What Enzyme is / is not

Enzyme sits between ordinary markdown and coding agents.

It’s not here to replace your knowledge base; rather, it makes the knowledge base you already have easier for agents to operate on.

Enzyme is

An agent-operable workspace compiler

Enzyme reads markdown structure — folders, tags, wikilinks, dates, frontmatter, and prose — and compiles it into a retrieval layer your agent can use.

The output gives the agent connected, temporally and structurally grounded context to help the agent decide what to read and cite.

An indexability assessment

Setup assesses the health of your agent’s workspace structure:

  • Which folders/tags/links are prime for indexing, because they expose real projects, people, meetings, or decisions?
  • What is noisy and should be excluded?
  • What small markdown repairs would make retrieval sharper?

Enzyme setup performs a deterministic scan and proposes an initial config that you can customize.

A conservative writeback loop

The most durable agent memory primitives are ordinary markdown artifacts: decisions, open loops, session notes, handoff briefs, research memos, and source-linked observations.

Then enzyme refresh makes those artifacts retrievable next session.

Enzyme is not

Not a hosted chat-memory database by default

Enzyme does not move your markdown into a hosted memory database as the default local workflow. The source of truth remains your workspace.

Catalyst generation uses Enzyme hosted credits/auth by default, or your own env provider when you intentionally pass --use-env-llm, but the local CLI path is built around compiled workspace retrieval, not hosted memory storage.

Not a proprietary hidden graph you must adopt

Enzyme may derive entities and catalysts internally, but it works best when the handles are visible in your files: folders, tags, wikilinks, dates, and frontmatter.

If the workspace does not expose people, projects, clients, or dates, Enzyme surfaces this in its setup diagnosis.

Use grep, editor search, or Obsidian search when you know the exact name, tag, link, date, filename, or phrase.

Use Enzyme when the useful question is associative or contextual:

  • “What changed between these two decisions?”
  • “What should I remember before the next meeting?”
  • “Where has this concern appeared under different names?”
  • “What older notes should inform this project now?”

Not a generic RAG/vector database

A vector database retrieves semantically similar passages. Although this is useful, it’s a fraction of the role of agent memory.

Enzyme uses the workspace’s visible structure and generated catalysts to retrieve source-grounded connections. It is strongest when the answer depends on project context, temporal traces, or recurring questions across notes.

Not magic over sparse or structureless corpora

Enzyme does not automatically understand every relationship from arbitrary text dumps. It works best when the workspace has some handles: dated captures, project folders, people pages, tags, wikilinks, or useful prose.

For raw imports, work with Enzyme setup as a wizard to collaborate on re-organizing into a better structure:

raw export
→ transformed into markdown capture documents
→ augment frontmatter with dates and entities
→ Enzyme indexes the resulting workspace

Not “everything local” unless configured that way

Retrieval over the compiled local index runs locally. Catalyst generation during init or refresh may send selected excerpts through Enzyme hosted credits/auth by default, or to your own env provider with --use-env-llm. Configure a local OpenAI-compatible provider and pass --use-env-llm if you want that stage local too.

Why markdown-first?

Memory products become hard to trust when they hide where memory lives. Enzyme’s position is simpler:

Your markdown workspace is the memory architecture. Enzyme helps the agent draw better connections between it.

Enzyme’s quality draws from the quality of the workspace’s handles, and Enzyme setup helps diagnose this.