pTrack · The HQ method

Give your AI agents a headquarters.

HQ is one versioned home for everything a team knows. AI agents work inside it. They learn your rules, write down why, and run in a sandbox that keeps them away from what they shouldn't touch.

I built my own HQ in May 2026. All my work runs through it, every day.

hq/
hq/
├── knowledge/      one area per kind of work
│   ├── clients/
│   ├── learning/
│   └── admin/
├── memory/         what the agents learned
│   ├── MEMORY.md   index, read every session
│   └── important/  kept long-term, on request
├── roles/          one skill per team role
├── rules.md        always-on rules
└── data_policy.md  what may live where
Running since
May 2026
Approved changes
123
Memories written
150
AI roles
13

My own HQ, counted on 20 October 2026.

The philosophy

Six ideas behind it

One home

Everything the team knows lives in one place, with one structure. Apps and tools live elsewhere; HQ points to them.

Plain files you own

Text files under version control. Anyone can read them without a special tool, nothing is locked in, and every change has a history.

Write it down, with the reason

Agents keep a memory. Each lesson records why it exists, so it can be judged later instead of followed blindly. Your notes and the agent's stay apart.

Ask before acting

Reading up on a task is not permission to start it. The agent summarises what it found, then asks.

Propose, then review

No change lands directly. Each one arrives as a proposal, gets reviewed, and a person approves it.

Layers, with honest limits

Any safeguard can fail on its own. Several overlap, and each one states plainly what it does not cover.

The system

How it fits together

Work flows in, the agent does it inside HQ, and every result passes review before it goes out. All of it runs inside the sandbox.

Select any part of the diagram to read what it does.

Comes in

HQ

Works

Review

Goes out

HQ

Memory

The agent writes down what it learns, in four kinds: who you are, how you want work done, what's going on, and where things live. Each entry holds one fact and the reason behind it. A short index loads at the start of every session.

Correct it once, and the lesson carries forward.

Memory

Correct it once

A new AI session knows nothing about yesterday. In HQ, every correction becomes a written rule that the agent reads at the start of the next session, so the lesson carries forward instead of being lost.

  1. 1

    Work

    The agent does a task.

  2. 2

    Feedback

    You correct it, or confirm an approach that worked.

  3. 3

    Write it down

    The rule, why it exists, and how to apply it. One fact per file.

  4. 4

    Next session

    The index loads at the start. The lesson applies from then on.

↺ and round again, every session

memory/confirm_before_acting.md
---
name: confirm-before-acting
type: feedback
---
Loading context is not permission to act.
Summarise what you found, then ask.

Why: an agent was asked to read up on a
project. It kept going, analysed the data
and rewrote a document nobody asked for.

How to apply: before the first action
that changes anything, ask. If unsure
whether it counts, it counts.
A real rule from my own HQ, shortened.

Four kinds of memory

Who you are
Role, expertise, how you like to work.
How to work
Corrections, and approaches that worked.
What's going on
Ongoing work, decisions, deadlines.
Where things live
Systems, dashboards, documents.

Kept clean

  • Updated instead of duplicated.
  • Deleted when it turns out wrong.
  • Checked against today before it's trusted.
  • Promoted to the long-term set only when you ask.

Safety

Five layers around the work

An agent with access can do damage quickly. So its work sits inside five layers, and each one stops a different kind of mistake.

SandboxThe operating system limits the agent to approved folders and approved websites. Everything else is refused outright, and the agent cannot change its own limits.
Secrets kept outPasswords and keys live where the agent can't read them, and a guard blocks attempts to display them. Tested: 10 of 10 checks passed.
Data policySix categories decide where each kind of data may be stored and which AI may read it. Being allowed to store something is not permission to send it.
Production lockedLive systems are off limits by default. When someone asks for a change there, the agent argues against it first.
A person decidesEvery change is a proposal. A second AI reviews it, and a person approves it.
The agent's work

Each layer can fail on its own. That's why there are five.

Quality

Fast, and done properly

The agent brings the speed. The rules it carries bring the quality.

The method lives in the role
A tester role carries risk-based test design and standard testing terminology. Every test plan follows it, whoever asks for one.
Check before claiming
Agents verify how something actually works before they judge it, and keep their sources' caution when they quote them.
Reviewed until clean
Findings from the second AI are fixed, and the proposal is reviewed again, as many rounds as it takes.
Written for the reader
Verdict first, short sentences, and every technical term explained the first time it appears.
Traceable
Every change has a history. Every rule has a date and a reason.

For a team

What it gives a team

This is how pTrack works, whether the job is building an HQ for a team, testing a product, or leading the people who build it.

info@ptrack.tech