AI systems

Two Loops for AI Memory

Live assistants should retrieve lightly; slower agents can synthesize the day after the work is done.

AI memory works better when it is split into two jobs.

During live work, the assistant should stay light. Its memory role is mostly retrieval: surface a prior decision, a useful constraint, or a note I would otherwise forget. If it tries to classify and archive everything at the same time, it becomes slower than the work.

The heavier job can happen later. A nightly agent can read the day, remove noise, connect notes, and turn residue into durable knowledge. It has time to think; the live assistant does not.

The rule I keep coming back to: during the day, ask the AI to remember things for you. At night, let it understand the day for you. Those are different agents.