Agents produce everything and remember nothing | Building Relinkia — the context layer you own | Author @ManningBooks, @packtpub, Eyrolles
France
Joined September 2009
- Tweets521
- Following494
- Followers254
- Likes897
Pinned Tweet
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.
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.
I'm building Relinkia because my own vault stopped working: 1,500 notes, and maybe 30 I still open.
The agent reads the notes, adds links between them and flags where two of them disagree. Which one I still believe stays my call.
A note you wrote in 2021 and one from yesterday come back side by side in search, with the same weight.
Some of what you wrote in 2021 you've since dropped, and nothing in the note says so. Search hands it back as if you still thought it.
In September 1838 Darwin read Malthus on population, an economics book, and the mechanism of natural selection fell into place.
The idea he needed was already written down, in a field he didn't work in. Spotting that kind of crossing is something software can be built to do.
Sort your notes app by last modified and scroll to the bottom. Most of what's down there was worth writing the day you wrote it.
Nothing in the app ever brought those notes back next to something new, and a note that never meets another one stops mattering.
Ask ChatGPT about a book you read last spring and you get the internet's average opinion of it.
Your margin notes and the objection you wrote next to chapter 3 never reach the answer. No model was trained on them, and they're the part that was yours.
You don't have a reading problem. You have a forgetting problem.
Most "learning" is just well-organized forgetting: highlight, save, archive, never touch again.
The bottleneck isn't input. It's what happens to ideas after you close the book.
Unpopular observation from running agent pipelines: the hard problem isn't giving agents access to knowledge. It's deciding what they don't get.
Everyone's building bigger context windows. Almost nobody's building the permission model that should sit in front of them.
Unpopular observation from running agent pipelines: the hard problem isn't giving agents access to knowledge. It's deciding what they don't get.
Everyone's building bigger context windows. Almost nobody's building the permission model that should sit in front of them.
Everyone is suddenly building the same thing: an Obsidian vault + Claude + MCP, with the agent reading and linking your notes.
Karpathy gave the pattern a name back in April: the LLM Wiki.
It's the best thing to happen to personal knowledge in years. And it has a blind spot almost nobody mentions.
Second gap: time.
Your 2023 position and last month's quiet reversal sit in the vault as equally true. An agent retrieving by similarity will hand you the stale one: confidently, fluently, wrong. A wiki that "gets smarter every day" should at least know which of your ideas are still yours.
So the LLM Wiki proves the demand: everyone wants a context layer they own.
Portable: solved. Structured and temporal: still wide open. That's the actual frontier.
Longer argument: relinkia.substack.com/p/your…
Your brain has about 86 billion neurons and holds far more concepts than that. So does an LLM, with its dimensions.
That's not a coincidence. It's geometry: concepts live as directions in a space, and a space can hold exponentially more directions than axes. It's called superposition, and it explains both why these models work and why they're so hard to read.
I made a video walking through it, no math degree required.
