Same-Evidence Reader-Budget Gating for Token-Efficient Agent Memory

Long-horizon agents can spend many reader tokens on redundant retrieved memories. We study a same-evidence reader-budget gate for the official open-source Mem0 package: the memory backend retrieves five records, and a deterministic risk rule chooses whether the frozen reader receives one or three of those exact records. On 311 LongMemEval questions in a low-accuracy constrained OSS setup, this adaptive gate answers 61/311 cases versus 62/311 for Mem0 top-5, while reducing reader tokens by 37.0\% and reader-visible retrieved-context tokens by 53.3\%. By comparison, fixed top-1 saves more tokens but falls to 50/311, fixed top-2 reaches 54/311, fixed top-3 reaches 60/311, and a matched-budget random top-1/top-3 route reaches 56/311. These results support a constrained token-savings claim for budgeted reader context, not a reproduction of Mem0 platform accuracy or a claim that the memory backend improves answer quality.