekly hygiene: ~40–80k

Track in memory-budget.md and memory-run-log.md.

Logging

Append one line per memory op:

| ts | op | tokens | notes |
|----|----|--------|-------|
| 2026-07-20T12:00Z | recall | 2.1k | durable+scratch |
| 2026-07-20T18:00Z | hygiene | 45k | report-only |

Pause / kill

• Env: MEMORY_PAUSE=1 → skills no-op writes
• Or banner at top of MEMORY-STATE.md: PAUSED: human freeze

Metrics that matter

Metric	Healthy
Memory Ready score	≥40 after week 1
Durable fact count	10–50 active
Recall pack tokens p95	under budget
Hygiene last run	< 14 days
Wrong-fact incidents	trending down

When to stop

If agents ignore memory files, fix recall skill wiring — do not grow the store.