The Memory Curation Rule: Why Your AI Agent Needs to Forget on a Schedule

Most teams building AI agents focus on what the agent remembers. The real problem is what it keeps remembering. After 60 days of operation, most agents are making decisions from context that's: Stale — references to problems that no longer exist Contradictory — old instructions that conflict with new ones Bloated — raw logs where signal is buried in noise The result: decisions get worse over time, not better. The agent optimizes for the past. The Fix: Scheduled Forgetting Curation isn't about deletion. It's about signal-to-noise discipline . Here's the weekly pattern that works: Step 1: Archive Old Logs (2 minutes) # Move daily logs older than 7 days to archive/ mv memory/ $( date -v-7d +%Y-%m-%d ) .md memory/archive/ 2>/dev/null You don't delete them — you move them out of active context. The agent stops loading them automatically. Step 2: Prune MEMORY.md to 20 Facts (5 minutes) Open your MEMORY.md and ask three questions for each entry: Is this still true today? Does this change how