Memory needs room to change.
A notebook that never updates eventually becomes a museum of old assumptions. Good memory needs to remember what changed, retire what no longer applies, and keep a route back to the evidence.
Source book by 递归客 (diguike) · Full attribution

Forgetting can be organization
Moving an old note out of the everyday index is different from destroying its record. An archive can preserve history while keeping the working desk current. Age, importance, and actual use can help decide what deserves attention.
The book explores retention strategies as design ideas. A frequently read note is not automatically correct, and a rarely read note may be crucial. Avoid turning a popularity score into a measure of truth.
Keep history without mistaking it for today
Use the courtyard for the event.
The courtyard is unavailable. Use Room B instead.
Working desk: two unqualified venue choices. The agent has a conflict to resolve; finding either note by itself could mislead it.
Fictional example. This experiment runs locally and does not call an AI or access your memory.
A thought can point to a picture
Work also produces sketches, audio, diagrams, and meetings. A short written observation can describe why an image matters and link to its original asset. The model or tool then needs the ability to inspect that asset when the task requires it.
Keeping the pointer is valuable; it does not make every modality interchangeable. A note about a diagram may omit precisely the spatial detail needed to understand it. The notebook should invite inspection of the original when words are not enough.
Memory is a foundation for learning from work
An agent can inspect previous attempts, notice repeated failures, and propose a better approach. That is reflection using recorded experience. It is not, by itself, proof of autonomous self-improvement or a change to the underlying model.
The book explicitly marks memory interoperability and some self-improvement ideas as future directions. Treat those as questions to explore. Our concrete starting point is simpler: make useful thoughts durable, find the relevant ones, and check them before they guide new work.
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.