Separate knowledge surfaces. Clear boundaries. Regular review.
Most software treats memory as a single bucket. Persek OS doesn't. Different kinds of knowledge live in different stores, each with its own rules for who writes, who reads, and how often things expire.
The reason is hygiene. Keeping knowledge accurate over years means not letting one kind of claim silently overwrite another. The system has six stores, one rule that says where each claim belongs, and regular review that keeps things from rotting in place.
Knowledge in Persek OS lives in six places. Each store is good at one job and bad at the others. Mixing them up is how systems quietly rot.
Canon behavior · hand-edited
Agent rules and standards. The instructions every agent loads on startup. Edited by hand, never written at runtime. This is the layer that says how each agent thinks.
gBrain entities · shared write
Live entities. People, projects, companies, concepts. Each entity has compiled truth (the synthesized current view) and a timeline (append-only evidence). Multiple agents can write here with source checks and review boundaries. Built by Garry Tan and integrated into the harness.
LLM Wiki research · append-only
External research. Compiled bibliographies, source ingestion, topic-level synthesis. The library Persek OS reads from when it needs to know what the world says about something. Built by nvk and integrated into the harness.
Agent memory topical · per-agent
Per-agent topical knowledge. Each agent has its own memory directory for things it has learned about its domain. Owned by the agent, refreshed on read.
Auto-memory learned · cross-session
Preferences and facts learned during sessions. The user says "don't do X" once; auto-memory remembers. Curated at session end, survives across sessions.
Session doc live · ephemeral
Current session context. Decisions made, work in flight, blockers. Gone when the session ends, but written to a continuity file so the next session picks up cleanly.
The rule that keeps the six stores from rotting into each other.
Without a boundary rule, the same claim ends up in three places. Two of them go stale. The agent reading the stale one acts on bad information. Multiply by years and the system collapses under its own weight.
The rule is simple: each kind of knowledge has one proper home. That keeps stale context from spreading and makes review possible when facts change.
The routing decision is enforced by discipline and source checks. When in doubt, the destination that can legitimately hold the claim is the right one.
Six steps from "I want to remember this" to "it is safely in durable memory." The fourth step is the trust layer. Other surfaces are lighter, but durable claims get the most review.
Regular review passes keep memory from drifting.
A live system accumulates drift. Stale entries pile up. Two agents can write contradicting facts. Sources move. Without maintenance, the memory rots from the inside. The system surfaces findings to me, not to itself, because nothing gets auto-resolved.
Review pass recurring
Checks for stale entries, broken sources, and anything that needs review. Advisory only. Nothing gets auto-deleted. Items wait for human review.
Consistency check recurring
Looks for contradictions across agents and flags them for me to review and resolve. The system surfaces; I arbitrate. Nothing gets rewritten autonomously.
Health report periodic
Reports on memory health: review volume, source quality, size trends, and entries flagged for attention.
Review briefing regular
The human-readable output becomes one review queue: what got flagged, what is stale, and what needs attention.