A developer works with Claude Code or Cursor for three hours on Monday. They establish twenty decisions along the way: why Postgres over DynamoDB, which auth pattern, which edge cases are out of scope, what broke when they tried the obvious approach.
On Tuesday they open a new session. The agent knows none of it. They spend the first fifteen minutes re-explaining.
Worse: mid-session, auto-compaction fires. The agent silently loses the constraint established two hours ago and confidently proposes the approach that was already rejected.
Worse still: a teammate picks up the work. The decisions live in a chat transcript that was compacted away, or in one person's head.