Second Brain

Scoping an AI-searchable store of Google Drive, TrinityMetrics and the rest

GenAI
working document
The working specification for the ai-second-brain project: the definition, the sources to feed in, the candidate tools, the open questions on purpose, and the first steps.
Published

September 30, 2026

Reminder, 2026-09-30. Nothing is built. The purpose in Section 1 is a first guess, and Section 5 lists the questions that have to be answered before choosing a tool.

1. Goal

Set up a second brain: one place holding the documents, notes and code accumulated over the years, searchable by an AI assistant that can answer a question from them and say which document the answer came from.

The first guess at the purpose has two parts.

  • Find past work. A question such as “where did I work out the EC50 identifiability argument?” should return the document, not a search of four places by hand.
  • Keep learning. Feeding the store and querying it is itself practice with the tools, and the store makes connections between old and new projects visible.

2. Out of Scope

  • Not published. The store holds private material. Nothing from Google Drive is copied into this repository, and this page describes the plan only.
  • Not Collected Wisdom. That project indexes the personal Google Docs on meaning, parenting and music as a site. This one is a private search tool over everything, work material included. The Google Drive inventory in Section 3 can serve both.

3. Sources

# Source What it holds Notes
1 Google Drive Documents, slides and spreadsheets from years of work Have an agent scour it and propose what is useful before loading anything
2 TrinityMetrics Every project, blog post and guide on this site Already in git and in Markdown, so the easiest source to load
3 Other GitHub repositories synpmx, xgxr and the rest Code and vignettes
4 Papers The PDFs and reference manager library behind the literature reviews
5 Email and meeting notes Decisions and the reasons for them Most sensitive; decide last
6 Claude conversations Sessions that worked something out and were never written up Check what can be exported
7 Bookmarks and read-later lists Things meant to be read

Rows 1 and 2 come first. The rest wait until the purpose in Section 5 is settled, since each one widens what has to be kept private.

For Google Drive, the first pass is an inventory rather than a load: a list of folders and documents, a sentence on each, and a mark for keep, skip or private. An agent with read access to Drive can draft it; Andy decides the marks.

4. Candidate Tools

A second brain in this sense is two parts: a store of documents, and a retrieval step that finds the passages relevant to a question and hands them to a language model to answer from. The second part is called retrieval-augmented generation (RAG). The tools below differ in where the store lives and who runs the retrieval.

Tool Store Fits
NotebookLM Google, reads Drive directly Google Drive with no copying; limits on source count per notebook
Claude Projects Uploaded files in claude.ai A few dozen documents per project
A folder of Markdown files, queried by Claude Code Local disk, in git TrinityMetrics and other repositories as they already are
Obsidian Local Markdown with links between notes Writing new notes as well as searching old ones

None has been tried for this yet. The term “second brain” comes from Tiago Forte’s Building a Second Brain (2022), which is about organizing notes by hand; the AI retrieval step is the newer part.

5. Open Questions

  1. What question should the store answer that cannot be answered now? Two or three real examples decide the tool better than a feature list.
  2. Work and personal together, or apart? Work material may have rules on where it can be uploaded.
  3. Load a copy, or point at the original? A copy goes stale; a pointer depends on the tool reading Drive or git directly.
  4. Who adds to it, and when? A store that is loaded once and never updated stops being useful within a year.

6. First Steps

  1. Write down three questions the second brain should be able to answer.
  2. Have an agent inventory Google Drive, as in Section 3.
  3. Try NotebookLM on a handful of Drive documents and Claude Code on TrinityMetrics, and compare the answers to the three questions.
Back to top