AgenticSTS proposes a bounded‑memory contract for long‑horizon LLM agents that assembles each decision from a fresh user message using typed retrieval, avoiding the jumbled mixture of raw past observations, tool calls, and reflections. This approach isolates the impact