Why does ChatGPT keep forgetting what I told it?

ChatGPT has two different kinds of memory, and both have limits. Inside a single chat, it remembers everything in that conversation — but only up to its context window (roughly 128k tokens on GPT-4-class models, more on newer ones). Across chats, it only remembers what its Memory feature has explicitly saved as a note. That feature is opt-in, has a cap of a few hundred items, and silently drops or overwrites older entries to make room. So even with Memory on, anything it didn't think was important enough to save is gone the next time you open a new chat.

Can I get my old ChatGPT conversations back if they've disappeared?

If you deleted them yourself or used Temporary Chat mode, no — OpenAI says those are permanently removed within 30 days and not recoverable. If your account was suspended or there was an outage, you can sometimes request a data export and recover history. Going forward, the only reliable fix is to export your data regularly (Settings → Data Controls → Export data), which gives you a downloadable ZIP of all your conversations as JSON and HTML files.

Does Claude have memory like ChatGPT?

Claude has a feature called Projects that lets you upload reference documents and set custom instructions that persist across chats within that project. It's useful but narrower than ChatGPT Memory — it doesn't auto-learn from your conversations the way OpenAI's Memory does. Anthropic has also rolled out a memory feature for Claude in 2025, but it's limited and varies by plan. For both tools, the safest assumption is: if you didn't manually save it, don't assume it's there next week.

Will my AI history transfer if I switch from ChatGPT to Claude or Gemini?

No. There is no standard format for AI chat history, no migration tool, and no provider has any incentive to make their memory portable to a competitor. If you want to switch tools — or use multiple in parallel — you have to either (a) re-onboard the new model from scratch, (b) keep a personal context document you paste into each one, or (c) use a third-party layer that holds your context separately from any one provider.

Is there a way to back up ChatGPT or Claude conversations automatically?

Not natively. OpenAI and Anthropic both require you to manually request a data export. There are a handful of browser extensions and third-party tools that scrape conversations on a schedule, but they break whenever the provider updates their UI and can violate terms of service. The most stable approach today is: set a calendar reminder to run a manual export monthly, store the ZIP somewhere durable (cloud drive, encrypted backup), and accept that there will be a small gap between exports.

Does using ChatGPT Memory mean my data is being used to train the model?

By default on consumer ChatGPT accounts, yes — your conversations (memory or otherwise) can be used to improve OpenAI's models unless you turn off 'Improve the model for everyone' in Data Controls. Enterprise, Team, and API usage are excluded by policy. Claude follows a similar pattern: free tier may be used for training under their updated policy; paid and API are not. If training is a concern, check each provider's data settings the day you sign up and check them again after any major policy update.

Is there a free way to give my AI a long-term memory?

Yes — the lowest-effort version is a personal context document. Keep a single page (in Notes, Google Docs, Obsidian, whatever) with the things you'd want any AI to know about you: how you write, what you're working on, your preferences, recurring projects, key people. Paste it into the first message of any new chat. It's clunky, but it's free, fully under your control, and works with every model that exists.

When is it worth paying for a dedicated memory tool?

It's usually worth it once any of these are true: you use AI more than a few times a week, you work across multiple models (ChatGPT, Claude, Gemini), you find yourself re-explaining the same context repeatedly, or you're producing work where consistency of voice and detail matters. Below that bar, a personal context doc plus periodic data exports is genuinely fine. Above that bar, the time you spend re-onboarding starts to dwarf the cost of a tool.

A guide · ~8 min read

Why Your AI Keeps Forgetting You — And What You Can Actually Do About It

If you've spent months training an AI assistant on how you think, how you write, what you're working on — and then watched it forget all of it overnight — you're not doing anything wrong. The tools themselves are built that way. Here's why, and what's actually worth doing about it.

The problem most people don't realize they have

Modern AI assistants like ChatGPT, Claude, and Gemini are excellent at understanding what you say inside one conversation. Where they fail is the moment you close that conversation and open a new one. Suddenly the model has no idea who you are, what you were working on yesterday, or that you asked it three weeks ago to stop using bullet points in everything.

This isn't a bug. It's the default. Each new chat is, by design, a blank slate. The features that try to bridge that gap — ChatGPT's Memory, Claude's Projects, custom GPTs, system prompts — are all partial. They save fragments. They cap out. They get cleared when you switch devices, sign in to the wrong account, or when the provider rolls out a model update.

The cost is invisible but real: you spend a few minutes re-onboarding the model every session, you accept worse output because the AI doesn't know your context, and you slowly stop trusting it with anything that requires continuity — long projects, ongoing writing voice, evolving plans.

Why this keeps happening

1. Context windows are finite

Every AI model has a maximum amount of text it can hold in working memory at once, called the context window. Even the biggest models (1M+ tokens on Gemini 2.5 Pro, 200k on Claude, 128k on GPT-4-class) fill up faster than you'd think. Once you exceed it, the oldest parts of the conversation are pushed out — and they don't come back.

2. Persistent memory features are opt-in and capped

ChatGPT's Memory feature stores a few hundred short notes about you. When it fills up, it silently drops or rewrites older entries. Claude's Projects feature only persists context within one project folder. None of these systems let you see the full set of stored memory in one place, edit it, version it, or move it somewhere else.

3. Providers reset things on their schedule, not yours

When a provider updates a model, deprecates an old version, or changes their memory architecture, your stored personalization is one of the first things to drift or vanish. There's no notification when this happens, and no rollback. (For more on this specific problem, see our companion article on model updates and memory loss, our guide to organizing the prompts you actually reuse, or our overview of the major AI personal assistants.)

4. There's no standard format for AI memory

Your ChatGPT memory cannot move to Claude. Your Claude project cannot move to Gemini. There is no MP3-equivalent for personalized AI context — every provider has its own walled garden, and none of them are incentivized to make their format portable.

Short answer

AI assistants forget you because their default behavior is a blank slate every new chat, their built-in memory features are small and opt-in, and there's no portable format that follows you across providers. The fix is some combination of regular data exports, a personal context document you paste in, and optionally a dedicated memory layer that lives outside any one provider.

Five things you can actually do about it

1. Turn on the built-in memory features and audit them weekly

ChatGPT Memory and Claude Projects are imperfect but free. Switch them on, then once a week open the memory panel (Settings → Personalization → Memory in ChatGPT) and delete anything wrong or outdated. Treat them like a sticky-note drawer — useful for a handful of important facts, not a long-term archive. Best for: casual users who only use one tool.

2. Keep a personal context document and paste it into new chats

One page. Plain text. Things any AI should know about you: who you are, how you write, current projects, preferences, things you don't want it to do. Store it in Notes, Obsidian, or a doc. At the start of an important chat, paste it as the first message. It's clunky but free, fully under your control, and works with every model that exists. Best for: people who want zero dependencies and don't mind the friction.

3. Export your chat history on a schedule

ChatGPT: Settings → Data Controls → Export data. Claude: Settings → Account → Request data export. Both deliver a ZIP of your conversations by email within a few minutes to a day. Set a calendar reminder to do it monthly, drop the file in your cloud drive or encrypted backup. This won't make your AI smarter, but it means you never permanently lose a conversation you cared about. Best for: anyone using AI for work with any institutional memory value.

4. Use a local notes app as a manual memory log

After a useful chat, copy the key takeaways into Obsidian, Apple Notes, Notion, or whatever you use. Tag them. Now your memory lives in a tool you control, and you can search it across years even as your AI tools come and go. The downside is it's manual — you have to remember to do it. Best for: people who already live in a notes app and don't want another tool.

5. Use a dedicated persistent-memory layer

A handful of tools now sit between you and the AI providers and hold your context separately. Konshus (us) is built around this — it ingests your existing ChatGPT and Claude exports, distills them into a portable profile, and lets you carry that profile across any model. There are also broader multi-model wrappers (Poe, OpenRouter) that bring memory features of their own. The honest tradeoff: you're trusting one more company with your data, so the bar for encryption, export, and deletion policies should be high. Best for: heavy users, people who work across multiple models, anyone whose AI history has real ongoing value.

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus

How to evaluate any memory or persona tool

If you decide to try a third-party memory layer — ours or anyone else's — here's the short checklist worth running through before you trust it with anything important:

Can you export everything? In a format you can actually use somewhere else — JSON, Markdown, plain text. Not just a screenshot or PDF.
Can you hard-delete? Not "deactivate," not "archive" — fully erased on request, with a clear timeline.
Is your data used for training? Look for an explicit no, in writing, on paid tiers. Anything vague means yes.
Is it encrypted at rest? Bonus points if individual entries are encrypted, not just the whole database.
Does it work with more than one model? A memory tool that only feeds one provider has the same lock-in problem you're trying to solve.
Who's behind it and how do they make money? Free tools that don't sell anything are usually selling you. Tools with a clear paid plan have an honest business model.

If a tool fails two or more of those, don't use it for anything you'd be upset to lose.

Frequently Asked Questions

If you want one tool that does all of this

Konshus ingests your existing AI history, distills it into a portable persona, and lets you carry that context across ChatGPT, Claude, Gemini, and whatever comes next. Encrypted, fully exportable, never used for training.

Meet Konshus