Autonomy

AIZK has no review system and will not gain one. Agents decide what to remember, where it belongs, and when changed evidence requires a correction. A successful write becomes a source immediately. The autonomous engine maintains rebuildable projections and operational health rather than judging whether knowledge deserves acceptance. Human operators maintain infrastructure rather than process memory.

A write drops an immutable source into the store, the ingest queue projects it into the graph, and scheduled passes keep that graph correct and current. The loop has two entry points, the on-write queue and the cron scheduler, and every derived pass feeds the same graph that recall reads back.

flowchart TD
    W["agent remember"] --> SRC["Document + Chunks<br/>immutable source"]
    SRC --> IQ[["ingest queue"]]
    IQ --> CP["ChunkProjectionJob<br/>gate, extract, consolidate"]
    CP --> KG[("graph<br/>entities + fact claims")]
    CP -->|entity_dirty| PP["ProfileProjectionJob"]
    PP --> KG

    SCH[["cron scheduler<br/>one task per scope set"]]
    SCH --> SP["SessionPromoteJob"] --> IQ
    SCH --> PP
    KG -->|growth gate| CM["CommunitiesJob"] --> KG
    KG -->|growth gate| RA["RaptorJob"] --> KG
    SCH --> IN["InsightJob<br/>observations back as facts"] --> KG
    SCH --> DD["DedupJob<br/>merge duplicates"] --> KG
    SCH --> DC["DecayJob<br/>archive stale facts"] --> KG
    KG -.->|recall| W

Maintenance never runs in a caller request. A pgqueuer worker drains durable jobs and a cron scheduler fans scheduled passes out once per exact scope set. Each job binds authority for only that scope set before it opens an application session.

Application jobs declare a typed payload, a stable entrypoint, priority, concurrency, and retry budget through the shared queue package. The package owns serialization, deduplication, database connections, and PgQueuer registration. New background behavior therefore states policy in one job class instead of repeating queue plumbing.

The scheduler discovers its roster from stored documents and unpromoted working memory under the database administrator role. It does not need a user or organization table. Scope-keyed watermarks debounce passes whose corpus has not changed.

Graph consolidation ranks only a bounded current fact set in PostgreSQL. The final transaction locks each exact scope, subject, and perspective slot, reruns that ranking, and writes only when it still matches the model-time snapshot. A concurrent change replans the candidate instead of applying a stale ADD or UPDATE. Content rows mint in one batch on the normal path, while an actual deterministic ID race falls back to isolated savepoints because RLS-hidden content cannot safely use ON CONFLICT.

Pass	Work	Trigger
graph build	gate, extract, consolidate, and write each pending chunk	ingest queue
session promotion	copy old or overflowing working memory into the graph pipeline	schedule
dedup	merge duplicate entity content without losing any scoped claim	schedule
decay	archive stale and rarely accessed facts from default recall	schedule
communities	detect entity clusters and write thematic summaries	growth gate
RAPTOR	build recursive summaries above communities	growth gate
profiles	refresh rolled-up entity descriptions	schedule and write queue
insights	derive higher-level observations and write them back as facts	schedule
backup	write and prune database backups	schedule

Profile projections run at priority 100 with a small fleet-wide concurrency limit. Chunk graph projections run at priority 50. Scheduled maintenance runs at priority 10 and allows one active job of each kind. A job retries from the database up to five times. Exhausted jobs remain held with their deduplication key, which prevents a poison job from being silently recreated until an operator or agent deliberately resolves or requeues it.

An A-and-B job binds A and B as its read authority, so RLS supplies A, B, and bridge knowledge. It writes derived artifacts to the exact A-and-B scope set. Creator identity is provenance only and is not the maintenance partition.

Operations

aizk admin database setup migrates to head, installs the queue schema, and grants the application role. Compose runs it in the one-shot setup service before either long-lived Aizk process starts. The public server has AIZK_AUTO_SETUP=0 and no owner credential. The private worker owns queue execution, scheduled maintenance, and backups.

aizk admin health reports migration currency, row security drift, row counts, queue depth, model identity, per-scope projection progress, and one bounded recall. Operational commands are CLI-only and are not registered on the network MCP server. In Compose, run the command inside worker because only that private process carries owner maintenance authority.