Table of Contents

• How Mneme remembers
• The workflow
• Privacy by design
• Syncing the outside world
• Limitations
• Start your own Mneme

It started with a single Claude Project. A few instructions, a couple of artifacts — manageable. Then I added context about my keyboard build. Then my guitar setup. Then my diving gear. Then I started journaling in there, using Claude artifacts as a scratchpad for things I wanted to remember. The instructions file grew. The artifacts multiplied. I’d open the project and scroll for seconds just to remember where I left off.

Eventually I had nine Claude Projects for different slices of my life, cross-referencing each other through notes I copied back and forth manually. And none of it touched my Kindle highlights, my browser bookmarks, or the handwritten diving logbooks sitting on the shelf.

This is what personal knowledge management (PKM) actually looks like in practice. My highlights from a psychology book had no link to the concepts I was actively working through. My diving logs had no connection to my certifications or gear. My bookmarks were a write-only archive. Each tool was fine in isolation and completely blind to everything else.

Standalone Obsidian would have been the obvious choice, but it couldn’t pull in external sources automatically, had no AI for triage and enrichment, and most importantly, Obsidian Sync doesn’t offer the encryption guarantees I wanted. So I built Mneme on top of it — an end-to-end encrypted, AI-assisted vault with automated sync, named after the Greek goddess of memory.

How Mneme remembers

The vault is organized around a single principle: information should flow toward its natural home, not accumulate in inboxes.

Inbox is the frictionless entry point. Any thought, link, highlight, or log entry that doesn’t have an obvious home lands here as a timestamped markdown file. The rule is: capture first, decide later.

Areas are the long-term domains I actively maintain — Books, Finance, Guitar, Scuba Diving, Wellness, and a few others. Each area has a consistent structure: a summary note, relevant sub-notes, and a log for time-series data. An area exists because it reflects an ongoing, durable interest — not a one-off curiosity.

Some entries skip the inbox entirely. The journaling flows — Claude Code skills like /log-dive, /log-guitar-practice, /log-meditation, /log-therapy — write directly to the relevant area’s log. Structured enough that there’s nothing to triage.

Resources is the staging ground for everything that doesn’t yet have a clear home. A resource is a note that might eventually graduate to an area, or get discarded once it’s clear the interest was fleeting. The key insight here is that Resources is not an archive. It’s a waiting room with an eviction policy: when three or more resources on the same topic accumulate, that’s a signal that a new area is warranted. One lonely resource untouched for six months is a signal to delete it.

The workflow

The system runs on three modes: Capture, Process, and Review:

Capture is deliberately low-friction. When something is worth saving — a thought, a link, a log entry — it goes into the inbox immediately, without worrying at the moment of capture about where it belongs.

Processing is where the inbox gets triaged: each note either moves to an existing area, gets parked in resources if its home isn’t clear yet, or gets discarded if it wasn’t worth keeping. This is also where cross-links get added — connecting a new book note to a related psychology concept, or a gear purchase to the Guitar area’s signal chain.

The judgment calls in this stage aren’t mechanical. A note about sleep quality could belong in Wellness or be a fleeting observation worth discarding. A half-formed thought about a leadership model might fit Psychology or deserve its own resource. For this I use Claude skills — short, purpose-built prompts that review each inbox note in context and decide where it belongs. The goal isn’t to automate the triage so I don’t think about it. It’s to make the right call consistently without spending twenty minutes second-guessing every note. Claude brings semantic judgment; I bring the final say.

Triage isn’t the only thing Claude is good for here. Existing notes can be enriched on demand. I once asked Claude to flesh out the signal path of my guitar gear — the pedals, the amp, the routing — with technical specifications pulled from the web. It came back with detailed specs for each piece of gear, I reviewed and corrected where needed, and the note went from a personal list to an actual reference. That loop — Claude proposes, I verify — is what keeps the vault accurate without making enrichment a manual research project.

Review is the maintenance pass: surfacing orphaned notes, flagging stale content that hasn’t been touched in six months, identifying resources that have crossed the three-note threshold and deserve promotion to a full area. A health check, not a rewrite session.

The three-mode discipline is what keeps entropy from winning. Most PKM systems have good capture. Almost none have a real process and review loop. That’s where the rot starts.

Privacy by design

Mneme lives in a private Git repository, encrypted at rest with git-crypt. Every markdown file is opaque in the remote — only someone with the symmetric key can read the contents. The key lives in my password manager and a secured external drive. This isn’t paranoia; it’s just the right default for a vault that contains financial accounts, health logs, and therapy session notes.

The less obvious privacy consideration is commit messages. Git-crypt encrypts file contents, not metadata. A commit message like “Add dive: Red Sea with John” leaks personal details in plaintext, visible to anyone who gains read access to the repo — today or later. The convention I enforce is structural, never personal: “Add dive log entry” is fine, “Add dive: Red Sea with John” is not. It’s a small discipline that matters over years of commits.

Syncing the outside world

The template’s core sync step just commits and pushes. But the architecture is easy to extend — you can add skills that pull external data into the vault before the commit. Manually copying in highlights and reading logs defeats the point, so I added two integrations to my own setup. The Goodreads RSS feed syncs my reading log automatically — books move between shelves (To Read → Currently Reading → Read) and the vault reflects it without me touching anything. The Kindle integration goes further: it fetches my highlights directly from Amazon’s reading interface and embeds them into the relevant book notes. There’s no official API for this, so it uses session cookies. It runs idempotently though — re-running it replaces the highlights section without duplicating anything.

The effect is that when I finish a book, the vault already has my highlights waiting. I just add context, cross-links, and a rating. The raw data arrives automatically; the enrichment is mine to add.

This is the inversion that makes the system work. Instead of manually migrating highlights from Kindle or transcribing a dive from a paper logbook into a markdown table, the system meets the data where it lives and brings it home. The handwritten logbooks I still fill out on the boat — that’s a ritual I don’t want to give up — but the moment I’m back at my desk, they get a digital entry in the vault.

Limitations

Mneme doesn’t run on mobile devices. The entire workflow runs through Claude Code skills in the terminal — there’s no app, no mobile interface, no web UI. I interface with it from my Neovim setup and it can also be interfaced straight with Claude Code. If something is worth capturing while I’m away from my desk, I’m relying on a Claude chat on my phone until I get back.

The end to end encryption of Mneme is a single point of failure. The encrypted vault is worthless without it. I keep copies in my password manager and on a secured external drive, but if both are somehow lost, so is every note, log, and highlight I’ve stored. That’s not hypothetical risk — it’s the direct tradeoff of choosing local encryption over a cloud-managed key. I think it’s the right tradeoff, and that’s fine, but it demands taking the backups seriously.

Start your own Mneme

If you want to build something similar, Mneme is published as a GitHub template repository. Fork it, run the /setup skill, and Claude will ask which areas you want to track and generate the vault structure and skills for you. Feel free to strip out the parts that don’t fit and make it yours.

I’d love to hear how others are using Claude skills outside of work — this felt like an unusual use case when I started it, but it’s become the part of my setup I touch most.