Working on a way for local LLMs to share context without centralizing everything

A few of us have been messing around with a problem that keeps coming up with local LLMs the model can be private, the files can be local but the second multiple people or multiple machines need the same context everything starts getting messy. You either duplicate a bunch of embeddings or vector stores, manually move context around or introduce some central service which kind of defeats part of the reason we were running locally in the first place. What we’re experimenting with is basically a small portable context layer where the useful facts or decisions can move between local setups without shipping the entire underlying dataset with them. Still very early and we’re arguing about things like provenance, permissions, stale context and how much the model should actually be allowed to remember. There are only a few of us working on it right now so we’ve got a little Kibu pod where we throw around builds and ideas but before we go much further I wanted to ask people here: If you use local models across more than one machine or with a team, how are you handling shared context today? Do you even want models sharing memory or context between environments or is keeping every local setup completely isolated actually part of the appeal?