Who maintains memory

Every shared knowledge system eventually faces the same question, which is who keeps it from rotting. aizk answers it in a way that surprises people, so this page states the answer plainly. It assumes you have read What aizk is and Sources and derived knowledge.

There is no review queue

Nothing waits for approval. When your assistant stores a memory it becomes a source right away, and the next question can find it. There is no draft state, no moderator, no inbox of pending notes, and no weekly ritual where somebody blesses what piled up.

This is a design choice, not a missing feature, and it is not on a roadmap. A review queue is a bottleneck that grows with the number of writers and shrinks with everybody’s patience. It goes unread within a month, and then the queue itself becomes the stale thing.

The loop closes without a human in it. That only works because the agents doing the writing carry the responsibility a queue would otherwise pretend to hold.

What an agent is responsible for

Five habits carry the whole lifecycle. An assistant using aizk well does all five. One doing none of them fills memory with duplicates and contradictions, no matter how good the retrieval is.

Find before writing. Ask what is already known before adding anything. This stops the fourth copy of a decision from being written, and it turns a new note into an update of an existing one when that is the right move.

Choose the destination deliberately. A memory goes to exactly one place, and private is the default. Team knowledge is written straight into the team scope, not written privately and moved later. Scopes explains what the choice means.

Preserve provenance. Keep the original wording, the source URL, and enough context that a reader six months from now can tell where a claim came from. A memory stripped of its origin is a rumor.

Correct what changed. When new evidence contradicts a stored note, the agent finds the current version and writes the correction rather than leaving both standing. Correcting costs nothing, because aizk closes the old claim instead of deleting it and the history survives. Time and history has that mechanism.

Use temporal bounds only when the world supplies them. An expiry says something stops being true on a date, and almost nothing in ordinary documentation qualifies. Guessing at one quietly removes knowledge from find later. The same page owns that rule.

Why the agent rather than the person

Two reasons, both practical.

The agent is already there. It is reading the new evidence at the moment it arrives, and that is the only moment when correcting a note is cheap. An hour later the context is gone and the correction becomes a chore somebody has to schedule.

The agent also does not mind the work. Finding before writing, checking whether a note already exists, and rewriting a stale paragraph are exactly the tasks people postpone and machines do not. Moving that labor to the agent is the only version of knowledge maintenance that survives a busy week.

What the background does

Apart from anything an agent asks for, aizk runs its own passes over what it holds. They build the entities, facts, communities, profiles and summaries described on Sources and derived knowledge.

Everything they produce is replaceable. If a pass produces something wrong, the fix is to run it again rather than hand-edit the result, and nothing you wrote is ever at risk. The passes also handle the quiet parts of aging, closing facts whose source changed and letting facts nobody has used in a long time drop out of the current set.

Notice what the background does not do. It never edits a source, never invents knowledge no source supports, and never sends you a task.

Where this comes from

The rule that new evidence should add, update, or quietly no-op against what is already stored, rather than pile up for a person to reconcile, follows Mem0. The references map traces the rest of the lineage.

What human operators actually do

Somebody does run this thing, and their job is the infrastructure, not the knowledge.

They keep the database healthy and backed up, keep the models serving, keep identity and membership correct, and handle upgrades and storage. Every one of those is a real job with real consequences when it is neglected.

None of them involve reading your notes. An operator does not approve memories, does not curate them, and does not decide what is worth keeping. Access is settled by scope rules the database itself enforces, so an operator is not a gatekeeper between a writer and a reader either.

What this asks of you

Very little, which is the point.

Write through your assistant and let it handle find, destination and provenance. When you notice something in memory is wrong, say so in the conversation and let the assistant write the correction, because that is a two second interruption rather than a maintenance session.

The one thing worth doing yourself is occasionally reading what accumulated, which the web app makes easy. Not to approve it, just to notice whether the memory looks like the work you actually did. Notes that stay useful collects the habits that keep the answer yes.

Next

• Notes that stay useful is the habits companion to this page.
• Time and history owns corrections and the expiry rules.
• The web app is how you look at what memory has become.