REMORY is a neural memory mechanism designed to address information loss that occurs when long-running AI agents compress their interaction history into a text summary to fit within a limited context window. It appends a small, fixed-size set of trainable memory tokens after the summary, and is trained so that a frozen language model can approximate the output it would have produced using the full, uncompressed history. On the SummHay benchmark, REMORY improved source attribution while leaving insight coverage largely unchanged, approaching full-context performance while retaining only about 5% of the original input positions. On long-horizon agent benchmarks including BrowseComp and Terminal-Bench 2.1, models augmented with residual memory produced fewer repeated tool calls and fewer tool errors.
