Static, RAG, and the memory category nobody shipped

Every AI tool now claims “memory.” Watch what they actually do and they fall into two buckets.

Static memory writes down facts. A decision, a preference, a snippet — stored as a record and handed back later, word for word. Useful, until the day your judgment moves on. The note doesn’t move with it. Last year’s call outranks this year’s, forever.

RAG memory looks smarter. It embeds your notes and pulls back the ones nearest your question. The retrieval is dynamic — but the understanding underneath is frozen the instant it was stored. It remembers what you said. It has no idea you’d say something different now.

For a chat assistant, that’s a mild annoyance. For an agent writing code in a project you started a year ago, it’s a trap. Old projects are full of old assumptions: architecture taste you’ve outgrown, security standards you’ve raised, “temporary” decisions that calcified, patterns you now reject on sight. Feed an agent the frozen version of you and it will confidently rebuild your past mistakes.

There’s a third category, and almost nobody ships it.

Evolutionary memory applies what changed since. Not “here are old notes about this project” — but “here is the project as it was, here is what you’ve decided since, here are the patterns you now reject that this code still uses, and here are the defaults from your other work that should apply before you touch anything.”

Memory stops being a notebook and becomes migration history for your own judgment. When you level up, your old projects inherit the upgrade automatically.

That’s the category Curator is built for. Local, encrypted, yours — and exposed over the Model Context Protocol, so your agent reaches it the same way it reaches the rest of your tools.

We’ll go deeper on how it works in later posts. For now, the one-line version:

Stop restarting from old context. Resume with what you know now.

Get early access — we’ll email when the pre-alpha is ready for you.