Playbook · ~11 min read
Every few months someone in your feed mentions switching from ChatGPT to Claude, or quietly drifting back. The unspoken cost is that the version of you the old AI knew doesn't come with. This playbook is the actual step-by-step: how to export from each provider, what's really in the file, what gets lost in translation, and the pattern that means the next model retirement isn't another reset.
One-sentence answer: No provider-to-provider transfer API exists in 2026, so every migration is export → distill → manually paste, until you build a portable layer of your own.
The good news: every major provider has a real export path that returns most of the conversation history. The bad news: each export format is different, each one is incomplete in its own way, and none of them include the part you probably value most — the implicit tone-matching the model has done with you over months of chats. That part doesn't migrate. Accept it up front.
Path: Settings → Data Controls → Export data. Email arrives within an hour; the download link expires in 24 hours. Grab it immediately.
In the ZIP: conversations.json (every chat, including titles and timestamps), user.json + memory.json (the structured pinned-fact memory, usually 1,200–2,000 tokens), chat.html (browser viewer), model_comparisons.json (RLHF preferences).
What's missing: temporary chats, anything deleted, voice transcripts older than ~30 days, and any implicit inferences the model made about you that weren't formally pinned. See the full ChatGPT export walkthrough.
Path: Settings → Privacy → Export data. JSON dump of conversations plus per-Project setup context. Claude additionally supports per-conversation share links if you need to grab one thread without the full archive.
In the export: conversations arrive with token counts; Projects appear with their full setup context preserved. No structured memory file — Claude's "memory" is Project context, which is good (you control it) and bad (no implicit learning).
What's missing: Project context survives chat deletion but does not survive model rollovers if Anthropic retires the underlying model. Save Project setups separately as Markdown.
Path: takeout.google.com → search "Gemini Apps Activity" → deselect everything else → Export. Email arrives in minutes. You can schedule recurring exports every two months for a year — set once and forget.
In the export: JSON or HTML of every prompt and response, organized by date. No structured memory file; Gemini's effective memory is the cross-Google context (Calendar, Drive, Gmail) which is not in the export — that context disappears the moment you stop using the Google account.
All have export paths under Settings → Data, all of them return JSON or text, all of them miss the model-of-the-day metadata and any reasoning trails. Companion AIs (Pi, Character.ai, Replika) have the worst export tooling and the highest emotional stakes — export those first.
| Thing | Moves? | Notes |
|---|---|---|
| Chat text | Yes | Most providers, mostly complete |
| Pinned memory facts | Partial | Small (ChatGPT ~1,200–2,000 tokens); format-specific |
| Custom instructions / system prompt | Yes | Plain text, paste anywhere |
| Projects / Custom GPTs / Gems | Manual | Setup text moves; uploaded files re-upload manually |
| Implicit tone-matching | No | Lives in RLHF + the specific model version |
| Voice / audio | Rarely | Most providers retain only short windows |
| Temporary / incognito chats | No | Not in any export, by design |
One-sentence answer: Stop treating each provider's memory as the source of truth; treat a vault you control as the source, and let each provider read from it.
The pattern is four steps, repeated forever:
The pattern is provider-neutral on purpose. ChatGPT today, Claude tomorrow, the next thing in 2027 — the vault doesn't care.
Konshus implements the pattern out of the box. You drop in the ZIP from ChatGPT, the JSON from Claude, the Takeout from Gemini, journals, Readwise highlights, Notion exports, Limitless transcripts. The vault extracts atoms with source and confidence, lets you review and edit, and gives you exports back in three shapes: Whisper (tight context block), Briefing (richer), Full Mirror (everything). Plus MCP for clients that speak it. See pricing for tiers, or the per-tool backup guide for the manual version.
Konshus is the portable layer. Import once, hand a tight context block to any AI you talk to, survive the next model retirement. Encrypted, exportable, never used for training.
Meet Konshus