When Does Agent Memory Transfer? Coverage-Calibrated Evaluation of Memory Interventions

A memory retrieved from an agent's own past can improve its trajectory or quietly damage it, and how relevant the memory looks does not reveal which. In short-horizon settings transfer tracks tool-call structure: a memory sharing the workflow helps even when it concerns a different object, while a lexically closer memory naming the wrong action harms, and agents follow the closer-looking one regardless. Structure is not a universal label for harm, since a long trajectory with room to recover can make even a mismatched memory helpful. We therefore recast memory reuse as an evaluation problem. A deployment decision depends on the transfer gain of a specific request--memory pair, the change in full-trajectory return caused by injecting that memory, which we measure by running the same request with and without it and generalize across pairs by tool-call structure rather than surface similarity. On the measured gain an agent may use a memory, fall back to none, or decline to decide, and coverage licenses the choice: where practice supplies comparable structure the gain is estimable and the decision can carry a guarantee, and where it does not the transfer sign is unidentifiable, so a forced use-or-reject decision is provably costly and the unresolved region becomes a target for new practice. On our controlled benchmark only structurally incompatible memories are reliably harmful, a similarity score and an LLM judge rank the harmful cases no better than chance, and gating on the measured gain matches an oracle at zero observed harm where those evaluators pay a real cost. The harm grows with executor size, so scaling the agent does not remove it. A long-horizon environment far from ceiling and two organic-retrieval settings surfaced no harmful class, which bounds where the mechanism applies and is why we estimate transfer rather than classify memories.