What's the difference between memory and context window?

Context window is the working scratchpad inside one conversation — what the model can 'see' right now. Persistent memory is the much smaller bucket OpenAI carries across conversations. The context window can be 128K tokens; the memory bucket is closer to 1–2K. Confusing the two is the most common mistake people make about what ChatGPT 'knows' about them.

How many people actually use ChatGPT memory?

OpenAI hasn't published a precise number, but disclosures and product updates through 2024–2025 suggest memory is on by default for most ChatGPT Plus users and a meaningful share of free users. Adoption of the explicit Memory panel (where users review and edit entries) is much lower — most people never open it.

Can I increase ChatGPT's memory limit?

Not directly. You can't pay for a larger memory bucket inside ChatGPT. What you can do is keep the long version of your context outside ChatGPT — in a vault — and paste a compact summary into the start of each conversation. That's the pattern Konshus is built around.

Where can I see what ChatGPT remembers about me?

Settings → Personalization → Memory. You'll see a scrollable list of entries. Most users are surprised by how short it is. If the list looks thin, that's not a bug — it's the design.

A field note · ~6 min read

How much does ChatGPT actually remember about you?

Q: How much does ChatGPT actually remember?

ChatGPT's persistent memory holds roughly 1,200–2,000 tokens at any time — about 8–12 short facts. That's a few paragraphs of context, not a relationship. The much larger context window (8K–128K tokens depending on the model) only applies inside a single conversation and resets the moment you close the thread.

Q: Does ChatGPT remember everything I tell it?

No. The model silently decides what's worth saving. Most of what you say lives only inside the current conversation and is gone when you close the tab. OpenAI's own help docs describe memory as a curated set of 'key details' — not a transcript.

Short answer: less than you think. ChatGPT's persistent memory holds roughly 1,200–2,000 tokens — enough for about 8–12 short facts. A few paragraphs. Not a relationship. Here's the real shape of it, and why the gap between what people assume and what's actually stored matters.

The numbers

~1,200–2,000 tokens — persistent memory bucket. Roughly 900–1,500 words.
8–12 entries — typical count visible in Settings → Personalization → Memory.
8K to 128K tokens — context window for a single thread (varies by model). Resets when the thread closes.
0 tokens — what the next ChatGPT model remembers about you if OpenAI re-architects memory during a rollout (it has happened).

The two numbers people confuse are memory and context window. The context window is the working scratchpad inside one conversation — large, generous, impressive. Persistent memory is the tiny bucket OpenAI carries across conversations. The difference is roughly two orders of magnitude.

What actually gets saved

ChatGPT's model decides. Most of what you say lives only inside the current conversation and is gone when you close the tab. The model silently picks a handful of things it thinks are worth keeping — your name, your work, a few preferences — and writes them to the memory panel.

Try this: open Settings → Personalization → Memory. Most people are surprised by how short the list is, even after months of heavy use. That's not a bug. That's the design.

Why the limit exists

Every memory entry has to be injected into the prompt of every conversation. At 2,000 tokens, that's already meaningful inference cost across hundreds of millions of users. A 20,000-token persistent memory per user would be financially and architecturally untenable for OpenAI today. So the limit isn't laziness — it's economics.

Which is fine, as long as you know it. The problem is that most users assume ChatGPT is quietly building a deep model of them and act surprised when it forgets something they mentioned three conversations ago. It didn't forget. It never stored it.

Short answer

ChatGPT remembers about a paragraph of you — 8–12 curated facts living in roughly 1,200–2,000 tokens. The bigger numbers you've seen (128K context windows) are per-conversation, not cross-conversation. If you want continuity that survives the next tab close or the next model update, the answer isn't a bigger memory inside ChatGPT — it's a memory outside it.

How to work around it

Three honest options:

Curate ruthlessly. Open the Memory panel, delete entries that don't matter, and force-add the ones that do by stating them clearly ("Remember that I…"). Cheap, free, and limited by the 8–12 slots.
Paste context at the top of every thread. Keep a personal summary in a notes app and paste it in. Works, but maintenance is on you, and it doesn't travel to Claude or Gemini.
Keep a vault outside ChatGPT. Import your ChatGPT export, let something else distill it, and export a compact persona you can paste into any model. This is what Konshus is built to do — see also how to back up ChatGPT.

Frequently Asked Questions

A paragraph isn't a memory. Build a real one.

Konshus ingests your full ChatGPT history, distills it into structured atoms, and gives you a portable persona that fits in any model's prompt — owned, encrypted, yours.

Meet Konshus