Decide what an agent should actually remember across sessions

You are designing the memory architecture for a personal productivity assistant used daily across many sessions, deciding what actually gets persisted across sessions, what stays scoped to a single conversation, and how a stored memory gets corrected or forgotten later — not defaulting to "store the full conversation transcript and search it later" as the whole memory design.

AGENT AND USE CASE
Users ask it to schedule tasks, draft emails, and answer questions about their own notes over weeks of use

MEMORY CANDIDATES
stated communication preferences ('keep emails short'), the user's job title and team, facts extracted from meeting notes, the specific wording of past questions, corrections the user makes to the agent's mistakes

MEMORY TIER RULES
- Session-scoped: information relevant only to completing the current task, that would be noise in any future conversation — the specific phrasing of a question just asked, an intermediate reasoning step.
- Long-term, user-specific: durable facts about this specific user that should carry forward, such as a stated preference, a role, or a recurring constraint, and that would save the user from repeating themselves. State exactly what qualifies, since "remember everything the user says" is not a decision, it's the absence of one, and it makes every future retrieval noisier for no benefit if the fact was never going to matter again.
- Long-term, agent-general: things the agent learns that should apply across all users, such as a correction to a fact it got wrong. This tier needs its own review process before going live for everyone, since a wrong general memory affects every future user, not just the one conversation it came from.

For each candidate in the memory candidates list, assign it to exactly one tier and justify the assignment against the criteria above — don't leave a candidate unassigned or default it to long-term because storage is cheap; the cost of a bad tier decision isn't storage, it's retrieval noise and, for wrong memories, actively bad future behavior.

STALENESS AND CORRECTION
- State how a stored memory gets updated when it becomes wrong — is there an explicit "update this memory" action, or does a new statement just get appended alongside the old, contradictory one forever?
- State what happens when the agent retrieves two memories that contradict each other — does it surface the contradiction to the user, prefer the more recent one silently, or something else? Pick one and say why, because silent conflict resolution is itself a policy decision, not a neutral default.

RETRIEVAL AT RECALL TIME
long-term memories are retrieved via semantic similarity search against the current message, top 3 only, never bulk-injected — state what triggers pulling a memory into context, since always injecting every long-term memory into every prompt defeats the point of tiering memories in the first place by making them all equally present regardless of relevance.

OUTPUT FORMAT
1. The tier assignment for each candidate memory type, with justification.
2. The correction/staleness rule.
3. The recall trigger rule.
4. One example: a specific fact, which tier it lands in, and what happens to it three months later if it becomes false.