the engineering team at Bybit had a problem. their knowledge was scattered across docs, meetings, and email. when people changed roles or left, the context they built up walked out with them.
so they built a knowledge compiler and open-sourced it.
no embeddings. no vector database. no $30 memory app.
just notes, documents, and transcripts compiled into a single searchable markdown wiki.
plain text. git-managed. version controlled. editable by hand.
Claude Code and Codex can query the entire knowledge base through MCP and build the structure automatically.
the result: scattered information becomes a wiki that the whole team can search, edit, and maintain.
but here is what the Bybit team figured out that most people haven’t.
you don’t need embeddings to give AI persistent context. you need structure.
raw information compiled into linked articles can be more useful to an AI than vector search across unstructured notes.
the model reads what matters, when it matters.
that is the same principle behind the architecture in the article above.
a few folders. one index file. a system that compounds every day you use it.