Your AI isn't forgetting things because it's getting confused — it's running out of room. Every model has a fixed "context window," the total amount of text it can hold and reason over at once. Once that window fills up, something has to give: older messages get crowded out, buried under newer ones, or dropped entirely.
Last updated: July 2026 · By the KissMySkills team
What a context window actually is
Every time you send a message to Claude, ChatGPT, or Gemini, the model doesn't just see that one message. It sees the entire conversation so far — everything you've typed, everything it's replied, plus any documents, images, or files you've attached — bundled together into one input. That whole bundle is the "context." The "window" is the ceiling on how much of it the model can process in a single pass.
Context windows are measured in tokens, which are small chunks of text — roughly ¾ of a word each. A model with a 100,000-token window can hold something like 75,000 words of combined conversation, documents, and instructions before it hits the wall. That sounds like a lot until you remember that a single 40-page PDF, a spreadsheet export, or a long meeting transcript can eat tens of thousands of tokens on its own — before you've even asked your actual question.
The important thing to understand: the context window isn't a database the model looks things up in. It's closer to short-term memory. Nothing outside that window exists to the model at all, and everything inside it competes for the model's limited attention.
Why it feels like the AI "forgets"
People describe this as the AI "forgetting" the start of a long chat or losing track of details from an uploaded document. That's a reasonable way to describe what you observe, but it's not quite what's happening under the hood. Two different things are going on, and they compound each other.
1. Soft forgetting: dilution
As a conversation grows, older content doesn't disappear — it just becomes a smaller and smaller fraction of what the model is looking at. Early instructions, decisions, or facts get proportionally less "weight" relative to everything said since. The model technically still has access to that content, but it has to work harder to locate and prioritize it among everything else crammed into the window, and in practice, quality on long threads visibly degrades: the model contradicts earlier answers, repeats questions you already answered, or loses the thread of a nuanced instruction from ten exchanges ago.
2. Hard forgetting: truncation
Once the conversation actually exceeds the window's hard limit, something must be cut to make room for new input. Most chat products handle this by silently dropping or summarizing the oldest turns. At that point the content isn't just diluted — it's genuinely gone. The model literally cannot see it anymore, no matter how important it was.
Neither of these is a bug exactly. It's the physical consequence of a fixed-size window filling up. But it means two ordinary habits — pasting in a huge document, and letting a chat run long and meandering — are the two fastest ways to tank output quality, often without you realizing why the answers got worse.
The hidden cost of "just upload the document"
Uploading a big file feels free. It isn't. A large contract, a research report, or a full dataset can consume a huge share of the available window before the model has processed a single word of your actual request. If your model has, say, a 150,000-token window and your document eats 90,000 of those tokens, you're left with 60,000 for the system instructions, the conversation history, and the model's own reasoning and response — on a task that may genuinely need to synthesize across the whole document.
This is exactly the situation that trips people up with heavy analytical work: marketing reports full of tables, financial models, long research corpora. The bigger and messier the input, the less room is left for the AI to actually think well about what you're asking it to do with that input.
Rafa turns Claude into a Senior Marketing Analytics Specialist for exactly this problem — parsing dense reports, dashboards, and campaign data without the standing instructions eating into the room your document needs.
View Rafa — Marketing Analytics Specialist →Why long, meandering chats make things worse too
The other trap is the marathon chat — the thread you never close because it's "got all the context already." Every back-and-forth, every abandoned tangent, every "actually, let's try it differently" stays in the window, taking up space and diluting what actually matters. By message fifty, the model is holding a pile of half-relevant history alongside your current, real ask — and giving all of it roughly equal billing.
The fix isn't "never have long conversations." It's being deliberate about what's actually taking up space in that window, and making sure the things you want the model to weight heavily — your current task, your current document — aren't drowning in things that don't matter anymore.
Where skill files come in
This is the practical reason skill files (or system prompts, custom instructions, project instructions — same idea, different name depending on the platform) exist. A skill file separates two very different kinds of content that normally get mashed together in the context window:
- Standing instructions — who the AI should act as, what standards to apply, what format to output in. This doesn't change task to task.
- Task content — the document you're analyzing, the data you're working with, the actual conversation about the actual problem. This changes every time.
When you paste a long, freeform explanation of "act like a senior analyst, follow this structure, use this tone, always do X and never do Y" into the chat itself, that explanation sits in the window competing for space and attention right alongside your document and your questions — every single time, in every conversation. A skill file loads that same instruction set once, compactly, as a defined role the model steps into. It's written tighter than a typical back-and-forth explanation, and it doesn't get re-litigated or re-explained mid-conversation the way ad hoc instructions often do.
The result is straightforward: more of the window is left for what actually varies — the report you uploaded, the data you're asking about, the twenty follow-up questions in this session. Quality stays higher for longer because the model isn't splitting its limited attention between "remembering how it's supposed to behave" and "understanding what you're actually asking right now."
Ryan configures Claude for the exact kind of ongoing, high-context executive work — board packs, briefings, decision memos — where a compact standing role matters more as the conversation grows.
View Ryan — Chief of Staff →Practical habits that actually help
- Start fresh threads for new tasks. A new conversation means a clean window — no diluted history competing with your current ask.
- Keep standing instructions separate and small. A skill file loaded once beats re-explaining your expectations in every chat.
- Trim what you upload. If you only need chapters 3 and 4, don't upload the whole book — every unnecessary page is tokens the model can't spend on your question.
- Summarize before you continue. On a long-running project, periodically ask the model (or do it yourself) to compress the key decisions so far into a short recap, then start a new thread with that recap plus your skill file.
- Watch for the warning signs. Repeated questions, contradicted earlier answers, or vague responses to specific asks are often symptoms of a crowded or truncated window — not a smarter prompt problem.
In summary:
Every AI model has a fixed context window, and once it fills up, older content gets diluted or dropped — that's the "forgetting" you're noticing, not confusion. Big document uploads and sprawling chat histories both eat into that same finite space. Keeping your standing instructions compact and separate, as with Rafa (Marketing Analytics Specialist) or Ryan (Chief of Staff), frees up room for the document and task content that actually needs it — so answers stay sharp for longer.
Plug in how you (or your team) actually use AI chats and get a real monthly token-waste and dollar estimate for re-prompting vs. a persistent skill file. No signup, no email — just the math from this post applied to your numbers.
Try the calculator →Frequently asked questions
What exactly counts toward the context window?
Everything in the current conversation: your messages, the model's replies, any system prompt or skill file instructions, and the full text extracted from any attached files or documents. It's all measured together in tokens against one shared limit.
Does starting a new chat actually fix the problem?
Yes, in the sense that it clears the window completely — no diluted history, no risk of truncation from an earlier point in the conversation. The tradeoff is you lose whatever context from the old thread you didn't carry forward, so it's worth bringing a short summary plus your skill file into the new chat rather than starting from nothing.
Why not just use a model with a bigger context window?
A larger window helps, but it doesn't eliminate the underlying dynamics — dilution and prioritization still happen at scale, and huge inputs still cost real tokens whether the ceiling is 100,000 or 1,000,000. A bigger window buys you more room; it doesn't buy you a model that automatically knows what to prioritize inside that room.
How is a skill file different from just writing good instructions in the chat?
Both can work, but instructions typed directly into a chat sit in the same crowded window as everything else and often get restated, contradicted, or diluted as the conversation grows. A skill file is loaded once as a defined role — compact and separate from the task content — so it doesn't compete for space with the document or data you're actually working on.
Related: Best Claude skills for executives & chiefs of staff · Best Claude skills marketplaces (compared)