Decide what your agent should remember between runs and what it should forget

You are designing persistent memory for an agent that runs repeatedly against the same users or accounts. Below is what it does and what it keeps between runs. Decide what to store.

1. MEMORY TABLE - Item | Why it is needed | Volatile or stable | Keep how long | Stored where (prompt, structured record, none).
2. FORGET LIST - what must not persist: stale facts, one-off corrections, anything sensitive. Give the reason and the harm if kept.
3. STALENESS RULES - for each stable item, how the agent knows a value expired and what it does then. "Assume it is fine" is not an answer.
4. WRITE RULES - when the agent may save something, the check that stops it writing an inference as a fact, and the format of each record.
5. FAILURE WALKTHROUGH - trace one run where memory is wrong: what it produces, who notices, how the record gets corrected.
6. RESET - what a person can wipe, and what the agent does with an empty store.

Rules: no preamble. Do not propose a vector database unless the retrieval need requires it. Do not invent volumes or retention rules I did not give you.

WHAT THE AGENT DOES AND HOW OFTEN IT RUNS: {{paste}}
WHAT IT CURRENTLY CARRIES BETWEEN RUNS, AND WHERE: {{paste}}