kalomaze: agents write useless memory files — narrow RLVR never taught in-context provability

kalomaze · x · 2026-09-13

kalomaze mocks a common agent behavior: unilaterally writing a useless .MD memory file documenting one specific mistake, assuming the note generalizes to future scenarios without the user asking.

His diagnosis: models have mostly learned, via schlocky narrow RLVR, to run epistemic self-checks only on tangible particular knowledge claims — never trained as agents to assert claims based on in-context provability. Memory becomes performative record-keeping rather than genuine generalization.

Original post →

kalomaze: agents write useless memory files — narrow RLVR never taught in-context provability

More from coding & agent