# Turning a personal knowledge graph into a daily AI co-pilot

> How Mason Hall built a personal knowledge graph his agents can query so 'what's the highest-leverage thing I can do today?' actually gets a useful answer — useful for personal AI memory / context-engineering questions.
> by Mason Hall (Merit Systems)
> group: personal
> event: building-ai-native-startups-002
> source: cyber•OS — https://os.cyber.fund/case-studies/personal-knowledge-graph-daily-copilot

## Problem

The context an agent needs to be genuinely useful is scattered across dozens of tools, and most people start every hard question in a fresh Claude/Claude Code chat that knows nothing about them. Mason Hall's framing: "context is king" — there's no reason not to give your agent all the context you have if it's indexable and searchable, yet almost nobody does.

## Approach

Mason built a personal knowledge base as an Obsidian-style knowledge graph and exposed it to Claude Code (and Hermes / OpenClaw) via a set of open-source skills (`npx skills add fm-hall/knowledgebase`). He ingested data from wherever he'd spent time — Twitter and LinkedIn GDPR data exports, bookmarks (Pocket → Linkwarden, plus the Obsidian Web Clipper), iMessage, Notion, Google Drive, Discord and Telegram — into a raw folder. Background agents then synthesize the raw dump into linked wiki pages with backlinks, and scheduled tasks clean up and enrich short pages over time. Data is enriched via agent-payable APIs (Agent Cache over x402). He drives it daily with commands like `/wiki, what should I work on today?` and other verbs (ingest, absorb, query, cleanup, breakdown); when the agent needs his judgment it asks what it's missing and he rambles answers in via Whisper Flow.

## Results

After about a month and a half of tuning it became a reliable daily co-pilot — surfacing the highest-leverage task, refining Merit's ICP, ranking BD/YC targets, and ideating side projects grounded in his actual interests. Setup took roughly two five-hour Claude sessions on the $200 Max plan (~10% of a week's usage limits), and querying stays token-efficient because LLMs traverse the backlinked file structure well. A company version runs on OpenClaw over work-channel context, with a one-directional design to feed the company agent into personal agents while preserving privacy. The whole system is open-sourced.