You need a pattern that only you recognise/understand, not just remember. It takes only one leak of your password from service FooBar that looks like "f....b" to know what to try on other sites. Patterns easy to remember are mostly easy to understand.

With LLM that sort of approach can be attacked at scale