fudan-generative-vision/remory is a Python repository covering residual memory learning for context compression. According to the repository description, it appends learned soft memory after compressing context into a summary, letting an agent continue its work without retaining the full original context. FACTS records no release version, so changes for any specific version cannot be confirmed. The repository is MIT licensed and has 268 stars.
Summaries produced by the agent serve as the condition for generating residual memory. As a result, the same summary can end up with different residual memory attached, and the original model is used frozen. The approach injects continuous embeddings instead of retrieving relevant documents. The paper compared residual memory usage on long-horizon agent benchmarks such as BrowseComp and Terminal-Bench 2.1, using Qwen3.8-27B and GLM-5.3-Flash as backends. On SummHay it reports scores close to the full-context setting while using only 5.2% of input positions. Deployment connects to SGLang, and it states that Qwen3.8-27B requires two or more GPUs of 80GB or more.
The performance numbers live in the graphs and the paper, and the README body does not present them as a table. The README also states that the released GLM encoder checkpoint is 1.24B, while the encoder described in the paper is 1.502B, so evaluations using the released checkpoint cannot be assumed to match the conditions behind the paper's numbers. With the GPU requirements also in mind, the approach is worth referencing, but reproduction and comparison of results are constrained.