The Second-Brain Scorecard
3 architectures, 1 corpus, 7 hard questions — which way to actually clone a knowledge base.
Updated: 2026-06-19 · Distilled from the full lab report — no signup required to read the article.
What's inside the one-page scorecard
- The head-to-head scoreboard: who won how many of the 7 synthesis questions
- Why a plain folder of markdown + Claude Code beat a production RAG pipeline 5–2
- Three failure modes: RAG confabulation, long-context budget exhaustion, brittle-but-honest grep
- The working-memory trap that breaks most second-brain systems in 2026
The bottom line: Basic read/grep tools mechanically enforce faithfulness — stick to the corpus, flag limits — while a RAG pipeline's generative step optimizes for fluency at the cost of truth. For a knowledge system, a tool that can say "I don't know" beats one that sounds confident and is wrong.
Frequently asked questions
Is RAG or long-context better for a second brain?
Neither won our test. A file-based setup (markdown + an agent with read/grep) beat both, 5 of 7 questions. Production RAG won 2 (and confabulated a fact that never happened); Gemini long-context won 1 — the single hardest question — but exhausted its token budget on the rest.
Why does RAG hallucinate on a knowledge base?
RAG's generative step optimizes for fluency. When semantically adjacent but factually disconnected chunks are retrieved, the model bridges the gap using its pretraining — in our test it invented an ElevenLabs shutdown that exists nowhere in the corpus.
How much does it cost to test a second-brain setup?
Our entire head-to-head — three architectures across seven synthesis questions plus a multi-turn working-memory probe — cost $4.30 in API calls.