Agent memory is useful only when it does not quietly outrank current truth.
Memory helps agents avoid starting from zero. It can remember preferences, recent work, open threads, recurring formats, and prior decisions. But business facts age. Account owners change. Contracts renew. Tickets close. Pricing exceptions expire. A remembered fact can become wrong while still feeling familiar.
The context layer should separate durable preferences from current operating truth. "This customer prefers concise weekly updates" may be useful memory. "This customer is on plan X" should probably be checked against the system of record before action. "Legal approved this clause" needs even stricter freshness and source rules.
Memory also needs provenance. Where did the remembered fact come from? Who said it? When? Was it observed, inferred, or approved? Can it be reused for this type of task? Should it be visible to this user?
A good pattern is memory with expiration and revalidation. Some memories never drive action without a fresh lookup. Some can inform tone but not facts. Some are scoped to a project or workflow and deleted when the work closes.
This is not anti-memory. It is memory with manners. Agents that forget everything are annoying. Agents that remember stale truth are dangerous.
The operating rule should be clear: memory can help the agent work, but authoritative sources win when the agent is about to decide, write, recommend, or act.
This is part 6 of 10 in The AI Context Layer.
