gbrain
A hybrid knowledge brain: markdown as the source of truth, PGLite for embeddings and search, testing whether retrieval quality justifies leaving flat files.
CREATED MAY 15, 2026 · UPDATED OCT 31, 2026
CAUSE OF DEATH
The market built it better. Claude and ChatGPT shipped memory, dedicated tools did the rest, and a markdown-plus-PGLite brain was a worse version of things I already pay for. The flat files survived; the index did not.
PROBLEM
Notes accumulate faster than they can be reread. Plain search finds what you already know the words for; embeddings find what you meant but lose the exactness that makes a note worth keeping.
HYPOTHESIS
Keeping markdown authoritative and treating the database as a disposable index gets both - and if the index adds nothing over grep at this corpus size, that is a result worth publishing rather than a failure to hide.
STACK
RESULT
Dead before it earned a benchmark. The question it asked, whether an index beats grep over my own notes, stopped mattering once the assistants I use every day started remembering things on their own.
LEARNINGS
- Before building infrastructure for yourself, check whether the tools you already pay for are about to ship it.
- Markdown as the source of truth was the right call. It is the only part that outlived the experiment.