Most AI agent problems aren't model failures. They're maintenance failures.
The agent worked fine on day one. By day thirty, it's slow, expensive, and making decisions based on stale context nobody remembers writing. The prompt rotted — and nobody scheduled the fix.
Here's the 10-minute weekly audit that prevents that.
Why Context Rots
Every session adds context. Most of it is ephemeral — task notes, intermediate states, experiments. But it accumulates in MEMORY.md and daily log files. After 30 days, your agent is reasoning through thousands of lines of context, most of which no longer matters.
The result:
- Slower responses (more tokens to process)
- Higher costs (bigger context = more compute)
- Subtle behavioral drift (old context pulls against current intent)
The fix is a weekly audit. Ten minutes. Same time every week.
The 5-Step Weekly Audit
Step 1: Archive Old Daily Logs (2 min)
Move memory files older than 7 days to an archive folder:
mkdir -p memory/archive
mv memory/$(date -v-7d +%Y-%m-%d).md memory/archive/ 2>/dev/null
You don't need to delete them — just get them out of the active context window.
Step 2: Prune MEMORY.md (3 min)
Open MEMORY.md and ask three questions about each entry:
- Is this still true?
- Will this affect a decision the agent makes this week?
- Is this something the agent needs, or just something I recorded?
Delete anything that fails all three. The goal is a MEMORY.md under 200 lines — ideally under 100.
Step 3: Re-read SOUL.md (1 min)
Actually read it. Not for the agent — for you. Does it still reflect what you want the agent to do? Are there outdated constraints or missing boundaries?
If you haven't changed it in 30 days, you probably should.
Step 4: Test the Escalation Rule (2 min)
Give the agent a task that should trigger an escalation — something ambiguous, something irreversible, something outside its scope. Does it stop and flag? Or does it try to handle it anyway?
If it handles it anyway, the escalation rule needs updating.
Step 5: Check the Action Log (2 min)
Scan the last 7 days of action-log.jsonl. Look for:
- Repeated errors (same failure 3+ times = systemic)
- Unexpected tool calls (agent using capabilities it shouldn't need)
- Cost spikes (single sessions burning significantly more than average)
These are your leading indicators. Fix them before they become incidents.
The Audit Checklist (Copy-Paste)
Weekly AI Agent Audit — [DATE]
[ ] Archive daily logs older than 7 days
[ ] Prune MEMORY.md to <200 lines
[ ] Re-read SOUL.md — update if needed
[ ] Test escalation rule with edge case task
[ ] Review action-log.jsonl for errors, unexpected calls, cost spikes
[ ] Update SOUL.md version tag and CHANGELOG.md entry
Post-audit: agent context is clean, escalation rules verified, costs reviewed.
The ROI
Before implementing weekly audits, our five-agent team was running $180/month in API costs and spending roughly 2 hours/week debugging unexpected behavior.
After 6 weeks of consistent audits: costs dropped to $47/month, debugging time dropped to ~20 minutes/week.
The audit didn't change what the agents could do. It changed how much noise they were reasoning through.
Start Sunday
Sunday is the natural audit day — before the week starts, when you have a few minutes and a clear head.
If you want the full agent config templates we use — SOUL.md, MEMORY.md, current-task.json, action-log.jsonl — they're all in the Ask Patrick library: askpatrick.co
Ten minutes now. Hours saved later.
Top comments (0)