Agent Memory Has a Sweet Spot: 10 Lines Optimal, Code Beats Rules for Tracking

rohanpaul_ai · x · 2026-10-05

Hands-on experiments on personal agent memory show that more notes stop helping past a point — extra memory lines make agents miss rules.

Key findings:

Rules vs code: written rules work for style preferences (e.g., signing with a first name), but for anything the agent must count or track, a stated rule failed 44% of the time while code keeping the total failed 0%.
Memory sweet spot: with Claude Haiku 4.5, violations dropped from 77% (no memory) to 20% at 10 lines, then rose back to 25% with longer memories.
Self-rewriting memory stalls: agents rewriting their own memory from user complaints improved early but plateaued near 48% violations, versus 7.1% when every preference was simply stated upfront.

Takeaway: use code for anything the agent must count or track, and keep memory short.

More from coding & agent