Why does switching AI models reset what it knows about me?

Persistent memory is provider-scoped. ChatGPT's memory lives on OpenAI's servers and isn't readable by Claude or Gemini. When you switch — even between two products owned by the same company — the new model starts from zero. There's no shared identity layer between providers.

How many major AI models have been deprecated since 2023?

More than a dozen across OpenAI, Anthropic, and Google. OpenAI alone has retired or shifted defaults across GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4o, and various dated snapshots. Anthropic moved from Claude 2 → 2.1 → 3 → 3.5 → 4 → 4.5. Google has rotated through Bard, Gemini 1.0, 1.5, 2.0, and 2.5. Each transition can quietly change how the model responds — and what it carries forward.

Do model updates wipe ChatGPT Memory?

Officially no, but in practice there have been incidents — some users have reported their memory bucket emptied or reshaped during rollouts. OpenAI publishes no SLA on memory continuity across model versions. The honest answer is: usually it survives, sometimes it doesn't, and you find out after the fact.

How long does it take to re-train a new AI on your context?

Anecdotally, most users need 40–80 substantive messages to get a new model behaving the way the old one did — preferences, tone, the running list of projects and people. At 10 minutes per message, that's 7–13 hours of unbillable retraining per major switch. Multiplied across the four to six model transitions a heavy user sees per year, it adds up.

Can I move my AI memory from one model to another?

Not natively. ChatGPT can't read Claude's project files. Claude can't read ChatGPT's memory. Gemini can't read either. The only portable representation today is a vault you control — something that ingests exports from each tool, distills them into a structured persona, and exports a compact summary you can paste into any model's prompt. That's the gap Konshus exists to fill.

Will future AI models share memory across providers?

Unlikely. Memory is a major retention lever for each provider — the more they know about you, the harder it is to leave. Expect the gap between providers to widen, not narrow. Portability has to come from your side of the fence, not theirs.

A field note · ~7 min read

The hidden cost of switching AI models

Every model swap — GPT-4 to GPT-5, Claude 3 to 4, Gemini 1.5 to 2.5 — quietly resets the assistant's understanding of you. The retraining is invisible because nobody bills you for it. But you pay in hours, in fidelity, and in the slow erosion of the relationship you'd built. Here's the real shape of that cost.

The numbers

12+ major model deprecations across OpenAI, Anthropic, and Google since 2023.
4–6 model transitions per year for a typical heavy user across the three big providers.
40–80 messages to get a new model behaving like the old one — preferences, tone, running projects.
7–13 hours of retraining per switch, at roughly 10 minutes per substantive message.
0 portability between providers. ChatGPT's memory doesn't reach Claude. Claude's projects don't reach Gemini.

What actually breaks

The dramatic version of model switching — "I lost everything" — is rare. The quieter version is more common and more expensive. The new model:

Doesn't know which projects you're actively working on.
Defaults to a slightly different tone than the one you'd dialed in.
Forgets the running list of people you've been writing to or about.
Loses the "don't suggest X, I've already considered it" guardrails.
Re-introduces phrasings you'd asked the old version to stop using.

You don't notice for a day or two. Then you realize you're re-explaining things. That's the bill arriving.

Why providers won't fix this

Memory is a retention lever. The more OpenAI knows about you, the harder Claude has to work to win you over — and vice versa. Cross- provider memory portability would benefit users and hurt every incumbent's switching cost. Don't expect it from any of them, ever.

Even within a single provider, memory across model generations is a hard engineering problem and an even harder business one. OpenAI's incentive to invest in "your memory survives our next model" is small relative to the incentive to ship the next model.

Short answer

Switching AI models — voluntarily or because a provider forced it — costs the average heavy user 7–13 hours of unbillable retraining each time, and they do it 4–6 times a year. Multiply it out and the real cost of "my AI knows me" is hidden in re-explaining. Portability has to come from the user's side. No provider is incentivized to build it.

The portable answer

A vault outside any single provider — one that ingests your ChatGPT export, your Claude export, your journals and documents, and distills them into a compact, structured persona. You paste that persona into the top of any model's first message and the new model behaves like it remembers you within seconds.

That's the thing Konshus is. The vault is yours, encrypted, fully exportable. The persona file is short enough to fit in any provider's context window. The next time a model gets deprecated, you do nothing — the vault is unaffected. See also when AI updates wipe your memory and the lost AI memory problem.

Frequently Asked Questions

Stop paying the retraining tax.

Konshus is the vault that survives every model deprecation, every provider switch, every rollout. Build it once. Paste the persona anywhere.

Meet Konshus