Other Minds — Godfrey-Smith "The octopus has 350 million neurons in its arms. The arms think for themselves." #distributed-cognition A Mathematical Theory of Communication "Shannon's key insight: meaning is irrelevant to the engineering problem of communication" #information-theory Definition, captured "stigmergy — organisms coordinating through traces left in the environment, not direct signals" #emergence The Use of Knowledge in Society "Hayek, 1945: 'the knowledge of which we must make use never exists in concentrated or integrated form'" #local-knowledge Design review "client keeps choosing the denser option. said they want minimal. the saves say otherwise" #taste User research "three user interviews mentioned 'trust' — each time meaning something different" #discovery Lecture note "Gödel: any formal system rich enough to express arithmetic contains true statements it cannot prove" #incompleteness Janeway's Immunobiology "The immune system recognizes self. Everything else gets a response." #recognition Team retro "the onboarding decision from Q2 resurfaced in the retention retro. nobody connected them" #decisions Arbesman — The Half-Life of Facts "half-life of a fact in physics: ~13 years. in surgery, closer to 7" #epistemology A Pattern Language "Alexander, 1977: 'a pattern language is a network of patterns that call upon one another'" #design-patterns gwern.net "Gwern on spaced repetition: the real cost isn't forgetting, it's never encoding in the first place" #memory Women, Fire and Dangerous Things "Lakoff: human categories organize around prototypes, not necessary and sufficient conditions" #cognition Dev journal "we keep re-deriving things the codebase already solved. git blame is not knowledge management" #infrastructure

Other Minds — Godfrey-Smith "The octopus has 350 million neurons in its arms. The arms think for themselves." #distributed-cognitionA Mathematical Theory of Communication "Shannon's key insight: meaning is irrelevant to the engineering problem of communication" #information-theoryDefinition, captured "stigmergy — organisms coordinating through traces left in the environment, not direct signals" #emergenceThe Use of Knowledge in Society "Hayek, 1945: 'the knowledge of which we must make use never exists in concentrated or integrated form'" #local-knowledgeDesign review "client keeps choosing the denser option. said they want minimal. the saves say otherwise" #tasteUser research "three user interviews mentioned 'trust' — each time meaning something different" #discoveryLecture note "Gödel: any formal system rich enough to express arithmetic contains true statements it cannot prove" #incompletenessJaneway's Immunobiology "The immune system recognizes self. Everything else gets a response." #recognitionTeam retro "the onboarding decision from Q2 resurfaced in the retention retro. nobody connected them" #decisionsArbesman — The Half-Life of Facts "half-life of a fact in physics: ~13 years. in surgery, closer to 7" #epistemologyA Pattern Language "Alexander, 1977: 'a pattern language is a network of patterns that call upon one another'" #design-patternsgwern.net "Gwern on spaced repetition: the real cost isn't forgetting, it's never encoding in the first place" #memoryWomen, Fire and Dangerous Things "Lakoff: human categories organize around prototypes, not necessary and sufficient conditions" #cognitionDev journal "we keep re-deriving things the codebase already solved. git blame is not knowledge management" #infrastructure

Memory retrieval tuned to your knowledge base.

Your users bring collections. Enzyme makes your agent fluent in what they care about — faster and at a fraction of the tokens.

42,000+ downloads · Local CLI · Hosted workflows

$ curl -fsSL enzyme.garden/install.sh | bash

$ cd ~/workspace/notes && enzyme install codex
# or: enzyme install claude / hermes / openclaw
Updated AGENTS.md
Installed .agents/skills/enzyme/SKILL.md

> set up Enzyme for this vault

$ cat AGENTS.md

<!-- enzyme:start -->
## Enzyme Workspace Context

This workspace uses Enzyme for artifact-native agent memory.
Keep durable memory as appendable markdown: session notes,
decisions, observations, preferences, and project logs.
Run all `enzyme` commands from the vault root.

### Working memory

`enzyme petri` is working memory: it returns current
entities and catalysts, which are source-grounded questions
compiled from the artifact trail.


Use catalyst phrases as vocabulary for `enzyme catalyze`
searches. They connect to precomputed content that the user's
raw words may not find.

### Search

- `enzyme catalyze "query"` searches by concept/theme.
  Compose queries from petri catalyst vocabulary.


<!-- enzyme:end -->

$ enzyme scan --write-config
Suggested config saved to ~/.enzyme/config.toml
Found folders, tags, links, logs, and expandable page clusters

$ enzyme init
Indexed: 1,247 markdown files
Catalysts: 84 generated
Ready for petri and catalyze

$ cat ~/.enzyme/config.toml

[vaults."/Users/you/obsidian"]
excluded_tags = []
excluded_links = []
excluded_folders = []
min_top_catalysts = 40
entities = [
  "folder:inbox",
  "folder:people",
  "folder:readwise",
  "folder:projects",
  "#pkm",
  "#travel",
]

scroll to discover

Helping thinkers compound their agent's knowledge at

and more

Your agent already leaves a memory trail.
Enzyme makes it usable next session.

Session notes, decisions, meeting captures, inbox files, and AGENTS.md are already memory artifacts. Enzyme treats that artifact trail as the source of truth. It reads the structure you can inspect—tags, folders, links, timestamps—and compiles associative context from it.

Other memory layers begin with a database of conversation facts. Enzyme begins with files and logs your agents can keep appending to.

SCAN → INIT

Point it at a markdown folder.

› enzyme init

Indexed: 1,247 discovered, 1,247 new

Catalysts: 84 generated

Discovering entities [14/14]

~ #user-interviews → What does the recurring mention of ‘trust’ across interviews reveal…

~ / investor-updates → How does the framing of retention shift between Q2 and Q3…

~ [[retention]] → What assumption about churn is contradicted by the onboarding…

~ #product-decisions → What does the gap between stated priorities and actual velocity…

Embedded: 980 docs (1,948 docs/sec)

Similarities: 84 catalysts × 980 documents

Done! (14.2s)

Tags, folders, links, timestamps, append-only logs—whatever structure you already have. No memory database.

PETRI → CATALYZE

The right context, before the first question.

User

“Why is retention dropping if onboarding scores are up?”

agent ▸ enzyme catalyze "retention onboarding"

matched catalyst · [[retention]]

“What assumption about churn is contradicted by the onboarding data in the Q3 investor update?”

→ investor-updates/q3.md

→ interviews/user-12.md

→ journal/aug-04.md

result

The Q3 investor update frames churn as an activation problem. But user 12’s interview says the opposite—they activated fine, then left because the product didn’t match what onboarding promised.

Enzyme connected an investor update and a user interview that never reference each other. After the agent writes a note back, refresh makes it part of the next session.

How did it find that? →

Memory without a memory database.

Connections are pre-computed at init from your artifact trail. Query-time retrieval runs against the compiled local index—not a live LLM call. Catalyst generation may use your configured/default provider during init or refresh; retrieval stays source-grounded in your files.

<20s 1k+ doc init

8ms query benchmark

0 query LLM calls

~30MB single binary

The loop is simple: scan the structure, initialize the index, inspect petri, catalyze for context, refresh after new artifacts land.

Claude/Codex supported · Pi via generic AGENTS.md · Hermes/OpenClaw experimental

How is this so fast? →

Early adopters

Thinkers using Enzyme to surface what matters in their notes.

"Enzyme can surface connections I'd only make if I had instant recall... freeing me to use that energy to create and write... I think this is just the beginning."

life coach (NYC)

"Enzyme helps me build richer, more cohesive themes and think more deeply."

HCI researcher, ex-Spotify (Zurich)

"I have a workflow that I love with Readwise, Snipd, and Obsidian. Enzyme beats inertia by quickly drawing me deeper into the meaning of what I'm doing."

songwriter and pastor (NYC)

Build products on what your users have captured.

Your users bring collections. Enzyme gives your agent their conceptual landscape from the first session.

Reading & highlights app Kindle highlights, podcast clips, article saves

Generate writing prompts from two years of reading — grounded in connections the user was already building.

Design tool Exploration sessions, chosen vs. rejected options, revision history

A conversational agent that knows the user's taste — helps them build with their own preferences, not generic defaults.

Engineering assistant ADRs, retros, decision logs, Slack digests

Onboard new engineers with answers that carry the full institutional reasoning, not just the current state.

Talk about your corpus →

Try it on your own vault.

$ curl -fsSL enzyme.garden/install.sh | bash

Local-first indexing with hosted memory workflows when you need them.

Building a product on imported content?

Let's talk about your corpus →

Not ready for the CLI? More integrations coming.

Memory retrieval tuned to your knowledge base.

Your agent already leaves a memory trail. Enzyme makes it usable next session.

Memory without a memory database.

Early adopters

Build products on what your users have captured.

Your agent already leaves a memory trail.
Enzyme makes it usable next session.