Personal agents may miss more rules when memory gets too long

A post argues that personal agents should keep memory short and use code for things they must track. It reports Claude Haiku 4.5’s rule violations fell from 77% without memory to 20% with 10 lines, then rose to about 25% with longer memories. In a spending-total test, a written rule failed 44% of the time, while code tracking the total failed 0%. The best methods for rewriting memory from complaints ended near 48% violations, versus 7.1% when the agent was told every preference.