Konshus.ai
Meet Konshus

Hub guide · 2026 edition

The Complete AI Backup Guide.

Every major AI you use is one product decision away from forgetting you. Here's how to back up everything ChatGPT, Claude, Gemini, Limitless, Plaud, Notion AI, and the rest actually know — before the next model update, account issue, or quiet sunset takes it with them.

A vault door opening, with translucent AI chat glyphs streaming in to be archived.

The standard advice for important things is to back them up. We back up photos. We back up code. We back up our taxes. Almost nobody backs up their AI — and almost nobody realizes that the AI they've been talking to for two years has more honest data about them than their journal.

That's a problem for two reasons. The first is that every provider has already, at some point, quietly trimmed memory, swapped a model, or changed defaults in ways that erased part of what their users had built. The second is that the version of you each AI has learned is stuck in that AI. ChatGPT can't hand it to Claude. Gemini can't read your Plaud transcripts. The whole modern stack is non-portable by default, and every export you don't do today is a gap in a future model's understanding of you.

This guide is the unsentimental, tool-by-tool walkthrough. What to click. What's actually in the file. What's missing. And what to do with the export so it isn't just a JSON file rotting in your downloads folder.

Three layers — most exports give you one

"Backing up your AI" sounds like one thing. It's actually three, in ascending order of value:

  1. Raw chats. Every message you've sent and every reply you've gotten. Most exports give you this well. High volume, medium value — it's the source material, not the understanding.
  2. Structured memory. The compact list of facts the model has "pinned" about you. ChatGPT exposes this. Most others don't. Low volume, high value — it's the model's stated belief about you in maybe 2,000 tokens.
  3. Derived voice and tone. How the model has learned to talk to you. This is never exported. It re-emerges only if you train a new model on the raw chats. It's the most expensive layer to recreate and the one that makes a switch feel like a death.

Provider exports cover layer 1 well, layer 2 partially, layer 3 never. That's the gap that vaults — including Konshus — exist to close. But first: get the exports.

Tool by tool

Each section has the path, what's actually in the file, what's missing, and (when we have one) a link to the deeper guide.

01 · Export the ZIP monthly. The memory file is tiny. The chats are everything.

ChatGPT

Path: Settings → Data Controls → Export data. OpenAI emails you a download link within an hour. The link expires in 24 hours, so download immediately.

In the ZIP: conversations.json (every chat you've ever had), user.json/memory.json (the structured memory ChatGPT keeps — usually only 1,200–2,000 tokens), chat.html (a clunky browser viewer),model_comparisons.json (your RLHF feedback).

What's missing: temporary chats, anything you deleted, voice transcripts older than 30 days, and any inferences the model made about you that it didn't formally pin to memory. The export is generous but incomplete.

Full ChatGPT export walkthrough →

02 · Projects context survives chats. But model rollovers don't survive Projects.

Claude

Path: Settings → Privacy → Export data. You get a JSON dump of conversations plus, separately, the system-prompt context for each Project you've created.

In the export: Claude is unusually well-organized. Projects appear with their full setup context preserved. Chats include token counts. No structured "memory" file like ChatGPT — Claude's memory lives entirely in 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 Claude version rollovers if Anthropic deprecates the underlying model. Back up the Project setup separately as Markdown.

Claude Projects & model deprecation →

03 · Takeout works. The hard part is finding the right product slug.

Gemini (Google)

Path: takeout.google.com → search "Gemini Apps Activity" → deselect everything else → Export. Google emails you a download link, usually within minutes. You can also schedule recurring exports (every 2 months for a year) — set this once and forget it.

In the export: JSON or HTML of every prompt and response, organized by date. No structured memory file — Gemini's "memory" is currently inferred from your Google account context, which is not separately exportable.

What's missing: the cross-product context Gemini uses (Calendar, Drive, Gmail) is not in the export — that's "live" memory, and it disappears the moment you stop using a Google account.

04 · Threads export. Spaces export. Reasoning trails don't.

Perplexity

Path: Account → Settings → Data → Download my data. Returns a JSON file with every thread and every Space you've created.

What's missing: the citation trails, the model-of-the-day metadata, and any Pro-search reasoning steps. You get the question and the answer, not the path. For research workflows that matters less; for understanding why the answer was what it was, it matters a lot.

05 · Companion AIs are the most fragile. Export them first.

Pi.ai & Character.ai

The companion-AI category has the highest emotional investment and the worst export tooling. Pi.ai offers a basic text export via Settings → Account → Download data. Character.ai lets you download per-character chat logs but not your personalization data.

What's at stake: the entire relationship is the data. When Replika reset their tuning in 2023, users described it as a death. Don't wait for your provider to do the same. Export monthly.

06 · Notion is the export. AI prompts and pages live together as Markdown.

Notion AI

Path: Workspace settings → Export all workspace content → Markdown & CSV. Notion AI's "memory" is really just your workspace pages, which are first-class data and export cleanly.

What's actually portable: every page, every database, every database property. The AI-generated summaries are baked into the pages they were written on. The AI's "understanding" isn't separately exportable, but the source material is.

Notion as AI memory →

07 · Meeting notes live in the cloud. Export early, export often.

Granola

Path: per-meeting export to Markdown or Notion is one click. Bulk export isn't a first-class feature — write a small script if you have hundreds of meetings, or use the Notion integration as your durable archive.

What's missing: the audio is gone the moment Granola processes it. The transcript and summary are all you have. Treat the Markdown export as the source of truth and feed it into a vault that can search across years of meetings.

08 · Continuous capture is the dream. The export window is the nightmare.

Limitless Pendant

Path: app → Settings → Export → choose date range. Recent data exports cleanly. Older transcripts may be aggregated or summarized rather than raw. Run the export weekly while the data is still fresh.

What you actually own: the transcripts you've downloaded and stored locally. Everything still in their cloud is one product decision away from changing shape.

How to back up your Limitless pendant →

09 · Transcripts in their cloud. Audio in their cloud. Subscription dependent.

Plaud

Path: per-recording export to text or Markdown. No bulk export. The audio files are the source of truth but live behind the subscription — cancel and you lose access. Download monthly to a local archive.

The trap: Plaud's value is the summary, not the recording. The summary is computed once and stored. If you re-record over capacity or change plans, summaries can be regenerated with different prompts, producing different output. The version you exported is the version that survives.

Limitless vs Plaud vs Granola: the ambient capture stack →

10 · Every meeting bot eventually changes terms. Export the transcripts now.

Otter, Fireflies & meeting bots

Otter: per-conversation export to TXT, DOCX, or PDF. Bulk via the API on paid plans.
Fireflies: Settings → Data → Export. Returns JSON with transcripts, summaries, and action items.
Read.ai, Krisp, Tactiq: all have per-meeting export paths; none have a one-click "back up everything I've ever recorded" button.

What to actually keep: the transcript and the summary. The audio is rarely re-listened to. The action items rarely survive contact with reality. The summary is what you'll grep against in a year.

What no provider offers

Run every export above and you'll have a folder of ZIPs. You'll be ahead of 99% of AI users. You'll also notice three things missing that no individual export can give you:

  • Portability across providers. ChatGPT's export can't be uploaded into Claude. Claude's export can't be read by Gemini. Each format is bespoke. Without translation, the exports just sit there.
  • Continuity across model versions. When GPT-5 ships and GPT-4o sunsets, your old memory file might be "compatible" but the tone-fit isn't. New model + old facts = a stranger using your data.
  • A portable voice. The way your AI speaks to you is the part you'd miss most, and it's the part that's never in the export. To preserve it you have to distill it — pull patterns from years of conversations into a compact persona that any model can read.

That's the gap we built Konshus to close. Drop the exports into one vault, distill them into atoms (decisions, projects, preferences, recurring themes), and export a portable persona you can paste into any AI — and which survives the next model swap. The exports above are the input. The vault is what makes them useful.

Short answer

Export ChatGPT, Claude, Gemini, and any meeting/voice tool you use, monthly. Store the ZIPs outside the provider. The exports cover chats well, structured memory partially, and voice/tone not at all — close that last gap with a vault that distills everything into a portable persona.

Frequently Asked Questions

The naming ritual

Backed up everything. Now name yours.

Once your AI's history is safe, it's worth giving the version of it that survives a name. Every Konshus starts with one — and from there, your memory becomes something portable.

Name Your Konshus
or see pricing →