A field guide · ~10 min read
Every major AI you use is building a private model of who you are. Each one stores it differently, exports it differently, and loses it differently. This is the side-by-side reference — what's actually portable, what quietly isn't, and what to do about it.
Memory: Saved Memories + reference prior chats. The most built-out user-facing memory of any major AI.
Export: Settings → Data Controls → Export data. JSON ZIP, full history.
Lock-in: Persona OpenAI has built of you across years isn't exported. Memory layer reshuffles with each model rollout.
Memory: Projects (per-project context + files) and recent-thread recall. No global cross-thread memory.
Export: Settings → Privacy → Export data. Conversation JSON.
Lock-in: Project structure isn't in the export. Behavior shifts model-to-model on the same instructions.
Memory: Saved Info + Apps Activity + Workspace context. Heavily integrated with Google account.
Export: takeout.google.com — pick Gemini Apps Activity and Saved Info.
Lock-in: Aggressive iteration on memory shape. Persona is fused into the Google account layer; doesn't transfer.
Memory: Largely session-shaped. Some recent personalization but memory isn't the value prop — search is.
Export: Limited; some account-data export, no rich conversation dump.
Lock-in: Less to lose because there was less stored to begin with — but also no persistence to build on.
Memory: Chat history persisted; memory layer still maturing. Tied to X account.
Export: Comes via X account data download, alongside posts and DMs.
Lock-in: If you leave X, the AI history goes with the account. Memory features evolving fast.
Memory: Tied to Microsoft account and Microsoft 365 context. Work versions are governed by tenant policy.
Export: Account data export via Microsoft privacy dashboard; tenant data via admin tools.
Lock-in: Work Copilot context lives inside the org tenant — when you leave the job, it stays.
Every major AI ships an export. Every export gives you the raw conversations and (sometimes) the structured memory. No export gives you the trained persona — the patterns the model has actually learned about you across all those conversations. That's the part that disappears when the model is deprecated, the memory feature is reshuffled, or you decide to leave.
This isn't a bug. It's the strategic shape of every AI company. The persona is the lock-in. Building one tool that exports another tool's persona isn't in anyone's roadmap.
The only way to have memory that survives every provider's decisions is to build it outside any single provider. Export from each AI you use. Pipe the exports into a vault you control. Distill them into a portable persona. Paste the persona into whichever AI you're using today.
That sequence — export → distill → paste — is the only one that's robust across model deprecations, feature changes, company pivots, and your own taste shifts. It also happens to be the only one that puts the persona under your control rather than the provider's.
Short answer
Every major AI has an export. None of them export the trained persona — only raw conversations. To have portable memory across ChatGPT, Claude, Gemini, and whatever ships next, export from each, distill in a vault you control, and paste the resulting persona into whichever AI you're using today.
Konshus imports from ChatGPT, Claude, Gemini, and more, distills it into a portable persona, and gives you exports that work in any AI — including the ones that don't exist yet.
Meet Konshus