That makes sense. We took a different route in v7, moving away from extracting fine-tuning data as validated subgraphs. Instead of managing a growing graph structure, our reasoning is handled dynamically using a mathematical model. This allows us to compute relationships on the fly rather than relying on a predefined, fixed graph.
Your approach of keeping the graph small as short-term memory and then offloading to long-term storage (or even specialized models like CNNs for certain tasks) aligns well with the idea of optimizing performance at scale. If you’re heading toward a multi-GraphRAG system where instances exchange subgraphs to balance load, that could be a solid way to manage reasoning distribution while still benefiting from structured memory.
For us, avoiding the need for predefined subgraphs has helped maintain efficiency without bottlenecks, and the mathematical reasoning model allows us to handle contextual relationships in real time. It’s a different path to the same challenge ensuring structured, scalable reasoning without the trade-offs of large, static graphs.
In one of the projects in V6 we did a multi graph model for domains. allowing smaller models to dynamically switch on the fly based on the input intent. it worked really good as well.
There are so many ways to build Ai system haha hence V7 …
Like Lynda says here lol
Some understanding of what the V for Version Stands for in Kruel.ai Each version was a full rework of how the memory was designed.
Each version through 4 years gave me a form of what I wanted of a system but if you look through time you will see the costs were huge back in SQL days than Graph brought that down, than Math brought it down even more.
Than I built it for offline running as well so that if you had the hardware and a good local model you could have an offline learning ai that fits in the palm of your hands.
Cloud version with openai = extreme intelligence with memory. Local models can offer similar but still nothing I have seen beats openai models. Close is not beating, and when it comes to accuracy that is very important as that is what ensures the truth in the data and understanding.