jphorism

Cool. Do you find it useful for human review too, or is this mainly about the agents having the right abstractions to work with?

actually, I think agents have a lot to design after Zettelkasten yet. It's the concept of having an append-only policy and gardening to consolidate knowledge. IMO we've seen some, but there is still so much potential

Agents respond well to keeping instructions simple -- I've noticed that elaborate knowledge bases are hard for agents to consistently grok, so by simplifying the agent abstraction Enzyme improves its ability to navigate

Yeah! You're highlighting something interesting in that taste is observed over compounded judgment (ie design decisions). The reason this works well is by observing questions, Enzyme turns compounded curiosity into an agent lens

QMD is great! Concept is similar, relies on having an existing Markdown store for your agent's memory.

Enzyme is 100x smaller and precomputes questions that help agents to write better queries (which theyre getting great at these days), instead of relying on a finetuned model. So it is overall faster

Enzyme produces local artifacts, so technically it could easily run on each team member's machine (it's cheap enough), provided there's another way to sync the knowledge base. Markdown + team kb's is an interesting area and I am also curious to see what standards will take shape -- probably an adjacent problem

Hey all!

I built Enzyme to help agents and people compound knowledge, together. This week I decided to open-source several years of my work (it began as an Obsidian plugin, then became the first MCP for Obsidian).

Enzyme is like an LSP for ideas; it orients agents by emulating your working memory from a local knowledge base (sqlite or .md). I decided that a language was the best way to represent several design elements of agent memory at once: append-only log locations, temporal distribution config, profiles for "dreaming", and more. Your agent can setup Enzyme, compile the Enzyme program, and maintain a local-first, self-updating lens over its wiki. It's very fast.

I use it daily to maintain my 8000+ note Zettelkasten that also acts as a personal CRM (meeting transcriptions), Readwise collection, journal, and home for agent rollout summaries.

Core to its design are questions that are generated from recent material and projected over the corpus. Instead of "dreaming" .md files over fragments, it generates questions to connect their ideas using embedding similarity (static embeddings developed by Flower Computer). What these questions notice is configurable.

The binary, including the embedding weights baked in, sits under 30 mb.

I would LOVE teenage engineering to release something open source for us to hack with.

Nice, what are you hoping to accomplish with this project?

Really cool project, thank you for sharing!! I've added it to the pipeline for my local first meeting recorder called aside: https://github.com/jshph/aside/

Cool to see that it pays off to have split things up by season!

having worked on the AI Playlist feature at Spotify (which is agentic) -- one thing you can try to prompt is "give me songs that sound <vibe> from my <liked songs | playlist called "X">.

An extension of this is "give me a sequence of N tracks from each of <list of playlist names for each past year> that all evoke <vibe / memory / activity>". Could be a fun prompt to play with

That's cool! Have you thought about some docstring augmentation loop -- either by having the agent log the tool uses or by auto querying (e.g. perplexity API) for example use cases in the wild?

A friend of mine uses Apple Reminders to rank his most favorite sentences of all time. It struck me that authority emerges through which ones continue to stick around (as forwarded ideas).

I wonder if the premium on consistent writing quality is different than the premium we place on consistent novelty.

I'd hazard a guess that from the writer's perspective, novelty scales with volume of thought / connections, which is (at present) a fragmented process and not that well-assisted by AI. OTOH, can "writing quality" be better approximated by LLMs?

It's possible to do both. I write in a small field journal to think better, then periodically use Wispr Flow to quickly transcribe it to Obsidian (where I can use LLMs on the writing).

The plumbing metaphor is interesting. The most adjacent place I've heard this metaphor is in theology (Mako Fujimura)

The goal of plumbing is to fix / repair; certainly it's possible to enjoy fixing and repairing. But is the joy of writing in "repairing ideas"? How is that a separate concept from creating new ones?

Awesome stuff. Seems to me like one great use case of an MCP for Ableton would be to develop muscle / workflow memory for music production workflows. Adjacently, I've started using Perplexity a bunch for this sort of thing because it indexes YouTube tutorial transcripts. Any thoughts on how to design MCPs for learning Ableton better?