Mnemosyne OS
New The documentation explains every engine, step by step. docs.mnemosyne-os.io →

Choosing your memory

Your memory will last for years. Choose it on what you can check.

A memory software keeps your notes, your documents and your conversations. You will trust it with years of work. Check three things before you choose it: a score you can recompute, respect for your machine, and search that is accurate and fast.

01 · An audited score

A score is worth something only if you can recompute it.

Every memory software publishes a percentage. That number depends on the test, its variant, the number of questions and the model that grades the answers. Changing the grader is enough to change the number.

Our own answers show it. On the 48 questions we used to tune our engine, the official LongMemEval judge grades them at 81.3%.

Our strict judge grades the same answers at 85.4%. Neither is wrong. They are two instruments.

So a number alone tells you little. Ask how it was measured, and whether you can measure it again.

For us, all five answers are public. The kit recomputes the score in one command.

How the others publish theirs →

02 · Your machine

How a memory manages your hardware matters as much as how it searches.

A local memory runs models on your computer. One reads your documents. Others listen to your voice, read aloud, look at your images. Each one takes memory, sometimes several gigabytes.

Software that starts them all at launch and forgets them on exit slows your whole computer. You then blame the computer.

Mnemosyne OS starts the image and voice engines when you use them. Before the voice assistant loads, the app compares your free memory with what its models need. If memory is short, it tells you first, and lists each running engine with a button to stop it.

When you quit the app, it stops every engine it started. Search runs on the processor. Our speed numbers are measured without a graphics card.

03 · Retrieval

The way it searches decides what you find.

Most AI memories search by meaning. They turn your question into a vector and take the closest passages. That works well for an idea. It often misses a name, a number or an exact reference.

Mnemosyne OS searches two ways at once. One channel searches by meaning. The other searches the exact words, with BM25. The two rankings are merged. The second channel runs on your machine, with no API call. The 77.1% was measured with it.

This double search has a cost. It takes about 5.3 times a meaning-only search. On 10,200 memories it answers in 108 ms, on a Intel Core Ultra 7 155H with no graphics card.

Why a second channel →

04 · Your files

Your memories must stay readable without the software.

You may change tools one day. Your vaults are ordinary files on your disk, readable without us. Your Obsidian or Markdown notes are read where they are.

Your turn

Ask these questions to every memory software, ours included.