Skip to content
aizkaizk

What aizk is

aizk is a memory you and your AI assistants share. You tell it things worth keeping, it keeps them, and later any assistant you use can ask for what is relevant and get it back in the original words, with a note about where it came from.

This page assumes nothing. Read it first.

The problem it solves

Section titled “The problem it solves”

An assistant forgets everything when the conversation ends. You compensate by pasting the same background in again, or by keeping notes in files that only you can find, or by explaining a decision for the fourth time because the fourth assistant was not there for the first three.

Meanwhile the useful things are scattered. Some live in a document, some in a chat, some only in somebody’s head. When a teammate needs the reason a choice was made six months ago, there is no one place to ask.

aizk is that one place, and an assistant can use it without you doing anything by hand.

You and a teammate each work through your own assistant, and both assistants share one aizkkeep, findkeep, findyouyour assistanta teammatetheir assistantaizk

What it stores

Section titled “What it stores”

Two different things, and the difference matters.

The first is what you actually wrote, kept exactly as you wrote it. A note, a decision, a paper, a PDF. aizk never edits this. It is the record.

The second is what aizk worked out from it. Reading your notes, it pulls out the things being talked about, people, projects, tools, results, and the statements connecting them, such as “the team chose GLiNER2 for extraction”. These are derived, they can be thrown away and rebuilt from your text at any time, and they exist only to help find the right source faster.

Only the first kind is authoritative. If the two ever disagree, your words win. Sources and derived knowledge goes into this properly.

Who can read it

Section titled “Who can read it”

Everything you store is private to you unless you say otherwise. There is no shared pool that notes drift into.

When you do want to share, you name the organizations a note belongs to. Naming one puts it in that team’s memory. Naming two produces a note that only people who belong to both can read, which is how knowledge about one collaboration stays inside that collaboration. Reading requires standing in every organization the note names, and the database itself enforces that rule rather than the application asking nicely. Scopes explains it fully.

What you get back

Section titled “What you get back”

Asking aizk a question does not produce an answer. It produces evidence, which is a short ranked list of the most relevant things it holds, each labeled with where it came from.

That distinction is deliberate. An answer hides its reasoning and cannot be checked. Evidence can be. Your assistant reads the evidence and forms the answer itself, and because each item names its source and its scope, both you and the assistant can tell a note you wrote last week from something the engine inferred.

  you ask   "why did we drop the LLM extractor?"
                     │
                     ▼
  ┌──────────────────────────────────────────────────────┐
  │ ## Evidence                                          │
  │                                                      │
  │ - Source excerpt · scope Book Club                     │
  │     We moved extraction to GLiNER2 because the LLM   │
  │     lane cost 4.1 s per chunk and GLiNER2 costs      │
  │     0.3 s at the same grounding rate.                │
  │                                                      │
  │ - Derived memory · scope Book Club                     │
  │     [aizk, world] (uses) aizk uses GLiNER2.          │
  └──────────────────────────────────────────────────────┘
                     │
                     ▼
  your assistant writes the answer, and can show its work

The response is also budgeted. aizk returns the most relevant items that fit inside a token limit, so it never floods your assistant’s context window with everything it knows.

It keeps when things were true

Section titled “It keeps when things were true”

A memory is not just text, it also carries time. aizk records when something was true in the world and, separately, when it was told. Correcting an old note does not erase it. The old version keeps its dates and the new one starts its own, so you can still ask what the team believed in June and get June’s answer. Time and history covers this.

Where it runs

Section titled “Where it runs”

AIZK has two supported deployment profiles. Both keep documents, claims, vectors, temporal state, scopes and durable work in one SQL backend. Both expose the same five MCP tools.

Profile Best fit Storage and compute
CockroachDB Cloud a compact distributed service on AWS CockroachDB stores memory, S3 stores original files, and Lambda runs the public server and worker
PostgreSQL a private self-hosted installation PostgreSQL with VectorChord stores memory, SeaweedFS stores original files, and local services run the model lanes

The profiles make different operational tradeoffs. The CockroachDB profile uses C-SPANN and a portable durable queue. The PostgreSQL profile uses VectorChord, BM25 and PgQueuer. Read Database profiles before choosing one.

What it is not

Section titled “What it is not”

It is not a search engine for your files. Plain text search is faster and simpler when you know the exact string you want, and aizk is not trying to replace it.

It is not a chatbot. It stores and retrieves, and the assistant does the talking.

It is not a document manager. Original files are kept safely and immutably, but the point is the knowledge in them, not the filing.

Next

Section titled “Next”