Why ChatGPT and Claude Forget Your Business (And How to Actually Fix It)

Renee Kwan built a custom GPT for her 12-person PR agency in Portland in about twenty minutes. She fed it the agency's style guide, a list of active clients, and a paragraph on how she likes pitches structured — tight, no adjectives in the subject line, always a specific reporter beat. For two weeks it wrote client update emails that sounded like her.

Then a client asked for a pitch angle on a story the agency had covered eight months earlier, and the custom GPT had no idea it had ever touched that account. The style guide was still there. The history wasn't. Renee spent the next twenty minutes re-explaining a client relationship her custom GPT should have already known.

That's not a bug specific to her setup. It's the accurate, documented behavior of the tools she's using — and almost everyone running a business through ChatGPT or Claude has hit some version of it.

The confusing part is that these tools genuinely do remember things now. OpenAI shipped a memory feature that persists facts across conversations. Custom instructions carry your preferences into every new chat. Claude Projects keep a knowledge base and a chat history scoped to one workspace. None of that is nothing. The problem is that "remembers something" and "operates with continuity" are different claims, and the gap between them is exactly where businesses like Renee's lose the twenty minutes back, over and over, without ever being told that's what's happening.

Here's what actually persists, what doesn't, and where the honest ceiling sits.

What Do ChatGPT and Claude Actually Remember About Your Business?

At a glance:

Custom instructions — What it holds: Standing preferences you write once — tone, format, rules to always follow · What it doesn't: Anything that happened; not a record of events, just personality

Memory — What it holds: Specific facts picked up in conversation, recalled later · What it doesn't: A reliable, permanent store — can conflict with custom instructions or get crowded out over time

Projects — What it holds: Files, instructions, and history scoped to one workspace (up to ~200K tokens on Claude) · What it doesn't: Anything outside that one workspace — walled off from every other project

Custom GPTs — What it holds: A reusable bundle of instructions and file context · What it doesn't: The memory feature's saved facts — each session still starts without relationship history

Uploaded files — What it holds: Content for the conversation or project it's attached to · What it doesn't: A standing understanding of your business — doesn't carry into an unrelated chat

Custom instructions. Both ChatGPT and Claude let you write standing preferences that get pulled into every new conversation — tone, format, things to always or never do. On ChatGPT, free and Go accounts get up to 1,500 characters for this; Plus, Pro, Enterprise, Business, and Education accounts get up to 5,000, per OpenAI's own help documentation. That's real, durable, and genuinely useful — it's why Renee's custom GPT could sound like her agency from message one. What it isn't is a record of anything that happened. Custom instructions are a personality, not a memory.

Memory. This is the feature that actually reaches toward what Renee needed: ChatGPT can save specific facts it picks up in conversation — a client's name, a preference you mentioned once — and recall them later without you repeating yourself. It's a real capability and a real improvement over a completely stateless chat window.

It also has real, documented edges. When a saved memory conflicts with a custom instruction — memory says you prefer a casual tone, custom instructions say formal — the two can override each other unpredictably depending on the conversation, rather than resolving in a defined, reliable way. And independent testing suggests there's a practical ceiling on how many saved memories stay accessible before older ones get crowded out — meaning what got remembered in March isn't guaranteed to still be there in August, with no visible warning that it dropped.

Projects. Both platforms now offer a version of this: a workspace that holds its own files, its own instructions, and its own conversation history, separate from your general chat. Claude's version gives each project a 200,000-token context window — Anthropic's own comparison is a roughly 500-page book — and when a project's uploaded material exceeds that, Claude automatically shifts to a retrieval mode that searches the material instead of holding all of it in view at once.

That's a genuinely large amount of working context for one workspace. It is also, deliberately, walled off from every other workspace. A project scoped to one client doesn't know anything about the project scoped to a different client, even if they're both accounts inside the same 12-person agency.

Custom GPTs. These bundle instructions, a bit of file context, and sometimes specific capabilities into something you can reuse. What they don't reliably do is carry the memory feature's saved facts the same way a normal chat does — a custom GPT can be built to sound exactly like your business and still start each new session without the running history a person in the same role would have accumulated. Renee's custom GPT is a real example of exactly this: excellent inherited voice, zero inherited relationship history.

Uploaded files. Whether it's a project's knowledge base or a one-off attachment, an uploaded document is read for the conversation or project it's attached to. It isn't synthesized into a standing understanding of your business that carries into an unrelated chat next week. Upload your client roster to one project on Monday and start a fresh chat on Wednesday, and that chat has no idea the roster exists.

None of this is a criticism of the engineering. Each of these features does what it's built to do, reliably. The confusion is that four different mechanisms — instructions, memory, projects, uploads — each hold a different slice of "knowing your business," and none of them was built to add up to the other three.

Why AI Still Forgets What You Told It

Derek Ansah runs operations at a 20-person specialty insurance brokerage in Hartford. He set up ChatGPT's custom instructions to always format underwriting summaries the same way — carrier, coverage limits, flagged exclusions, in that order — and relied on memory to hold onto which of the brokerage's dozen carrier partners had which quirks: one that's slow on workers' comp renewals, one that requires a specific disclosure form the others don't. For a few months this worked well enough that Derek stopped double-checking it.

Then a summary came back with the disclosure-form flag missing for the one carrier that actually requires it — the memory that should have carried that detail simply wasn't surfaced in that conversation, with no indication anything had been dropped. Nobody caught it until a renewal got returned by the carrier's compliance desk. Derek's business didn't have a policy for what happens when a memory silently doesn't apply, because nothing about the interface tells you when that's happening — it doesn't fail loudly, it just quietly doesn't retrieve what it should have.

Talia Munro, who runs a nine-person e-commerce brand agency in Raleigh, hit the other side of the same wall. She built a Claude project per client, uploaded each client's brand guidelines, past campaign performance, and a running log of what messaging had already been tried. It's genuinely good within a project — Claude can reference a campaign from four months ago inside that client's workspace without being reminded.

What it can't do is notice that two different clients are running nearly identical seasonal promotions in the same week, because that comparison would require connecting two projects Claude is architecturally kept from connecting. The isolation that makes each project reliable is the same isolation that makes it unable to see across the business as a whole.

Grant Oyelaran coordinates freight bookings for a 15-person logistics firm in Minneapolis, running a separate custom GPT for each of the firm's four biggest shipping clients. Every Monday, before the week's routing exceptions come in, Grant spends the first twenty minutes of his day re-pasting the prior week's open issues into each GPT — a customer's dock hours changed, a carrier got flagged for late pickups — because none of it survived the week on its own.

That's not a workflow he chose. It's the workflow the tools require, because nothing in any of the four GPTs can check the state of the actual bookings, notice on its own that Monday has arrived, or update the other three GPTs when something relevant to all of them changes.

Why Can't Any of It Act, Schedule, or Hand Off Work?

Here's what all three of those failures have in common, and it's not a memory limitation you can prompt your way around: none of it can be scheduled, none of it can act on its own, and none of it can hand a task to anything else.

• Memory can hold a fact. It can't decide, on the morning something is due, to bring that fact forward without being asked.

• Custom instructions can shape a summary's format. They can't go check whether the underlying data changed since the last time you looked.

• A project can hold a client's entire history. It can't reach into another project, another tool, or another person's inbox to act on what it knows.

Every one of these mechanisms is retrieval — something that answers when prompted. None of them is a standing presence that notices, initiates, or follows through.

That's the actual shape of the gap Renee, Derek, Talia, and Grant are each running into from a different angle. It isn't that ChatGPT or Claude "isn't smart enough" to remember their businesses. It's that remembering and operating are different capabilities, and every mainstream chat interface — however good its memory gets — was built to do the first, not the second.

Will ChatGPT and Claude Fix This as They Improve?

It's a fair question, since these tools improve constantly. But context-window growth and memory launches are exactly what's already happened, and the ceiling is still there. GPT-4's original context window held about 8,000 tokens; today's mainstream ChatGPT and Claude models handle context windows in the hundreds of thousands of tokens, and Memory itself only shipped in 2024. None of that expansion changed what kind of thing memory is — a bigger box for facts to sit in until asked about, not a system that decides on its own when to act on them.

A model upgrade can make retrieval more accurate — better at surfacing the right saved fact, less likely to contradict itself. It can't make a chat interface initiate a scheduled task, watch a connected system, or hand work to a specialist without being prompted first, because that isn't a capability gap in the model — it's an architectural choice about what a chat product is for. Renee's custom GPT could get a smarter underlying model tomorrow and it still wouldn't notice, on its own, that a client's contract was up for renewal next week. Nobody asked it to check.

That's worth being explicit about, because "wait for the next update" is a genuinely common response to this problem, and it's the wrong one. The fixes that matter here — standing context that doesn't need re-loading, and something that acts on a schedule without a prompt — aren't features that arrive as a side effect of a better language model. They're a different category of product, built around persistence and initiative rather than a bigger, faster answer engine.

What Does Re-Explaining Yourself to AI Actually Cost?

There's no study yet that measures the specific cost of re-briefing an AI tool the way there's research on other kinds of context loss — this is too new a behavior for that research to exist. But the underlying mechanism, a person's attention getting reset and having to rebuild context from scratch, is exactly what's been measured in adjacent research for years, and it's worth being honest about what that research does and doesn't say.

A Harvard Business Review study tracking 137 people across three Fortune 500 companies for up to five weeks found they switched between applications roughly 1,200 times a day, and spent just under four hours a week — about 9% of their total work time — reorienting themselves after each switch. A separate study from Qatalog and a Cornell researcher found it takes people about 9.5 minutes on average to get back into a productive rhythm after toggling to a different app, with 43% of respondents describing the constant switching as mentally exhausting.

Those numbers are about switching between human-facing tools, not about re-explaining context to an AI system — that specific number doesn't exist yet, and this piece won't invent one. What the research does establish is the general cost of a reset: attention and context don't resume instantly, they get rebuilt, and rebuilding has a measurable price even when the information being rebuilt is something the person already knew perfectly well five minutes earlier.

Every time Grant re-pastes last week's exceptions, or Renee re-explains a client relationship her custom GPT should already have, that's the same reset — just with an AI tool standing in for the app being toggled to.

How to Give AI Context About Your Business That Actually Sticks

Fixing this isn't a matter of a bigger memory limit or a larger context window, though both help at the margins. It requires something that behaves less like a retrieval feature and more like a standing colleague: one place that holds what's true about the business — not a fact here and a project there — that stays current without being re-uploaded, that every part of the work can draw on, and that can act on a schedule instead of waiting to be asked.

This is the specific gap Kuvai is built to close. Company Context is a single, org-level knowledge base — not a project, not a per-GPT instruction set — that every teammate a business hires draws on, so a fact grounded in one function is available to every other function without anyone re-explaining it. Each teammate also keeps its own memory on top of that shared foundation, accumulating what it personally learns from doing the work.

And because a teammate is a standing role rather than a chat window — closer to the difference between renting a tool and hiring a role — it can run on a schedule, up to 25 recurring or one-time tasks per teammate, and pull in another teammate mid-task when a job needs a different specialty, the way Grant would want his logistics coordinator to loop in someone from billing without him re-explaining the client from scratch.

None of that makes the underlying models smarter. It gives what they already know somewhere permanent to live, and a way to act on it without a person carrying the context back and forth by hand.

The tools getting real use out of AI right now aren't the ones with the largest context window. They're the ones that stopped treating "remembering" and "getting work done" as the same problem — because they aren't, and no amount of memory alone closes that gap.

Tired of re-briefing a tool that should already know your business? Sign Up for Free — no credit card required, free to start, cancel anytime.