Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Main idea
A workspace that accumulates for weeks has to lighten what it presents, or it becomes unusable: a current state that swells, an illegible list of actions, files reloaded without being read. The reflex is to erase what no longer serves. That misreads what you were after.
What you want to lighten is what gets presented at the moment you pick the work back up — not what exists. The two separate cleanly: a piece of information that no longer serves drops out of the default load, gets condensed or archived, and stays on disk. An idea replaced by another isn't overwritten: it is marked as superseded, keeps its identifier, and points to the one taking over. You genuinely delete in only two cases — an obvious error, an unambiguous duplicate.
The flip side of that precaution is the one thing that gives the setup its value. A conviction has to be defensible two months later, even though it has been superseded since: you need to be able to trace back to what you thought at the time and to what grounded it. A line of reasoning whose history has been erased can no longer be defended — all that is left is a position, without the path that led to it.
Why it matters
This separates two gestures that get confused under the word "cleaning". Lightening a load is reversible and gets decided often; deleting is irreversible and should be decided almost never. Treating them as a single gesture means paying the price of the second every time you only wanted the first.
It also explains why a file system shouldn't imitate the brain on this point. Human forgetting is a strength because it frees; here, what frees is selection at the entrance, and erasure brings nothing but a loss.
Nuances and limits
What has dropped out of the default load is, in practice, forgotten: nobody goes looking for it, save in exceptional circumstances. The difference from a deletion is only real on the day someone traces the chain back, and that day is rare.
And accumulation has a cost of its own. Superseded versions piling up make searching harder, and keeping an old idea doesn't in itself say which of the two was authoritative on a given date.
Open questions
- How do you find again what has dropped out of the default load, when you no longer know it exists?