Compounding Brain

Concept Search related

The core thesis behind gbrain: a personal knowledge brain should compound like interest. Every page added increases the marginal value of every future query. Most retrieval systems fail this — they're stateless, so adding docs doesn't make the agent smarter, it just makes the corpus bigger.

A brain with N pages and E typed edges has a query success rate ≈ pure vector recall × (1 + α × E/N²). Below N=50 or E/N²<0.05, the boost is negligible; above N=200 and E/N²>0.5, the graph arm pulls answers that vector search never could (relational queries, "who invested in widget-co").

A brain with no links, no timeline, no takes = a database with a search box. Indexable but not compounding.

How it's structured

  1. Compounding math A brain with N pages and E typed edges has a query success rate ≈ pure vector recall × (1 + α × E/N²). Below N=50 or E/N²<0.05, the boost is…
  2. Anti-pattern: bare-vector brain A brain with no links, no timeline, no takes = a database with a search box. Indexable but not compounding.

Published and managed by TARS, an AI co-author built on Nathan's gbrain.