You built a custom GPT so you would stop re-explaining yourself, and it worked for about a week. Then it started forgetting who you are, giving generic answers, and quoting a fact you fixed two months ago. The setup was the easy part. Getting a custom GPT to remember your context over the long run is the part almost no tutorial covers, and it is the part that actually matters.
This guide walks the full build, then goes past it: how a custom GPT actually stores context, where its memory quietly breaks, and how to design one that stays current instead of drifting back to generic.
Key takeaways
- A custom GPT is a saved, reusable version of ChatGPT with its own instructions, knowledge files, and abilities. Building one takes no code and about 20 minutes.
- The catch nobody mentions: a custom GPT has its OWN separate memory and does not inherit your main ChatGPT memory, so it starts out not knowing you.
- Its three context slots are all capped or static: instructions cap at 8,000 characters, you attach up to 20 knowledge files, and those files never update themselves.
- Custom GPTs drift in long sessions because they lean on a fixed system prompt and a narrow context window. Structure beats volume.
- The durable fix is a live context layer that updates itself from your real work and feeds any AI, so your custom GPT is never working from a stale snapshot.
What is a custom GPT?
A custom GPT is a version of ChatGPT you configure once and reuse. Instead of pasting the same background, tone, and rules into every new chat, you bake them into a named assistant that opens already knowing its job. OpenAI shipped them so anyone, no code required, could turn a repeated prompt into a permanent tool.
Under the hood a custom GPT is three things stacked together:
- Instructions: a system prompt that defines its role, voice, rules, and what to do or avoid. This is the single most important field.
- Knowledge: files you upload (brand guidelines, product specs, style guides, past work) that the GPT can retrieve from when a question is relevant.
- Capabilities: toggles for web browsing, image generation, code interpreter, and Actions that call external APIs.
That is the whole model. A custom GPT is not fine-tuned and not trained on your data. It is a wrapper: a fixed set of instructions plus a retrievable pile of files on top of the same base model everyone uses. Understanding that tells you exactly where its memory holds and where it cracks.
How to create a custom GPT, step by step
You need a ChatGPT Plus, Team, or Enterprise plan to build one. From there it is quick.
- Open the builder. In ChatGPT, click your profile or the sidebar, choose Explore GPTs, then Create. You land on a split screen: a Create tab that builds conversationally, and a Configure tab where you set everything by hand. Use Configure. You want control, not a chat guessing at your intent.
- Name and describe it. Give it a clear name and a one-line description. This is cosmetic, so keep it short.
- Write the instructions. This is the real work. Spell out the role, the audience, the tone, the format you want back, and the hard rules. Be specific. "Write like a helpful assistant" produces generic output. "Write in second person for non-technical B2B buyers, short paragraphs, no jargon, always end with one concrete next step" produces something usable.
- Upload knowledge. Add reference files the GPT should pull from: your style guide, product one-pager, pricing, examples of your best work. It retrieves relevant chunks when they matter, so more useful files means grounded answers.
- Set capabilities. Turn on web browsing, image generation, or code interpreter only if the job needs them. Leaving all three on for a writing GPT just adds noise.
- Test in the preview. Run your real, hard prompts in the right-hand preview, not softball ones. Watch for where it goes generic or ignores an instruction, then tighten that line.
- Save and share. Save it to yourself, or share via link, or publish to your workspace or the GPT Store.
That is a working custom GPT. Now for the part the tutorials stop before.
Custom GPT examples that are worth building
The GPTs that earn their keep are narrow and repeated. A few that hold up:
- On-brand writer: loaded with your voice guide and 5 examples of your best writing, so drafts land in your tone instead of ChatGPT house style.
- Support reply drafter: loaded with your product docs and refund policy, so replies are accurate instead of confidently wrong.
- Meeting-to-action GPT: paste a transcript, get a clean list of decisions and owners in your format every time.
- Positioning sparring partner: loaded with your ideal customer and competitors, so it pressure-tests messaging against your actual market, not a generic one.
Notice what every good example has in common: it only works because it holds specific, current context about you. Which is exactly the thing that quietly decays.
The catch: a custom GPT has its own separate memory
Here is the single most misunderstood fact about custom GPTs, and it is worth saying plainly.
A custom GPT does not automatically inherit your main ChatGPT memory. It gets its own, separate memory. So if you have spent months teaching regular ChatGPT that you run a B2B SaaS company and write for non-technical buyers, your shiny new custom GPT knows none of it on day one. You are starting over.
A custom GPT is not the same assistant that already knows you. It is a fresh one wearing a name tag.
That surprises people because the whole promise was "stop repeating yourself." And it explains the most common complaint: the custom GPT that keeps forgetting who you are. It was never told. It cannot read your main chats, and unless memory is enabled and building, it only knows what is in the instructions box and the files, both of which you have to maintain by hand.
The three ways a custom GPT stores context (and where each one breaks)
Every custom GPT holds context in three slots. Each has a hard ceiling or a decay problem.
| Context slot | What it holds | Real limit | How it goes stale |
|---|---|---|---|
| Instructions | Role, tone, rules, permanent facts | Caps at 8,000 characters | You edit it by hand; it never updates on its own |
| Knowledge files | Documents it retrieves from | Up to 20 files, pulled via embeddings | Static snapshots. Fix a price or a policy and the old file still answers |
| Built-in memory | Facts it learns during use | Separate per GPT, limited and passive | Often misses, sometimes remembers the wrong thing, easy to drift |
Read that table and the failure mode is obvious. The instructions box is small and hand-written. Knowledge files are frozen the moment you upload them. Built-in memory is thin and does not reliably capture what changed. All three depend on you being the maintenance worker, and none of them notices when a fact about you is no longer true.
That is why a custom GPT feels sharp in week one and generic by week six. Your reality moved. The GPT's context did not.
Why custom GPTs drift and forget in long sessions
Beyond stale context, there is a second failure: within a single long conversation, a custom GPT starts ignoring its own instructions.
This happens because the GPT leans on a fixed system prompt that loses influence as the conversation grows, and on a context window that can only hold so much at once. Early instructions get crowded out by later turns. The model literally has less room to remember the rules you set. It is not broken, it is running out of attention.
The practical fix is structure, not volume. A few habits that hold up:
- Write instructions as a modular framework, not a wish list. Numbered sections (Role, Audience, Rules, Output format) survive long sessions far better than one long paragraph.
- Put the non-negotiables first and last. Models weight the start and end of the instructions most.
- Keep it lean. Custom GPT memory and instructions are for stable preferences and context, not verbatim content. If you need an exact structure every time, paste it in when you need it.
- Audit the memory. Open Settings, Personalization, Manage Memories periodically and delete what is wrong or outdated. Memory is only useful if it is accurate.
These help. But they are all workarounds for the same root problem: the context lives in slots you have to keep filling and cleaning by hand.
The real fix: give your custom GPT a context layer that updates itself
Step back and the pattern is clear. A custom GPT is only as good as the context you feed it, and every built-in way to feed it is capped, siloed, or static. The instructions box is 8,000 characters. The files are frozen. The memory is thin and separate. You are the update mechanism, and you will not keep up, because keeping a knowledge base current by hand is a full-time job nobody signed up for.
The durable answer is to stop treating the custom GPT as the place your context lives, and instead give it a live memory layer outside the box, one that updates itself and serves any AI you use.
That is exactly what Locul does. Locul is a local-first second brain that builds itself from what you already produce, your local files, PDFs, Notion pages, dictation notes, and your LinkedIn profile, and keeps itself current as those things change. When you fix a price or pivot your positioning, the old fact is marked superseded and the new one takes over, so your AI stops quoting the version of you that no longer exists. It then serves that context to your AI tools over MCP, so the same current knowledge reaches Claude, ChatGPT, and the workflows around them, instead of being trapped in one custom GPT's separate memory.
The difference is the maintenance loop. A custom GPT's knowledge file is a snapshot you have to re-upload. A Locul brain is a live model of your work that stays right on its own. Feed a custom GPT from a source that updates itself and the whole "it keeps forgetting me" problem stops being your job to solve.
If dictation is how you think out loud, that is the fastest way to keep the brain current: tools like Contextli turn what you say into notes Locul folds into your second brain automatically, so the context behind every answer reflects what you decided this week, not last quarter.
Custom GPT vs custom instructions vs memory: which does what
People conflate these three, and mixing them up is why setups feel messy. Here is the clean split.
- Custom instructions are global. They apply to every regular ChatGPT chat and are the right home for your permanent, universal context (who you are, how you want answers). See our guide to ChatGPT custom instructions that make AI sound like you.
- Memory is the evolving layer. It accumulates facts across chats, but it is passive and imperfect, and each custom GPT keeps its own. If you are weighing platforms, our breakdown of ChatGPT memory vs Claude memory covers how each handles it.
- A custom GPT is task-specific. It layers its own instructions and files on top for one recurring job, and starts from its own blank memory.
The strong setup uses all three: custom instructions for permanent context, memory for evolving context, and a custom GPT per recurring task. And underneath all of them, a live brain that keeps the facts they draw on correct. For the full playbook on making any AI hold your context long term, see how to give your AI a memory that actually lasts.
FAQ
What is a custom GPT?
A custom GPT is a saved, reusable version of ChatGPT that you configure with your own instructions, uploaded knowledge files, and enabled capabilities. It opens already knowing its role, so you do not re-explain yourself each chat. It runs on the same base model as regular ChatGPT; the customization is the instructions and files wrapped around it, not a retrained model.
How do I create a custom GPT?
In ChatGPT, open Explore GPTs, click Create, and use the Configure tab. Name it, write detailed instructions covering role, tone, rules, and output format, upload reference files as knowledge, and turn on only the capabilities the task needs. Test it against your real prompts in the preview, then save or share. You need a Plus, Team, or Enterprise plan.
Does a custom GPT remember previous conversations?
Only partially, and not automatically. A custom GPT has its own separate memory that a builder can enable, and it does not inherit your main ChatGPT memory. Even when enabled, that memory is passive and can miss things. For reliable recall, front-load stable facts in the instructions and connect an external memory layer that updates itself and works across your tools.
What is the character limit for custom GPT instructions?
The instructions field caps at 8,000 characters. That is enough for a focused role but not for a full knowledge base, which is why anything large or frequently changing belongs in knowledge files or an external, self-updating context layer rather than crammed into instructions.
How many files can I upload to a custom GPT?
You can attach up to 20 files as knowledge in the GPT builder. The GPT retrieves relevant chunks from them via embeddings rather than loading them all at once. The real limitation is not the count, it is that those files are static: when a fact changes, the old file keeps answering until you manually replace it.
Why does my custom GPT keep forgetting my context?
Two reasons. First, it started from its own blank memory and only knows what you put in its instructions and files. Second, in long sessions its fixed system prompt loses influence and its context window fills up, so early instructions get crowded out. Structure your instructions as a modular framework and feed the GPT from a context source that stays current, rather than relying on its built-in memory to hold everything.
Build the GPT once, keep its context alive
A custom GPT is genuinely useful, and building one is the easy 20 minutes. The hard part is everything after: keeping its context accurate as your work, your pricing, and your positioning move. The built-in slots, an 8,000-character box, 20 static files, a thin separate memory, all put that maintenance on you, and that is the job that never gets done.
The move is to stop being the update mechanism. Let a live second brain hold your real, current context and feed it to every AI you use, custom GPTs included, so the answer you get today reflects who you are today. See how Locul builds and keeps that brain current for you, free to start with 500 memories and local AI.
FAQ
