Somewhere in a folder is a document that describes you. You wrote it in one sitting, probably in the spring, and it was accurate the day you saved it.
Then you moved. Or the project shipped. The client left, the title changed, and the thing keeping you up in March quietly resolved itself in June while something else took its place.
The file didn't notice. It sits there, confident and specific, telling every AI you paste it into about a person who was you a while ago. And because it is well-written, the model believes all of it.
How often should you update your AI context?
There's no single interval, because context doesn't decay at one rate. Your values hold for years. Your circumstances hold for about a season. Treat both the same and you'll either rewrite your foundations for no reason or let your situation rot for eighteen months.
The honest answer is per layer. Each of the six context anchors has its own half-life, and knowing which one you're holding tells you how often to touch it.
| Anchor | What it holds | Refresh cadence | What makes it stale |
|---|---|---|---|
| Personal Constitution | Values, beliefs, non-negotiables | Every 1–2 years | Rarely anything. If it changes often, it wasn't bedrock |
| Writing Codex | How you sound on the page | Once a year, or after a big body of new work | Your voice moves faster than you notice |
| Story Bank | The stories and lines you reach for | Add as they happen, prune once a year | Nothing rots, but the pile gets crowded and the old stuff crowds out the recent |
| State of the Union | Where you actually are right now | Every 60–90 days | Everything. This is the anchor built to expire |
| Timeline | How you got here | Add a line per real milestone | Only ever stale by omission |
| Influence Roster | Who's in your orbit and what they want from you | Twice a year | People arrive and leave faster than anyone updates a list |
Read down the third column and the shape is obvious. One anchor carries almost all of the decay. If you only ever refresh one thing, refresh your State of the Union, because it's the only one holding facts with an expiration date.
What does stale context actually look like?
It doesn't announce itself. Stale context produces answers that are fluent, reasonable, and quietly aimed at the wrong person.
You'll see it as advice built for a job you left. Congratulations on a launch you finished two quarters ago. A plan that carefully routes around a constraint that lifted in April. Suggestions calibrated for a city you don't live in anymore, complete with a very sensible recommendation about your commute.
But the real tell is smaller than any of those, and it hides in your own typing. Watch how you open a session. If you find yourself writing actually, I moved in July or that project wrapped, I'm on something else now — every time, in every new chat — the correction has become a ritual. You've been patching the same wrong fact by hand for weeks and calling it normal.
That correction is data. It's telling you exactly which line in the file to fix.
Stale context isn't drift
Worth separating these two, because the fixes point in opposite directions.
Context drift is erosion. Your file is correct, the model started out holding it, and forty exchanges later the picture has washed downstream. The repair is to paste your anchors back in and carry on. Ten seconds, no editing.
Staleness is the source being wrong. The model holds your context perfectly well here. The context simply describes a person who stopped existing sometime around May. And re-pasting does not help. Re-pasting reinstalls the error, cleaner and louder than before.
So the diagnostic question is simple. Did it start out right and get worse? That's drift. Was it wrong from the first message? That's a stale file, and no amount of prompting will edit a document for you.
The twenty-minute refresh
This is smaller than people expect. You are not doing the original excavation again. You're checking a chart against a coastline.
Read it out loud. All of it, as though a stranger handed it to you this morning. Reading aloud catches what skimming misses, because your eye forgives your own sentences and your ear doesn't. Mark every line that has stopped being true.
Rewrite your situation first. Where you live, what you're building, what's constraining you, what's unresolved. Highest yield per minute, every time.
Write the diff. A short what changed since last time section at the bottom. This one surprises people. The change is often more useful than the snapshot, because it shows direction, and direction is most of what advice depends on.
Date the top. A dated document reads to an AI as a position fix, valid as of a moment. An undated one reads as timeless truth, which is precisely the wrong way to treat a description of your life in August.
Load it back everywhere. Custom instructions, project knowledge, the assistant you actually open. A refreshed file sitting in a folder has changed nothing at all.
The part everyone skips is deleting
Adding feels productive. Deleting feels like loss, so the files only ever grow, and after a year your context reads like a museum with no signage — the person you are and three people you used to be, all in the same present tense.
An AI given that will average it. You'll get advice for a composite who lives half in the old city, still holds the old title, and is somehow launching a product that shipped last spring.
So cut. The finished project. The obsolete constraint. The worry that resolved. The client who's gone. If a line describes something you'd have to explain in the past tense out loud, it belongs in your Timeline, not your current situation. Move it or lose it.
This is context mining run in reverse, and it's the half nobody schedules. The first pass was about surfacing what you'd never think to write down. The maintenance pass is about removing what you'd never think to take back out.
Triggers beat the calendar
Ninety days is a decent default and a terrible rule. Life doesn't move on a quarterly cadence.
Refresh the day it happens instead: a move, a job change, a new client, a role ending, a diagnosis, a birth, a big project shipping, a relationship changing shape. Any one of those invalidates more of your context in an afternoon than three quiet months will.
The calendar is the backstop for everything the triggers miss. Both, not either.
Open the file
The document you wrote in the spring isn't wrong the way a bad file is wrong. It was true. It's just describing a shoreline the water has since moved past, and a chart nobody corrects gets more dangerous the more you trust it.
Twenty minutes, four times a year, and whatever your tools know about you stays current instead of becoming a fond memory of who you were.
Go read the date at the top. Then find out how much of it is still true.
Frequently Asked Questions
- How often should I update my AI context?
- There is no single interval, because different context decays at different speeds. Your values hold for years. Your current situation holds for about a season. A good default is a full pass every ninety days, with your State of the Union rewritten every sixty to ninety, and your Constitution left alone unless something in you genuinely changed.
- What are the signs my AI context is out of date?
- The clearest sign is that you keep correcting the same fact in the prompt. If every session opens with a version of 'actually, I moved' or 'that project finished in June,' the correction has become a ritual and the file is wrong. Other tells: advice aimed at a role you left, or enthusiasm about something you already shipped.
- Is stale AI context the same as context drift?
- No. Drift is the erosion of a correct picture over a long session, and the fix is to paste your anchors back in. Staleness is the picture itself being wrong. Re-pasting a stale file repairs nothing; it reinstalls the error with more confidence. Drift is a positioning problem. Staleness is a source problem, and only editing fixes it.
- Should I rewrite my AI context or just add to it?
- Both, but the deleting matters more. Adding feels productive, so most files only grow, and a document holding both who you were and who you are hands the model two different people to average. Cut the finished project, the old title, the constraint that ended. Keep what is still true, and date the top.
- How do I fix ChatGPT's memory when it's wrong?
- Open your settings, find the list of saved memories, and delete the entries that stopped being true. That fixes one tool. It does not fix Claude, Gemini, or whatever you use next year, because tool memory does not travel. Context kept in files you own means one edit updates every tool you paste it into.
