How to Build a Two-Layer AI Memory System for You and Your Whole Company

By Treynor Tetik, Marketing Events Associate at SafelyYou · June 5, 2026

What happens to everything your team knows when someone quits? In most companies, it walks out the door. Treynor Tetik built a system to stop that. In this GTM Clawllective session, he walked through Nix, an internal AI memory system that gives every employee a private, searchable memory and rolls verified knowledge up into a shared company brain. He covered the full pipeline, a privacy gate built for healthcare data, how the AI "hops" across linked notes to connect ideas, and the build-versus-buy realities of running a system like this. Here is everything he shared.

What is Nix?

Nix is an internal AI memory system that Treynor built and has been running personally for several months. Think of it as an in-house wiki that your AI maintains on its own. It collects information from your tools, turns it into individual wiki pages, links those pages together, and lets your AI search and synthesize across all of it.

The design borrows heavily from Obsidian, the note-taking app known for its visual graph of linked notes. Treynor was also inspired by a piece from Andrej Karpathy on building an LLM wiki system, essentially a database your AI can hop around to build up and reuse memories over time. That was the seed of the whole project.

The core problem Nix solves is information sprawl. Your knowledge is scattered across Zoom, Slack, email, and meeting transcripts, and it leaves with people when they go. Nix captures that knowledge in one place and keeps it usable.

How does the two-layer architecture work?

Nix runs on two separate layers that do not touch each other.

Treynor estimates people will spend roughly 80% of their time in the personal layer, which is exactly why he did not bake collaboration features into it. Personal memory notes are not things you usually want to share or co-edit. Knowledge only crosses from personal to organizational when it is reviewed and promoted (more on that below).

What happens to data before it enters your wiki?

Capture is baked into your CLAUDE.md file, which acts as your AI's system prompt. As you work across Gmail, Slack, and other connectors, the instruction to collect information and create wiki pages is always running in the background. From there, captured information moves through a pipeline before it ever lands in your wiki.

The PHI gate matters because SafelyYou is a healthcare company that occasionally handles patient data. Treynor was clear that the gate is fully programmatic, with no AI involved. It scans the text for patterns that look like protected information and quarantines anything risky. He built this custom layer specifically to safeguard the company against audits and prevent non-compliant data from leaking into the system.

What is "wiki hopping" and why does it beat plain RAG?

When you ask your AI a question, it reaches into Nix, runs a RAG search, and lands on a relevant wiki page. At the bottom of each page is a set of links to related pages, each with a short description. The AI reads those descriptions and decides where to "hop" next.

It usually does one hop, occasionally two, expanding the radius around your original search before pulling everything together and synthesizing a response. Tags, shown as green dots in the graph view, give the AI another way to steer its search, surfacing every page that shares a tag like "competitors."

Treynor contrasted this with standard RAG, which he described as great at specific one-source retrieval: you search, it pulls back one piece of information, and that is mostly it. Nix connects the surrounding context. He explained that wiki hopping lets the AI link thoughts and ideas it normally would not, giving it a more cohesive view of what you are actually getting at. In his words, it is "much smarter at conceptualizing things than it would be just retrieving a single bit of information."

How does personal knowledge get promoted to the organization?

This is the part that lets the org grow smarter without forcing anyone to manually publish their notes. Your AI watches your personal layer and, when it spots something valuable to the company as a whole, recommends pushing it up.

The promotion path has a human-in-the-loop safeguard:

1. The AI flags a candidate and adds it to a review queue.
2. Once a week, an AI review pass scores each candidate against a confidence threshold.
3. If confidence is above 75%, it gets pushed to the organizational wiki automatically.
4. If it is below that threshold, it routes to a human admin (Treynor or another reviewer) who decides whether to push it or whether something looks off.

The goal is a wiki that updates itself. As Treynor put it, the idea is that "you don't really have to think about this," because capture and promotion are wired into the CLAUDE.md prompt and happen on their own. It directly addresses the classic failure of a traditional wiki that nobody ever bothers to update. Self-updating is the whole point.

How does the system stay up to date?

Nix never deletes information. When a piece of knowledge changes, the system appends the new version, dates and tags it, and marks the old version as retired. That way the AI works from current information but stays aware of what was replaced, so an outdated copy buried elsewhere does not cause confusion.

On top of that, two background processes keep the graph healthy:

• The dream cycle. Every 24 hours, the system looks for dropped connections, links the AI failed to make between related pages, and fills them back in.
• Scheduled sub-agent crawls. Treynor periodically spins up a batch of sub-agents to crawl the entire wiki, find anything outdated or broken, fix dropped links, and correct errors.

What data sources can it ingest?

Treynor seeded Nix by pointing sub-agents at his existing files and connectors to auto-generate a starting base of wiki pages, then layered in ongoing automated ingestion.

One concrete payoff: when Treynor has an upcoming meeting, Nix pulls the profiles of everyone attending, the relevant product and project context, and adjacent topics, then assembles a pre-meeting brief he called "way, way better than anything I'd had prior to setting up this wiki system."

How is this different from Notion or Confluence?

Treynor chose Obsidian as his personal foundation largely because of the graph view and the fact that it works with plain markdown files on your machine. A practical advantage of that approach: markdown files are easy to view and edit in collaboration with Claude, whereas Notion locks content into its own document format.

The honest limitation is that Obsidian has no live collaboration. Treynor pointed out that Nix is a memory layer, not a collaboration tool, and most usage lives in private personal notes anyway. The bigger reason he built his own system instead of using Obsidian or Notion at scale: Obsidian has no enterprise account, Notion's enterprise tier is expensive, and neither could meet the HIPAA and PHI standards his company requires. So he cloned the Obsidian experience himself for the organizational layer. For authentication, he plans to use Google SSO.

What should you consider before building your own?

A large part of the session was a candid discussion of the realities, not just the upside. If you are weighing whether to build a system like this, Treynor's hard-won advice was consistent:

• Access control is a maintenance nightmare. Setting per-employee permissions across many shared folders is exactly the kind of ongoing upkeep he deliberately avoided baking in early.
• Build versus buy is a real decision. Low-maintenance internal tools are no-brainers to build. High-maintenance systems with dependencies need constant monitoring, and you can end up as a one-person team keeping it all alive.
• Token costs are coming. Subsidized AI usage will not last forever, and using a top-tier model for routine work like writing emails gets expensive fast. Treynor expects more companies to set per-employee token caps (he pointed to reports of caps at large enterprises) and to invest in internal compute as the subsidy era ends. Training people to use the right model for the right task matters.
• Plan for the builder leaving. Build in safeguards so that if the person who built the system departs, someone can pick it up and keep it running. (Notably, preventing knowledge loss when people leave is the same problem Nix itself is designed to solve.)

Treynor also shared where Nix is headed: an organizational chart layer with profiles of who reports to whom and who to ask about what, an "agent in a box" onboarding kit, and a department-based plugin system (a marketing plugin, a sales plugin) bundled with skills. He offered to share a starter prompt so anyone can build their own wiki system in Obsidian and seed it with sub-agents.

Key Takeaways for GTM Teams

• Separate personal and organizational memory. Keep individual notes private by default and promote only verified knowledge up to the shared layer. This is what makes people comfortable using it.
• Wiki hopping beats single-shot retrieval. Linking related pages lets your AI connect ideas and build cohesive context, not just return one snippet.
• Make it self-updating. Bake capture into your CLAUDE.md, append instead of delete, and run scheduled passes (a daily "dream cycle" and periodic sub-agent crawls) to repair dropped links and stale content.
• Add a human-in-the-loop gate for shared knowledge. A confidence threshold (Treynor used 75%) plus weekly admin review keeps the org wiki trustworthy.
• If you handle sensitive data, gate it programmatically. A deterministic PHI/PII filter that quarantines risky content protects you in a way an LLM filter cannot guarantee.
• Respect the maintenance cost. Access control, dependencies, token spend, and succession planning are the real reasons to think carefully about build versus buy.

Want to Attend the Next Session?

The GTM Clawllective is a community of go-to-market practitioners sharing how they actually build with AI, live and hands-on. Sessions are not recorded for public release, so showing up is the best way to see these builds in detail and ask questions directly. Next up, Michael Sarguia is demoing a competitive monitoring system he built with Claude. Join the community to get on the schedule.

About the Speaker

Treynor Tetik is a Marketing Events Associate at SafelyYou, a company spun out of UC Berkeley's Artificial Intelligence Research Lab that builds AI, hardware, and remote clinical expertise to make senior living safer and more person-centered. He holds a hybrid role, leading AI enablement work while also running the company's event marketing, and earned the latitude to build internal AI systems after early wins opened up broader access to company systems.

His background spans event and marketing roles at Epic Rides, Hensley Beverage Company, and the University of Arizona. Nix grew out of his own daily use of an AI memory system, which he is now extending across his organization.

Connect with Treynor on LinkedIn.