as always it depends on your hardware, the tiny models for embedding / reranking typically have low latency
qmd is focussed on local to the point of designing around single machine setups and this creates a gap where one runs agents+qmd on their laptop and LLMs on their ai box
yup, one of the reasons I built gmd to replace qmd, also that I wanted to use Typesense and incorporate llm-wiki concepts. Updating memory does depend on the nature of the change and may impact multiple memories or indexed files, and why llm-wiki concepts are needed too.
Right now, imo, the main reason to build something like this or a harness is to understand the quirks son you can evaluate production grade implementations as the space matures. We are all still very early in the curve.