I’ve been building FERNme , an open-source, brain-inspired memory engine for AI agents. While testing, I noticed something interesting. With the same FERNme memory, graph, and retrieval pipeline, a stronger API reasoning model performed very well in my initial tests, while a lightweight local 8B model occasionally made mistakes. The memory itself didn’t change, only the reasoning model did. This made me think memory and reasoning are separate problems. Human memory also isn’t useful just be...