The Real Test of a Memory Is How Often You Repeat Yourself
A memory benchmark asks whether the right note comes back when you ask for it. That is retrieval. The test that matters day to day is simpler: how often you have to tell your agents the same thing again. That is the measure Curator is held to.
There is a whole category of tests for AI memory, and most of them ask one question in different forms: given a long history and a question about it, does the right piece come back? That is a fair thing to measure. It is also retrieval, and retrieval answers a query and changes nothing. A memory is supposed to change what happens next.
Retrieval, recall, memory
It helps to keep three words apart. Retrieval answers a question you asked. Recall is the present bringing back what matters without being asked, at the moment it bears on the work. Memory is the past changing what the system does now. A store can retrieve perfectly and still leave you repeating yourself every morning.
The measure
So the test Curator is held to is the one a person feels: how often do you have to tell your agents the same thing again? A standing instruction re-given at the start of a session, a correction made again because last week’s did not stick, a rule pasted in once more before you step away. Each repeat is a moment a memory exists to carry.
We count those repeats, and a memory earns its place when that count falls. That is the bar, and it is a harder one than any score, because the only way to move it is to change what your agents actually do.
Benchmarks stay, in the back seat
This does not mean throwing benchmarks out. They are good at catching a regression, and Curator keeps them for exactly that. What they are not is a ranking. Readers, judges, and context sizes differ from one published result to the next, and several headline numbers in this space are vendors’ own claims, so no two rows are measured the same way.
Curator is not being tuned for a retrieval number. Nor is keeping everything a tool ever printed the same as remembering. What a tool’s output taught you is worth keeping, held as something you know. The raw printout is an archive.
What it changes
Holding Curator to this measure changes which work comes first. The question for any change becomes whether it means you say something one fewer time: whether a standing order reaches the session that needs it, whether a correction holds the next time it matters. A change that only lifts a retrieval score is not the goal.
It is the same instinct behind a memory you can measure beats one you just trust, pointed at the question that matters most to the person using it.
Where it stands
Curator is a working pre-release product: local, encrypted, and reached over MCP from the agents you already use. Follow release preparation.
Was this useful?