Mnemosyne OS
Journal
We publish what we measure, what we miss and what we refuted. Every number arrives with its protocol.
-
September 13, 2026 10 min read
AI memory compared: mem0, Zep, Letta, Cognee, Supermemory, us
Where each published AI memory score comes from: the benchmark, its variant, how many questions, which model graded it. Two of the highest do not say.
comparisonbenchmarkmemory Read →
-
September 12, 2026 4 min read
Your memory re-reads itself while you are away
No single note holds what you spent on the bike this year. So while the machine is idle, your memory re-reads itself and writes the total beside your own notes.
dream-stateconsolidationmemory Read →
-
August 18, 2026 7 min read
A multiplier cannot rescue a zero
Dense embeddings cannot hear a rare proper noun, and doubling a zero is still zero. A second, local ranking fixed it, and 48 new questions confirmed the gain.
longmemevalbenchmarkretrieval Read →
-
August 10, 2026 6 min read
Present is not the same as loadable
A package can be there, at the right version, and still be unreachable to the code that imports it. How v1.3.7 shipped with no local model.
electronpackagingesm Read →
-
August 4, 2026 4 min read
We gave personality control of memory. It cost 31 points.
The most seductive feature we built this year made memory recall worse. The 61-question measurement, the ablation, and why personality shipped as a costume.
personalityretrievalbenchmark Read →
-
July 21, 2026 4 min read
The memory my brain kept asking for
Why I built a local-first memory OS: a prosthesis for a brain that ran too fast for the world.
memoryadhdsovereignty Read →
-
July 16, 2026 5 min read
72.9%, and the three questions we miss
We ran LongMemEval on the full-haystack variant, the one nobody shows. The number, the protocol, the levers we refuted, and the three misses we own.
longmemevalbenchmarkmemory Read →