Solving the amnesiac problem for LLM agents. Research series on agents that compound knowledge across sessions — first measurement: +4.6 pp accuracy lift on Terminal-Bench 2.1 with an open-weight executor and a single failure-derived skill file.