The agentic era needs a memory layer. Here's what it looks like.
We gave agents tools, planning, and the ability to act. We forgot to give them a place to remember. That gap is why your agents feel brilliant and amnesiac at the same time.
The last two years gave agents everything except a memory. They can call tools, plan, reason, and act, and then a session ends and it's all gone. The next run reintroduces itself, re-asks what it already knew, and re-pays the model to reconstruct context it had an hour ago. We built agents that are simultaneously capable and amnesiac.
Durable, semantic, isolated, the layer agents were missing.
The requirements are specific and unglamorous, which is probably why it got skipped. It has to be durable, so a restart doesn't wipe a year of context. It has to be semantic, so 'how does Alice like to be contacted?' finds the fact you stored as 'Alice prefers email.' It has to scale to millions of memories without slowing down. And it has to be isolated, so one user's agent never reads another's. That's the layer, and it's the same engine that powers the cache.
The bottom line
Tools made agents capable. Memory makes them coherent. The agentic era doesn't need a smarter model as much as it needs somewhere to remember, durable, semantic, and strictly its own.