If you feel that you have forgotten what was said before after compression, Hermes Agent focuses on checking the context window and compression configuration of the compressed model. Compression is not a lossless archive, it summarizes the intermediate dialogue into shorter content, and information may be lost if the summary model is too small or the configuration is not properly configured.
Let's distinguish between two situations first
- Normal compression: Reduced detail, but task objectives, key decisions, and file paths remain.
- Abnormal loss: Even core constraints, user preferences, and completed steps are broken, and the follow-up answers are significantly deviated.
The most overlooked configuration
The official documentation reminds that the context window of the summary model should be at least as large as the main model. Because the middle part of the session is sent to the summary model when compressed, if the summary model window is smaller, the compression call may fail, resulting in the intermediate content not being reliably wrapmarized.
How to repair
- Check out the
compression.enabled,threshold,target_ratioinconfig.yaml. - Check the
auxiliary.compression.modeland provider to make sure it's not a cheap model with little context. - After completing each section of the long task, Hermes writes a "list of current statuses," including goals, revised documents, and open issues.
- Write important constraints into project documents or AGENTS.md, not just in early chat logs.
If you're looking for rigorous delivery, it's safest to proactively let Hermes summarize a "recoverable checkpoint" before compression. This makes it easier to retain key information even if the context is compressed later.
