The second brain that writes itself

Bartosz Sojka

Claude Community Meetup · Amsterdam · 15 September 2026

research-atlas

From an empty folder to 403 markdown files

17 Aug → 10 Sep 2026

Why markdown

Agents are trained on it

It's their native input and output

Files are how agents act

They live on a computer: read a file, write a file, run a command

No integration

A file appears as a side effect of working. No tool call, no schema.

You can read and fix it

Any sentence, in any editor

research-atlas

How to find information

[[wikilinks]] cites raw/
CLAUDE.md needs/ sources/ methods/ concepts/ guides/ raw/ summaries/ index.md log.md _meta/ templates ×

Demo 2 · recording

One change, propagated

replays/demo-2/

paused–Demo 2 · one change, propagatedP play · ↑↓ speed · . , step · R/E start/end

How a wiki decays, and what keeps it true

Decays

  • Contradicts itself
  • Goes stale
  • Only grows

Keeps it true

Loopsscheduled scan, check and grow
[[wikilinks]]a change finds everything that depends on it
Page templateseach page type has a fixed shape
lint.pymechanical checks on every change

Is the structure right?

Demo 1: a new guide, from the right reads

replays/demo-1/

paused–Demo 1 · is the structure right?P play · ↑↓ speed · . , step · R/E start/end

Live

An interface on top

readwritten

Why markdown, not MCP or an API

A folder of markdown

  • See whether the agent read the right things
  • Change the structure until it does
  • Links show what else a change must reach

Behind MCP, an API or a CLI

  • Hard to verify it read the right sources
  • You can't change how the data is stored or delivered
  • Update one record, and nothing says what else should change

Only on files can you verify, and change, what the agent reads.

Take home

  1. 1Make a custom structure for your use case
  2. 2Measure how your agent works
  3. 3Keep it up to date and check for problems
  4. 4Grow your wiki into a new structure when the current one gets inefficient

Bartosz Sojka