You switch chatbots, paste in the import prompt, and watch a tidy paragraph land in your new assistant. It feels like you moved your AI's mind. You moved a paragraph of text. When you import AI memory between chatbots, that text block is the entire transfer: a written briefing your old assistant wrote about you. The model weights, the provider's tuning, and the running sense of your projects all stay behind, and they were the part that made the old assistant feel like it knew you.
Claude and Gemini both shipped import features in early 2026, and both run on the same trick. A prompt asks your old chatbot to dump everything it remembers about you, and you paste that dump into the new one. The migration looks complete. Underneath, you copied a frozen description that starts going stale the second it lands. The new model reads it once, folds it into its own memory store, and from there it drifts down its own path. Here is what actually transfers, what physically cannot, why the copy rots, and how to keep one profile that survives every switch.
Short answer: import moves a text summary, not your AI's behavior
Importing memory into Claude or Gemini transfers a written summary of your stored preferences and facts, generated by your old chatbot on request. Both products use the same method. You copy an extraction prompt, run it in the old app, and it returns everything it has saved about you as one block of plain text. You paste that block into the new assistant, which reads it and updates its own memory. Reporting on both rollouts, including tech outlet The Decoder, describes the identical prompt-based handoff.
The word import does a lot of quiet work here. In normal software, import means moving a file in its native format with full fidelity. In AI memory, there is no native format and no file. There is a paragraph of English that the receiving model interprets however it likes. Hand two assistants the same summary and they will not behave the same, because each one decides what to keep, how to weight it, and when to surface it.
You are not moving your AI's mind. You are handing the new model a written briefing about yourself and hoping it reads carefully. The briefing is text. Everything that made the old assistant feel like it knew you stays in the old assistant.
What is actually new: Claude and Gemini import flows in 2026
In early 2026, Anthropic and Google both shipped one-click-style memory import, turning a power-user trick into an official feature. Anthropic added the memory import flow for Claude on March 2, 2026, and made it available to free users. Google followed on March 26, roughly three weeks later, with ChatGPT and Claude import tools for Gemini. Two of the largest model makers landed near-identical features in the same window, with the same goal: cut the cost of switching assistants.
Claude's flow runs on the summary method. You take an extraction prompt, run it in ChatGPT, Gemini, Grok, or any chatbot that stores preferences, and paste the resulting text into Claude's memory settings. Claude then processes that text into its own memory. Step-by-step guides for the feature walk through exactly this copy-from-old, paste-into-Claude sequence across multiple source chatbots.
Gemini went one step further with two paths. The first is the same prompt-to-summary import for preferences. The second lets you upload your old chat history as a ZIP export, reported as supporting files up to 5 GB with a cap of five ZIP uploads per day, so you can continue prior conversations inside Gemini. That history import was not available in the EEA, the UK, and Switzerland at launch, and it excluded business, enterprise, and under-18 accounts.
Now the part the headlines skipped. As of June 2026, the import only runs one way. Claude and Gemini accept memory from rivals, but ChatGPT does not import memory from Claude or Gemini at all. To bring context the other direction, you copy and paste it in by hand. So the "portability race" everyone cheered is really two challengers building an on-ramp away from the market leader, not an open road in every direction. Read that way, import is a customer-acquisition tool wearing a portability costume.
Memory summary vs full chat history: the difference that matters
A memory summary is a short distilled profile. A full chat history is the raw archive of everything you ever typed. Importing a summary hands the new model a compressed, opinionated digest written by the old model. Importing a chat history, where supported, hands it the source transcripts to read for itself. These are not the same transfer, and most people blur them together.
The summary is lossy by design. Your old assistant decided what counted as worth remembering, phrased it in its own words, and dropped the rest. Nuance, exceptions, and the reasoning behind a preference rarely survive that squeeze. A line like prefers concise answers tells the new model nothing about the dozen times you asked for more depth on a specific topic.
Full chat history carries more raw signal, but it pushes the work onto the new model and onto you. A multi-gigabyte archive of old conversations is not a profile. It is a haystack. The receiving assistant still has to read it, interpret it, and decide what to remember, and it will land on different conclusions than the model that originally lived through those chats. More data in does not mean the same behavior out.
| Aspect | Memory summary import | Full chat history import |
|---|---|---|
| What moves | A short text profile of preferences and facts | Raw transcripts of past conversations |
| Who writes it | The old model, compressed and opinionated | You and the old model, unfiltered |
| Fidelity | Lossy: nuance and reasoning dropped | High raw signal, but unstructured |
| Effort to use | Low: paste and go | Higher: new model must re-read and re-derive |
| Where supported | Claude and Gemini via extraction prompt | Gemini via ZIP upload (up to 5 GB, 5/day), with regional limits |
| Result | New model behaves differently from a frozen brief | New model re-interprets and still diverges |
Neither path moves behavior. The summary moves a description of behavior, and the history moves the evidence behind it. Either way the new model rebuilds its own understanding from scratch, which is why an imported assistant feels close to the old one at first and then slowly stops feeling like it at all.
The three things import cannot carry: weights, tuning, and time-aware state
Import cannot move model weights, provider-specific tuning, or the assistant's live, time-aware sense of you. These three are where most of the felt intelligence lives, and none of them fit in a text block. Once you see why, the decay stops being a surprise and becomes the obvious outcome.
1. Weights
The weights are the trained network running your old assistant. The provider owns them, they are far too large to export, and users never touch them. Switch to Claude or Gemini and you switch to an entirely different brain. The summary is the only thing crossing over, and a different brain reading the same notes reaches different conclusions. That is structural, not a bug waiting for a patch.
2. Provider-specific tuning
Each assistant ships with its own system prompts, safety layers, response formatting, and personality tuning, all baked in by the provider. That tuning shapes how it reads your memory. The same profile produces a terse, cautious reply in one product and a chatty, expansive one in another. None of that tuning is yours to move. It belongs to whichever model you happen to be using.
3. Time-aware state
Your old assistant built a running, recency-weighted picture: what you worked on last week, which preference you reversed yesterday, what is active versus archived. A one-time summary flattens all of that into a static snapshot with no timeline. The new model starts from a single frozen point with no idea what is current, so it treats a year-old preference and a fresh one as equally true.
Weights, tuning, and time-aware state are the parts that made your assistant feel like it knew you. Import carries none of them. It carries the footnotes, not the book.
Why it decays: no shared format, so re-curation is required
Imported memory decays because the major chatbots share no common format for AI memory, which makes every transfer a one-time lossy snapshot that goes stale on arrival. Open efforts to standardize memory exist, such as the OpenMemory MCP project, but Claude, Gemini, and ChatGPT do not speak a single agreed schema between them. Each provider stores memory its own way and exposes it as free text. The instant you import, the copy and the original start drifting apart, and nothing keeps them aligned.
Look at what happens after the paste. You keep using the new assistant. It learns new things, forgets the imported ones it judged unimportant, and rewrites the rest in its own words. Meanwhile your old assistant, if you still touch it, keeps evolving on its own. Now you hold two divergent profiles and no source of truth. A month later, neither one matches who you are.
Because the transfer is one-shot, keeping continuity means re-curating by hand. Every time you switch or add a tool, you run the extraction prompt again, paste again, and check what survived. The industry framed these import tools as portability. A copy you must manually regenerate on every move is closer to a recurring chore than to real portability.
The durable fix: own one portable, model-agnostic profile
The durable fix is to stop treating any single chatbot as the home of your memory. Keep one portable profile you control, then feed it into whatever model you use. When the canonical version of who you are lives outside every provider, switching stops meaning rebuilding. You brief the new assistant from your profile instead of extracting it from the last one.
This flips the decay problem on its head. With provider-locked memory, your truth lives inside Claude or Gemini, so leaving costs you loss and re-curation. With an external profile, the providers turn into interchangeable front ends. Each one gets a fresh, current briefing from the same source, and no single switch costs you continuity. The profile is the asset. The chatbots are disposable.
This is the gap MemX (memx.app) is built for. MemX is an external, model-agnostic AI memory layer: one structured profile you own, designed to be fed into different assistants rather than trapped inside one of them. It is private by architecture, with per-user isolation, encryption at rest, and on-device options. To be clear about what that does not mean: MemX does not claim end-to-end encryption or a zero-knowledge design. It is a place to keep one durable profile so a provider's import tool becomes a convenience, not your only lifeline.
A useful test before you trust any memory feature: if this provider shut down tomorrow, what would you lose? If the answer is everything it knew about you, your memory is locked in. If the answer is a sync you redo elsewhere, you own it.
A monthly maintenance habit so your profile survives switches
A short monthly review keeps your portable profile accurate and turns any future import into a five-minute task instead of a rebuild. Memory drifts in every system, imported or not, so a light recurring pass beats a painful migration later. Keep it simple enough that you actually do it.
- Export and read your current memory from whichever assistant you use most, using its built-in export so you see exactly what it thinks it knows.
- Correct the stale entries: reversed preferences, finished projects, old job titles, tools you dropped. These are what quietly mislead a new model after import.
- Update your single source-of-truth profile with the changes, so the canonical version stays ahead of any individual chatbot.
- Re-feed that profile to any new or secondary assistant you have started using, instead of chaining an extraction off your last chatbot.
- Prune hard. A short, current profile imports cleanly. A long, contradictory one confuses every model that reads it.
The habit matters more than the tool. Whether you keep your profile in a dedicated memory layer or a plain document, the discipline of reviewing and pruning monthly is what makes you portable. Import features are a nice convenience sitting on top of that discipline. They are a poor substitute for it.
The 2026 portability push is real and mostly good for users, but read the feature for what it is. Claude and Gemini made it easy to copy a summary between chatbots. They did not make it possible to move your AI's mind, because that was never sitting in a text box. Own the profile, keep it current, and let the models compete to read it.
01Can you really import AI memory between chatbots?
Yes, with a catch. As of June 2026, Claude and Gemini let you import memory by copying a text summary from one chatbot and pasting it into another. You move a written profile of preferences, not the model's actual behavior, tuning, or weights. The summary is the whole transfer.
02Can you import memory into ChatGPT from Claude or Gemini?
No. ChatGPT does not offer a tool to import memory from rival assistants. The import flow runs one way: Claude and Gemini accept memory from competitors, but moving context into ChatGPT means copying and pasting it by hand, one detail at a time.
03What is the difference between memory import and chat history import?
Memory import moves a short, lossy summary of your preferences written by the old model. Chat history import, supported by Gemini via ZIP upload, moves raw transcripts. The summary is quick but compressed; the history is fuller but unstructured. Neither moves the assistant's actual behavior.
04Why does imported AI memory get worse over time?
The major chatbots share no common format for memory, so every import is a one-time, lossy snapshot. After you paste it, the new model reshapes it and your old assistant keeps evolving separately. The two profiles drift apart, so you must re-run the import by hand whenever you switch.
05How do I keep my AI memory when switching models?
Keep one portable, model-agnostic profile you own outside any single chatbot, and feed it into each assistant you use. Review and prune it monthly so it stays current. Then any provider's import tool becomes a convenience rather than your only way to preserve continuity.
