Corrections Must Persist, Compound, and Be Visible

From erin-ahmed at cleric:

"A correction should do three things. It should persist, compound and then be visible."

• Persist — agent applies the knowledge in the same context. If a user teaches Cleric a customer-impact-assessment procedure after an outage, the agent should reuse it on the next outage.
• Compound — generalise that knowledge into a different context. E.g. guidance about distinguishing an internal instance from a production customer instance should transfer to unrelated investigations ("automatically excluding internal instances from an error rate analysis").
• Visible — "the agent needs to show its work. Display the use of knowledge and its actions and reasonings." Two benefits: users can see when the knowledge was applied (trust), and they can spot when it was applied incorrectly (safety).

Why this matters

This is a concrete operationalisation of "memory" — not just "store the conversation" but three behavioural tests a memory system must pass. Contrasts usefully with faye-zhang's memory collapse (agentic-posttraining-failure-modes): Ahmed's "compound" is what Zhang's post-training pipeline loses after 20–100 epochs.