
Count the AI tools you used this month. ChatGPT, Claude, Copilot, Gemini, Perplexity. Now: how many of them know what you were thinking last Tuesday? Zero. Every single one starts from a blank page.

Agents produce everything and remember nothing | Building Relinkia — the context layer you own | Author @ManningBooks, @packtpub, Eyrolles

Count the AI tools you used this month. ChatGPT, Claude, Copilot, Gemini, Perplexity. Now: how many of them know what you were thinking last Tuesday? Zero. Every single one starts from a blank page.

What "capture" looks like in the Claude app: Claude asks before it writes the updated map to Relinkia over MCP. My notes are in French, the routine is the same in any language.

7/ The whole routine is one Claude skill, for any subject: github.com/Relinkia/skills/t… It runs on a free Relinkia account, or without Relinkia on a folder of markdown files in Claude Code. You lose versioning and phone access there, the method stays the same.

6/ Next session, Claude rereads the map first and picks up from there, in claude.ai on my phone or in Claude Code. Both read the same notes, so it doesn't matter where the last session happened.

5/ When a descent ends I say "capture". Each concept gets a note with the problem it solves, a worked numeric example, why it's used there, common misconceptions, code and three recall questions. A map keeps what I know and where it stops.

4/ A session starts from an anchor. This week it was a post about JEPA. From my phone I go down one concept at a time: latent space, SIGReg, BatchNorm, LayerNorm, RMSNorm, and finally attention.

3/ Claude's memory and Projects helped, but pieces went missing and older material came back half rebuilt. Now the reference lives in Relinkia, the tool I'm building, over MCP: every note is versioned, and any session can read it.

2/ Each new chat started from zero. I re-explained what I already knew, and the reference docs I had asked for were scattered across dozens of old conversations or rebuilt from fragments.

1/ I'm learning deep learning with Claude, one concept at a time. The explanations are good. What I understood used to disappear when the conversation ended, so I changed how I work. Thread.

The questions worth asking an AI about your own reading are narrow. Where you contradict what you wrote in 2024, or which of your notes answer each other without you noticing. No prompt gets there without your notes and the links between them.

Read-later apps keep what you save and note apps keep what you write. The step between reading something and seeing your own field differently because of it is still done by hand, which is why it mostly doesn't happen.

You rarely notice the day you change your mind. It happens over months, in notes that never sit next to each other, until you argue for something your own notes argued against in March. A tool that reads your notes could show you that March note first.

Most AI second brain demos run on a corpus picked for the demo and a query tried a dozen times beforehand. For the Relinkia demo, the full corpus rebuilds from public sources with its dates, and the run isn't scripted. If the feature doesn't fire, there's no video.

Two notes you wrote three years apart may hold the same idea from different angles, and memory alone will never put them side by side. Comparing every note against every other is tedious for a person and cheap for software.

Paste fifty notes into ChatGPT and it answers from whichever ones happened to fit the question. Point RAG at the raw files and you get quotes back, not the reasons you linked them. What both miss is how your notes relate, and that's written nowhere in the files.

Most second brains store things very well. A brain keeps less, but it lets an old memory run into a new one, sometimes one that contradicts it. A folder tree never puts two notes in front of each other unless you open both yourself.