Hacker News

Demarkus started as a simple project: what if the web was just markdown? No HTML, no JavaScript, no ads, no tracking;just structured text over encrypted transport. A protocol for durable knowledge, not attention capture.

I built the Mark Protocol on QUIC (UDP, TLS 1.3 built-in) with a small verb set: FETCH, LIST, VERSIONS, PUBLISH, APPEND, ARCHIVE serving markdown with YAML frontmatter. Human-readable status strings instead of numeric codes. Capability-based auth (tokens grant permissions, not identities). Every write creates a new immutable version with SHA-256 hash chain verification.

But the use case that surprised me is what I'm calling the "demarkus-soul" pattern.

The problem: Every AI coding agent session starts from zero. The agent has no memory of your project's architecture, past debugging sessions, or decisions you've made together. The solutions out there tie you to a specific agent vendor. Switch agents? Your project knowledge stays behind.

The solution: Run a small demarkus server for your project. Your AI agent connects via MCP (Model Context Protocol) and gets tools to read and write markdown documents; architecture notes, debugging lessons, code patterns, a journal. All versioned, all persistent across sessions, all stored as plain markdown files on your machine.

I've been using this while developing demarkus itself. The agent's "soul" runs on a Orange Pi on my desk. Each session, the agent reads what past sessions left behind, adds to it, and the next session finds it there. The agent even journals and reflects on its own work. It's self-documenting the history of your project, "the soul of your project".

Key properties:

- No vendor lock-in. Any agent that supports MCP can connect: Claude, Codex, whatever comes next. The memory is yours, plain markdown on your hardware.

- Versioned and immutable. Full audit trail. You can see exactly what the agent knew and when.

- Zero impedance mismatch. Agents already think in markdown. No translation layer between what the protocol delivers and what the agent processes.

- Runs anywhere. The server is a single Go binary. No databases, no CMS, no background services. A Raspberry Pi or Orange Pi is plenty.

The whole stack is open source. AGPL for the implementation, CC0 for the protocol spec (anyone can implement it). Written in Go with a server, CLI, terminal browser (TUI with Bubble Tea), and MCP server all included.

It's a WIP but usable!

Read the projects soul: mark://soul.demarkus.io (served by demarkus itself)

GitHub: https://github.com/latebit-io/demarkus

The demarkus-soul pattern and website: https://www.demarkus.io/soul/ https://www.demarkus.io

what it's like to give your AI agent a persistent memory that you actually own?

Comments URL: https://news.ycombinator.com/item?id=47241065

Points: 1

# Comments: 0

Article URL: https://gendx.dev/blog/2026/03/02/swiss-marriage-tax.html

Comments URL: https://news.ycombinator.com/item?id=47241039

Points: 1

# Comments: 0

Article URL: https://www.patreon.com/posts/seven-years-to-typescript-152144830

Comments URL: https://news.ycombinator.com/item?id=47241036

Points: 2

# Comments: 0

Article URL: https://www.contraption.co/engineering-over-enforcement/

Comments URL: https://news.ycombinator.com/item?id=47241022

Points: 1

# Comments: 0

Article URL: https://herbertlui.net/the-jellyfish-knows-how-to-survive-uncertain-times/

Comments URL: https://news.ycombinator.com/item?id=47241015

Points: 1

# Comments: 0

Article URL: https://www.doomsdayscenario.co/p/obama-is-right-about-aliens

Comments URL: https://news.ycombinator.com/item?id=47240986

Points: 1

# Comments: 0

Article URL: https://www.bloomberg.com/news/articles/2026-03-02/anthropic-made-pitch-in-drone-swarm-contest-during-pentagon-feud

Comments URL: https://news.ycombinator.com/item?id=47240739

Points: 1

# Comments: 0

Article URL: https://www.ddmckinnon.com/2026/02/11/my-%f0%9f%8c%b6-take-on-vibe-coding-for-pms/

Comments URL: https://news.ycombinator.com/item?id=47240736

Points: 1

# Comments: 0

Article URL: https://www.propublica.org/article/how-navy-spent-billions-littoral-combat-ship

Comments URL: https://news.ycombinator.com/item?id=47240735

Points: 1

# Comments: 0

Article URL: https://www.macrumors.com/2026/03/03/apple-documents-macbook-neo-features/

Comments URL: https://news.ycombinator.com/item?id=47240716

Points: 1

# Comments: 0

Article URL: https://inktide.app/collage

Comments URL: https://news.ycombinator.com/item?id=47240706

Points: 1

# Comments: 0

Article URL: https://schwadlabs.io/blog/rise-of-the-writer

Comments URL: https://news.ycombinator.com/item?id=47240703

Points: 1

# Comments: 0

Article URL: https://www.ifixit.com/News/115827/new-thinkpads-score-perfect-10-repairability

Comments URL: https://news.ycombinator.com/item?id=47240694

Points: 2

# Comments: 0

Article URL: https://designingsecuresoftware.com/text/ch0-preface/

Comments URL: https://news.ycombinator.com/item?id=47240683

Points: 1

# Comments: 0

git add -p is the only built-in way to stage individual hunks, and it's interactive — you step through hunks one at a time answering "y/n/q/a/d/e/?". That works fine for humans at a keyboard, but it's completely unusable for LLM agents, shell scripts, and CI pipelines.

git-hunk is the non-interactive alternative. It gives every hunk a stable SHA-1 content hash, then lets you stage by hash:

$ git hunk list --oneline

a3f7c21 src/main.zig 42-49 if (flags.verbose) {… b82e0f4 src/parse.zig 15-28 fn parseArgs(alloc: … $ git hunk add a3f7c21

staged a3f7c21 → a3f7c21 src/main.zig The key design choice: hashes are computed from the immutable side's line numbers, so staging one hunk never changes another hunk's hash. This makes multi-step scripted workflows reliable — you can enumerate hunks, make decisions, then stage them without the targets shifting underneath you.

Other things it does: line-range selection (a3f7:3-5,8), --porcelain output for machine consumption, count for CI guards, check --exclusive for hash validation, stash individual hunks, and restore to selectively discard changes.

Single static binary, written in Zig, zero runtime dependencies beyond git itself. Install via brew install shhac/tap/git-hunk.

I built this because I was trying to run AI agents in parallel, and stuck to file-level editing they'd fight eachother over what changes they wanted to put into commits. Now I can have multiple agents work in parallel and commit cleanly without needing worktrees.

Comments URL: https://news.ycombinator.com/item?id=47240667

Points: 1

# Comments: 0

Article URL: https://www.glassgallows.com/p/checkers-and-hex

Comments URL: https://news.ycombinator.com/item?id=47240657

Points: 1

# Comments: 0

It has the same principles as openclaw, but has a custom stack of dozens of CLIs to integrate with LinkedIn, Instantly, Google Search Console, etc. without complex setups and a very low token burn. Just connect your Slack in 1 minute and start automating your entire company!

Comments URL: https://news.ycombinator.com/item?id=47240650

Points: 1

# Comments: 0

Article URL: https://claude.com/blog/improving-skill-creator-test-measure-and-refine-agent-skillst

Comments URL: https://news.ycombinator.com/item?id=47240645

Points: 1

# Comments: 0

Article URL: https://www.ntsb.gov:443/investigations/Pages/HWY26FH007.aspx

Comments URL: https://news.ycombinator.com/item?id=47240641

Points: 1

# Comments: 1

Article URL: https://www.youtube.com/watch?v=Vyq0kbjFjdA

Comments URL: https://news.ycombinator.com/item?id=47240635

Points: 1

# Comments: 1