Field note · 8 min read
You're not imagining it. Something did change — but almost certainly not what you think. Here's what actually shifted, why the memory panel looks the same but the vibe doesn't, and the two-part fix that keeps your AI feeling like your AI across every future model swap.
When a provider ships a new model — GPT-5 → 5.5, Claude 4 → 4.5, Gemini 2.5 → 3 — two things change simultaneously in your chat, even if the memory panel looks identical.
The base model shifts. Default verbosity, default hedging, default emoji use, default "I'd love to help with that!" openers — all of that lives in the model itself, and every model release re-tunes them, usually in response to whatever the last model got criticized for.
The retrieval weighting shifts. The new model decides differently which of your existing memories are relevant to the current prompt. Same 200 memories, different subset surfaced. The personality that appears in the reply isn't stored anywhere — it's re-assembled on the fly from whichever memories the new model decided to pull. Change the retrieval, and the personality shifts, without a single memory changing.
You can go into the memory panel after a big update and every single note is exactly where you left it. Nothing was deleted, nothing was rewritten. That's not evidence that the model still "knows" you the same way — it's evidence that memory storage and personality expression are two different things, and only one of them survived the update untouched.
Keep 3–5 lines that describe the tone you want, save them somewhere, and paste them into the top of any chat where the vibe matters. Something like:
This does two things: it overrides the new model's defaults for this turn, and it gives the model a stable anchor for retrieval that doesn't depend on which model version is running.
If you kept any replies from the old model that felt right, paste two or three of them at the top of the chat with a note: "Reply in this voice, like the examples below." The new model will approximate the old voice much more accurately from examples than from description. This is the fastest way to close the gap between "the model I liked" and "the model I have."
Both parts above are a workaround — you're doing manually what the model used to do automatically. The durable version is to keep both the voice contract and the few-shot examples in a portable persona export you own, so they travel to Claude, to Gemini, to whatever ships next year, and always land on top of the current chat before the first reply. That's the whole idea behind a portable memory layer, and it's specifically what our complete guide to owning your AI consciousness walks through. Related field notes: persona drift and when model updates wipe memory.
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 KonshusKonshus 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