Main idea

A long memory is written by addition: each day, what was judged durable gets appended. Nothing in that movement ever removes anything, and a file that only grows ends up as expensive to read as the history it was meant to replace.

An arbitrary ceiling — three hundred lines in long_memory.md — turns that movement into a cycle. Crossing the limit triggers a consolidation: you reread, you merge what says the same thing twice, you archive what the rest of the project has made false, and the file drops back under the bar. The limit doesn't need to be right; it needs to exist, because it is what creates the occasion to reread.

Without a ceiling, rereading stays a good intention nothing triggers. The size constraint is what makes it compulsory — and what makes the long memory a digest rather than a second journal.

Why it matters

This gives concrete shape to a requirement usually stated with no means of enforcing it: knowledge is only worth something if it is maintained. Maintenance can't be decreed, it has to be triggered.

And the ceiling says something about the content, not only the volume: a bounded memory forces you to arbitrate between two durable facts, and therefore to rank things that would never have been compared.

Nuances and limits

Consolidation destroys. Merging two entries erases the circumstance in which each was written, and the archive keeps the text without keeping the reason for having retained it.

And a ceiling set too low manufactures regular amnesias that no alert flags: what leaves the long memory leaves silently.

Open questions

• Who rereads the consolidation, given that it is done by the same agent that wrote the entries?