The State of AI Chat Data 2026: A Portability Report

Methodology, and what this report does not claim

Every number here comes from one of two places: a count of code that ships in a production browser extension, or a structural property of a published data format that anyone can verify by reading a file. Both are reproducible without taking our word for anything.

What this report is not:

• Not a census of the AI industry. The platform count reflects the platforms one tool has built capture support for. Real coverage of the space is larger. Treat our figure as a floor.
• Not a usage study. Be Mimos stores conversations locally on each user's own device by design, so there is no central pile of chat data to analyse — which is the point of the architecture, and also the reason we cannot tell you what people talk to AI about. We consider that a good trade.
• Not a benchmark of competing products. Where other tools appear, it is because a format or behaviour is worth naming, not to score anyone.

Where a claim depends on something we cannot check, it is not in the report. Where it depends on something we can, the way to check it follows the claim.

The capture surface: 41 platforms, and why that number keeps moving

Four years ago the honest version of this section would have been a single number: one. The interesting thing about 2026 is not that AI chat exists, it is that it has fragmented into dozens of independent silos, each with its own conversation model, its own URL scheme, and its own idea of what a "message" contains. Reasoning traces, tool calls, attached files, generated images, deep-research runs that produce a document rather than a reply — the shape of a "conversation" now differs meaningfully between platforms.

That number is also unstable in a specific direction: upward, in bursts. Four of those 41 platforms were added in a single fortnight this July. This matters more than it sounds, because it sets the half-life of any AI data strategy that assumes a fixed set of tools. A workflow built around exporting from two platforms is a workflow that quietly stops covering your work the moment your team adopts a third.

How to check this: the count is taken from the capture scripts registered in the extension's built manifest, recounted directly against the provider list in the shipped build. It is not a marketing figure carried forward from an old page — earlier counts of 26, 33 and 41 all correspond to earlier builds, and each was accurate on its date.

Fifteen formats, and four of them are not documents

Once a conversation is captured, it has to be written somewhere. In practice the destination falls into one of two very different categories, and conflating them is the most common mistake we see in how people plan their AI data.

Eleven of the fifteen formats are document or data formats — PDF, Word, Markdown, HTML, plain text, JSON, CSV, XLSX, image, ZIP archive, and direct-to-Google-Docs. Their job is to be read, filed, shared or analysed by a human or a spreadsheet.

Four are machine-training formats — ShareGPT, Alpaca, Tavern and Oobabooga. Their job is to be consumed by a fine-tuning pipeline. Nobody reads them.

The reason this split deserves its own section is that people increasingly want both from the same archive, and only discover the incompatibility at the point of use. A year of saved conversations is simultaneously a personal knowledge base (wants Markdown, wants search, wants to stay readable in ten years) and a potential training corpus (wants strict JSON with consistent roles). Those two futures place genuinely opposed demands on how you store things today.

What a format can and cannot carry

This is the finding we would most like people to take away, because it is invisible until it has already cost you something.

The four training formats are not interchangeable. They divide into two structurally different families, and the difference is not cosmetic:

Conversation-shaped formats

ShareGPT and OpenAI-style messages JSONL store an ordered list of turns, each tagged with a role. A system prompt is a turn. A follow-up question with no answer yet is a turn. Three assistant messages in a row are three turns. The structure can represent whatever actually happened.

Pair-shaped formats

Alpaca and Oobabooga store pairs: one user input, one assistant output. That is the entire vocabulary. Which means, as a matter of structure rather than implementation quality:

• A system prompt cannot be represented. It has no slot. It is dropped.
• A question you asked but never got an answer to — the last message in most real conversations — has no partner, so it is dropped.
• Two consecutive assistant messages collapse or lose one side.

Converting a real conversation from ShareGPT to Alpaca is therefore lossy by definition, not by accident. A six-message conversation with a system prompt and a trailing unanswered question comes out the other side as two clean instruction/output rows — and the tool that did it usually says nothing at all.

That silence is the actual problem. Loss during a format conversion is often acceptable; you may genuinely only want the pairs. Unreported loss is not, because it removes your chance to decide. Any converter worth using should count what it discarded and tell you — and print the number, not bury it in a log.

How to check this: take any conversation with a system prompt, export it as ShareGPT, convert it to Alpaca with any tool, and count the rows. The arithmetic is not subtle.

Where the data actually lives

A conversation you can read is not the same as a conversation you hold. In 2026 chat history is, for most people, a tenant on someone else's platform: it persists at the platform's discretion, under the platform's retention policy, and it is reachable only through the platform's interface. Accounts get closed, products get sunset, regions get restricted, and the archive goes with them.

There are three distinct storage postures worth naming, because they are frequently marketed as if they were one:

1. Platform-resident. The default. Your history exists where the conversation happened. Nothing to install, nothing to manage, and nothing you control.
2. Vendor-hosted mirror. A third-party tool copies your conversations onto its own servers so it can offer search and sync. Convenient, and it means your chat history now exists in a second place governed by a second privacy policy.
3. Local-first. Conversations are stored on your own device, with backup — if you want it — going to storage you already own rather than through a vendor's servers.

Be Mimos takes the third posture: conversations are kept in the browser's own storage, and cloud backup is optional and goes to the user's own cloud drive. We want to be precise about what that does and does not mean, because this is an area where marketing language routinely outruns engineering. It means chat content is not routed through our servers. It does not mean end-to-end encryption: bulk backup data is written to the user's own drive as plaintext JSON, and we would rather say so plainly here than let a stronger-sounding claim stand.

The practical consequence of local-first storage is the one we mentioned in the methodology: we cannot mine this data, and neither can anyone else who did not start with your device.

What to do with this if you are planning for 2027

Four things follow from the above, in rough order of how much time they save.

Capture at the source, not at the end. Every export path in existence can only export what was retained. Retention decisions are made silently, by platforms, before you know you needed the conversation. The gap between "I might want this" and "I want this" is usually weeks, which is exactly the window in which chat history becomes hard to reach.

Keep one lossless copy, then derive everything else. Pick a conversation-shaped archive as the master — something with roles and turn order intact — and generate the PDF, the Markdown note, or the training set from it as needed. The reverse direction does not work: you cannot reconstruct a system prompt from an Alpaca row that never had one.

Assume the platform list grows. Any process pinned to today's two or three tools has a shorter life than the work it is protecting.

Ask what a conversion dropped. If a tool cannot tell you, it does not know, which means neither do you.

Questions we expect