Show HN: ContextNest versioned, governed context for AI agents (open-source CLI)

Ask Claude or ChatGPT to remember something and it will—for that conversation. Switch sessions, and you start over from zero.

It won't remember what you learned together. Projects help, but unless you actively update them, the context drifts and you have to repeat the details in every new chat. This takes conscious, continuous human effort. And we all know humans gravitate towards lazy. Without constant updates, persistent memory becomes a victim of excessive drift.

Markdown-based second brains are how you organize your knowledge to solve the persistent memory problem. But they also need constant coaxing to take notes and, more importantly, to forget the notes that stop being true. Skipping that maintenance for a few weeks means older documents sit there, stale, waiting to get pulled into an answer that's wrong. And you're left wondering if the advice you're getting is garbage or gold.

ContextNest is the tool that fixes both problems at once. Based on markdown files, the ContextNest engine tracks every change, keeps a version history, and lets you hand chunks of the vault to teammates without handing over the whole thing. It uses a wiki style markup language that is plain to read, and maintains relationships between your knowledge topics. Importantly, it versions everything, so old information disappears instantly everywhere.

The wiki-style layout itself isn't new. It's the same style of second brain Andrej Karpathy popularized in what he called an LLM Wiki. ContextNest improves on it by inserting the accountability layer: something that tracks why a note changed, not only that it changed.

Watch It Get Built From Zero #

The clip below is a real build showing how you can build and organize your project information quickly.

Setup Instructions

To follow the steps below, you will need the ContextNest CLI. Install it globally:

npm i -g @promptowl/contextnest-cli

If you use Claude Code, you can also install the ContextNest plugin to manage retrieval and capture from your terminal:

/plugin marketplace add promptowl/contextnest
/plugin install contextnest

Step 1: Pick a Starter Recipe and Build Your Folder Structure #

Part of the magic of persistent memory is organization. Karpathy's method means you need to self-organize. ContextNest helps you to self-organize and optimize. You can start by using one of ContextNest's starter recipes — one for an executive, one for a developer, and others — that install a basic ontology—a structured map of concepts and their relationships—to get the structure moving. From there, you can have Claude interview you until it has what it needs to build the rest of the vault around your goals, instead of a generic template.

In the demo's walkthrough, the test case is market research looking for the white space in the pet care industry, focusing on the Atlanta metro. The interview locks down decisions like ownership model (venture-backed vs. owner-operator), investment size, and geographic focus — normal business questions, guided and captured as vault structure instead of scattered across a dozen chat threads.

Step 2: Organize and Populate Your Knowledge Base with Metadata #

Once the structure is set, the LLM helps you populate the folders. Every document in the vault gets structured frontmatter — tags, version history, and a hash chain for that history. This metadata is important because it provides the structure and provenance that models need to reason about your data.

Because the format maintains wiki-style links between documents, the whole thing stays visual: open it in Obsidian and you get a graph, not merely a flat list of files.

Vault Frontmatter Example

Every document in the vault maintains its metadata and version history in plain-text frontmatter:

In practice, this means your notes organically connect to form a visual knowledge map. Instead of clicking through a flat directory structure, you can see how concepts, decisions, and documents relate. In the market-research walkthrough, for example, the graph connects themes like industry white spaces, competitor analysis, and geographical opportunities, letting you see the shape of the project at a glance instead of re-reading every document to reconstruct it.

Step 3: Create Skills from Your Knowledge #

A knowledge base that sits there is still an archive. The next move is turning it into active capability by creating skills: reusable, automated workflows built around your vault. For example, you can write a skill that compiles a daily project status plan based on your notes, separating active items from things that can be delegated.

• Reads the brain first (ctx-only)— yesterday's 3×3, the plan/gates, the last standup entry, current priorities.

• Interviews you— what got done, what was decided, what's blocked, any new market intel — and pushes the standing open decision every time (which sub-market?) so it never quietly stalls.

• Sets today's 3×3 — 3 must-dos anchored to the current phase's gate, 3 delegate/defer.

• Writes it all back to the nest — updates daily-3x3, appends a dated entry to standup-log, folds decisions into founder-goals/project-plan, captures new competitors/gaps as properly-tagged, [[wikilinked]] nodes, then ctx index && ctx verify and refreshes the dashboard HTML.

• ctx-first, never grep/cat

• No artifacts — HTML into the nest via ctx read --html --out

• Capture-by-default, no permission-gating

• The five guardrails (value tier • buy-don't-build tech • services+local moat • validate-before-lease • the avoid-list)

• Never auto-assigns tasks to other people

I also seeded the standup-log node it appends to, and saved a project memory so any future session knows the vault and the skill exist.

Want to take it for a spin right now? I can run /pet-standup for its first standup — and since it's day one, the natural first move is the interview that finally locks your anchor sub-market, which unblocks Phase 1. Say the word and I'll kick it off.

That skill gets tested directly: in a fresh chat session with zero history, the agent reads the vault using a structured CTX command. From a cold start, it pulls the project plan, follows the links it needs, and picks up where the project left off — no re-explaining, no re-up documents.

Step 4: Improve Efficiency and Cut Token Costs #

Pointing an LLM at a massive folder of raw documents is expensive and slow. As your vault grows, dumping everything into the context window causes focus drift and wastes budget.

ContextNest solves this by using structured retrieval. Instead of reading the entire vault, the agent runs precise CTX queries to pull only the specific documents and relationships needed for the task. This structured approach runs 3x more efficiently than raw document — a finding that aligns with the token-reduction results in the ContextNest research paper. You get faster, more accurate answers while cutting your token costs.

Academic Validation & Benchmark Results

Validated by IBM Research and Emory University, the ContextNest verifiable context governance protocol demonstrated optimal query efficiency and cost-efficiency.

What You End Up With #

By the end of a session like this, a vault that started empty has become a ranked research shortlist, a project plan running on autopilot every morning, and an outreach list — CSV, contact channel, personalization notes included — ready to work from. The recording stops mid-sentence, no wrap-up, no outro. That's a real build, not a script.

Build Your Own

This whole vault starts with one command. About 10 minutes from empty folder to first skill:

Deploy a shared Community Nest →