How to find your own notes again

The real problem: why your own notes stop coming back

You kept the discipline. Six months, sometimes three years of notes taken seriously — in Obsidian, in Logseq, in a tidy Notion, or as a Zettelkasten. Four hundred notes, maybe a thousand. Then one day you look for something you are certain you wrote down, and it does not come back.

The failure is not in your memory. It sits between two kinds of memory. What survives six months is the shape of an idea; what the search box demands is the string you typed back then. The mismatch is structural: you remember meaning, but you have to search by words.

What you remember	What you would have to type
"Something about metrics turning against us"	Goodhart’s law
"The study on fragmented attention in open-plan offices"	task resumption cost
"The bit where he takes apart the multitasking myth"	attentional switching
"The article that changed my mind on remote work"	the exact title, forgotten

None of these queries land in a lexical engine. And the miss costs you twice: the note goes unused, and trust erodes. A knowledge base you no longer trust to give back what you put in stops being fed — you keep going for a few weeks out of habit, then the folder becomes a graveyard. That is how a second brain dies: not through a decision to quit, but through quiet disinvestment.

What doesn’t work, honestly

The three standard answers fail for the same reason: each assumes discipline sustained over years, against a base that keeps growing.

Tags. A taxonomy holds while it is small. Past a hundred notes you have #productivity, #focus and #deep-work covering the same ground. Above all, tagging assumes that at the moment of writing you already know how you will search later — which nobody does.

Folders. A tree forces one filing decision on a note that belongs in three places, and it expires: January’s hierarchy no longer describes your concerns in November, and nobody retroactively reorganises a thousand files.

Link discipline. The best of the three, the heart of the Zettelkasten, and the most expensive. Linking a new note requires knowing what is already there; past a few hundred notes that knowledge outgrows working memory. The mesh degrades silently: recent notes link to each other while older ones fossilise into unreachable islands. In practice the discipline lasts about three weeks after you finish the book.

None of this condemns the methods: they are excellent for thinking by writing, restating an idea to understand it, letting two distant ideas meet. Their weak point is not thought, it is retrieval — a separate technical problem you do not solve by asking the user to be more rigorous.

The workflow with Synthiz

1. Open a memoire per durable topic, not per project: a topic outlives the project, which is what makes a source read in March still useful in November.
2. Put sources in, not just your notes: articles by link, PDFs, YouTube videos, podcasts. Audio and video are transcribed automatically, bringing in the material a Zettelkasten leaves outside for lack of typing time.
3. Capture in one click with the Chrome extension while you read, skipping the "I’ll turn this into a note later" step that never happens.
4. Search with today’s words. The semantic layer finds the passage by meaning, the lexical layer catches acronyms, proper nouns and figures.
5. Search from the page you are on. The extension panel queries your base without switching tabs and shows the excerpt around the term, not a results list.
6. Go back to the exact source. Timestamped citations point into the transcript to the second, so you check what the source says rather than what your transcription retained.
7. Write your own thoughts against the sources. That is the part the machine does not do: deciding what the meeting of two ideas means.

What it changes

Search fuses two signals — a semantic layer on embeddings (bge-m3) and a full-text lexical layer — with both scores normalised then combined. More importantly, relevance is not decided by a fixed threshold but by a document’s distance from the distribution of its own query (mean + z*stddev), together with an absolute floor.

That choice came from measurement. On a production corpus of 948 embeddings, the query "féodalisme" — five genuinely relevant documents in the base — peaked at 0.510 similarity, against 0.424 for a query matching nothing: nine hundredths apart. A fixed cutoff at 0.65 returned nothing to a user who did hold those five sources; lowering it far enough to reach them let anything back in. There is no natural zero either: an article about Cleopatra still scores 0.350 against "féodalisme", two texts in the same language always sharing a floor of similarity.

In practice, you find a source by describing the idea in today’s words, without exhuming the vocabulary you used at the time. And when your base holds nothing on the subject, the search returns zero results instead of three adjacent documents — an honest answer beats noise you might act on.

The honest limits

Synthiz does not replace Obsidian for writing. Long-form drafting, a note worked over for weeks, the outline of a chapter: Obsidian is better at all of it, and should stay where it is.

It does not replace thinking through links. The graph you weave by hand encodes how you reason. Concepts link automatically in Synthiz: a different thing, useful, but not a Zettelkasten you built yourself.

There are no local Markdown files that you own. This is a hosted service. Obsidian leaves you a folder of plain text readable in ten years with any editor, whatever happens to the company behind it. Synthiz cannot offer that, and an export feature is not local-first.

It is not a note editor. No plugins, no themes, no CSS, no Dataview, no templates. If your need is to write, this is the wrong tool.

Quality depends on what you put in. A thin base stays thin: better search improves retrieval, it does not create material.