What a Context Window Actually Is — and Why a Bigger One Doesn't Mean a Better Memory

You spend an hour walking an AI through a proposal. Background, constraints, the three principles you agreed on at the start. Then you ask it to write the final section "according to the three principles we set", and not one of them comes through.

You remind it. It apologises, and gets a different one wrong.

It is not being careless. Those three principles are no longer in front of it.

A desk that cannot hold everything

A language model has no memory in the way you do. It does not file things away and retrieve them later. Every time it responds, it reads everything currently visible to it and predicts what comes next.

That "everything currently visible" is the context window.

Picture a desk. Every message you have sent, every document you pasted, every reply the model wrote — all of it laid out flat on that desk. Before answering, it scans the whole surface. What is on the desk, it can see. What is not on the desk does not exist as far as it is concerned.

The desk has edges. New material goes on, old material falls off. The three principles from an hour ago are on the floor. The model has not forgotten them; it cannot see them.

Tokens: how desk space is measured

Capacity is not measured in words. It is measured in tokens.

A token is the smallest unit the model splits text into — roughly four characters, or three-quarters of a word, in English. Chinese and other non-Latin scripts typically consume more tokens per character of meaning, which is one reason the same question can cost noticeably more in Chinese than in English.

What matters is that everything on the desk takes up space:

This explains three familiar symptoms of a long session. The model starts re-asking questions you already answered. Responses get slower. And if you are paying per use, the bill accelerates — because every new answer requires re-reading the entire desk.

Does a bigger window mean a better memory?

No. This is the misconception worth correcting in 2026.

Flagship windows have expanded enormously over the past two years; million-token capacity is no longer remarkable. Vendors like advertising the number because it is easy to understand and easy to compare. But fits and is read carefully are two different things.

Researchers documented the effect years ago and named it lost in the middle: place a critical fact in the exact centre of a very long input, and retrieval accuracy drops measurably compared with placing it at the beginning or the end. The performance curve is U-shaped — strong at both ends, sagging in the middle. Recent models have improved substantially on simple needle-in-a-haystack tests, but the tendency persists on tasks that require reasoning across many separated passages.

In desk terms: the desk is three times longer now, but attention still favours what is directly in front and what sits at the far edge. The pages in the middle are the ones that get skimmed.

The practical conclusion is blunt. A large window is not a licence to dump everything in. Eighty pages of unsorted material often performs worse than a three-page summary you prepared yourself.

Five habits that pay off immediately

1. Put critical instructions at the top or the bottom

Since the middle is the weak zone, do not bury your actual requirement inside a wall of pasted text. State the task before the material, then restate the requirement after it.

TASK: Below is a transcript of a meeting. Produce minutes with four columns:
decision, owner, deadline, follow-up.

(full transcript here)

RESTATED: Output the four-column table only. Do not infer anything that is
not explicitly in the transcript.

Saying it twice looks redundant. The difference in output is not.

2. One topic, one conversation

Do not discuss a quotation, then a leave schedule, then return to the quotation in the same thread. The unrelated middle section occupies desk space for the rest of the session and actively interferes — the model cannot tell which material is currently relevant.

The rule is simple: new topic, new conversation.

3. Summarise long documents before interrogating them

To work through a long report, resist pasting the whole thing and asking follow-up questions against it. Split it in two. First, ask for a structured summary: sections, key figures, conclusions. Then paste only the section you actually care about and go deep.

Step one produces a small artefact. Step two starts from a clean desk. The combined cost is usually lower than the single-pass approach, and accuracy is better.

4. The handoff summary: how to continue a long task

You are halfway through something substantial, the desk is nearly full, and the work is not done. Do not push through, and do not open a fresh conversation and re-explain from scratch. Ask the model to close its own books first:

Summarise progress so far as a handoff note: decisions already made, open
questions, and the next step. Bullet points only. Do not add new content.

Take that summary, open a new conversation, and paste it in as the starting point. You have compressed a desk full of loose paper into a single page and started again with room to work.

If you use a CLI tool such as Claude Code, this is exactly what /compact does — fold the process into its conclusions and free up the desk. Same principle, automated.

5. Check how much desk you have left

If your tool exposes context usage — /context in Claude Code, a progress indicator in some web interfaces — get into the habit of glancing at it. Past the halfway mark is the time to tidy up, not after the model starts losing track.

Without a readout, a rough heuristic works: more than twenty or thirty exchanges, or more than one long document pasted, and assume you are past halfway.

Three things people get wrong

1. Assuming it remembers the last conversation. Unless the product explicitly offers cross-session memory, every new conversation starts with an empty desk. What you explained last week is gone.

2. Assuming "please remember this" works. That sentence is just another page on the desk, and it falls off with everything else. Durable preferences belong in a system-level setting — a system prompt, project instructions, a CLAUDE.md-style rules file — not in a verbal aside mid-conversation.

3. Assuming clearing the conversation wastes progress. The opposite is usually true. A clean desk produces better work than an overflowing one, and costs less doing it.

Capacity is a spec; tidying is a skill

The context window is the hard boundary of this generation of AI, and the concept fewest people understand relative to how much it shapes their daily experience. Once it clicks, most "the AI got dumber" moments have both an explanation and a fix.

Models will keep getting stronger and desks will keep getting longer. But as long as only what is visible counts, tidying the desk remains a skill worth having.

One practical note: if you are in Hong Kong or China and want reliable access to long-context models like Claude or GPT, the obstacle is rarely technical. Official sign-up is not available in either market, and foreign card requirements block payment. That gap is what Essevin covers — one account with direct access to multiple official models, topped up through local payment methods, billed on actual usage.

Further reading: Claude Code Getting Expensive? Three Commands to Keep Your Context in Check https://essevin.com/blog/claude-code-context-compact-clear-en?inv=zd9km4gn&lang=en applies the same principles to CLI work, where /context, /compact and /clear map onto habits five, four and two above.

Information in this article is current as of 24 July 2026 and is provided for general reference only; it does not constitute advice of any kind. Third-party product features, pricing and policies are subject to their official announcements. Essevin service details are as shown on essevin.com and in the console.