AI memory systems

give an assistant a memory that outlives the session

Every assistant you have used so far forgets. The context window is not memory — it is a desk that gets cleared when you close the window, and everything you explained goes with it. This track is about the layer you build so that it does not. It is taught from one real, running system rather than from a diagram: Marvin, the memory the Heidelberg team runs behind its own AI sessions. You will see what writes, what reads, what deletes, how text becomes vectors, why the retriever fuses keyword search with meaning instead of choosing one, and how the whole thing is held to a number that fails loudly when recall drops. The track ends where most write-ups stop early: the promotion loop. Capture and recall are the easy half; a memory only compounds when an expiring conversation can become a reviewed, durable lesson in seconds. Build that step first, and the rest of the machinery is worth having.