Storing isn't remembering

With Agent M’s memory core functional, I started testing retrieval, but it wasn’t working the way it should. After measuring retrieval success, it turned out that I only got the right relevant memory card when I prompted the agent 10.9% of the time. Why? Well, it turns out there is a lot to learn about how to make memory retrieval and semantic search “good”. Most of the misses weren’t the search coming back empty — it found notes on the right topic that just didn’t hold the answer. Being about a thing and answering a question about it are not the same property, and a similarity score can’t tell them apart. The words I searched with mattered more than I expected, too: passing the actual question through to the search, instead of the keywords I’d pulled out of it first, was the single biggest gain of the whole effort. And keyword search and vector search miss different things, so neither one alone was ever going to be enough. So today I learned that there is a lot more to remembering than having the memory stored.