This is how you search across past Claude sessions without burning tokens on irrelevant details. The workflow is smart: search returns a lightweight index with IDs and titles, timeline shows you what happened before and after interesting moments, then you only fetch full details for the observations that matter. That three-step filter saves roughly 10x on tokens compared to fetching everything upfront. Use it when someone asks "didn't we already fix this?" or "how did we handle authentication last time?" The batch fetch with get_observations is a nice touch. If you want synthesized answers instead of raw records, there's a knowledge agent mode that reads your history and responds conversationally.
npx skills add https://github.com/thedotmack/claude-mem --skill mem-search