Are AI memory systems actually needed?

Memory systems for products like Claude Code and Codex might not be as reliable as they seem because they're heavily dependent on the harness, how the LLM is guided to retrieve and use them.

More customizable agents like OpenClaw and Hermes solve some of this with plugins and custom memory systems.

But then you need memory that works across devices, local sessions and cloud agents.

So I'm wondering : is AI memory actually a product, or just infrastructure agents will eventually handle themselves?