Not everything enters long-term memory. User messages scroll through a rolling history window and vanish after ~10-15 cycles. Only items the agent explicitly chooses to REMEMBER are embedded in LTM. PIN holds temporary task state. The decision gate asks: is this a reusable skill, a proven belief, a relationship fact, or a hard-won lesson? If yes, remember. If it is transient task chatter, let it scroll away.
Query retrieves stored beliefs. NAL/PLN reasoning derives new knowledge from retrieved premises. Derived conclusions are themselves stored via remember, closing an autonomous learning loop: memory feeds reasoning feeds memory.