Trimtab Labs

For reference only. Your brain will be different. This is just a sample brain.

go deeper

A brain built for someone else won't fit your brain; you've got to build it from the ground up so that it feels like home.

Plain files
Everything is markdown in one folder the owner controls

Capture layer

How new material flows in: phone voice memos and screenshots, podcast snippets, bookmarks, past chat history, email notes, older note-app archives, brain dumps typed or dictated

Ingestion archive
Raw sources kept after processing, so they can be processed again when models improve

Templates A, B, C

Every note is tagged by who wrote it

Each note is a small header of fields plus a few fixed sections. The same shape every time means an agent always knows where to look.

For efficient recall
  • The owner's exact words are kept, never paraphrased, because their own words are the keywords they will search with later.
  • Tags, links under Related, and the concept index connect each note to the rest of the brain.
  • The source on B and the model on C mean every claim can be traced to where it came from.

Privacy depth

Every file carries a depth from 1 (public) to 5 (sealed); outside services see only what the depth allows

A design philosophy that enforces privacy and content provenance.

Folder structure

Folders by kind of content, grown as content arrives; the brain is the root of everything, with projects living inside it

  • journals · ideas · writing
  • content: drafts, scripts, published posts, clips
  • client files · business operations: SOPs, contracts, finances
  • projects · research and sources · reference · lists
  • meetings and calls · people
  • AI-written synthesis · orientation · ingestion archive · scripts

Index

A master index of the brain, plus a concept index: one page per recurring theme, each listing the notes it was compiled from

Semantic search
Every note embedded; search by meaning, not just keywords

Router

A small always-on core tells each fresh agent session what to load, tier by tier, so it gets exactly the context a task needs, no more and no less

  1. 0Always on
    Loaded when
    every session, every request
    What it holds
    the rules, the router itself, a short profile of who the owner is, the memory index
  2. 1Navigation and status
    Loaded when
    "what's next", project status, "where does this live"
    What it holds
    the master index, current state, project board, to-do lists, the latest session handoff
  3. 2Deep context
    Loaded when
    strategy, life decisions, analysis or judgment on the owner's own material
    What it holds
    the owner's full profile and worldview, thinking patterns, keyword guide, content map, records from past chapters
  4. 3Voice
    Loaded when
    writing meant to sound like the owner
    What it holds
    the voice guide, on top of tier 2 for longer pieces
  5. 4The whole brain
    Loaded when
    a full deep dive across everything
    What it holds
    a snapshot of the entire brain, loaded only on request

The router reads each request and loads tiers in order, only as deep as the request needs. When unsure between two tiers, it takes the larger one: a few wasted tokens cost less than an answer that misses who the owner is.

  • A skill can declare the context it needs; invoking it loads exactly that.
  • A project with its own instructions has them read before any work inside it.
Orientation docs
The documents the router points to: master index, current state, project board, to-do lists, the owner's profile, voice guide

Skills

Named operations an agent can run, packaged once and reused

  • capture a brain dump · file a voice memo · clip and file a bookmark
  • draft in your own voice · repurpose a post into a thread · plan a content calendar
  • prep for a meeting · summarize a call · onboard a new client · write a proposal
  • research a topic · find what you already wrote about it · check for duplicates
  • weekly review · project status · wind down a session
Scripts
Tools the agents and skills call
Sub-agents
Specialist agents, for example an adversarial verifier of produced work

Memory

Rules and lessons the agents carry between sessions, one fact per file, in its own version-controlled repository; a short always-on index, the rest surfaced only when relevant

Hooks
Small programs that run before or after an agent's action and enforce rules no matter what the model does: guards, template validators, memory surfacing, checkpoints

Git

Every change is a commit, so nothing is ever lost and every change can be traced

Git, your settings
  • Keep it on your own machine only, or push it to a private remote.
  • Commit after every change, or in batches at the end of a session.
  • Work on branches to try a big change, then merge it or throw it away.
  • Choose what git ignores: large media, secrets, things you can regenerate.

Backups

Copies of the brain kept somewhere else, on whatever schedule suits you

Backups, your settings
  • Schedule backups however you see fit: daily, weekly, monthly.
  • To an external drive, a private cloud, or both.
  • Encrypted before they leave your machine, or not.
  • Or have the whole thing cloud based from day one.

Scheduled jobs

Jobs that run on their own on a schedule (the "cron jobs"): syncs, exports, maintenance, watchers, reminders

Scheduled jobs, your settings
  • Pick what runs on its own and when: every few minutes, nightly, weekly.
  • Human in the loop or not: a job can act on its own and report, or prepare the work and wait for your approval.
  • Alerts only when something changes that you would act on.
  • Pause or switch off any job at any time.
Weekly maintenance
Every week: counts and dates refreshed, three review agents run in parallel (document coherence, new file quality, project pulse), mechanical fixes applied, judgment calls brought to the owner

The seed, if well tended, becomes your second brain. Your second brain, if well tended, connects your capabilities to the growing power of artificial intelligence.

seed repo (GitHub, opens in a new tab) book a session (opens in a new tab)