AI memory is not a retrieval problem. It is an authority problem.
This is a problem I keep running into with long-horizon AI work.
In extended projects, LLMs often remember past decisions and fail to prioritize them. Exploratory “what if” discussions slowly start to behave like approved decisions. Constraints lose their force. Recency starts to win over intent.
I have hit this repeatedly in software architecture and research workflows. So I stepped back and tried to design a theoretical solution.
The Hub-Spoke idea
The framework takes the Hub-Spoke idea and applies it to AI conversation. The idea is simple: separate canonical decisions from exploration, and make state changes explicit.
- A central Hub holds approved decisions and constraints.
- Spokes allow exploration without polluting that state.
- Nothing updates the Hub without a user-approved merge.
This is not a memory system. It is a governance model for conversational state.
If you have seen long-running AI workflows slowly break despite “good memory,” I would value your critique.