traqr.dev·working document·rev. 2026.10
Nothing is real to a model until it lives outside the model.
TraqrOS is the system I built on top of Claude Code to run my own work: research, my week, and the apps I’ve shipped. It started as a memory database and an agent harness. After months of running it every day, it’s really one idea applied everywhere: whatever a model learns, decides or claims gets written down outside it, in pieces small enough to check. It’s open source, and the pieces install separately.
§ 00·thesis
Two rules. Everything below is one of them, built.
Outside the walls.
A model starts every session with nothing. What it learned last time only exists if it was written somewhere it can read back: a database row, a file, a commit, a ticket. The same goes for what it claims. Merged isn't live and done isn't delivered until something checks the thing itself. So every part of TraqrOS is either a place for work to live outside the model, or a check against one.
Strip it down to primitives.
The fewer moving parts between the model and the truth, the better the work. A memory is one row. A claim on a task is one key. A session's notes are one markdown file. A change is one pull request. Small, plain pieces the model can read whole and put together itself beat one clever system it has to understand first.
§ 01 is where what it learns lives. § 02 is where the rules live. § 03 is the procedures. § 04 is where the work lands.
§ 01·memory
It remembers what it learned.
Every session, your AI rediscovers your codebase from scratch. TraqrDB is a memory database in your Postgres — agents store what they learn as they work, and load it back before the same mistake happens twice.
fig. 1 — the same gotcha, stored in session 1, recalled in session 47. Real tool calls, replayed.
§ 02·harness
One config file. The rest is generated.
Write .traqr/config.json, run npx traqr render. The CLI generates everything your AI needs to work on the project like it’s been there for months.
{
"project": {
"name": "my-saas",
"type": "code"
},
"memory": {
"supabaseUrl": "...",
"embeddingProvider": "openai"
},
"vcs": "github",
"skills": ["ship", "einstein"],
"worktrees": {
"slots": ["feature1", "bugfix1"]
}
}CLAUDE.md
The project intelligence file — conventions, app routing, memory triggers, workflow rules. Your AI reads it before touching anything.
Slash commands
/ship, /einstein, /bethesda, /daily-brief, /call — 50-odd skills, generated from your config and tuned to your project.
Memory MCP server
Twelve tools: store, search, context, browse, enhance, pulse, archive, audit, read, correct, forget, purge.
Worktree slots
Parallel branches with port assignments, shell aliases, and a pre-push hook that keeps everyone off main.
Guardian
The automated PR lifecycle — rebase, build, merge, sync the other slots, write what it learned back to memory.
Obsidian vault
Research docs, agent diaries, decision records, daily briefs. The knowledge that doesn't fit in a commit message.
§ 03·skills
The index.
Skills are slash commands — each one a documented procedure the AI follows. These are the ones I use most; the full set is generated from your config.
/shipcommit, build, push, open the PR, notify Slack/resyncpre-flight rebase with impact analysis/syncsync every worktree slot after a merge/simplifyreview changed code for reuse and efficiency/nextphasepick up the next phase of a Linear project/einsteindeep research with parallel subagent waves/bethesdaautonomous exploration — agents find their own work/debatestructured devil's advocacy on a decision/deepreflectmeta-analysis of how the agents are performing/coschief-of-staff awareness across all slots/daily-briefmorning intelligence from every data source/callprocess a meeting transcript into structured notes/checkpointmid-work save to memory and the vault/deciderecord a decision with its full context/traqr-initinitialize TraqrOS on any project/traqr-updateupgrade in place, preserving your sections/traqr-system-slackwire Slack routing and Block Kit/traqr-system-guardianconfigure the automated PR lifecycle/traqr-system-cronsstandardized cron auth and escalation…and ~30 more. Each one is a markdown file you can read, edit, or delete.
§ 04·agents
PRs merge while I sleep.
Agents work in parallel git worktrees, one slot per branch. When one ships a PR, Guardian takes it from there: rebase, build, merge, sync the other slots, and write what it learned back to memory.
fig. 2 — the four most recent agent-merged PRs in this repo, read from the GitHub API when this page was built (June 10, 2026). Titles as the agents wrote them.
§ 05·install
Take the piece you want.
The memory server stands alone — most people start there. The CLI generates the full harness. Both run on your infrastructure: any Postgres 15+, GitHub or GitLab (auto-detected from your remote), any embedding provider, Slack or none.
path a
Just the memory
The MCP server, pointed at your Postgres. Twelve tools, works with Claude Code, Cursor, Codex, or any MCP client.
path b
The full harness
Memory plus the generated CLAUDE.md, skills, worktree slots, and Guardian. Everything this page describes.
published packages · versions checked 2026-10-17
traqr-memory-mcpv0.2.41The MCP server. Twelve memory tools, an interactive setup wizard, works with any MCP client.@traqr/memoryv0.2.62The TypeScript library underneath — multi-strategy retrieval, entity extraction, bring your own Postgres.@traqr/cliv0.1.13traqr init, traqr render, traqr status, traqr verify.@traqr/corev0.1.53Config schema, template engine, VCS abstraction for GitHub and GitLab.