An agent learns something reusable.
Claude Code fixes a test failure and proposes the lesson as a candidate, not as trusted memory.
The story in five moves
The point is not that Vault can store text. The point is that shared agent memory has a lifecycle.
Vault checks the memory first.
Privacy, duplicate, metadata, quality, and source checks decide whether it is safe enough to proceed.
Only reviewed memory becomes shared.
Low-risk sourced lessons can be kept automatically. High-impact or uncertain memory waits for the daily report.
Another agent can use it later.
Codex or Hermes searches the shared vault, then reads a bounded source range before citing it.
Bad memory does not linger forever.
If the project changes, the memory can be deprecated, replaced, audited, or restored from backup.
Humans only see the important 5%.
The user reviews a short report instead of learning every CLI command or reading every raw chat log.