Hermes' memory, simply put, is "to remember who you are, what you have done, and what you prefer". By default, Hermes holds persistent memory with local MEMORY.md and USER.md, which work together with project context files. It is not a rough stuffing of the entire chat history into context, but a persistent memory that has been screened.
MEMORY.md, USER.md, and Honcho Difference
| Levels | Function | Who is it for? |
|---|---|---|
| MEMORY.md / USER.md | Local Lasting Memories, Preferences, Project Notes | Stand-alone use, personal assistant |
| Honcho | Deep user modeling across sessions and platforms | Multi-device, long-term companion application |
| Memory Providers | External memory plug-in | People who want to put their memories in independent service |
If you're just using Hermes locally, the default memory is usually sufficient; If you're going to make a product that "remembers users for a long time", Honcho would be more suitable. The official documentation also highlights that Honcho and built-in memory can work in parallel, with the former being more like a deep portrait and the latter being more like local notes. Many people confuse memory, skill, and context files, but they actually have different divisions of labor: memory is "what to remember", context files are "what to do", and skills are "what to load when encountering a specific task".
Who is it not suitable for? If you just want to ask and answer once without cross-session memorization, don't add Honcho to avoid configuration complexity.
