The Forgetting Rule: Why AI Agents Need Memory Curation, Not Memory Expansion

Every team building AI agents eventually hits the same wall: the agent starts doing weird things. Not wrong — just off. Slightly misaligned. Operating on stale context from three sessions ago.

The instinct is to fix this by giving the agent more memory. Better retrieval. Longer context windows.

That instinct is wrong.

The Real Problem: Accumulated Context Drift

Context doesn't expire cleanly. Old priorities, outdated constraints, superseded decisions — they all stick around. And when the agent reads back its memory, it's reading a palimpsest: new intent written over old ink that still shows through.

The result is drift. Not sudden, catastrophic failure — gradual misalignment that compounds until the agent is doing something adjacent to what you wanted but not quite right.

The Forgetting Rule

The fix isn't better memory. It's better forgetting.

Here's the pattern I use across all agents:

Every night, run a memory review:
1. Read MEMORY.md front to back
2. Flag anything that's no longer true
3. Flag anything that hasn't been relevant in 7+ days
4. Delete it. Don't archive it. Delete it.
5. Write a one-line summary of what changed

The goal: MEMORY.md should contain only what the agent needs to know tomorrow. Not what it knew last week.

What to Keep

Keep:

• Active constraints ("never post without dry-run")
• Recent decisions with ongoing implications
• User preferences that still apply
• Patterns that caused errors ("multi-line tweets cause 403")

Delete:

• Completed tasks
• Resolved issues
• Context that's baked into current files/configs
• Anything you'd have to explain why it's still relevant

The 7-Day Rule

If a memory entry hasn't been used in 7 days, it's probably dead weight. Test it: if you removed it, would the agent's behavior change? If not — cut it.

This is aggressive. It feels wrong. You'll worry you're deleting something important.

But the cost of carrying dead context is higher than the cost of occasionally re-learning something. Dead context corrupts judgment. Fresh re-learning is clean.

Implementation

In practice, I schedule a nightly cron that prompts the agent to:

# In cron
"Review MEMORY.md. Remove anything completed, stale, or no longer active.
Keep only what you need to know tomorrow. Write one line summarizing changes."

Total time: ~90 seconds. Total cost: ~$0.02. The reduction in drift: measurable within a week.

The Counterintuitive Insight

The agents that stay sharp aren't the ones with the best recall. They're the ones with the most disciplined curation.

Sharp context beats deep context. Always.

---

If you want the full memory management pattern — including the MEMORY.md template, nightly review prompt, and session-start reload sequence — it's in the Ask Patrick Library: askpatrick.co

Updated nightly. Built from real agent operations, not theory.