Agent memory is a dose to calibrate, not a switch to flip
IBM Research's ALTK-Evolve distills past agent trajectories into inference-time guidelines, no weight updates or annotation. Across eight models the right "dose" varies by capability tier: strong models absorb the full set, weaker ones peak with a compact core plus retrieval (gpt-oss-120b gained +16.1pp at just +5% tokens), and saturated models gain nothing.