Delay-tolerant networking (DTN) provides the mechanisms needed for both single and multihop space communications when cyclical bandwidth is available between the nodes. Because node contacts can be infrequent, achieving high bandwidth utilization during contacts or extending their duration as possible are topics of special relevance. In this work, an online approach for the Licklider Transmission Protocol (LTP) that is suitable for small size, weight, and power (SWaP) devices is suggested. The method dynamically searches for the optimal segmentation, i.e., the maximum payload size to be used to transmit data blocks. It is contrasted with a Q-learning-based method, which achieves similar performance with larger space complexity. Either option attains reliable bundle transmissions with acceptable delay and throughput levels well beyond the range achievable with static payload lengths and adverse conditions of high loss rates. The approach has minimal information requirements and does not require knowledge of the current bit-error-rate, transmission rate, modulation type, power, gains, nor the one-way light times. The method also requires minimal changes to the sending engine and is fully compatible with standard LTP receive engines. The technique can help to improve the effective DTN capacity by extending the usable contact times and achieving transmissions with improved performance when signal fluctuations occur.