AI as a tool. How to turn it into a permanent agent.

A machine that predicts text

• Predicts the next word across a vast corpus — on the statistics of language.
• By default it returns the average: the most typical and expected.
• The ceiling of the answer is set by your context — higher context, higher answer.

Each time — a little differently

• It outputs a distribution and draws from it: one question — different answers.
• With no fact it completes something plausible: confident on the surface, empty inside — a hallucination.
• Cost — no guarantee of a repeat. Benefit — language, flexibility, novelty.

Two different natures

That's why AI is verified against reality.

Meaning is geometry

• Every word is a point in a space of thousands of dimensions; close in meaning sit close together.
• On the way in, words become points; on the way out, meaning unfolds back into words — in any language.
• Learned in one language, it's available in another: the meaning is shared.

The window is an active slice of that space

• Only what's loaded right now lands in the window, on your desk.
• Every request carries everything anew — the chat has no memory between them.
• The desk is large, but finite.

You tuned the chat, it got good — and you want to keep it as an agent.

The better you tuned it — the faster it dies.

Move knowledge into a base

• The agentic approach replaces prompt engineering: knowledge lives in a base outside the chat.
• For each question it pulls the pages closest in meaning — the desk stays fresh.
• A prompt lasts one session. A base lasts forever.

Not a prompt. A working loop

The agent works through a system: takes the task, reads the base, builds a plan, acts, checks against reality, and records the lesson back into the vault.

That creates process memory, not one lucky answer.

A permanent agent and independence

• Memory in the base: edit the base — the agent gets smarter, fixes accumulate.
• One base — many agents.
• The asset is yours: a better or cheaper model ships — you switch freely.

This system is half a century old — Zettelkasten

• Scholars organized knowledge this way long before AI.
• A network of linked notes — an external second brain.
• Now your base is read by AI too.

The Obsidian graph is a context map

A dot is a page, a line is a link in meaning. In the demo we open this graph: the agent does not “remember magically,” it finds the right corner of the base and puts it into the window.