Why I built a knowledge arbitration engine instead of just another memory layer
If you've spent any real time building with LLMs, you've probably run into the same wall I did. Memory systems today are very good at storing things and surprisingly bad at explaining why they believe what they believe. Most of the popular options, things like Mem0, Zep, and similar tools, tend to boil the whole problem down to a single model call. The model looks at a handful of facts, picks a winner, and the system moves on with its life. There's no audit trail, no way to reason about why one claim beat another, and honestly, no real concept that a fact could be true in one place and false in another.
That last part is what actually bothered me enough to start building something new.
Think about it for a second. A pricing rule can be perfectly legal in Germany and illegal in the United States. A regulation can be accurate today and obsolete in six months. Two sources can both look credible on paper and still flatly contradict each other. Collapsing all of that nuance into a single black box LLM judgment felt like exactly the wrong foundation for anything used in insurance, legal, or banking contexts, domains where you eventually have to explain your reasoning to a regulator, not just satisfy a user with a plausible sounding answer.
