Keep the thought. Shrink the paperwork.
You spend an afternoon learning that the event cannot start after sunset because the venue closes at eight. Tomorrow, you do not need to relive the entire afternoon. You need that discovery, its reason, and where it came from.
Source book by 递归客 (diguike) · Full attribution
A thought is a useful piece of meaning
Here, a thought-sized note means one coherent discovery, decision, change, or constraint. It should make sense when you encounter it later. ‘We changed it’ is short, but it is not a useful memory. ‘Move the event earlier: the venue closes at 8 p.m.; confirmed in the booking email’ gives a future reader something to act on.
Claude-Mem calls its structured notes observations. An observation can include a title, a narrative, facts, concepts, and links to files. The title is a handle for recognition; the details preserve what the title cannot carry.
Which note keeps the useful thought?
Read the booking email. Check sunset. Sketch a 7:30 start. Notice the venue closes at 8. Ask the host about an earlier start. Set the plan for 6 p.m.
Fictional example. This experiment runs locally and does not call an AI or access your memory.
Distillation has a cost
Summarizing means deciding what to keep. A note can omit an exception, flatten uncertainty, or turn a tentative idea into an apparent fact. Good distillation preserves the reason, the scope, and the evidence that a future task might need.
The point is not to make every note as tiny as possible. It is to make it as small as it can be while remaining useful. Think of packing a suitcase: leaving the air behind helps; leaving your passport behind does not.
Human-scale meaning, machine-scale handling
@thedotmack’s central idea is that useful meaning can be packaged at a scale a person recognizes, then handled across many such packages by an AI. We can write one thought at a time; a model can receive many written thoughts together and look for relationships among them.
How many fit depends on their token size, the model’s window, and the space reserved for other work. The calculator on the course home page shows the arithmetic. Fitting 100,000 notes is a capacity scenario, not evidence that every model will reason well over all of them. Claude-Mem also retrieves notes selectively instead of loading the entire library by default.
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.