Questions and answers

The questions that come up in the first week, answered plainly. This page assumes you have read What aizk is. Where the honest answer is a limitation, it says so.

Does it read my files?

No. aizk never scans a disk, a repository, or a folder. It holds exactly what was handed to it through keep and nothing else.

Your agent can read files and decides what to send, so the practical boundary is in its instructions. Claude Code provides a safe starting point, including the rule against secrets. AIZK cannot prevent an agent from sending sensitive text.

What happens if I delete something?

There is no delete tool, and that is a real limitation. Correction is the mechanism instead. You store an updated statement, the new version becomes current, and the old one keeps its dates and stops appearing in ordinary find. Time and history covers it.

If something genuinely has to go, an operator can retract what a source contributed and remove the source at the database. That is a real capability, but it is an operator command, not something you or your agent can do from a chat.

Can I get my data out?

Yes. Everything you can see exports to a plain JSONL file, including sources, chunks and derived knowledge. Preserved originals come back byte for byte. The active SQL backend holds the memory, while the configured object store holds original file bytes. Export is an operator command rather than a user interface button, so a copy means asking whoever runs the deployment.

Why did find not find my note?

Four usual causes, most common first.

  find missed it
    │
    ├─ stored in the last few minutes ─▶ still processing, check status
    │
    └─ older than that
          ├─ other items came back       ─▶ it lost on ranking or the token budget
          │
          └─ nothing came back
                ├─ it lives in someone else's scope ─▶ you are not in that organization
                └─ it is yours, with an expiry       ─▶ past its expiry, history keeps it
                    that has passed

Writing returns before the note is searchable, so status gives a range for when new material becomes findable. Check it before concluding anything is wrong. If other things came back but not the one you wanted, it was found and beaten, since find returns the best items that fit a token budget and a thin note can lose to a rich one. A more specific question usually fixes it, and Asking memory well has the phrasing advice. If nothing relevant came back at all, check whether the note lives in an organization you belong to, and whether it carried an expiry that has since passed.

Why does it return evidence instead of an answer?

Because an answer cannot be checked and evidence can. A single sentence hides which note produced it, when it was written, and whether it was your words or an inference. Evidence keeps all three, so your agent can prefer a source excerpt, notice that two items disagree, and tell you which is newer. It also keeps aizk out of the reasoning business, a job the assistant already does well. Evidence and provenance goes into the labels.

How is it different from putting notes in a repository?

Four differences that matter, and one honest concession. A repository note is findable by exact string, while aizk is findable by meaning, which is what you have when you half keep a decision from six months ago. A repository has no notion of who may read what beyond the whole repository, while aizk carries a scope on every row and the database enforces it. A repository has one timeline, the commit log, while aizk tracks when a file changed and when a statement was true separately. A repository note is trapped in one project, while aizk answers across everything you can see at once.

The concession is that for a small project, files plus grep are simpler and faster. aizk earns its setup once memory has to cross projects, people, or years. See Scopes.

Does a model provider see my private memory?

It depends on the deployment profile. The self-hosted PostgreSQL profile runs its model lanes on the deployment hardware by default. The AIZK AWS demo uses hosted endpoints for text embedding, captions and extraction, so accepted content reaches those providers. Ask the operator which profile and providers are active before storing sensitive material.

Your assistant is a second boundary. When your agent calls find, the evidence enters that agent’s context, and if the agent is hosted then that text travels to its provider like everything else in the conversation. aizk does not change what your assistant does with what it reads, so if that matters for a note, the decision belongs at the point of asking.

What happens when two people write the same thing?

Both sources are kept. Nobody’s words overwrite anybody’s words. The derived knowledge converges instead, so when two notes produce the same statement, one shared copy exists and each scope holds its own claim with its own dates.

Statements tied to a speaker behave differently on purpose. An opinion, a preference, an observation, and a personal experience stay attached to whoever said them, so two people can hold opposing views without one erasing the other. Statements about the shared world do not get that, and a later one supersedes an earlier one. Entities, facts, ontology has the detail.

Is there a review step?

No, and there is not going to be one. Nothing sits in a queue between keep and the moment that memory can be found. There is no approval state and no human gate.

That is deliberate. A review queue only works if somebody drains it, and a memory that fills faster than a person can approve becomes a backlog. The correctness work moves to the agents instead. They find before writing, update a maintained note rather than adding a competing one, and correct what changed. Who maintains memory explains the arrangement, and Notes that stay useful is the habit list.

Next

• Glossary defines every term used here.
• Notes that stay useful is how to keep memory worth reading.
• Sign-in troubleshooting covers connection problems.