Blog
Development notes and updates on Enzyme.
-
Compile-time vs runtime memory: what your agent knows before the first message
Most agent memory tools build understanding through conversation. Enzyme extracts it from content that already exists. Two architectures, different tradeoffs, and when each one is the right choice.
-
The preprocessing layer RAG is missing
RAG retrieves passages. Enzyme builds thematic structure before anyone asks a question. Why retrieval alone isn't enough for accumulated content — and how the two work together.
-
Search isn't the hard part
Your team can build semantic search in a week. The preprocessing that makes it useful takes months. Here's what that layer actually involves.
-
An LSP for your notes
CLAUDE.md tells the agent how to search. It can't tell it what 7,000 notes are about. Enzyme builds a conceptual index so it stops grepping blind.
-
The case against tagging your notes
Tags are useful clustering signals — but they're not the retrieval mechanism you think they are
-
When worse embeddings give better results
The case for approximate search in personal knowledge