How GPTree memory works

GPTree remembers important facts, decisions, and patterns from your conversations. Not transcripts, and not everything you ever typed: the things worth keeping. Mention in March that your newsletter goes out on Thursdays and your readers are mostly small business owners, and a conversation in June already writes for Thursday and for them, without you saying it again.

That's the whole idea in one sentence: you stop re-explaining yourself. Every AI chat tool makes you restate who you are, what you're working on, and how you like things done, and it forgets the moment the tab closes. GPTree keeps the parts worth keeping.

The half of this we care about most is control. Automatic memory you can't inspect is creepy, so the whole system is built around the opposite idea: you can see, edit, disable, or delete every single thing GPTree has remembered.

Kinds, layers, and scopes

A memory has three properties worth knowing.

Its kind: a fact ("the launch is in October"), a decision ("we're going with the shorter name"), a pattern (you always want the summary first and the details after), or an identity ("I run a small bakery and do all our marketing myself"). Identity is the one to know about, and it gets its own section below.

Its layer: Working, Episodic, or Semantic. In practice you'll meet layers mostly as filters when you're looking through what's been stored.

Its scope: Account, Project, or Conversation. Scope is the one to actually internalize, because it controls where a memory applies. Conversation-scoped memories stay with that one conversation. Project-scoped memories travel with the project, which is why a brand-new conversation inside your "Kitchen renovation" project already knows the budget and the contractor's name you settled on last week. Account scope follows you everywhere: your name, your job, the fact that you want plain English and no jargon.

Pinning things on purpose

GPTree picks up memories on its own as you work. When something matters enough that you don't want to leave it to chance, pin it yourself: the Pin to Memory dialog lets you edit the content before it's saved, then choose a Memory Type, a Memory Layer, and a Scope. Scope defaults to the current conversation, which is the conservative choice. Raise it to Project or Account when the fact deserves the wider reach.

And here's the honest part: the automatic side doesn't get it right every time. Sometimes it saves what you'd call noise, and sometimes it misses the thing you thought was obviously important. That's exactly why pinning exists, and why nothing in memory is untouchable.

Who you are, and how it gets corrected

Most memories add up. Identity memories replace each other instead, and that difference is the whole point of them.

When you describe yourself in a chat, who you are, what you do, who you do it for, where you work, GPTree stores it as an identity memory on your account, so it reaches every project and every conversation rather than the one you happened to be in. Identity carries more weight at retrieval than an ordinary fact, because almost every answer depends on it.

The part that matters more: when you say something new about the same part of your identity, the old version is retired rather than kept alongside. Say "I'm a high school teacher now" today and the "I'm a graduate student" from March stops being used. Reject a wrong one and the same thing happens: "I don't run a restaurant, I run a catering company" retires the belief it names, so it stops shaping answers from that turn on. Updating your job title doesn't disturb what GPTree knows about your family or your side project, though, because the replacement is per topic rather than wholesale.

Retired identity memories are not deleted. They stay in the Inspector marked as contradicted, so you can see what GPTree used to think and when it changed its mind. Pinned memories are exempt from all of this: if you pinned it, only you retire it.

Only your own words count. A description GPTree writes about you in an answer never becomes an identity memory, no matter how confidently it is phrased.

Saved is not the same as used

Storing a memory and using it in a reply are two different steps. Every turn, GPTree pulls the memories that look relevant to what you just asked, up to a budget, so a long memory list never floods the conversation. On top of that, one class of automatic memory is held back: a very short fact the extractor was not confident about, the kind of half-sentence that gets clipped out of a longer answer. Those fragments read as noise to the model and used to push replies off target, so they stay stored and visible in the Inspector without being fed into new answers.

Pinning overrides all of it. A pinned memory is always eligible, whatever its length or confidence, which is the reliable way to make sure something you care about is in front of the model every time.

The short profile from your setup chat

Memory is not the only thing GPTree carries between conversations. Your setup chat also leaves behind a short profile: your role or focus, and the problem you said you were trying to solve. That profile rides along into every new conversation so a fresh chat can pick up where you are instead of starting cold.

It only helps while it is right. If GPTree describes your work in a way you would not, correct it in the chat the way you would correct a colleague: "I'm not a graphic designer, I'm a wedding photographer." A correction like that updates the stored profile, and if you state the accurate version in the same breath, that replaces it. Reject a description without offering a new one and the field is cleared instead, because no assumption is better than a wrong one. The change applies to the conversations you open afterwards, so it sticks rather than lasting only until you close the tab.

Seeing and editing what's saved

The Memory Inspector is the management surface. It shows everything GPTree has stored, filterable by project, search text, layer (Working, Episodic, Semantic), kind (fact, decision, pattern), and scope (Account, Project, Conversation). Each row can be edited, pinned, disabled, or deleted. Disable is the useful middle option: the memory stops being used, but it's still there if you change your mind later.

A quick pass through the Inspector every few weeks keeps it honest. Delete the stale decisions (the launch moved to November, and the old October date is still in there), fix the facts that drifted, pin the ones you'd be annoyed to lose.

Two escape hatches exist. If you don't want a session remembered at all, use a temporary conversation: it saves no memories, and the conversation itself is deleted after 24 hours of inactivity. And for questions about data retention or training, the live Privacy Policy is the authoritative source; this article covers the controls, not the policy.

Memory next to projects

GPTree has two context systems, and they split cleanly. Projects hold instructions and reference files the AI uses: deliberate context you author up front. Memory accumulates as a byproduct of working. Your brand voice guide or the vendor's price list belongs in project instructions or an attached file. "We decided on Tuesday to skip the spring sale" belongs in memory, because nobody thinks to write that down ahead of time.

One more relative worth knowing about. When a branch has grown to 40 messages and you want to carry it forward without hauling the whole history, "Summarize this branch" compresses it into a usable summary. Cheaper than hoping memory caught everything.