Do I lose my memories when a model is deprecated?

Not the stored list — saved memory entries and chat history are account-level, so they survive a model retirement. What you lose is the behaviour: how the retired model interpreted and used those memories. Same facts, different assistant.

Why does my AI feel like a stranger after an update?

Because the part of the relationship that lived in the model rather than the database was never stored anywhere. Tone, the level of explanation you'd settled into, the running jokes, the things it had stopped asking — all of that is reconstructed each time, and a new model reconstructs it differently.

How much warning do providers give?

It varies, and it has generally been measured in weeks to a few months for consumer surfaces. Enterprise API customers typically get longer notice than consumer app users, who often find out when something feels different rather than when an email arrives.

Is there a 'here's what this model knew about you' export?

No provider offers one. You can export your data; you cannot export a model's working understanding of you. That gap is the whole problem, and nobody has announced plans to close it.

What should I do before a model I rely on is retired?

Export your conversations, write down the standing rules you'd have to re-teach ('never summarise unless I ask', 'I'm technical, skip the basics'), and keep both somewhere outside the provider. Re-teaching from notes takes an afternoon; re-teaching from memory takes months.

Does switching providers cost the same as a model retirement?

Broadly yes — both reset the learned layer. The difference is that a retirement happens on the provider's schedule rather than yours, which is why having your own copy matters more than picking the 'right' provider.

Field note · August 26, 2026 · ~7 min read

What happens to your AI memory when a model is deprecated

The official answer is reassuring and technically true: nothing happens. Your saved memories are stored at the account level, your chat history is intact, and the new model can read all of it. The lived answer, which thousands of people posted some version of after the GPT-4 retirement, is that it doesn't feel like nothing happened at all.

Two layers, only one of them saved

Everything an AI knows about you sits in one of two layers.

The stored layer is the part you can see: saved memory entries, chat history, uploaded files, custom instructions. It lives in a database attached to your account, and it is genuinely unaffected by a model change.

The learned layer is everything that was never written down. How much explanation you want. Which topics need care. The fact that you're technical, so it stopped defining terms in week two. The register it settled into with you. This layer is reconstructed from the stored layer every single time you talk — and a different model reconstructs it differently, from identical inputs.

A deprecation deletes the second layer and nothing else. Which is why the honest description of the experience is not "I lost my data." It's "it doesn't know me anymore," said by someone whose data is entirely intact.

What the record actually shows

Retirements are not rare events you can plan around individually. OpenAI has sunset multiple GPT-4-era models; Anthropic has retired Claude 3-series versions with more notice but no migration path for what the old version understood; Google has cycled Gemini versions steadily enough that "which Gemini" is a real question. Our running deprecation list tracks the dates.

The pattern across all of them is the same. Enterprise API customers get long notice, versioned endpoints, and a migration guide. Consumer users get a changelog entry, and usually notice because something feels off before they read anything.

The export that doesn't exist

No provider ships a "here's what this model understood about you" export, and it isn't an oversight — the understanding isn't stored in a form anyone could hand you. It's an emergent property of a specific model reading your specific history.

Which leaves exactly one durable strategy: keep the inputs, and keep them somewhere that outlives any one model. Not because your provider is careless, but because even a perfectly run provider retires models. That's the job.

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus

A short pre-deprecation checklist

Export conversations. Every provider offers this. Do it quarterly so you're never more than a quarter behind.
Write down your standing rules. The ten or so instructions you'd be annoyed to re-teach. These are the highest-value thing you own and they take fifteen minutes to record.
Note the corrections. Anything you've had to correct more than once — a name, a role, a thing you don't do anymore — will need correcting again on the new model.
Keep it outside the provider. A backup stored inside the thing you're backing up isn't one.
Re-teach deliberately, not gradually. After a switch, spend one session establishing context on purpose. It's faster than six weeks of drip-feeding.

Frequently Asked Questions

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus