Why Claude Slows Down, Forgets Things, and Cuts You Off — and How to Fix It

If you use Claude for real work, you've hit the wall: a long chat that gets sluggish, an assistant that seems to forget what you said 20 messages ago, or the dreaded "usage limit reached" right as you're about to finish. None of that is random. It all comes down to one thing — tokens — and once you understand how they're spent, you can stretch a Pro plan a lot further than most people do.

This isn't about the API or writing code. It's about the habits that separate someone who burns through their allowance by lunch from someone who works all afternoon without a hiccup.

The one concept that explains everything: the context window

Think of the context window as Claude's working memory for a single conversation. Every model on every paid plan gets 200,000 tokens of it — roughly 150,000 words. A token is a chunk of text, a bit smaller than a word on average. Your messages, Claude's replies, and any files in the conversation all sit inside that window.

Here's the part almost nobody explains: Claude re-reads the entire conversation on every single reply. When you send message number 30, Claude isn't just reading your new message — it's reprocessing all 29 that came before it, plus every file you attached along the way. That's why a long thread gets slower and eats your usage faster the longer it runs. The conversation isn't just long; it's getting more expensive with every turn.

There's a floor and a ceiling here. The length limit is how big one conversation can get before it's full (that 200K window). The usage limit is how much you can do across all your conversations before your plan needs to reset. They're different meters, and both are driven by tokens.

Why the "forgetting" happens now — and why it's mostly fine

When a conversation approaches the top of the context window, Claude now summarizes the earlier parts to keep going instead of just stopping. If you've ever seen a note that Claude is "organizing its thoughts" mid-chat, that's this automatic compaction working. Your full history is preserved so Claude can still reference it, but the working version it reasons over is compressed.

This is genuinely helpful — it means marathon chats no longer hit a brick wall. The tradeoff: conversations long enough to trigger compaction consume more of your usage, and fine details from early in a very long thread can get fuzzy after they're summarized. If precision on something you said an hour ago matters, that's your cue to start fresh rather than push a bloated thread further.

Use Projects for anything document-heavy

This is the single biggest lever for prosumers, and it's the one most people use wrong. When you drop a PDF into a normal chat, that document's full text lives in the conversation and gets re-read on every message. Attach a 100-page spec and ask five follow-up questions, and you've paid for that spec five times over.

Projects work differently. Files you add to a Project are stored separately and pulled in through retrieval — Claude loads only the relevant chunks for each question rather than re-reading the whole document every turn. For reference material you'll be asking about repeatedly (a contract, a style guide, product docs, meeting notes), a Project is dramatically more efficient than pasting into a chat.

Two things keep a Project lean: keep the project instructions short — general context and Claude's role, not task-specific detail that belongs in the individual chat — and delete files you're no longer using. Every active file is part of what retrieval has to work against.

Chat vs. Project, in one line

Pasting a document into a chat = Claude re-reads all of it on every reply. Putting it in a Project = Claude retrieves only the relevant parts when needed. For anything you'll reference more than once, use a Project.

Start a new chat more often than feels natural

Because every reply reprocesses the whole thread, a fresh conversation is the fastest reset there is. New chat, and the history counter goes back to zero — Claude has nothing to re-read, so your first message costs only itself. When you switch to an unrelated task, don't carry the old context along for the ride; open a new chat.

A useful habit: one conversation per topic or project, not one endless conversation for everything. If you've been troubleshooting a spreadsheet formula for 40 messages and now want help drafting an email, that email doesn't need the formula history dragging behind it. The instinct to keep everything in one place feels tidy, but it's what quietly drains your allowance.

Turn off what you're not using

A few toggles are token-intensive whether or not you need them for a given task. Extended thinking makes Claude reason more deeply before answering — great for hard problems, wasteful for "reword this paragraph." Turn it off when the task is simple. Connectors and tools also carry a cost: each connected tool adds to what Claude has to account for, so disconnecting ones you're not actively using both frees up context and stretches your usage.

None of these are things you set once and forget. Match them to the task in front of you — heavy tools for heavy work, stripped down for quick jobs.

A note on plans and limits

Usage on Pro resets on a rolling five-hour window that starts from your first message, and — worth knowing — your usage is shared across Claude.ai, Claude Desktop, and Claude Code. Burn tokens in one and you've got less headroom in the others. Anthropic doesn't publish an exact message count because it isn't one; how far your allowance goes depends on how big your conversations are, which is exactly what the habits above control.

This is why "just upgrade to Max" isn't always the answer. Max gives you more headroom, but a single bloated conversation — a full document uploaded into a chat, extended thinking left on, 60 messages of accumulated history — can eat a surprising chunk of even a Max allowance in one operation. The person on Pro with lean habits often gets more done than the person on Max with messy ones.

Where to start

Pick the document you paste into Claude most often — the one you keep dropping into new chats to ask about — and turn it into a Project this week. Add the file once, keep the project instructions to a few lines, and ask your questions inside that Project instead of pasting the document again. It's the change that pays back fastest, and it makes the difference between the two limits obvious the moment you feel a heavy chat stop dragging.

Next in the series: Part 2 — Cutting Your Claude API Bill. For developers: prompt caching, model routing, and the Batch API, with real numbers. If you're not writing code against the API, you can safely skip it and jump to Part 3 on Claude Code.