Contents
"But I already told it." I hear this a lot. Someone explained their business to ChatGPT a month ago and today it's drafting as if that never happened. Or the opposite, which is worse: it's confidently using a fact that has since changed. Both come from the same mix-up. Context isn't memory.
Two different things
Context is everything available for this answer. What's in the chat, what you've attached, what your instructions say. Memory is a product feature that can carry selected things forward into later work. In Claude and ChatGPT it's selective. It depends on your settings and your plan. It doesn't go back through every past conversation for every answer.
What memory is good for
Preferences and tone. How I like things written. That I want short paragraphs. The words I never use. Those are stable, so carrying them forward saves me repeating myself.
What belongs in the brief
Anything that can change. Say a property manager talked through a lease renewal with the AI a few weeks ago, including the end date. Since then the owner has agreed to a different term. Ask for the renewal letter today without supplying the current details and you may get the old date back. Memory didn't fail. The fact changed and nothing told it.
The rule I work to is simple. Facts that change live in a maintained source: the property system, the procedures, the fee schedule. The relevant version goes into the brief, even if you've discussed it before. That makes the answer easier to check, because you know exactly what it was built from.
When the chat has wandered
You'll know the moment. The job has changed, an old assumption keeps coming back or the useful decisions are buried under experiments. Don't keep pushing. Write a five-line handover and start fresh.
Goal: the renewal letter to the owner. Reader: an owner who prefers plain language. Source: the current lease details, pasted in from the system. Decisions: no figure that isn't in the source, no promise about timing. Open question: whether the owner wants a fixed term or periodic.
Read that summary yourself before you paste it. A long chat can carry an abandoned idea into the handover as if it were a decision. Then give the new chat the handover and the source. Ask it to flag anything missing before it drafts.
One more thing. A new chat isn't a blank account. It clears the conversation, not your memory settings, your custom instructions or the project the chat sits in. Those still shape the answer. When you want a clean comparison, check what your product carries forward first.
What a person still does
Decides what's stable enough to remember and what has to be looked up every time. Keeps the source maintained. I run one shared memory across all my tools and I still put the current facts in the brief. The tools I run this on, memory included, are on the stack.