Docs · Getting started
Introduction
Mnemetric is an AI memory benchmark observatory and the home of the Whole-Memory Benchmark Suite. It runs memory systems through pinned public harnesses and publishes every result together with the evidence behind it.
Explore the site
OverviewHow the systems compare, as a table and as charts, plus what we measured.Published claimsEvery figure these projects publish, with the source behind each one.Benchmark catalogFourteen benchmark families and how far each one is implemented.Coverage mapTwenty-four memory capabilities mapped onto twenty test modules.Development evidenceExperiments, including the failures, with downloadable raw files.Attempt historyEvery recorded attempt from the signed, append-only ledger.
Three rules
Recall is not accuracy
Finding the evidence and answering correctly are different things. Separate columns.
Missing is not zero
Not measured means an empty bar, never a low one.
We compete here too
Mnemosyne is our own system. It follows the same rules as everyone else.
Where the project stands
Research preview
One signed result is on the leaderboard. Nothing is publishable yet, and no other system has been run through our harness. Numbers for other systems are quoted from what they publish.