Ryan “RyMac” McKinney · USMC · 15 August 2026 · 5 minute read
When the context window fills, the oldest material drops out of view. The model does not announce it, cannot tell you what it lost, and keeps answering in exactly the same confident voice it used when it still had everything. That gap between confidence and completeness is the whole problem.
Older content falls out of what the model can see. Depending on the tool, it either silently truncates or compacts the earlier part of the session into a summary.
Either way you lose detail, and nothing flashes red.
You have been working for 40 minutes. It is going well.
Then it suggests something you already ruled out. You correct it. Ten minutes later it does it again.
It is not being stubborn. The decision is simply no longer in front of it, and it has no way of knowing that it used to be.
Compaction is genuinely clever and it buys real room.
But a summary is lossy by definition. The specific number. The exact wording the client used. The reason you rejected option B.
Those are exactly the details a summary throws away as unimportant.
It keeps the shape of the session and loses the parts you would have wanted.
It helps. It does not solve it, and this is the part people resist.
Research into long-context performance keeps finding that models get worse at using information as the amount in front of them grows. A bigger window filled with everything you own can produce worse results than a small window holding the right 3 things.
The fix is not more room. It is deciding what sits on the desk and what waits in a drawer.
1. Write decisions into a file, not the chat. The moment you decide something, it goes in a file. The chat is not storage.
2. Keep the always-read file short. Mine is 11 lines. It does no work, it just says where to go.
3. Start fresh sessions on purpose. Not as a panic move when it degrades. As the normal rhythm, because the state lives in files and a new session loses nothing.
~/business/ ├── CLAUDE.md <- the map. 11 lines. always read first ├── Context.md <- the desk. what we're doing today └── ops/ ├── clients/ ├── offers/ ├── sops/ └── archive/
Once the state is in files, a full context window stops being a crisis. You clear the desk and the work is still there.
Open a new session and ask where a project stands. 2 moves or fewer and your filing is sound. If it cannot, the information was never written down.
More on the mechanic in what a context window actually is, and on the difference between this and forgetting across sessions in AI agent memory.
The method: Interpretable Context Methodology, the paper. Breakdown: Interpretable Context Methodology in plain English.
CLAUDE.md, Context.md, the folder skeleton, a 1-page setup sheet, and a teardown checklist for whatever you are already paying for. 20 minutes, start to finish. No card, no call, no course to buy.
Send me the folder system