Why memory is the real blocker for AGI

Every conversation with an LLM starts from zero.

No memory of what you discussed last week. No sense of context built over time. Just a context window — expensive, short, and empty at the start of each session.

That's the core problem I keep coming back to when people ask why we don't have AGI yet.

Humans don't remember word-for-word what was in an email from last week. But they remember the essence of it. The gist. The relationship context. That accumulated understanding shapes how they think and respond.

LLMs don't have that. Every session, they have to search for the right context before they can do anything useful. It's a retrieval problem masquerading as an intelligence problem.

That's just not how superintelligence should work.

Until models have continuous memory — learning and retaining across sessions, not just within a context window — we're going to keep hitting the same ceiling. Smarter use cases require remembering. And right now, that's the missing piece.