There is a lot of good writing right now about agent memory. Vector stores, episodic recall, context compaction, memory layers, mem0. The framing is consistent: the model can reason but it cannot remember, so we bolt on a memory system and close the gap.
I think a large share of what gets called a memory problem is not one. It is a machine that turned off.
Those two failures feel identical from the chat window and they have nothing to do with each other.
Two different things we call memory
Semantic memory is the one everyone is building. What did the user tell me three weeks ago, what is in this codebase, what did we decide about the schema. Facts and context. Embeddings help. This is real work and I am not dismissing it.
