Your mind is a small desk.
Picture a desk with room for a few cards. One says what you are doing. Another holds something you just noticed. A third reminds you what to do next. Bring in another card and something has to move.
Source book by 递归客 (diguike) · Full attribution
A moment is a working space
@thedotmack describes his own attention as a few thoughts at a time, moving in roughly three-second bursts of moment-to-moment understanding. In this course, that is our picture of the desk: a small place where something becomes meaningful right now. It is a personal metaphor, not a scientific measurement of everyone’s brain.
An AI model also has a bounded working space. Its context window is the material available for one request: instructions, conversation, retrieved notes, tool results, and room for its response. The window is measured in tokens, the pieces of text the model reads. A token is not a thought; one thought may take many tokens to express.
Make room for the next thought
The venue closes at 8 p.m.
Move the event before sunset.
Keep a backup room for rain.
Bring a thought onto the desk:
Choose another thought. Only three fit on this toy desk.
Notebook · 0 saved thoughts
No thoughts have moved off the desk into the notebook yet.
Fictional example. This experiment runs locally and does not call an AI or access your memory.
A bigger desk still needs arranging
A larger context window can hold more cards. That does not guarantee the model will connect the right cards, notice a contradiction, or use an old fact correctly. Available space and reliable understanding are different questions.
Context engineering means arranging this working space deliberately. You choose what belongs on the desk for the current task, what can wait in a notebook, and what needs checking against the original evidence. A huge pile of yesterday’s paperwork is rarely the best starting point.
The notebook changes what a fresh start means
Starting a new conversation clears the desk. Without a memory system, useful discoveries may only exist in the old conversation. With one, a new session can retrieve notes and regain its bearings.
Claude-Mem adds that notebook outside the model’s context window. It captures supported activity, turns useful work into observations, and makes those observations available for later recall. Storing the notebook does not itself use the next request’s tokens. Reading selected pages does.
Behind this lesson
Related chapters in 递归客’s original book:
The source examines Claude-Mem v12.6.2. Links are pinned to the book revision used for this course; current implementation details may differ.