Why I stopped putting LLMs in my agent memory retrieval path

Every agent pipeline I've touched in the last eighteen months reinvents memory, and most of them do it badly. Planner decisions never reach the executor. Giant prompts get passed between agents as "context." Tokens burn on stale data. An LLM call sits in the retrieval path, so the same query returns different ranked results on different runs — which makes the system impossible to reason about and impossible to unit-test. The fix, once I'd seen it happen enough times, is boring: treat memory as...