Enzyme

Enzyme compiles your notes and other knowledge sources into a local index that:

1. Uses temporally grounded context sampling that captures the contextual use of tags and wikilinks, or natural accumulation in folders (i.e. pseudo-Zettelkasten)
2. With this context, it "dreams" questions in the background (called catalysts) and embeds them. At query time, each question finds its nearest docs, so it avoid needing to "dream" full summaries. (see note)
3. Keeps refreshing, but separates new doc ingestion (local, fast) from catalyst evolution (cheap, periodic)

The catalysts help orient the agent to the wiki. Enzyme saves up to 50% tokens vs. plain agentic search with grep.

(Note: your system may already generate lots of memories in different buckets; Enzyme catalysts help thread them together)

Get started

Open your Markdown notes folder with your agent and paste this prompt.

Set up Enzyme for the Markdown notes folder I have open. If you cannot identify it, ask me for its path.

Install Enzyme if needed using https://raw.githubusercontent.com/byenzyme/enzyme/main/install.sh, and install the instructions for this agent. Read and follow the installed enzyme-workspace-setup skill.

Use it to suggest settings from my notes and prepare search. Show me what you propose to include or exclude, and let me correct it before building the local search index.

Then answer one useful question using my notes, with links to the sources, and suggest a question I can ask next. Keep my existing notes in place and unchanged. Ask separately before any cleanup, model download, or recurring task.

After setup, keep working with your agent as usual. The new guidance tells your agent how to use Enzyme and when to refresh it.

Prefer to install it yourself first? Run this, then paste the same prompt:

curl -fsSL https://raw.githubusercontent.com/byenzyme/enzyme/main/install.sh | bash
cd /path/to/your/notes
enzyme install claude # or: codex, hermes, openclaw

What it's for

Enzyme was built for knowledge bases that grow quickly:

• Obsidian vaults that accumulate dated notes in a few folders
• Meeting notes and AI-native, Markdown CRM setups
• Agent memory written as Markdown

At setup time, it proposes a .enzyme program by reading your wiki and existing agent instructions, like:

profile relationships {
  seek "what matters between people"
  notice ["meaningful exchanges", "shared interests", "unfinished conversations"]
}

workspace "notes" {
  source markdown "notes" { path "~/notes" }

  question budget 40
  sample across time
  favor recent periods

  learn questions from folder "people"
    including linked pages
    about relationships
  learn questions from folder "meetings" about operational
  learn questions from tags ["startup", "design"] about decisions
  learn questions from folder "inbox"

  project questions into "~/notes/Readwise"

  // Guidance compiled for your agent, not an enforced hook.
  when asked {
    "Use grep for names, titles, and exact phrases."
    retrieve passages through learned questions
    answer with sources
  }
}

And you can give feedback to modify the program.

Enzyme helps most when a question spans many notes written over time: context for a decision on a project, or a broader theme arising from conversations with one person. It's designed to support knowledge captures that might be later be important, even if they don't serve a current task. Enzyme is focused on doing one thing well: giving agents the tools to make ideas compound. And it ships with a set of profiles designed around a personal knowledge base were refined over 2 years of personal use:

Read...	For...	Profile
Project notes	what's stuck and what keeps blocking	operational
Decision records	why a choice won, what would change it	decision_trace
Saved articles	connections to what you're already working on	resonance_trace
Journals	what keeps returning across entries	reflective
People notes	what matters in these relationships	relational
Feedback / activity logs	what works for you, under what constraints	preference_evidence
(default)	costs, assumptions, live tensions	tension_trace

Reference

Setup without an agent

enzyme compile -v
# Review the .enzyme settings file at the path printed by compile.
enzyme init --quiet
enzyme doctor
enzyme catalyze "a question about your notes"

compile proposes settings and saves them; init builds the local index from them; doctor reports what was indexed and any problems.

Other sources

For Apple Notes, a preset builds a named workspace program from your Notes database. Review the saved .enzyme program before running init:

enzyme --workspace apple-notes compile --preset apple-notes \
  "$HOME/Library/Group Containers/group.com.apple.notes/NoteStore.sqlite"
enzyme --workspace apple-notes init

For Are.na, use enzyme --workspace <name> compile --preset arena <channel-url> and then enzyme --workspace <name> init. Run refresh with the same --workspace name to pick up later changes (--collection is an accepted alias).

Retrieval commands

enzyme petri shows what Enzyme found worth asking about. In a terminal it renders a tree; piped, it emits JSON.

enzyme petri -n 2

Petri
├─ #working-with-others · tag · active · 11 mentions
│  └─ catalysts
│     ├─ Why did every auth change still route through me after I kept promising to write it down?
│     ├─ Why did the third auth extraction attempt stall when nobody wanted to own the session model?
│     ├─ Which reasons for handing the session model to Maya had I stopped believing by then?
│     ├─ Did extraction stay paused once Maya's team owned the session model?
│     └─ …
└─ #system-design · tag · active · 10 mentions
   └─ catalysts
      ├─ Why did I pick a monolith with inner module boundaries over splitting services yet?
      ├─ Did I ever settle who owns the session model once auth moves out?
      ├─ Which part of auth extraction kept adding layers instead of removing one?
      ├─ Does my view that independence means who can change what still hold?
      └─ …

An agent reads this map first, then searches. enzyme catalyze ranks notes by wording and meaning together, and each result carries the catalyst that best connects it to the query:

enzyme catalyze -n 3 "why we keep rewriting the auth layer" \
  | jq '{search_strategy, results: [.results[] | {file_path, similarity, via_catalyst_text, content}]}'

{
  "search_strategy": "hybrid",
  "results": [
    {
      "file_path": "retros/2025-q3-platform-retro.md",
      "similarity": 0.817,
      "via_catalyst_text": "Why did I keep scoping auth extraction as a two-week project, and what kept stalling it?",
      "content": "# Q3 platform retro\n#system-design #working-with-others\n\nScoped auth extraction as a two-week project for the third time. The real blocker was not the token service; nobody wanted to own the session model. Every proposal added a layer instead of removing one. Action: decide ownership before touching code again."
    },
    {
      "file_path": "journal/2025-05-12.md",
      "similarity": 0.693,
      "via_catalyst_text": "Why did every auth change still route through me after I kept promising to write it down?",
      "content": "#working-with-others\n\nReviewed eleven PRs today and wrote none. Every auth change still routes through me because I wrote the session code and nobody else wants to touch it. I keep saying I will write it down so someone else can own it. I have not."
    },
    {
      "file_path": "1on1s/2026-01-maya.md",
      "similarity": 0.579,
      "via_catalyst_text": "What keeps pulling me back to rewriting when Maya's smaller version already worked?",
      "content": "# 1:1 with Maya\n#working-with-others\n\nMaya replaced the token refresh path with something half the size of mine. She asked me twice and then stopped asking, which is the point. I want to do this with billing before I start rewriting it myself."
    }
  ]
}

Message archives (experimental)

Enzyme indexes SQLite tables alongside Markdown — iMessage, WhatsApp, Mail, or any table of dated rows.

Declare a source sqlite inside a named workspace program, then run init and periodic refresh. The SQLite source setup skill shows the source declaration and source-scoped readings.

None of these is a special case. A handle_id column is repeated person values across dated rows, the same way [[links]] are repeated person values across dated notes. Both collapse to entity occurrences with effective dates, and nothing downstream knows which one it came from: the same profiles, budgets, and catalysts apply to a message thread and a folder of meeting notes.

A source names columns by the role they play rather than by the app they came from:

Role	Means
id	the row's identity
who	participants; scalar, JSON array, or delimited
when	the row's timestamp
what	the text to read
where	the container the row belongs to (optional)
weight	numeric significance per occurrence (optional)

API keys

Enzyme needs an API key only for catalyst generation and for Markdown enzyme compile's selection step. Source preset compilation does not call a model. By default enzyme init uses Enzyme's hosted bootstrap and ignores inherited OPENAI_* variables so it does not spend your personal key. Credential resolution is explicit key → configured local model → anonymous brokered free config, with no additional configuration.

The first configured vault on a machine initializes without login. Refresh, publishing, account credits, and additional vaults require enzyme login.

To bring your own OpenAI-compatible key, pass --llm env (alias --use-env-llm), which reads only OPENAI_API_KEY plus optional OPENAI_BASE_URL and OPENAI_MODEL. init and refresh also take --llm local (the installed local model; fails if none is installed), --llm hosted, and --llm none (index, embed, and select without generating catalysts). Without --llm and without any hosted or env key, catalyst generation is skipped and indexing, embedding, and local search still work.

A model served on your own machine (LM Studio, Ollama, a llama.cpp server) works through --llm env with OPENAI_BASE_URL pointing at it; OPENAI_API_KEY must be non-empty even if the server ignores it. Enzyme treats localhost addresses like its built-in 135M model: a compact prompt, at most 10 excerpts per entity, one call per entity, and a 512-token output cap. That includes a proxy on localhost in front of a remote provider. For a larger model, set ENZYME_LOCAL_MODEL=0 to send it the same prompts and job splitting as hosted generation (ENZYME_LOCAL_MODEL=1 does the reverse for a small fine-tune hosted elsewhere). That path runs up to 8 calls at once and each request must finish within 120 seconds, so on a slower model set ENZYME_LLM_CONCURRENCY to what the server handles and lower ENZYME_CATALYST_MAX_QUESTIONS_PER_CALL to keep each call short.

Markdown enzyme compile is an explicit OpenRouter Decisions operation. It reuses the hosted lease from enzyme login and the free-config broker; an explicit OPENAI_API_KEY with OPENAI_BASE_URL=https://openrouter.ai/api/v1 takes precedence. Catalyst generation uses OPENAI_MODEL; Decisions uses ENZYME_JEV_MODEL (default typesafe/jev-1.13).

License

Enzyme is licensed under the Apache License 2.0.