Explainers · measured on 4.47M files
What is the context window? It's the model's RAM.
Every model (ChatGPT, Claude Code, all of them) thinks inside a fixed budget of text called the context window. It is working memory: the system prompt, your messages, the model's own replies, and every tool result all share the same space. When it fills, the oldest context falls away and the model forgets. So the real question is not how big the window is: it is what you let fill it. On a real 4.47-million-file machine we measured what an agent actually loads into that window, and most of it was junk from looking for files.
The concept, plainly
One screen of working memory. Everything shares it.
Text is fed to a model as tokens (roughly word-pieces). The context window is the maximum number of tokens the model can hold in mind at once, a couple hundred thousand on today's frontier models. It is one shared budget, and three things compete for it:
Here is the part that costs you: the model re-reads the entire window on every single turn. Bigger windows help, but they don't change the rule: whatever you load in, you pay to carry, again and again, until it scrolls off the top and is forgotten. Which makes tool output the dangerous line item. And the tool an agent reaches for most is the one that looks for files.
The centerpiece
Watch a file hunt eat the window. Same budget, two ways to spend it.
Both agents below have the identical context window. The one on the left shells out to ls, grep and find to locate a file; every listing it reads gets pasted into the window and stays there. The one on the right asks an index and gets the exact path back. Same task, very different memory pressure.
The window is a fixed box; the split is what changes. The ~58% file-hunting share is measured, not drawn (see below); the segment sizes are illustrative. Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4.47M files.
The two anchors
You've met the context window twice. ChatGPT and Claude Code, same law.
The clearest way to feel a finite window is the two AI tools most people have already used. They fill the same budget in different ways.
Paste a long document, keep chatting, and eventually it "forgets" the top of the thread. You have hit the edge of the window. It has no access to your machine, so its budget fills with conversation, pasted text and retrieved chunks. The lesson is the shape: the window is finite, and what you pour in is what runs out.
To answer "where is X," it runs directory listings and greps, and each result lands in the window. On a real machine we watched this fill roughly 58% of the session's tokens with file-hunting output before the model got to the actual work. Output it re-reads every turn, then forgets.
of a measured Claude Code session's tokens went to finding files, not thinking, then were forgotten a turn later.
Same finite window, two famous places you have felt it. For the class of agents that read your machine (Claude Code, Cursor, Codex and the rest), file search is the single biggest thing crowding it out.
One real task, measured
The worst kind of memory pressure. Junk you re-read every turn, then forget.
Because the model re-reads its whole window on every turn, a directory dump does not cost you once. It costs you on turn one, and on every turn after, right up until it scrolls off the top and the model forgets it ever saw it. You are paying to carry context the agent cannot even use. An indexed lookup returns one line: the path. Nothing to re-read, nothing to forget, nothing crowding out the reasoning you actually wanted.
The fix
Keep the window for thinking. Hand search to an index.
Interlinked keeps an always-warm index of every file on the machine: names and contents, across every repo and the roughly 90% of the disk that was never in git. An agent asks; it answers in the microsecond range and returns just the path. The window stays empty for reasoning. Search stops being a budget the agent has to ration and becomes a fact of the environment, like RAM. If the model is the CPU and the window is the RAM, this is the fast storage they were both missing.
MEASURED ON A RYZEN 9 9950X3D · 64 GB DDR5 · NVMe · WINDOWS 11 · 4.47M FILES.
Do this today
Give your agent a clean window.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Codex, Windsurf, Zed, Cline and the rest. No JSON editing.
Let it index. Filename search works immediately; content search fills in behind it. At rest the whole engine sits around 44 MB, less than a browser tab.
Ask in plain language. Whole-machine when you don't know where something lives; scoped to a folder when you do. One call, one path back: nothing pasted into the window.
Watch the window stay open. The tokens that used to go to directory archaeology now go to reasoning. Same model, same budget: spent on the work instead of the search.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.