Expand description
No-learn mode: measure without changing what is being measured.
A recall is not a read. It records activations, strengthens co-activations,
emits resonance pulses, refines attention weights, rewrites the narrative and
stores that narrative as a memory. That is the learning, and it is the point
of the system – but it also means the act of measuring changes the state the
next measurement starts from.
Measured: replaying one query returned its top hit at L2=0.45051, then 0.25906, 0.20475, 0.17290 and 0.16108 across consecutive runs, while the pre-fetch confidence on that query climbed 62% -> 81% -> 98%. A single recall rewrote 17 files in the output directory. A benchmark built on that reports how often the system has been asked rather than what it knows, which is why the recorded recall figures were not reproducible: three runs of one unmodified binary gave R@1 of 37, 34 and 34.
Setting MICROSCOPE_NO_LEARN=1 returns from the recall as soon as the
ranking is printed, before any learning state is written. The search is
untouched – the same blocks are scored, boosted and ranked, and the same
numbers are printed – so only the learning writes are skipped, and repeated
runs against one index stop feeding each other.
This is about the memory system only, and is unrelated to [hooks] read_only
in the config, which governs the hook manager. Opt-in and off by default: a
server that forgets a query it just served is broken, so nothing changes
unless a measurement asks for it.
Functions§
- enabled
- True when the process must not write learning state.