Your Context Should Outlive Your Chatbot

Every assistant now remembers you, but that memory lives on its servers, in its format. This week: how to pull your context into one document you own, fix it, and prove it works in a second tool.

Last week I told you to set up one task so you never start from a blank prompt. Here's the catch nobody mentions: the longer you use one assistant, the more of your working context piles up inside it — your role, your preferences, the corrections you've made a dozen times. That's useful right up until you need to switch tools, and then you find out you don't actually have a copy of any of it. Fix that before you need it.

On Monday I wrote about Salesforce launching its own model, Koa, and leading with the fact that it controls the weights. You're not going to train a model. But you can own the layer that makes any model useful to you. The vendors have made this easier than it sounds: Anthropic now documents a way to import memory from another assistant and export Claude's memory of you, including a plain prompt that asks Claude to write out what it remembers, verbatim. Switching is a paste now, if you have something to paste.

Four sections, two pages at most. First, who you are: role, company, who you write for. Second, how you want work back: length, format, tone, and the things you never want to see — the phrases you keep deleting. Third, standing facts: product names, prices, team names, the terminology your company uses and the words it doesn't. Fourth, one or two finished examples you were happy with, pasted in full. Plain text or a Google Doc — anything you can open without the assistant. If a new hire could read it and draft your Monday memo, it's done.

Two honest reasons. First, memory feels automatic. Claude, ChatGPT and Gemini all save things about you as you go, so you never had to write any of it down — which also means you've never read it. Some of what's in there is stale, and some of it is just wrong. Second, one tool feels like enough. It is, until your company standardizes on a different one, the price changes, or a better model ships somewhere else. Then you start from zero and spend a month re-teaching the new one what the old one knew.

Block 30 minutes before Friday. Ask your main assistant to write out everything it remembers about you, word for word — Claude's help page gives the exact prompt, and in ChatGPT you can review saved memories under Settings, Personalization, Memory. Paste it into a document. Delete what's wrong, add the four sections above, and save it somewhere that isn't the chatbot. Then the test: paste it into a second assistant and run one real task you did last week. If the draft comes back close, you're portable.

Skip this and your context stays in a box you can't read. The day you need to leave — or the day the tool quietly forgets something important — you'll be rebuilding it from scratch. Thirty minutes now buys you the one thing Salesforce was bragging about on Monday: the option to walk.