Your Second Brain Is Lying to You — Unless You Build Trust Into the Graph

A practitioner walkthrough of schema-first extraction, confidence scoring, and refusal-to-invent pipelines with myKG, Claude Code, and Obsidian

The Agent That Cited a Partnership That Did Not Exist

Last month I watched a personal-knowledge agent answer a partner question with total confidence. It named a vendor relationship, attached a tidy edge label, and spoke as if the claim sat in the vault the way a signed contract sits in a filing cabinet.

I opened the source notes. The vault mentioned an informal advisory contact and explicitly said there was no active commercial contract. The graph had flattened hedge language into a hard fact. The agent had not “reasoned poorly.” It had trusted a graph that never earned trust.

That failure is the real second-brain problem in 2026. Building a graph from Markdown is easy. Pointing a language model at a folder and collecting entities is a weekend project. The hard part is deciding which edges deserve to drive decisions, which edges belong in exploration mode only, and which “facts” should never have been written down as facts at all.

This article is a workshop-style rebuild of that problem and one practical answer: myKG, an open-source extractor…