← All posts · 2026-04-26
Moss vs ChatGPT memory: what each one actually does
ChatGPT got memory in 2024. By 2026 it's the most-used AI memory in the world simply because more people use ChatGPT than any other model. So when someone asks "is Moss really different from ChatGPT memory?", they're asking a fair question. Both products say the same word. Both promise to remember.
The mechanics are different enough that the daily-use feel is different. Here's the honest comparison.
How ChatGPT memory works
When you tell ChatGPT something it considers worth remembering ("I'm vegetarian", "I'm building a SaaS in Cairns"), it writes a short fact to a list attached to your account. Next time you start a chat, that list of facts is appended to the system prompt. The model sees them, can act on them.
That's it. It's a key-value bag of facts the user manually curates (or that ChatGPT auto-extracts and the user can prune). The list is short — capped at maybe a few hundred entries — and the facts are short. ChatGPT calls this "memory across chats."
It's useful. It's not a memory system in the way a human has memory. There is no recall, no association, no resurfacing the relevant past when you change topic. The next chat starts with a clean slate plus your facts list.
How Moss memory works
Moss takes every conversation you have and runs a scribe over it. The scribe extracts semantic units — decisions, findings, preferences, context — and stores them in a personal knowledge graph with embeddings, entity links, and timestamps. When you start a new conversation, Moss runs a retrieval pass that finds units relevant to what you're actually saying right now, not just what you've previously flagged as "remember this."
When you change topic, Moss re-runs retrieval. When a contradiction lands ("I was building in Brisbane" → "actually I moved to Cairns"), the curator agent resolves it. When something genuinely surprising happens ("I just shipped my first paying customer"), it surfaces in next-day morning briefs.
The tagline for the difference is "carried, not fetched." ChatGPT memory waits for you to hit the right keyword and then hands you back the matching fact. Moss carries the relevant past with you into every new conversation.
When the difference shows up in daily use
For a casual user — the person who tells ChatGPT they're vegetarian once and benefits from it for dinner suggestions for the next year — ChatGPT memory is fine. Possibly better than Moss, since the bar for "good" is just "remember the vegetarian thing."
The difference shows up the moment you have an ongoing project, a recurring relationship, or a long-running thread of thinking:
- You're building a product. ChatGPT remembers the product name and the stack. Moss remembers the architectural decisions you made three weeks ago, the customer feedback that came in last Tuesday, the contradiction between two design choices you haven't reconciled yet.
- You're researching a topic over months. ChatGPT remembers you care about the topic. Moss surfaces the article you read in February when a related claim shows up in April.
- You're managing a relationship — co-founder, client, team member. ChatGPT remembers their name. Moss remembers the conversations you've had about them, the patterns of disagreement, what their last decision was.
For these use cases, the saved-facts model isn't enough. It's the difference between a friend who can answer "what's my dietary restriction?" vs. a friend who actually knows you.
The 7M-token claim and why it matters here
ChatGPT's memory list lives inside the model's context window each turn. The model reads your facts, then your message, then generates. That works fine when the relevant past fits in the window.
Moss's memory doesn't fit in any model's context window. We have users with 1M, 3M, 7M+ tokens of conversation history. Moss retrieves the relevant slice on every turn and feeds only that slice into the LLM. Recall stays sharp at scale because we're not asking the model to find the needle in the haystack — we're handing it the needle.
You can see the recall numbers on the benchmark page. The short version: at 7M+ tokens we still hit the right context, while frontier LLMs degrade past 1M.
What ChatGPT does better
Two things, in fairness:
- It's already where you are. If you live in ChatGPT, the friction of switching to Moss is real. Moss has an import flow for ChatGPT exports, but it takes a few minutes and you have to remember to re-do it occasionally.
- It's free, with caveats. ChatGPT's free tier includes memory. Moss has a real free tier too — 15 exchanges per day with full persistent memory, indefinitely (not a 7-day trial). Paid tiers from $9/mo unlock higher caps and document upload.
If your use case is "remember a few facts about me", ChatGPT is enough. If your use case is "be a thinking partner that knows my work the way I do", you'll feel the gap quickly.
Trying Moss
Try Moss on the free tier — 15 exchanges per day with full persistent memory, no credit card. The free tier exists for genuine everyday memory, not a 7-day trial that locks you out. Paid tiers start at $9/mo for higher caps and unlock document upload, longer context, and the premium model lineup.
Import your ChatGPT export to seed memory from your existing history (document upload feature, paid tiers). The first time Moss surfaces something from three months ago that you'd forgotten you said, you'll know.
See the benchmark for the recall numbers. Pricing here.
Try Moss · Blog · Home · Benchmark