--- title: "My second brain is an ADHD accessibility system an agent can navigate" url: "https://toddpaulbrownjr.com/writing/my-second-brain-is-an-accessibility-system/" author: "Todd Paul Brown Jr." description: "The Brain Vault: 1,765 documents in two layers, 50 pillars, 17 playbooks. An ADHD accessibility system designed so an agent can navigate it and act on real…" kind: "article" updated: "2026-09-26T02:50:36+00:00" --- # My second brain is an ADHD accessibility system an agent can navigate I built the Brain Vault for one user. Me. It's a local Obsidian vault holding 1,765 documents: 869 synthesized notes, timelines and concept pillars, and 822 raw imports. It carries 7,075 links in the synthesized layer, median five per document. Fifty concept pillars span eleven domains. Seventeen agent playbooks define named routines — imports, interlinking, health checks, weekly review — with a routing layer that tells Claude Code where to start and which sources are stale. A Python script ranks the synthesized layer with PageRank; a 3D visualization plugin I call Cosmos renders domains as stars, pillars as planets, evidence as moons. The plugin is built, not shipped: 31 tests, one feedback round still open. **Status: shipped**, for its one user. I've been using it every day since August 2024. This is largely designed as a support for my ADHD. It's an accessibility tool, and a highly effective one. A notes app is built for retrieval — where did I put that? This is built for connection and action: what does an agent need to know right now to act on real context? ## Why I built it Working memory doesn't hold the entry points I need. A cap of five active priorities keeps me from spinning out. Finished work evaporates from memory unless I write it down the same day it ships. An energy map lets me schedule against real focus capacity instead of pretending every hour is equal. The vault started around March 2026 as a set of instructions I wrote for the agents that would build it. From my working notes that month: the purpose is not to create a chat archive, but a map of my brain that I can continue to update intuitively. This neural map should be intuitive for AI agents, as they will be continually helping me organize and derive insight from this documentation. A roadmap, a README, a clear index — these would all benefit future agents. The seed material was an estimated 1,600 chats covering December 2023 through March 2026, about 1,300 of them from ChatGPT. The rule I set from the start: it is crucial to understand that the important part of these chats is what I said, and not the input from ChatGPT. That same rule applies to every vault item today. Writing the brief for agents first — a roadmap, a README, a clear index — is the same move I made with LEGION's agent-first README in 2025. The idea that a map beats an archive is mine; it's the reason the vault works. ## How it works Two layers, on purpose. A raw layer holds imported source material and is never edited. A synthesized layer — concept pillars, notes, timelines — sits on top and links back. Importance scoring counts only the synthesized layer. An import pipeline splits a batch into one file per item, removes near-duplicates at a 0.80 similarity threshold with every decision logged, then scores each item for signal strength. Triage happens before extraction, not after. Ninety-six percent of documents carry structured frontmatter. Undated material gets an inferred date and a stated confidence level, not a guess dressed as fact. The seventeen agent playbooks define how the vault maintains itself. A routing layer tells the agent where to start; a playbook tells it what to do once it's there. Operating documents that answer ADHD failure modes directly: a live priorities document (capped at five); operating assumptions held as confidence-scored hypotheses rather than invisible beliefs; the energy map for scheduling; and an accomplishment ledger built as a deliberate counterweight to undercounting my own output. The PageRank script is standard-library Python. It builds the link graph, ranks the synthesized layer, and writes the scores back to frontmatter. The same scores drive the Cosmos plugin, which renders a sparse version of the dense analytical graph. Domains become stars, concept pillars become planets, evidence notes become moons. The plugin uses Three.js for the scene and renders additive sprites against a black sky. I'm on Mountain time. My Claude Code usage report covering mid-July through early September 2026 shows 855 messages over 262 hours. Nothing between midnight and 6:00 a.m. Peaks around 10:00, 15:00 and 19:00. Sessions usually open from a substantial pre-written brief — a PRD, a roadmap, a handoff file, or an "execute this prompt" pointer — then run for hours, sometimes overnight. I often watch running sessions from my phone. Since late August 2026, I write my own entries and an agent files them. My method: I'm gonna try to add bullets first for coverage and then fill in the details on a second pass. A weekly check-in with written wins. Anything shipped gets a ledger line the same day. The stack: Obsidian; Syncthing for device-to-device sync with no cloud; Claude Code as the maintenance agent. ## What broke and how I fixed it The nebula that never rendered. Scene fog was washing the additive sprites to black. No test caught it; I found it by looking at the output. Five fix rounds in one day, and the record drifted. The feedback document's round numbers don't match the coding session's, because three rounds happened in chat and were never written down. I know better than that. An earlier headline number was wrong. An earlier summary said about 1,850 documents and 56 pillars. The measured counts when I finally inventoried the vault properly: 1,765 total, 50 pillars. I was working from an estimate I hadn't checked. ## What built around it In March 2025, about the same time I started the Brain Vault's import pipeline, I started building LEGION — AI APIs talking to each other on my computer, then a web app. The vision was simple to say: I wanted to be able to have a discussion with the VP about strategic initiatives while the research, problem solving, asset acquisition and fulfillment were handled by the AI organization without further input from me. OpenAI launched the Responses API and Agents SDK that same month (11 March), citing customers struggling with custom orchestration logic. Google Cloud blogged about "agentic orchestration" emerging as a new paradigm on 24 March. Google Cloud Next followed on 9 April with ADK, Agent Engine, and the line "Every enterprise will soon rely on multi-agent systems." I was building around the same architectural insight in the same weeks the industry was converging on it, before it had a common name. My 2025 attempts failed on the problems that later harness writing is about: state that doesn't survive a session, agents ignoring their instructions, agents editing things they shouldn't. The April attempt ran Docker with FastAPI and a LangChain task runner. Agents lost work in wiped containers, edited the UI into black-on-black, refused tasks, ignored the handbook until a default prompt forced it. The May attempt was LEGION v2.0 in Lovable; the GitHub connection failed. On 10 September 2026, roughly 18 months after I first described it, LEGION was actualized. It finally all came together into a system powerful enough to rebuild the entire website within a couple of hours. The Brain Vault is where LEGION reads context. The routing layer, the playbooks, the priorities document, the accomplishment ledger — those are the files an agent opens to know what matters right now and what to do about it. ## Habits that work with the brain, not against it I have ADHD. What that means for me is I'm constantly chasing novelty. I get bored fast. Repetitive work makes me miserable. When I started my own company, I figured out that ADHD wasn't the problem — the environment had been the problem. In a job where every day is different, where every month there are new tools to learn and new problems to solve, my brain isn't a liability. It's the reason I'm good at this. Novelty-seeking is a superpower in a career built on doing things that haven't been done before. A lot of what looks like a deficit in one setting is a strength in another. The trick is finding the setting — or building it. I will not commit every day to solving small problems, but rather commit every month to solving large problems. My job is to do things that haven't been done before. I don't want to feel like anyone expects anything from me other than my presence and contribution. The Brain Vault is the system I built to work with that brain. A map, not an archive. A routing layer because working memory won't hold the entry point. A hard cap of five active priorities. A completed-work ledger because finished work evaporates from memory. An energy map for scheduling against real focus capacity. Architecture and operating documents designed around ADHD failure modes so an agent can do the navigating while I do the work only I can do. It's shipped. I use it every day. It works.