Pinky the agent runs a CRM of 1,000+ people: a Second Brain built on md files and git instead of Notion

A familiar pain: notes, tasks, contacts and goals are scattered across a dozen services, and time goes into keeping Notion tidy rather than into actual work. The author of the channel "Tochki nad II" (Dots over AI) solved it radically: he set up a folder of markdown files, connected an agent, and a year later says he no longer understands how he lived without it.

What the system is

The agent is called Pinky. It reads and writes to the base on its own, remembers where you left off and, according to the author, argues more often than he would like. Inside lives a CRM of 1,000+ people: calls are recorded in MacWhisper, and the agent sorts them into contact cards. Alongside is a task planner that the author and the agent vibe-coded, a camp analytics dashboard (payments, funnel, corporate deals), lab results, finances, goals for the year, an insight journal and a psychological profile that important decisions are run through.

The main effect comes from having everything in one memory. Business, health, people and thoughts can be cross-referenced, and it is at these intersections, the author says, that the most interesting insights are born.

The agent has roles: doctor, camp mentor, financial analyst, editor, English tutor. It switches them on by itself when the topic of conversation matches, so a single decision can pass through several "lenses". There are also protocols, meaning actions taken without being asked. At the start of a session it shows the loose ends from last time, at the end it records insights and commits changes, and it also reminds the author to review goals and audit the base.

How it works technically

Data lives in markdown files in git, organized by the Johnny Decimal system. Each entity has one owner file, and every other place links to it. This protects against duplicates and inconsistencies, the main disease of any homemade database.

Working rules are written in CLAUDE.md, which the agent reads at the start of every session. Details are moved into separate specs so the main file doesn't bloat. Roles and skills live in their own folder, also as md instructions and scripts. There are about 70 scripts in total: they pull data from databases, finances, news and calendars.

Search runs through a custom MCP server. It combines ChromaDB with multilingual embeddings and keyword-based BM25, and merges the results via RRF (reciprocal rank fusion). The index is updated every half hour. This combination catches both semantic matches and exact titles, names and numbers, where embeddings alone often miss.

Background tasks on the Mac run through launchd: syncing the CRM with the camp database, call processing, task mirroring. The planner is deployed on Railway, opens on a phone as a PWA with Face ID, and from an Apple Watch dictation goes straight to the API. Memory between sessions is kept in short summaries with open loose ends and in a cheat-sheet file with pointers to the relevant files.

Where to start and the catch

The author's advice: don't design the system from scratch, start with an interview. Open the LLM that has accumulated the most of your history and ask it to come up with 50 questions about your work, projects, habits and what you constantly forget. Set the goal right away: at the end the model assembles a PRD for your Second Brain. You hand that PRD to the agent, and it builds the first version. After that the system grows gradually, with regular feedback.

There are limitations too. This is one person's account, with no metrics and no comparison with off-the-shelf tools. Building it requires comfort with git, scripts and the terminal. And keeping health, finances and a psychological profile in one base accessible to an agent means thinking seriously about security: where the index runs, what data goes to cloud models and what happens if a token leaks. The post ends with an ad for a paid camp, which should be taken into account when evaluating it.

The practical takeaway for a developer or product manager: a working architecture for a personal agent is fairly simple today. It is markdown and git as the source of truth, one owner per entity, CLAUDE.md with rules, hybrid search via MCP and a couple of background tasks. It can be built over a weekend and then refined as real use cases appear.