Sources and derived knowledge

aizk stores two kinds of thing and treats them very differently. Getting that difference straight early makes everything else easier to reason about. This page assumes you have read What aizk is and nothing more.

A source is what somebody actually wrote

A source is the record. It is the note you typed, the brief your assistant wrote on your behalf, the page you pointed at by URL, or the contract you uploaded. aizk stores it as given and never rewrites a word of it. Nothing in the engine edits a source, and nothing quietly merges two of them into a tidier third one.

A source comes in one of two shapes.

Most are text, plain or Markdown, and text is the shape to prefer. Everything downstream reads text, so a note written as text is the cheapest thing to store and the most useful thing to get back.

The other shape is a preserved original file. aizk keeps the bytes exactly as they arrived and converts a readable text version alongside them, so the same document can be found by meaning and still handed back in full later. Preserve an original when the exact document may genuinely be needed again, a signed contract, a form, a paper, or a presentation. Files, PDFs and web sources walks through that path and its size limits.

Either shape carries the scope set deciding who may read it, covered on Scopes, and the times deciding when it applies, covered on Time and history.

Derived knowledge is what aizk works out afterward

Once a source lands, background work reads it and builds a second layer on top. None of that layer is something you wrote. All of it is a projection of what you wrote.

Derived thing	What it is	Name in the web app
Entity	A named thing the sources talk about, a person, a project, a paper	Subject
Fact	One statement linking named things, carrying the quote it came from	Finding
Community	A cluster of related entities the engine noticed	Theme
Profile	A rolling summary of one entity across everything said about it	Profile
Summary	A summary written over a group of sources rather than one	Summary

Entities, facts, ontology explains how prose turns into the first two, which is where most of the value sits.

Where this comes from

The derived layer, the entities and facts aizk pulls from your text, follows the temporal entity and fact graph of Zep and Graphiti. Keeping the raw source authoritative echoes Does Memory Need Graphs. The rule that a source always outranks a derived reading is aizk’s own. The full map of prior art traces every mechanism.

The two layers are not equal

  ┌──────────────────────────────────────────────────┐
  │  SOURCES     what you wrote or uploaded          │
  │              authoritative, never edited         │
  └───────────────────────┬──────────────────────────┘
                          │ read by the engine
                          ▼
  ┌──────────────────────────────────────────────────┐
  │  DERIVED     entities and facts                  │
  │              communities, profiles, summaries    │
  │              rebuildable, never the record       │
  └──────────────────────────────────────────────────┘

    delete the lower box and aizk rebuilds it from the upper one
    delete the upper box and the memory is actually gone

That asymmetry is the whole point. Derived knowledge is disposable by design. When the extraction models improve, or a bad batch of facts turns up, the fix is to rebuild the layer rather than to repair it by hand. Nothing of yours is at risk in that operation because nothing of yours lives there.

Why the split matters when evidence disagrees

Find returns both layers mixed into one ranked list, and every item says which layer it came from. Sooner or later two items will disagree.

When they do, the source wins. A source is a human statement that somebody committed to memory on purpose. A derived fact is a machine reading of some source, and it can be a stale reading, a partial one, or simply a wrong one. So a source excerpt saying the team moved off the LLM extractor outranks a derived memory still claiming the team uses it.

This is exactly why aizk labels each item instead of blending everything into one confident paragraph. Your assistant can see which layer a claim came from and weigh it accordingly, and so can you. Evidence and provenance covers the labels in full.

The same rule settles a subtler case. A derived fact often reads as more authoritative than it is, because it arrives as a clean assertion while the sentence behind it was hedged or conditional. Extraction now refuses a fact that overstates its own sentence, and a claim it keeps but the source qualified says so and brings that source excerpt along. How settled a claim is covers those marks.

What happens when a source changes

Correcting a note does not leave the old derived claims standing. When a source is replaced, the facts built from its previous text stop counting as current and the facts built from the new text take over. The old ones are not deleted, they are closed, so the record of what aizk believed stays intact.

A later, different note disproving an earlier one closes the old claim the same way, and that is how you fix a derived claim that is simply wrong. How settled a claim is shows how to write one.

That is the same machinery that handles a fact simply going out of date, and Time and history owns the explanation.

What this means in practice

Three habits follow directly from the split.

Write the source you would want to read. Derived knowledge is only as good as the prose under it. A note with a real title and real sentences produces useful entities and facts. A note that is three keywords produces almost nothing.

Do not try to write facts by hand. There is no tool for adding an entity or a fact directly, and that is deliberate. You write sources and the engine derives the rest, which keeps every derived claim traceable to a real quote in a real document.

Do not treat the derived layer as a backup. It is an index, not an archive. If the text matters, the text belongs in a source.

Next

• Entities, facts, ontology shows how prose becomes the derived layer.
• Writing memory well is the practical guide to authoring sources.
• Time and history explains the clocks every source and fact carries.