Glossary
Memory
Memory is a product feature that stores facts about you between conversations and puts them back in front of the model next time, so an assistant appears to know you rather than starting from nothing.
In plain terms
The model forgets everything the moment a conversation ends, and always has. What this feature does is keep a short file about you at the side and paste the relevant parts into the top of each new conversation. It feels like being remembered. It is closer to somebody reading a note about you just before they say hello, and the difference matters as soon as the note is wrong.
Why it matters
It is the feature that turns a capable stranger into something that knows how you work, and it is the feature where a single mistake becomes permanent. An assistant that misreads one remark and files it as a standing preference will apply that preference to everything afterwards, silently, for as long as the entry survives. Nothing prompts you to read what has been stored about you. Ten minutes doing so is one of the highest-value things available in any assistant, and the panel is easy to go years without opening.
How it works
The model itself holds nothing between requests. Every answer is produced from what is in front of it at that moment, and when the conversation closes that material is simply gone. Anything that looks like continuity is the product around the model deciding what to place in front of it next time, which is a filing decision rather than a change in the model.
Entries arrive in two ways, and the second is where the surprises live. Sometimes you ask directly, which produces exactly what you intended. Sometimes the product watches the conversation and extracts what it takes to be durable facts about you, without asking, which is convenient and occasionally captures something you said once in passing as though it were a rule.
Not everything stored goes into every conversation. The product picks what seems relevant to what you are asking, which is why an assistant sometimes raises a detail with no bearing on the question and sometimes ignores something you know it has. Both are the selection step, not the storage, and neither is fixed by adding more entries.
Whatever is inserted occupies part of the same fixed space your documents and your conversation are competing for. A large store quietly reduces the room available for the material you actually want considered, which is a reason to prune it rather than to let it accumulate as a matter of principle.
Nothing expires by itself. An entry that was true when it was written stays exactly as it was, so a project that finished, a role you left or a preference you have changed goes on shaping answers until somebody deletes it. Most products let you view and remove entries, and the control is usually somewhere nobody looks.
Four entries, and only one you chose
Seen in the wild
Open the memory panel in an assistant you use daily and read what it has decided to keep, which is usually shorter and stranger than you expect.
ChatGPTAsk inside a body of work that has been running a while what it already knows about it, then compare that against what you would have chosen to tell it.
ClaudeBuild an assistant step that keeps its own record of past exchanges, where you can see precisely what is carried forward because you decided what goes in.
n8n
Common misconceptions
People assume
The model has learnt about me.
In fact
It has not changed at all. Storage sits beside it and text gets inserted ahead of your message, which produces the same experience by entirely different means. The distinction matters because it tells you where to go when something is wrong: to a list you can edit, rather than to a vendor you would have to persuade.
People assume
It remembers everything I have said.
In fact
It keeps a small extracted set, chosen by the product rather than by you, and often not the parts you would have picked. A remark made once can become a standing preference while something you consider central is never captured at all. Reading the actual list is the only way to know which happened.
People assume
Switching memory off deletes what it has stored.
In fact
Usually those are two different controls. Turning it off generally stops the feature adding and using entries; removing what is already there is a separate action. Where this matters, check both, and check whether anything remains recoverable in an export after you have done so.
Telling them apart
Memory vs Context window
Memory
What survives between conversations. A small stored set, chosen by the product, inserted when it seems relevant.
What fits inside one conversation. Everything currently in front of the model, and gone when the conversation ends.
Forgetting something from an hour ago in the same thread is the second. Forgetting something from last week is the first, and only the first has a list you can go and edit.
Memory vs Knowledge base
Memory
Facts about you and how you work, usually short, usually extracted automatically, usually personal to your account.
Documents you have deliberately supplied for the assistant to answer from, usually shared, and versioned like any other material.
One remembers that you prefer short replies; the other holds the refund policy. Confusing them leads to organisations trying to run company knowledge through a personal memory feature, which does not end well.
Questions
- Does the assistant actually learn from me?
- No. The model is unchanged by anything you type, and would give an identical answer to a stranger asking the same question with the same material in front of it. What changes is the material, because the product has stored some of your earlier conversation and puts it back. That is storage, not learning.
- How do I see what it has stored?
- Most assistants offer a settings panel listing the entries, where they can be read, edited and deleted individually. It is worth doing once properly and then occasionally, because entries accumulate quietly and the list is frequently both shorter and odder than people assume before they look.
- It has remembered something wrong. What happens now?
- It shapes every answer it seems relevant to, indefinitely, because nothing expires on its own. Correcting it in conversation may or may not update the entry. Deleting the entry in the settings panel is the reliable fix, and it is the reason a wrong memory is worth treating as a small maintenance job rather than an irritation.
- Is what it remembers private?
- That depends on the product and the plan, and it is answered in the terms rather than by the feature. The questions worth asking are where the store lives, who else can see it, whether it is included in exports, and whether its contents may be used for training. Anything you would not want recorded should not be said in a conversation with this feature switched on.
Key takeaways
- The model holds nothing between conversations; memory is the product storing and reinserting text.
- Entries are often extracted automatically, so the list is chosen by the product rather than by you.
- Only some of what is stored goes into any given conversation, which explains both odd mentions and odd omissions.
- Nothing expires by itself, so a fact that was true once keeps shaping answers until deleted.
- Reading and pruning the stored list takes ten minutes and almost nobody does it.
Last checked July 2026