# How it is built underneath: files, parser, modules, retrieval.

> How IllAIra is built underneath: Markdown and parser, Nuclei and modules, search by meaning, provenance, Sphere Grid, compatibility, documentation and integrations — and the two names of every thing.

*designed version: https://illaira.com/en/voglio-capire · index: https://illaira.com/en/llms.txt*

---

[Skip to content](#contenuto)

I want to understand it

No magic: a Markdown file with a declared structure, a parser that reads it, an app that draws it and searches it. Every card leads to the page that explains that piece in full; then the four roads to give an AI memory, side by side; at the bottom, the table of the two names of every thing.

- ## Markdown and parser

  The Reminder is a .md file with a first line that declares the type, permission zones [P1]…[P5] that say who may touch what, and headings with a role. The parser reads it line by line and says what does not add up, before the model reads it.

  [Anatomy of a file, line by line](https://illaira.com/en/approfondimenti/anatomia-di-un-file.md) · [The permission zones](https://illaira.com/en/approfondimenti/zone-di-permesso.md)
- ## Nuclei and modules

  Modules are the abilities (functional) and the memories (mnemonic); Nuclei group them and switch them on together. Two homogeneous files merge without losing a line: what is missing is grafted in switched off, what you already have is not touched.

  [The Nuclei, in How it works](https://illaira.com/en/come-funziona.md#nuclei)
- ## Search by meaning (retrieval)

  An embedding model downloaded once, an index built on the device, and a search that finds the passage even in different words. In the browser and on the App, with the same catalogue of models: the texts never leave.

  [How it is switched on, and what changes](https://illaira.com/en/clm.md#funzioni)
- ## Provenance of memories

  Every journal entry carries a date and a topic, and the model writes it in the right form with the Logger. With Deferred Neural Tracing the model emits a trace of which modules it used, to paste into the Sphere Grid; the “Links” screen, on the App and in the WebApp, shows where each log was filed and which block names which thing.

  [The one-tap log, and “Links”](https://illaira.com/en/clm.md#funzioni) · [Logger and Neural Tracing on the Bazar](https://illaira.com/en/bazar.md)
- ## The Sphere Grid

  Not a note graph: it draws the structure declared in the file — modules, sections, logs — with the pulse on what is switched on. In 3D the whole archive in shells; in “Links” the links.

  [Why it is not the graph you have already seen](https://illaira.com/en/approfondimenti/non-e-un-grafo-di-note.md)
- ## Compatibility

  Any model that accepts a file or a permanent instruction: Claude, ChatGPT, Gemini, Grok, Copilot, local models. Where to upload it, service by service, and the line to add so it is not treated as an attachment.

  [Where to upload it, model by model](https://illaira.com/en/approfondimenti/dove-caricare.md)
- ## Documentation

  The user guide as a PDF in five languages, the protocol in full, the technical deep dives — and the site itself in markdown, for those who read it with a model instead of with their eyes.

  [The Standard and the documentation](https://illaira.com/en/documentazione.md) · [The deep dives](https://illaira.com/en/approfondimenti.md) · [The site in markdown (AIndex)](https://illaira.com/en/llms.txt)
- ## Agents, toolkit and integrations

  The file framework gives an agent who it is, what it may touch and what it remembers. The Toolkit takes your data out of the providers, and the migration guide turns chats, memory and custom instructions into Reminder, modules and journals.

  [Agents + IllAIra](https://illaira.com/en/agenti.md) · [Free your AI: the migration](https://illaira.com/en/libera-la-tua-ia.md)

## Compared with the other roads to give an AI memory

Four roads are in use today, and three of them work. We put them in a row with what each does better than us, because a comparison that lists only the other side’s limits is not a comparison, it is an advertisement. The line that separates us is not «more memory»: it is who decides what stays, where it lives, and what it takes to move it.

| The road | Who writes it, who decides | Where it lives | Changing model | Its strong point, and its price |
| --- | --- | --- | --- | --- |
| The provider’s memory (ChatGPT, Claude, Gemini) | The system: it notes what it judges worth keeping, in its own words. You can read the list, delete a line or ask for it to be reworded, and add a few lines of your own in the box it gives you. The form stays its own. | On the provider’s servers, in a form that is theirs. | An export where a law requires it: a copy, to be reworked by hand before another model can use it. | Strong point: it asks nothing of you, from day one. Price: what goes in is chosen for you unless you say otherwise, contexts bleed into each other, and leaving means redoing it by hand. |
| Memory frameworks for developers (Mem0, Letta, Zep, Cognee) | A pipeline: a model extracts facts from the conversation and stores them in a vector or graph database, automatically and at scale. | In a database — hosted by them, or on a server of yours — behind an SDK or an API. | The engine can change; the memory stays in that store and moves with it. | Strong point: scale and automation — thousands of users, facts that carry a date, retrieval measured on benchmarks. Price: it is infrastructure for whoever builds an agent, not a tool for whoever talks to one; it needs an API key or a server of yours, and the memory is a store to query, not a document to reread. |
| Notes and local files (Obsidian, a personal wiki, a folder of .md, MemPalace) | You, or a script that files everything verbatim. Nothing is chosen in your place — and nothing is chosen. | On your disk, in plain text or in a local database. | Attach or paste: the files go anywhere, and you carry them by hand every time. | Strong point: searching, and bringing things back up — an index finds one line among ten thousand, and you start immediately. Price: a note says what you wrote, not who is reading it; nothing declares what may not be rewritten, and the model reads only what you paste each time. |
| IllAIra | The model writes the entry with the Logger; you decide what stays, with one tap on the MemoKen. Identity, modules and zones are declared in the file, and the app enforces the zones. | In .md files on your device: a folder, or the browser’s storage. No server of ours in between. | The same file in another engine: paste it, or put it where that model keeps its documents. The surface changes, the identity does not. | Strong point: you decide what stays, who it is stands in the file before the question, and the file travels by itself. Price: tokens at every conversation, a format to learn, a yes or a no that nobody can say in your place — and inside the model the zones are text. |

None of the four rules out the others. A developer building on Mem0 or Letta can still hand the agent a Reminder for who it is; a personal wiki fits inside a memory module; the provider’s memory is where the migration guide starts. What we do not do, on purpose, is the automatic pipeline and the store: the entry passes under your eyes, and it lands in a file you can open without us.

[The objections, one by one](https://illaira.com/en/come-funziona.md#obiezioni-risposte) · [Free your AI: the migration](https://illaira.com/en/libera-la-tua-ia.md)

## The same things, with two names

In Simplified mode the WebApp calls things by what they do for you; in Intermediate and Advanced by their technical name, and in Intermediate with the explanation beside it. It is the same thing and the same file: this table is the map, and it is the very one the app uses.

| In Simplified mode | In Intermediate and Advanced mode | What it is |
| --- | --- | --- |
| Your AI | Reminder | Reminder: the file that says who your AI is, how it speaks and what it must not touch. |
| What your AI remembers | Memory modules | Memory modules: the journals that collect your AI’s memories and dated notes. |
| How you want it to behave | Functional modules | Functional modules: the abilities and rules you switch on and off. |
| Save this memory | MemoKen | MemoKen: in the WebApp the browser extension, in the App the button that sits over the other apps; it carries a memory from the chat into the right journal. |
| Memories useful for this conversation | Search by meaning | Search by meaning (retrieval): a model downloaded once finds the passages close to what you describe, even in different words. |

## In the WebApp: Advanced mode

Markdown editor, block encryption, model catalogue with sizes and licences, search scores, functional modules and merges: everything Simplified mode keeps one click further away, and Intermediate brings halfway. You choose at first launch, switch from Settings › Mode, and nothing is lost.

[Open the WebApp](https://illaira.com/en/webapp)

Two paths, one base

- ### [I want to use it](https://illaira.com/en/voglio-usarla.md)

  Your AI remembers you, from one conversation to the next and from one model to another. What changes for you, how a memory is saved, where your files stay, and how to try it in one click.

  [The simple path](https://illaira.com/en/voglio-usarla.md)
- ### I want to understand it You are here

  Markdown and parser, Nuclei and modules, search by meaning, provenance, Sphere Grid, compatibility: how it is built underneath, with the documentation and the table of the two names of every thing.

Same files, same engine: only the words we tell it with change. In the WebApp too: Simplified, Intermediate and Advanced mode, and you switch whenever you like.