The Hermes Agent local model memory failure, often not because the memory system is not turned on, but because the model does not stably call the memory tool. Someone in the community has a similar problem with a small Qwen local endpoint: memory_enabled is on in the configuration, but the model is unstable on memory ops.
First, distinguish between two types of problems
Configuration is not open: memory files, providers, char limits, and flush policies are not set.
Model doesn't use: The model can chat, but it doesn't reliably initiate tool calls, replace memories, or organize user personas.
The most practical investigation method
- Test the same conversation with a cloud model that is determined to support tool calling.
- If the cloud model can remember but the on-premises model can't, the problem is that the model capability or endpoint is compatible.
- Reduce the amount of information written at once, and don't let the small model summarize too many facts at the same time.
- Break down memory into clear facts, such as "users use Mac + Termux," instead of a long story.
Small models are not impossible to memorize, but don't use them to undertake too complex tool planning and long-term portrait sorting.
In a word: If the local model memory is invalid, verify the tool calling first, and then adjust the memory configuration.
Official open source address: https://github.com/NousResearch/hermes-agent; Official document entrance: https://hermes-agent.nousresearch.com/.
