Transcripts are not knowledge

Session transcripts feel like documentation because they contain every answer. They fail as memory for four reasons:

1. They are not queryable by intent. The answer to "how do we handle rate limits" is buried in a Tuesday afternoon about an unrelated bug.
2. They are per-person and per-tool. Your Cursor history says nothing to a teammate's Claude Code session on another machine.
3. They have no trust signal. Nothing marks which statements are still true, which were guesses, and which were reverted the next day.
4. They cannot be revoked. When a decision changes, the old transcript still teaches the old way.

What survives

Only statements that were promoted - pulled out of the flow of chat, given a title, a type, an owner, and a verification date - compound across sessions and teammates. Everything else is exhaust.

The promotion step is the whole game. Tools like okfshare exist to make that step one command instead of a documentation project.