Collectors without readers rot. The Learning Loop is the mechanism that turns observation into durable change.
Always close the loop. Any system that collects data must define how and when that data gets reviewed and applied back. The full path: data, then review, then apply, then the originating system improves. A collector with no reader is drift-in-waiting.
Persek OS has four learning surfaces, each with its own review path, owner, and feedback loop. Every collection surface has a consumer and a real moment when review happens. Signal collection alone is trivial; what runs after is what matters.
This page covers the four surfaces, why each runs at a different rhythm, how repeated evidence becomes durable guidance, and the one boundary rule that ties the whole loop together.
One mechanism wouldn't work. Different signals need different readers and different rhythms.
Surface 01 · Intel weights fast feedback
The fastest feedback surface in the system. When I rate a brief item up or down, the filter adapts around which sources have earned trust. A source that publishes noise sinks. A source that publishes signal rises. Source quality changes fast enough that slow review is too blunt. Over weeks, the firehose gets quieter and more pointed without hand-curation.
Surface 02 · Memory cycles periodic review
Recurring review checks for stale context, contradictions, and items that need human attention. It produces reports for me to act on. It does not mutate anything. It is advisory.
Surface 03 · Pattern capture pattern review
After work sessions, review surfaces anything worth remembering: a preference learned, a workaround that worked, a mistake to avoid next time. A deeper pass clusters recurring observations over time. Patterns that cross a repeated-evidence threshold become candidates for durable guidance. The slower rhythm matters: patterns take time to become patterns.
Surface 04 · Dev review project review
Every code review produces findings. Each finding gets routed: fix inline now, open a tracked issue, or flag as guidance candidate when the same pattern keeps appearing. Periodic distillation rolls lessons into a shared ledger so the lesson isn't trapped inside one project's history.
Forcing every signal into the same rhythm over-reacts on slow signals and under-reacts on fast ones.
Intel feedback moves quickly because source quality changes fast. A podcast that was great six months ago can drift into noise by Q2. The filter has to adapt quickly enough to keep the brief useful.
Memory cycles move more slowly because stale-detection doesn't need to be real-time. A memory entry that goes stale on Tuesday afternoon doesn't hurt anything if it gets reviewed soon after.
Pattern capture moves slowly because patterns take time to become patterns. Repeated observations are the earliest signal that something is structural, and that takes more than a single day of evidence.
One observation is a note. Repetition is the first signal that something may be structural.
When the same pattern appears repeatedly across distinct sessions, it becomes a candidate for durable guidance. A human makes the final call.
The same principle applies across surfaces. Recurring findings, repeated session patterns, and repeated preferences all deserve review before they become durable system guidance.
No job in this system changes durable system guidance without a human touching the promotion.
Automated rewriting has a silent failure mode. Degraded output with no error signal. You don't find out the loop has been lying to itself until a decision made three weeks later turns out to be based on a hallucinated pattern. Keeping the promotion step explicit means every structural change to the system carries attribution and can be audited.
The one exception is intel source-weights. Those rewrite automatically, but the output is consumed only by the next brief's filter, never by durable rules or memory. The automation is scoped strictly downstream. It never propagates upstream into system guidance.
This is the most important constraint in the whole loop. Everything else is mechanism. This is the rule that keeps the mechanism honest.